File: | llvm/lib/Target/PowerPC/PPCISelLowering.cpp |
Warning: | line 9265, column 36 Although the value stored to 'SplatBits' is used in the enclosing expression, the value is never actually read from 'SplatBits' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- PPCISelLowering.cpp - PPC DAG Lowering Implementation -------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the PPCISelLowering class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "PPCISelLowering.h" |
14 | #include "MCTargetDesc/PPCPredicates.h" |
15 | #include "PPC.h" |
16 | #include "PPCCCState.h" |
17 | #include "PPCCallingConv.h" |
18 | #include "PPCFrameLowering.h" |
19 | #include "PPCInstrInfo.h" |
20 | #include "PPCMachineFunctionInfo.h" |
21 | #include "PPCPerfectShuffle.h" |
22 | #include "PPCRegisterInfo.h" |
23 | #include "PPCSubtarget.h" |
24 | #include "PPCTargetMachine.h" |
25 | #include "llvm/ADT/APFloat.h" |
26 | #include "llvm/ADT/APInt.h" |
27 | #include "llvm/ADT/ArrayRef.h" |
28 | #include "llvm/ADT/DenseMap.h" |
29 | #include "llvm/ADT/None.h" |
30 | #include "llvm/ADT/STLExtras.h" |
31 | #include "llvm/ADT/SmallPtrSet.h" |
32 | #include "llvm/ADT/SmallSet.h" |
33 | #include "llvm/ADT/SmallVector.h" |
34 | #include "llvm/ADT/Statistic.h" |
35 | #include "llvm/ADT/StringRef.h" |
36 | #include "llvm/ADT/StringSwitch.h" |
37 | #include "llvm/CodeGen/CallingConvLower.h" |
38 | #include "llvm/CodeGen/ISDOpcodes.h" |
39 | #include "llvm/CodeGen/MachineBasicBlock.h" |
40 | #include "llvm/CodeGen/MachineFrameInfo.h" |
41 | #include "llvm/CodeGen/MachineFunction.h" |
42 | #include "llvm/CodeGen/MachineInstr.h" |
43 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
44 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
45 | #include "llvm/CodeGen/MachineLoopInfo.h" |
46 | #include "llvm/CodeGen/MachineMemOperand.h" |
47 | #include "llvm/CodeGen/MachineModuleInfo.h" |
48 | #include "llvm/CodeGen/MachineOperand.h" |
49 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
50 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
51 | #include "llvm/CodeGen/SelectionDAG.h" |
52 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
53 | #include "llvm/CodeGen/TargetInstrInfo.h" |
54 | #include "llvm/CodeGen/TargetLowering.h" |
55 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
56 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
57 | #include "llvm/CodeGen/ValueTypes.h" |
58 | #include "llvm/IR/CallingConv.h" |
59 | #include "llvm/IR/Constant.h" |
60 | #include "llvm/IR/Constants.h" |
61 | #include "llvm/IR/DataLayout.h" |
62 | #include "llvm/IR/DebugLoc.h" |
63 | #include "llvm/IR/DerivedTypes.h" |
64 | #include "llvm/IR/Function.h" |
65 | #include "llvm/IR/GlobalValue.h" |
66 | #include "llvm/IR/IRBuilder.h" |
67 | #include "llvm/IR/Instructions.h" |
68 | #include "llvm/IR/Intrinsics.h" |
69 | #include "llvm/IR/IntrinsicsPowerPC.h" |
70 | #include "llvm/IR/Module.h" |
71 | #include "llvm/IR/Type.h" |
72 | #include "llvm/IR/Use.h" |
73 | #include "llvm/IR/Value.h" |
74 | #include "llvm/MC/MCContext.h" |
75 | #include "llvm/MC/MCExpr.h" |
76 | #include "llvm/MC/MCRegisterInfo.h" |
77 | #include "llvm/MC/MCSectionXCOFF.h" |
78 | #include "llvm/MC/MCSymbolXCOFF.h" |
79 | #include "llvm/Support/AtomicOrdering.h" |
80 | #include "llvm/Support/BranchProbability.h" |
81 | #include "llvm/Support/Casting.h" |
82 | #include "llvm/Support/CodeGen.h" |
83 | #include "llvm/Support/CommandLine.h" |
84 | #include "llvm/Support/Compiler.h" |
85 | #include "llvm/Support/Debug.h" |
86 | #include "llvm/Support/ErrorHandling.h" |
87 | #include "llvm/Support/Format.h" |
88 | #include "llvm/Support/KnownBits.h" |
89 | #include "llvm/Support/MachineValueType.h" |
90 | #include "llvm/Support/MathExtras.h" |
91 | #include "llvm/Support/raw_ostream.h" |
92 | #include "llvm/Target/TargetMachine.h" |
93 | #include "llvm/Target/TargetOptions.h" |
94 | #include <algorithm> |
95 | #include <cassert> |
96 | #include <cstdint> |
97 | #include <iterator> |
98 | #include <list> |
99 | #include <utility> |
100 | #include <vector> |
101 | |
102 | using namespace llvm; |
103 | |
104 | #define DEBUG_TYPE"ppc-lowering" "ppc-lowering" |
105 | |
106 | static cl::opt<bool> DisablePPCPreinc("disable-ppc-preinc", |
107 | cl::desc("disable preincrement load/store generation on PPC"), cl::Hidden); |
108 | |
109 | static cl::opt<bool> DisableILPPref("disable-ppc-ilp-pref", |
110 | cl::desc("disable setting the node scheduling preference to ILP on PPC"), cl::Hidden); |
111 | |
112 | static cl::opt<bool> DisablePPCUnaligned("disable-ppc-unaligned", |
113 | cl::desc("disable unaligned load/store generation on PPC"), cl::Hidden); |
114 | |
115 | static cl::opt<bool> DisableSCO("disable-ppc-sco", |
116 | cl::desc("disable sibling call optimization on ppc"), cl::Hidden); |
117 | |
118 | static cl::opt<bool> DisableInnermostLoopAlign32("disable-ppc-innermost-loop-align32", |
119 | cl::desc("don't always align innermost loop to 32 bytes on ppc"), cl::Hidden); |
120 | |
121 | static cl::opt<bool> UseAbsoluteJumpTables("ppc-use-absolute-jumptables", |
122 | cl::desc("use absolute jump tables on ppc"), cl::Hidden); |
123 | |
124 | static cl::opt<bool> EnablePPCPCRelTLS( |
125 | "enable-ppc-pcrel-tls", |
126 | cl::desc("enable the use of PC relative memops in TLS instructions on PPC"), |
127 | cl::Hidden); |
128 | |
129 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"ppc-lowering", "NumTailCalls" , "Number of tail calls"}; |
130 | STATISTIC(NumSiblingCalls, "Number of sibling calls")static llvm::Statistic NumSiblingCalls = {"ppc-lowering", "NumSiblingCalls" , "Number of sibling calls"}; |
131 | STATISTIC(ShufflesHandledWithVPERM, "Number of shuffles lowered to a VPERM")static llvm::Statistic ShufflesHandledWithVPERM = {"ppc-lowering" , "ShufflesHandledWithVPERM", "Number of shuffles lowered to a VPERM" }; |
132 | STATISTIC(NumDynamicAllocaProbed, "Number of dynamic stack allocation probed")static llvm::Statistic NumDynamicAllocaProbed = {"ppc-lowering" , "NumDynamicAllocaProbed", "Number of dynamic stack allocation probed" }; |
133 | |
134 | static bool isNByteElemShuffleMask(ShuffleVectorSDNode *, unsigned, int); |
135 | |
136 | static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl); |
137 | |
138 | // FIXME: Remove this once the bug has been fixed! |
139 | extern cl::opt<bool> ANDIGlueBug; |
140 | |
141 | PPCTargetLowering::PPCTargetLowering(const PPCTargetMachine &TM, |
142 | const PPCSubtarget &STI) |
143 | : TargetLowering(TM), Subtarget(STI) { |
144 | // On PPC32/64, arguments smaller than 4/8 bytes are extended, so all |
145 | // arguments are at least 4/8 bytes aligned. |
146 | bool isPPC64 = Subtarget.isPPC64(); |
147 | setMinStackArgumentAlignment(isPPC64 ? Align(8) : Align(4)); |
148 | |
149 | // Set up the register classes. |
150 | addRegisterClass(MVT::i32, &PPC::GPRCRegClass); |
151 | if (!useSoftFloat()) { |
152 | if (hasSPE()) { |
153 | addRegisterClass(MVT::f32, &PPC::GPRCRegClass); |
154 | addRegisterClass(MVT::f64, &PPC::SPERCRegClass); |
155 | } else { |
156 | addRegisterClass(MVT::f32, &PPC::F4RCRegClass); |
157 | addRegisterClass(MVT::f64, &PPC::F8RCRegClass); |
158 | } |
159 | } |
160 | |
161 | // Match BITREVERSE to customized fast code sequence in the td file. |
162 | setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); |
163 | setOperationAction(ISD::BITREVERSE, MVT::i64, Legal); |
164 | |
165 | // Sub-word ATOMIC_CMP_SWAP need to ensure that the input is zero-extended. |
166 | setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom); |
167 | |
168 | // PowerPC has an i16 but no i8 (or i1) SEXTLOAD. |
169 | for (MVT VT : MVT::integer_valuetypes()) { |
170 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
171 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand); |
172 | } |
173 | |
174 | if (Subtarget.isISA3_0()) { |
175 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Legal); |
176 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Legal); |
177 | setTruncStoreAction(MVT::f64, MVT::f16, Legal); |
178 | setTruncStoreAction(MVT::f32, MVT::f16, Legal); |
179 | } else { |
180 | // No extending loads from f16 or HW conversions back and forth. |
181 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); |
182 | setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand); |
183 | setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand); |
184 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); |
185 | setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand); |
186 | setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand); |
187 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); |
188 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); |
189 | } |
190 | |
191 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
192 | |
193 | // PowerPC has pre-inc load and store's. |
194 | setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal); |
195 | setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal); |
196 | setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal); |
197 | setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal); |
198 | setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal); |
199 | setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal); |
200 | setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal); |
201 | setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal); |
202 | setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal); |
203 | setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal); |
204 | if (!Subtarget.hasSPE()) { |
205 | setIndexedLoadAction(ISD::PRE_INC, MVT::f32, Legal); |
206 | setIndexedLoadAction(ISD::PRE_INC, MVT::f64, Legal); |
207 | setIndexedStoreAction(ISD::PRE_INC, MVT::f32, Legal); |
208 | setIndexedStoreAction(ISD::PRE_INC, MVT::f64, Legal); |
209 | } |
210 | |
211 | // PowerPC uses ADDC/ADDE/SUBC/SUBE to propagate carry. |
212 | const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 }; |
213 | for (MVT VT : ScalarIntVTs) { |
214 | setOperationAction(ISD::ADDC, VT, Legal); |
215 | setOperationAction(ISD::ADDE, VT, Legal); |
216 | setOperationAction(ISD::SUBC, VT, Legal); |
217 | setOperationAction(ISD::SUBE, VT, Legal); |
218 | } |
219 | |
220 | if (Subtarget.useCRBits()) { |
221 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
222 | |
223 | if (isPPC64 || Subtarget.hasFPCVT()) { |
224 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Promote); |
225 | AddPromotedToType(ISD::STRICT_SINT_TO_FP, MVT::i1, |
226 | isPPC64 ? MVT::i64 : MVT::i32); |
227 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Promote); |
228 | AddPromotedToType(ISD::STRICT_UINT_TO_FP, MVT::i1, |
229 | isPPC64 ? MVT::i64 : MVT::i32); |
230 | |
231 | setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote); |
232 | AddPromotedToType (ISD::SINT_TO_FP, MVT::i1, |
233 | isPPC64 ? MVT::i64 : MVT::i32); |
234 | setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote); |
235 | AddPromotedToType(ISD::UINT_TO_FP, MVT::i1, |
236 | isPPC64 ? MVT::i64 : MVT::i32); |
237 | } else { |
238 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i1, Custom); |
239 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i1, Custom); |
240 | setOperationAction(ISD::SINT_TO_FP, MVT::i1, Custom); |
241 | setOperationAction(ISD::UINT_TO_FP, MVT::i1, Custom); |
242 | } |
243 | |
244 | // PowerPC does not support direct load/store of condition registers. |
245 | setOperationAction(ISD::LOAD, MVT::i1, Custom); |
246 | setOperationAction(ISD::STORE, MVT::i1, Custom); |
247 | |
248 | // FIXME: Remove this once the ANDI glue bug is fixed: |
249 | if (ANDIGlueBug) |
250 | setOperationAction(ISD::TRUNCATE, MVT::i1, Custom); |
251 | |
252 | for (MVT VT : MVT::integer_valuetypes()) { |
253 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote); |
254 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote); |
255 | setTruncStoreAction(VT, MVT::i1, Expand); |
256 | } |
257 | |
258 | addRegisterClass(MVT::i1, &PPC::CRBITRCRegClass); |
259 | } |
260 | |
261 | // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on |
262 | // PPC (the libcall is not available). |
263 | setOperationAction(ISD::FP_TO_SINT, MVT::ppcf128, Custom); |
264 | setOperationAction(ISD::FP_TO_UINT, MVT::ppcf128, Custom); |
265 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::ppcf128, Custom); |
266 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::ppcf128, Custom); |
267 | |
268 | // We do not currently implement these libm ops for PowerPC. |
269 | setOperationAction(ISD::FFLOOR, MVT::ppcf128, Expand); |
270 | setOperationAction(ISD::FCEIL, MVT::ppcf128, Expand); |
271 | setOperationAction(ISD::FTRUNC, MVT::ppcf128, Expand); |
272 | setOperationAction(ISD::FRINT, MVT::ppcf128, Expand); |
273 | setOperationAction(ISD::FNEARBYINT, MVT::ppcf128, Expand); |
274 | setOperationAction(ISD::FREM, MVT::ppcf128, Expand); |
275 | |
276 | // PowerPC has no SREM/UREM instructions unless we are on P9 |
277 | // On P9 we may use a hardware instruction to compute the remainder. |
278 | // When the result of both the remainder and the division is required it is |
279 | // more efficient to compute the remainder from the result of the division |
280 | // rather than use the remainder instruction. The instructions are legalized |
281 | // directly because the DivRemPairsPass performs the transformation at the IR |
282 | // level. |
283 | if (Subtarget.isISA3_0()) { |
284 | setOperationAction(ISD::SREM, MVT::i32, Legal); |
285 | setOperationAction(ISD::UREM, MVT::i32, Legal); |
286 | setOperationAction(ISD::SREM, MVT::i64, Legal); |
287 | setOperationAction(ISD::UREM, MVT::i64, Legal); |
288 | } else { |
289 | setOperationAction(ISD::SREM, MVT::i32, Expand); |
290 | setOperationAction(ISD::UREM, MVT::i32, Expand); |
291 | setOperationAction(ISD::SREM, MVT::i64, Expand); |
292 | setOperationAction(ISD::UREM, MVT::i64, Expand); |
293 | } |
294 | |
295 | // Don't use SMUL_LOHI/UMUL_LOHI or SDIVREM/UDIVREM to lower SREM/UREM. |
296 | setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); |
297 | setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); |
298 | setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); |
299 | setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); |
300 | setOperationAction(ISD::UDIVREM, MVT::i32, Expand); |
301 | setOperationAction(ISD::SDIVREM, MVT::i32, Expand); |
302 | setOperationAction(ISD::UDIVREM, MVT::i64, Expand); |
303 | setOperationAction(ISD::SDIVREM, MVT::i64, Expand); |
304 | |
305 | // Handle constrained floating-point operations of scalar. |
306 | // TODO: Handle SPE specific operation. |
307 | setOperationAction(ISD::STRICT_FADD, MVT::f32, Legal); |
308 | setOperationAction(ISD::STRICT_FSUB, MVT::f32, Legal); |
309 | setOperationAction(ISD::STRICT_FMUL, MVT::f32, Legal); |
310 | setOperationAction(ISD::STRICT_FDIV, MVT::f32, Legal); |
311 | setOperationAction(ISD::STRICT_FMA, MVT::f32, Legal); |
312 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal); |
313 | |
314 | setOperationAction(ISD::STRICT_FADD, MVT::f64, Legal); |
315 | setOperationAction(ISD::STRICT_FSUB, MVT::f64, Legal); |
316 | setOperationAction(ISD::STRICT_FMUL, MVT::f64, Legal); |
317 | setOperationAction(ISD::STRICT_FDIV, MVT::f64, Legal); |
318 | setOperationAction(ISD::STRICT_FMA, MVT::f64, Legal); |
319 | if (Subtarget.hasVSX()) { |
320 | setOperationAction(ISD::STRICT_FRINT, MVT::f32, Legal); |
321 | setOperationAction(ISD::STRICT_FRINT, MVT::f64, Legal); |
322 | } |
323 | |
324 | if (Subtarget.hasFSQRT()) { |
325 | setOperationAction(ISD::STRICT_FSQRT, MVT::f32, Legal); |
326 | setOperationAction(ISD::STRICT_FSQRT, MVT::f64, Legal); |
327 | } |
328 | |
329 | if (Subtarget.hasFPRND()) { |
330 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f32, Legal); |
331 | setOperationAction(ISD::STRICT_FCEIL, MVT::f32, Legal); |
332 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f32, Legal); |
333 | setOperationAction(ISD::STRICT_FROUND, MVT::f32, Legal); |
334 | |
335 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f64, Legal); |
336 | setOperationAction(ISD::STRICT_FCEIL, MVT::f64, Legal); |
337 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f64, Legal); |
338 | setOperationAction(ISD::STRICT_FROUND, MVT::f64, Legal); |
339 | } |
340 | |
341 | // We don't support sin/cos/sqrt/fmod/pow |
342 | setOperationAction(ISD::FSIN , MVT::f64, Expand); |
343 | setOperationAction(ISD::FCOS , MVT::f64, Expand); |
344 | setOperationAction(ISD::FSINCOS, MVT::f64, Expand); |
345 | setOperationAction(ISD::FREM , MVT::f64, Expand); |
346 | setOperationAction(ISD::FPOW , MVT::f64, Expand); |
347 | setOperationAction(ISD::FSIN , MVT::f32, Expand); |
348 | setOperationAction(ISD::FCOS , MVT::f32, Expand); |
349 | setOperationAction(ISD::FSINCOS, MVT::f32, Expand); |
350 | setOperationAction(ISD::FREM , MVT::f32, Expand); |
351 | setOperationAction(ISD::FPOW , MVT::f32, Expand); |
352 | if (Subtarget.hasSPE()) { |
353 | setOperationAction(ISD::FMA , MVT::f64, Expand); |
354 | setOperationAction(ISD::FMA , MVT::f32, Expand); |
355 | } else { |
356 | setOperationAction(ISD::FMA , MVT::f64, Legal); |
357 | setOperationAction(ISD::FMA , MVT::f32, Legal); |
358 | } |
359 | |
360 | if (Subtarget.hasSPE()) |
361 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); |
362 | |
363 | setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom); |
364 | |
365 | // If we're enabling GP optimizations, use hardware square root |
366 | if (!Subtarget.hasFSQRT() && |
367 | !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTE() && |
368 | Subtarget.hasFRE())) |
369 | setOperationAction(ISD::FSQRT, MVT::f64, Expand); |
370 | |
371 | if (!Subtarget.hasFSQRT() && |
372 | !(TM.Options.UnsafeFPMath && Subtarget.hasFRSQRTES() && |
373 | Subtarget.hasFRES())) |
374 | setOperationAction(ISD::FSQRT, MVT::f32, Expand); |
375 | |
376 | if (Subtarget.hasFCPSGN()) { |
377 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Legal); |
378 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Legal); |
379 | } else { |
380 | setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand); |
381 | setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand); |
382 | } |
383 | |
384 | if (Subtarget.hasFPRND()) { |
385 | setOperationAction(ISD::FFLOOR, MVT::f64, Legal); |
386 | setOperationAction(ISD::FCEIL, MVT::f64, Legal); |
387 | setOperationAction(ISD::FTRUNC, MVT::f64, Legal); |
388 | setOperationAction(ISD::FROUND, MVT::f64, Legal); |
389 | |
390 | setOperationAction(ISD::FFLOOR, MVT::f32, Legal); |
391 | setOperationAction(ISD::FCEIL, MVT::f32, Legal); |
392 | setOperationAction(ISD::FTRUNC, MVT::f32, Legal); |
393 | setOperationAction(ISD::FROUND, MVT::f32, Legal); |
394 | } |
395 | |
396 | // PowerPC does not have BSWAP, but we can use vector BSWAP instruction xxbrd |
397 | // to speed up scalar BSWAP64. |
398 | // CTPOP or CTTZ were introduced in P8/P9 respectively |
399 | setOperationAction(ISD::BSWAP, MVT::i32 , Expand); |
400 | if (Subtarget.hasP9Vector()) |
401 | setOperationAction(ISD::BSWAP, MVT::i64 , Custom); |
402 | else |
403 | setOperationAction(ISD::BSWAP, MVT::i64 , Expand); |
404 | if (Subtarget.isISA3_0()) { |
405 | setOperationAction(ISD::CTTZ , MVT::i32 , Legal); |
406 | setOperationAction(ISD::CTTZ , MVT::i64 , Legal); |
407 | } else { |
408 | setOperationAction(ISD::CTTZ , MVT::i32 , Expand); |
409 | setOperationAction(ISD::CTTZ , MVT::i64 , Expand); |
410 | } |
411 | |
412 | if (Subtarget.hasPOPCNTD() == PPCSubtarget::POPCNTD_Fast) { |
413 | setOperationAction(ISD::CTPOP, MVT::i32 , Legal); |
414 | setOperationAction(ISD::CTPOP, MVT::i64 , Legal); |
415 | } else { |
416 | setOperationAction(ISD::CTPOP, MVT::i32 , Expand); |
417 | setOperationAction(ISD::CTPOP, MVT::i64 , Expand); |
418 | } |
419 | |
420 | // PowerPC does not have ROTR |
421 | setOperationAction(ISD::ROTR, MVT::i32 , Expand); |
422 | setOperationAction(ISD::ROTR, MVT::i64 , Expand); |
423 | |
424 | if (!Subtarget.useCRBits()) { |
425 | // PowerPC does not have Select |
426 | setOperationAction(ISD::SELECT, MVT::i32, Expand); |
427 | setOperationAction(ISD::SELECT, MVT::i64, Expand); |
428 | setOperationAction(ISD::SELECT, MVT::f32, Expand); |
429 | setOperationAction(ISD::SELECT, MVT::f64, Expand); |
430 | } |
431 | |
432 | // PowerPC wants to turn select_cc of FP into fsel when possible. |
433 | setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); |
434 | setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); |
435 | |
436 | // PowerPC wants to optimize integer setcc a bit |
437 | if (!Subtarget.useCRBits()) |
438 | setOperationAction(ISD::SETCC, MVT::i32, Custom); |
439 | |
440 | if (Subtarget.hasFPU()) { |
441 | setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Legal); |
442 | setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Legal); |
443 | setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Legal); |
444 | |
445 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Legal); |
446 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Legal); |
447 | setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Legal); |
448 | } |
449 | |
450 | // PowerPC does not have BRCOND which requires SetCC |
451 | if (!Subtarget.useCRBits()) |
452 | setOperationAction(ISD::BRCOND, MVT::Other, Expand); |
453 | |
454 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); |
455 | |
456 | if (Subtarget.hasSPE()) { |
457 | // SPE has built-in conversions |
458 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Legal); |
459 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Legal); |
460 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Legal); |
461 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal); |
462 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal); |
463 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal); |
464 | } else { |
465 | // PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores. |
466 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); |
467 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); |
468 | |
469 | // PowerPC does not have [U|S]INT_TO_FP |
470 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Expand); |
471 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Expand); |
472 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand); |
473 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); |
474 | } |
475 | |
476 | if (Subtarget.hasDirectMove() && isPPC64) { |
477 | setOperationAction(ISD::BITCAST, MVT::f32, Legal); |
478 | setOperationAction(ISD::BITCAST, MVT::i32, Legal); |
479 | setOperationAction(ISD::BITCAST, MVT::i64, Legal); |
480 | setOperationAction(ISD::BITCAST, MVT::f64, Legal); |
481 | if (TM.Options.UnsafeFPMath) { |
482 | setOperationAction(ISD::LRINT, MVT::f64, Legal); |
483 | setOperationAction(ISD::LRINT, MVT::f32, Legal); |
484 | setOperationAction(ISD::LLRINT, MVT::f64, Legal); |
485 | setOperationAction(ISD::LLRINT, MVT::f32, Legal); |
486 | setOperationAction(ISD::LROUND, MVT::f64, Legal); |
487 | setOperationAction(ISD::LROUND, MVT::f32, Legal); |
488 | setOperationAction(ISD::LLROUND, MVT::f64, Legal); |
489 | setOperationAction(ISD::LLROUND, MVT::f32, Legal); |
490 | } |
491 | } else { |
492 | setOperationAction(ISD::BITCAST, MVT::f32, Expand); |
493 | setOperationAction(ISD::BITCAST, MVT::i32, Expand); |
494 | setOperationAction(ISD::BITCAST, MVT::i64, Expand); |
495 | setOperationAction(ISD::BITCAST, MVT::f64, Expand); |
496 | } |
497 | |
498 | // We cannot sextinreg(i1). Expand to shifts. |
499 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
500 | |
501 | // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support |
502 | // SjLj exception handling but a light-weight setjmp/longjmp replacement to |
503 | // support continuation, user-level threading, and etc.. As a result, no |
504 | // other SjLj exception interfaces are implemented and please don't build |
505 | // your own exception handling based on them. |
506 | // LLVM/Clang supports zero-cost DWARF exception handling. |
507 | setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom); |
508 | setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom); |
509 | |
510 | // We want to legalize GlobalAddress and ConstantPool nodes into the |
511 | // appropriate instructions to materialize the address. |
512 | setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); |
513 | setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); |
514 | setOperationAction(ISD::BlockAddress, MVT::i32, Custom); |
515 | setOperationAction(ISD::ConstantPool, MVT::i32, Custom); |
516 | setOperationAction(ISD::JumpTable, MVT::i32, Custom); |
517 | setOperationAction(ISD::GlobalAddress, MVT::i64, Custom); |
518 | setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom); |
519 | setOperationAction(ISD::BlockAddress, MVT::i64, Custom); |
520 | setOperationAction(ISD::ConstantPool, MVT::i64, Custom); |
521 | setOperationAction(ISD::JumpTable, MVT::i64, Custom); |
522 | |
523 | // TRAP is legal. |
524 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
525 | |
526 | // TRAMPOLINE is custom lowered. |
527 | setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom); |
528 | setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom); |
529 | |
530 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex |
531 | setOperationAction(ISD::VASTART , MVT::Other, Custom); |
532 | |
533 | if (Subtarget.is64BitELFABI()) { |
534 | // VAARG always uses double-word chunks, so promote anything smaller. |
535 | setOperationAction(ISD::VAARG, MVT::i1, Promote); |
536 | AddPromotedToType(ISD::VAARG, MVT::i1, MVT::i64); |
537 | setOperationAction(ISD::VAARG, MVT::i8, Promote); |
538 | AddPromotedToType(ISD::VAARG, MVT::i8, MVT::i64); |
539 | setOperationAction(ISD::VAARG, MVT::i16, Promote); |
540 | AddPromotedToType(ISD::VAARG, MVT::i16, MVT::i64); |
541 | setOperationAction(ISD::VAARG, MVT::i32, Promote); |
542 | AddPromotedToType(ISD::VAARG, MVT::i32, MVT::i64); |
543 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
544 | } else if (Subtarget.is32BitELFABI()) { |
545 | // VAARG is custom lowered with the 32-bit SVR4 ABI. |
546 | setOperationAction(ISD::VAARG, MVT::Other, Custom); |
547 | setOperationAction(ISD::VAARG, MVT::i64, Custom); |
548 | } else |
549 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
550 | |
551 | // VACOPY is custom lowered with the 32-bit SVR4 ABI. |
552 | if (Subtarget.is32BitELFABI()) |
553 | setOperationAction(ISD::VACOPY , MVT::Other, Custom); |
554 | else |
555 | setOperationAction(ISD::VACOPY , MVT::Other, Expand); |
556 | |
557 | // Use the default implementation. |
558 | setOperationAction(ISD::VAEND , MVT::Other, Expand); |
559 | setOperationAction(ISD::STACKSAVE , MVT::Other, Expand); |
560 | setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom); |
561 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom); |
562 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom); |
563 | setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i32, Custom); |
564 | setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, MVT::i64, Custom); |
565 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i32, Custom); |
566 | setOperationAction(ISD::EH_DWARF_CFA, MVT::i64, Custom); |
567 | |
568 | // We want to custom lower some of our intrinsics. |
569 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
570 | |
571 | // To handle counter-based loop conditions. |
572 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i1, Custom); |
573 | |
574 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i8, Custom); |
575 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i16, Custom); |
576 | setOperationAction(ISD::INTRINSIC_VOID, MVT::i32, Custom); |
577 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
578 | |
579 | // Comparisons that require checking two conditions. |
580 | if (Subtarget.hasSPE()) { |
581 | setCondCodeAction(ISD::SETO, MVT::f32, Expand); |
582 | setCondCodeAction(ISD::SETO, MVT::f64, Expand); |
583 | setCondCodeAction(ISD::SETUO, MVT::f32, Expand); |
584 | setCondCodeAction(ISD::SETUO, MVT::f64, Expand); |
585 | } |
586 | setCondCodeAction(ISD::SETULT, MVT::f32, Expand); |
587 | setCondCodeAction(ISD::SETULT, MVT::f64, Expand); |
588 | setCondCodeAction(ISD::SETUGT, MVT::f32, Expand); |
589 | setCondCodeAction(ISD::SETUGT, MVT::f64, Expand); |
590 | setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand); |
591 | setCondCodeAction(ISD::SETUEQ, MVT::f64, Expand); |
592 | setCondCodeAction(ISD::SETOGE, MVT::f32, Expand); |
593 | setCondCodeAction(ISD::SETOGE, MVT::f64, Expand); |
594 | setCondCodeAction(ISD::SETOLE, MVT::f32, Expand); |
595 | setCondCodeAction(ISD::SETOLE, MVT::f64, Expand); |
596 | setCondCodeAction(ISD::SETONE, MVT::f32, Expand); |
597 | setCondCodeAction(ISD::SETONE, MVT::f64, Expand); |
598 | |
599 | if (Subtarget.has64BitSupport()) { |
600 | // They also have instructions for converting between i64 and fp. |
601 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); |
602 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Expand); |
603 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); |
604 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Expand); |
605 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); |
606 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand); |
607 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); |
608 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); |
609 | // This is just the low 32 bits of a (signed) fp->i64 conversion. |
610 | // We cannot do this with Promote because i64 is not a legal type. |
611 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); |
612 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); |
613 | |
614 | if (Subtarget.hasLFIWAX() || Subtarget.isPPC64()) { |
615 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); |
616 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); |
617 | } |
618 | } else { |
619 | // PowerPC does not have FP_TO_UINT on 32-bit implementations. |
620 | if (Subtarget.hasSPE()) { |
621 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Legal); |
622 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal); |
623 | } else { |
624 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Expand); |
625 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand); |
626 | } |
627 | } |
628 | |
629 | // With the instructions enabled under FPCVT, we can do everything. |
630 | if (Subtarget.hasFPCVT()) { |
631 | if (Subtarget.has64BitSupport()) { |
632 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom); |
633 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom); |
634 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom); |
635 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom); |
636 | setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom); |
637 | setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom); |
638 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom); |
639 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom); |
640 | } |
641 | |
642 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); |
643 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); |
644 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom); |
645 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom); |
646 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); |
647 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); |
648 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); |
649 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); |
650 | } |
651 | |
652 | if (Subtarget.use64BitRegs()) { |
653 | // 64-bit PowerPC implementations can support i64 types directly |
654 | addRegisterClass(MVT::i64, &PPC::G8RCRegClass); |
655 | // BUILD_PAIR can't be handled natively, and should be expanded to shl/or |
656 | setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand); |
657 | // 64-bit PowerPC wants to expand i128 shifts itself. |
658 | setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom); |
659 | setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom); |
660 | setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom); |
661 | } else { |
662 | // 32-bit PowerPC wants to expand i64 shifts itself. |
663 | setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom); |
664 | setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom); |
665 | setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom); |
666 | } |
667 | |
668 | // PowerPC has better expansions for funnel shifts than the generic |
669 | // TargetLowering::expandFunnelShift. |
670 | if (Subtarget.has64BitSupport()) { |
671 | setOperationAction(ISD::FSHL, MVT::i64, Custom); |
672 | setOperationAction(ISD::FSHR, MVT::i64, Custom); |
673 | } |
674 | setOperationAction(ISD::FSHL, MVT::i32, Custom); |
675 | setOperationAction(ISD::FSHR, MVT::i32, Custom); |
676 | |
677 | if (Subtarget.hasVSX()) { |
678 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f64, Legal); |
679 | setOperationAction(ISD::FMAXNUM_IEEE, MVT::f32, Legal); |
680 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f64, Legal); |
681 | setOperationAction(ISD::FMINNUM_IEEE, MVT::f32, Legal); |
682 | } |
683 | |
684 | if (Subtarget.hasAltivec()) { |
685 | for (MVT VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) { |
686 | setOperationAction(ISD::SADDSAT, VT, Legal); |
687 | setOperationAction(ISD::SSUBSAT, VT, Legal); |
688 | setOperationAction(ISD::UADDSAT, VT, Legal); |
689 | setOperationAction(ISD::USUBSAT, VT, Legal); |
690 | } |
691 | // First set operation action for all vector types to expand. Then we |
692 | // will selectively turn on ones that can be effectively codegen'd. |
693 | for (MVT VT : MVT::fixedlen_vector_valuetypes()) { |
694 | // add/sub are legal for all supported vector VT's. |
695 | setOperationAction(ISD::ADD, VT, Legal); |
696 | setOperationAction(ISD::SUB, VT, Legal); |
697 | |
698 | // For v2i64, these are only valid with P8Vector. This is corrected after |
699 | // the loop. |
700 | if (VT.getSizeInBits() <= 128 && VT.getScalarSizeInBits() <= 64) { |
701 | setOperationAction(ISD::SMAX, VT, Legal); |
702 | setOperationAction(ISD::SMIN, VT, Legal); |
703 | setOperationAction(ISD::UMAX, VT, Legal); |
704 | setOperationAction(ISD::UMIN, VT, Legal); |
705 | } |
706 | else { |
707 | setOperationAction(ISD::SMAX, VT, Expand); |
708 | setOperationAction(ISD::SMIN, VT, Expand); |
709 | setOperationAction(ISD::UMAX, VT, Expand); |
710 | setOperationAction(ISD::UMIN, VT, Expand); |
711 | } |
712 | |
713 | if (Subtarget.hasVSX()) { |
714 | setOperationAction(ISD::FMAXNUM, VT, Legal); |
715 | setOperationAction(ISD::FMINNUM, VT, Legal); |
716 | } |
717 | |
718 | // Vector instructions introduced in P8 |
719 | if (Subtarget.hasP8Altivec() && (VT.SimpleTy != MVT::v1i128)) { |
720 | setOperationAction(ISD::CTPOP, VT, Legal); |
721 | setOperationAction(ISD::CTLZ, VT, Legal); |
722 | } |
723 | else { |
724 | setOperationAction(ISD::CTPOP, VT, Expand); |
725 | setOperationAction(ISD::CTLZ, VT, Expand); |
726 | } |
727 | |
728 | // Vector instructions introduced in P9 |
729 | if (Subtarget.hasP9Altivec() && (VT.SimpleTy != MVT::v1i128)) |
730 | setOperationAction(ISD::CTTZ, VT, Legal); |
731 | else |
732 | setOperationAction(ISD::CTTZ, VT, Expand); |
733 | |
734 | // We promote all shuffles to v16i8. |
735 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Promote); |
736 | AddPromotedToType (ISD::VECTOR_SHUFFLE, VT, MVT::v16i8); |
737 | |
738 | // We promote all non-typed operations to v4i32. |
739 | setOperationAction(ISD::AND , VT, Promote); |
740 | AddPromotedToType (ISD::AND , VT, MVT::v4i32); |
741 | setOperationAction(ISD::OR , VT, Promote); |
742 | AddPromotedToType (ISD::OR , VT, MVT::v4i32); |
743 | setOperationAction(ISD::XOR , VT, Promote); |
744 | AddPromotedToType (ISD::XOR , VT, MVT::v4i32); |
745 | setOperationAction(ISD::LOAD , VT, Promote); |
746 | AddPromotedToType (ISD::LOAD , VT, MVT::v4i32); |
747 | setOperationAction(ISD::SELECT, VT, Promote); |
748 | AddPromotedToType (ISD::SELECT, VT, MVT::v4i32); |
749 | setOperationAction(ISD::VSELECT, VT, Legal); |
750 | setOperationAction(ISD::SELECT_CC, VT, Promote); |
751 | AddPromotedToType (ISD::SELECT_CC, VT, MVT::v4i32); |
752 | setOperationAction(ISD::STORE, VT, Promote); |
753 | AddPromotedToType (ISD::STORE, VT, MVT::v4i32); |
754 | |
755 | // No other operations are legal. |
756 | setOperationAction(ISD::MUL , VT, Expand); |
757 | setOperationAction(ISD::SDIV, VT, Expand); |
758 | setOperationAction(ISD::SREM, VT, Expand); |
759 | setOperationAction(ISD::UDIV, VT, Expand); |
760 | setOperationAction(ISD::UREM, VT, Expand); |
761 | setOperationAction(ISD::FDIV, VT, Expand); |
762 | setOperationAction(ISD::FREM, VT, Expand); |
763 | setOperationAction(ISD::FNEG, VT, Expand); |
764 | setOperationAction(ISD::FSQRT, VT, Expand); |
765 | setOperationAction(ISD::FLOG, VT, Expand); |
766 | setOperationAction(ISD::FLOG10, VT, Expand); |
767 | setOperationAction(ISD::FLOG2, VT, Expand); |
768 | setOperationAction(ISD::FEXP, VT, Expand); |
769 | setOperationAction(ISD::FEXP2, VT, Expand); |
770 | setOperationAction(ISD::FSIN, VT, Expand); |
771 | setOperationAction(ISD::FCOS, VT, Expand); |
772 | setOperationAction(ISD::FABS, VT, Expand); |
773 | setOperationAction(ISD::FFLOOR, VT, Expand); |
774 | setOperationAction(ISD::FCEIL, VT, Expand); |
775 | setOperationAction(ISD::FTRUNC, VT, Expand); |
776 | setOperationAction(ISD::FRINT, VT, Expand); |
777 | setOperationAction(ISD::FNEARBYINT, VT, Expand); |
778 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Expand); |
779 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand); |
780 | setOperationAction(ISD::BUILD_VECTOR, VT, Expand); |
781 | setOperationAction(ISD::MULHU, VT, Expand); |
782 | setOperationAction(ISD::MULHS, VT, Expand); |
783 | setOperationAction(ISD::UMUL_LOHI, VT, Expand); |
784 | setOperationAction(ISD::SMUL_LOHI, VT, Expand); |
785 | setOperationAction(ISD::UDIVREM, VT, Expand); |
786 | setOperationAction(ISD::SDIVREM, VT, Expand); |
787 | setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Expand); |
788 | setOperationAction(ISD::FPOW, VT, Expand); |
789 | setOperationAction(ISD::BSWAP, VT, Expand); |
790 | setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand); |
791 | setOperationAction(ISD::ROTL, VT, Expand); |
792 | setOperationAction(ISD::ROTR, VT, Expand); |
793 | |
794 | for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) { |
795 | setTruncStoreAction(VT, InnerVT, Expand); |
796 | setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand); |
797 | setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand); |
798 | setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand); |
799 | } |
800 | } |
801 | setOperationAction(ISD::SELECT_CC, MVT::v4i32, Expand); |
802 | if (!Subtarget.hasP8Vector()) { |
803 | setOperationAction(ISD::SMAX, MVT::v2i64, Expand); |
804 | setOperationAction(ISD::SMIN, MVT::v2i64, Expand); |
805 | setOperationAction(ISD::UMAX, MVT::v2i64, Expand); |
806 | setOperationAction(ISD::UMIN, MVT::v2i64, Expand); |
807 | } |
808 | |
809 | for (auto VT : {MVT::v2i64, MVT::v4i32, MVT::v8i16, MVT::v16i8}) |
810 | setOperationAction(ISD::ABS, VT, Custom); |
811 | |
812 | // We can custom expand all VECTOR_SHUFFLEs to VPERM, others we can handle |
813 | // with merges, splats, etc. |
814 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom); |
815 | |
816 | // Vector truncates to sub-word integer that fit in an Altivec/VSX register |
817 | // are cheap, so handle them before they get expanded to scalar. |
818 | setOperationAction(ISD::TRUNCATE, MVT::v8i8, Custom); |
819 | setOperationAction(ISD::TRUNCATE, MVT::v4i8, Custom); |
820 | setOperationAction(ISD::TRUNCATE, MVT::v2i8, Custom); |
821 | setOperationAction(ISD::TRUNCATE, MVT::v4i16, Custom); |
822 | setOperationAction(ISD::TRUNCATE, MVT::v2i16, Custom); |
823 | |
824 | setOperationAction(ISD::AND , MVT::v4i32, Legal); |
825 | setOperationAction(ISD::OR , MVT::v4i32, Legal); |
826 | setOperationAction(ISD::XOR , MVT::v4i32, Legal); |
827 | setOperationAction(ISD::LOAD , MVT::v4i32, Legal); |
828 | setOperationAction(ISD::SELECT, MVT::v4i32, |
829 | Subtarget.useCRBits() ? Legal : Expand); |
830 | setOperationAction(ISD::STORE , MVT::v4i32, Legal); |
831 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4i32, Legal); |
832 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4i32, Legal); |
833 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i32, Legal); |
834 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i32, Legal); |
835 | setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal); |
836 | setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal); |
837 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal); |
838 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal); |
839 | setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal); |
840 | setOperationAction(ISD::FCEIL, MVT::v4f32, Legal); |
841 | setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal); |
842 | setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal); |
843 | |
844 | // Without hasP8Altivec set, v2i64 SMAX isn't available. |
845 | // But ABS custom lowering requires SMAX support. |
846 | if (!Subtarget.hasP8Altivec()) |
847 | setOperationAction(ISD::ABS, MVT::v2i64, Expand); |
848 | |
849 | // Custom lowering ROTL v1i128 to VECTOR_SHUFFLE v16i8. |
850 | setOperationAction(ISD::ROTL, MVT::v1i128, Custom); |
851 | // With hasAltivec set, we can lower ISD::ROTL to vrl(b|h|w). |
852 | if (Subtarget.hasAltivec()) |
853 | for (auto VT : {MVT::v4i32, MVT::v8i16, MVT::v16i8}) |
854 | setOperationAction(ISD::ROTL, VT, Legal); |
855 | // With hasP8Altivec set, we can lower ISD::ROTL to vrld. |
856 | if (Subtarget.hasP8Altivec()) |
857 | setOperationAction(ISD::ROTL, MVT::v2i64, Legal); |
858 | |
859 | addRegisterClass(MVT::v4f32, &PPC::VRRCRegClass); |
860 | addRegisterClass(MVT::v4i32, &PPC::VRRCRegClass); |
861 | addRegisterClass(MVT::v8i16, &PPC::VRRCRegClass); |
862 | addRegisterClass(MVT::v16i8, &PPC::VRRCRegClass); |
863 | |
864 | setOperationAction(ISD::MUL, MVT::v4f32, Legal); |
865 | setOperationAction(ISD::FMA, MVT::v4f32, Legal); |
866 | |
867 | if (Subtarget.hasVSX()) { |
868 | setOperationAction(ISD::FDIV, MVT::v4f32, Legal); |
869 | setOperationAction(ISD::FSQRT, MVT::v4f32, Legal); |
870 | } |
871 | |
872 | if (Subtarget.hasP8Altivec()) |
873 | setOperationAction(ISD::MUL, MVT::v4i32, Legal); |
874 | else |
875 | setOperationAction(ISD::MUL, MVT::v4i32, Custom); |
876 | |
877 | if (Subtarget.isISA3_1()) { |
878 | setOperationAction(ISD::MUL, MVT::v2i64, Legal); |
879 | setOperationAction(ISD::MULHS, MVT::v2i64, Legal); |
880 | setOperationAction(ISD::MULHU, MVT::v2i64, Legal); |
881 | setOperationAction(ISD::MULHS, MVT::v4i32, Legal); |
882 | setOperationAction(ISD::MULHU, MVT::v4i32, Legal); |
883 | setOperationAction(ISD::UDIV, MVT::v2i64, Legal); |
884 | setOperationAction(ISD::SDIV, MVT::v2i64, Legal); |
885 | setOperationAction(ISD::UDIV, MVT::v4i32, Legal); |
886 | setOperationAction(ISD::SDIV, MVT::v4i32, Legal); |
887 | setOperationAction(ISD::UREM, MVT::v2i64, Legal); |
888 | setOperationAction(ISD::SREM, MVT::v2i64, Legal); |
889 | setOperationAction(ISD::UREM, MVT::v4i32, Legal); |
890 | setOperationAction(ISD::SREM, MVT::v4i32, Legal); |
891 | setOperationAction(ISD::UREM, MVT::v1i128, Legal); |
892 | setOperationAction(ISD::SREM, MVT::v1i128, Legal); |
893 | setOperationAction(ISD::UDIV, MVT::v1i128, Legal); |
894 | setOperationAction(ISD::SDIV, MVT::v1i128, Legal); |
895 | setOperationAction(ISD::ROTL, MVT::v1i128, Legal); |
896 | } |
897 | |
898 | setOperationAction(ISD::MUL, MVT::v8i16, Legal); |
899 | setOperationAction(ISD::MUL, MVT::v16i8, Custom); |
900 | |
901 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom); |
902 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom); |
903 | |
904 | setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom); |
905 | setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom); |
906 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom); |
907 | setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom); |
908 | |
909 | // Altivec does not contain unordered floating-point compare instructions |
910 | setCondCodeAction(ISD::SETUO, MVT::v4f32, Expand); |
911 | setCondCodeAction(ISD::SETUEQ, MVT::v4f32, Expand); |
912 | setCondCodeAction(ISD::SETO, MVT::v4f32, Expand); |
913 | setCondCodeAction(ISD::SETONE, MVT::v4f32, Expand); |
914 | |
915 | if (Subtarget.hasVSX()) { |
916 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f64, Legal); |
917 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal); |
918 | if (Subtarget.hasP8Vector()) { |
919 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Legal); |
920 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal); |
921 | } |
922 | if (Subtarget.hasDirectMove() && isPPC64) { |
923 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Legal); |
924 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Legal); |
925 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Legal); |
926 | setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i64, Legal); |
927 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Legal); |
928 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Legal); |
929 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal); |
930 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal); |
931 | } |
932 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Legal); |
933 | |
934 | // The nearbyint variants are not allowed to raise the inexact exception |
935 | // so we can only code-gen them with unsafe math. |
936 | if (TM.Options.UnsafeFPMath) { |
937 | setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal); |
938 | setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal); |
939 | } |
940 | |
941 | setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal); |
942 | setOperationAction(ISD::FCEIL, MVT::v2f64, Legal); |
943 | setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal); |
944 | setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal); |
945 | setOperationAction(ISD::FRINT, MVT::v2f64, Legal); |
946 | setOperationAction(ISD::FROUND, MVT::v2f64, Legal); |
947 | setOperationAction(ISD::FROUND, MVT::f64, Legal); |
948 | setOperationAction(ISD::FRINT, MVT::f64, Legal); |
949 | |
950 | setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal); |
951 | setOperationAction(ISD::FRINT, MVT::v4f32, Legal); |
952 | setOperationAction(ISD::FROUND, MVT::v4f32, Legal); |
953 | setOperationAction(ISD::FROUND, MVT::f32, Legal); |
954 | setOperationAction(ISD::FRINT, MVT::f32, Legal); |
955 | |
956 | setOperationAction(ISD::MUL, MVT::v2f64, Legal); |
957 | setOperationAction(ISD::FMA, MVT::v2f64, Legal); |
958 | |
959 | setOperationAction(ISD::FDIV, MVT::v2f64, Legal); |
960 | setOperationAction(ISD::FSQRT, MVT::v2f64, Legal); |
961 | |
962 | // Share the Altivec comparison restrictions. |
963 | setCondCodeAction(ISD::SETUO, MVT::v2f64, Expand); |
964 | setCondCodeAction(ISD::SETUEQ, MVT::v2f64, Expand); |
965 | setCondCodeAction(ISD::SETO, MVT::v2f64, Expand); |
966 | setCondCodeAction(ISD::SETONE, MVT::v2f64, Expand); |
967 | |
968 | setOperationAction(ISD::LOAD, MVT::v2f64, Legal); |
969 | setOperationAction(ISD::STORE, MVT::v2f64, Legal); |
970 | |
971 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Legal); |
972 | |
973 | if (Subtarget.hasP8Vector()) |
974 | addRegisterClass(MVT::f32, &PPC::VSSRCRegClass); |
975 | |
976 | addRegisterClass(MVT::f64, &PPC::VSFRCRegClass); |
977 | |
978 | addRegisterClass(MVT::v4i32, &PPC::VSRCRegClass); |
979 | addRegisterClass(MVT::v4f32, &PPC::VSRCRegClass); |
980 | addRegisterClass(MVT::v2f64, &PPC::VSRCRegClass); |
981 | |
982 | if (Subtarget.hasP8Altivec()) { |
983 | setOperationAction(ISD::SHL, MVT::v2i64, Legal); |
984 | setOperationAction(ISD::SRA, MVT::v2i64, Legal); |
985 | setOperationAction(ISD::SRL, MVT::v2i64, Legal); |
986 | |
987 | // 128 bit shifts can be accomplished via 3 instructions for SHL and |
988 | // SRL, but not for SRA because of the instructions available: |
989 | // VS{RL} and VS{RL}O. However due to direct move costs, it's not worth |
990 | // doing |
991 | setOperationAction(ISD::SHL, MVT::v1i128, Expand); |
992 | setOperationAction(ISD::SRL, MVT::v1i128, Expand); |
993 | setOperationAction(ISD::SRA, MVT::v1i128, Expand); |
994 | |
995 | setOperationAction(ISD::SETCC, MVT::v2i64, Legal); |
996 | } |
997 | else { |
998 | setOperationAction(ISD::SHL, MVT::v2i64, Expand); |
999 | setOperationAction(ISD::SRA, MVT::v2i64, Expand); |
1000 | setOperationAction(ISD::SRL, MVT::v2i64, Expand); |
1001 | |
1002 | setOperationAction(ISD::SETCC, MVT::v2i64, Custom); |
1003 | |
1004 | // VSX v2i64 only supports non-arithmetic operations. |
1005 | setOperationAction(ISD::ADD, MVT::v2i64, Expand); |
1006 | setOperationAction(ISD::SUB, MVT::v2i64, Expand); |
1007 | } |
1008 | |
1009 | if (Subtarget.isISA3_1()) |
1010 | setOperationAction(ISD::SETCC, MVT::v1i128, Legal); |
1011 | else |
1012 | setOperationAction(ISD::SETCC, MVT::v1i128, Expand); |
1013 | |
1014 | setOperationAction(ISD::LOAD, MVT::v2i64, Promote); |
1015 | AddPromotedToType (ISD::LOAD, MVT::v2i64, MVT::v2f64); |
1016 | setOperationAction(ISD::STORE, MVT::v2i64, Promote); |
1017 | AddPromotedToType (ISD::STORE, MVT::v2i64, MVT::v2f64); |
1018 | |
1019 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Legal); |
1020 | |
1021 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i64, Legal); |
1022 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i64, Legal); |
1023 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2i64, Legal); |
1024 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2i64, Legal); |
1025 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal); |
1026 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal); |
1027 | setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal); |
1028 | setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal); |
1029 | |
1030 | // Custom handling for partial vectors of integers converted to |
1031 | // floating point. We already have optimal handling for v2i32 through |
1032 | // the DAG combine, so those aren't necessary. |
1033 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i8, Custom); |
1034 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i8, Custom); |
1035 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2i16, Custom); |
1036 | setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i16, Custom); |
1037 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i8, Custom); |
1038 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i8, Custom); |
1039 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2i16, Custom); |
1040 | setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i16, Custom); |
1041 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i8, Custom); |
1042 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom); |
1043 | setOperationAction(ISD::UINT_TO_FP, MVT::v2i16, Custom); |
1044 | setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom); |
1045 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i8, Custom); |
1046 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Custom); |
1047 | setOperationAction(ISD::SINT_TO_FP, MVT::v2i16, Custom); |
1048 | setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom); |
1049 | |
1050 | setOperationAction(ISD::FNEG, MVT::v4f32, Legal); |
1051 | setOperationAction(ISD::FNEG, MVT::v2f64, Legal); |
1052 | setOperationAction(ISD::FABS, MVT::v4f32, Legal); |
1053 | setOperationAction(ISD::FABS, MVT::v2f64, Legal); |
1054 | setOperationAction(ISD::FCOPYSIGN, MVT::v4f32, Legal); |
1055 | setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Legal); |
1056 | |
1057 | if (Subtarget.hasDirectMove()) |
1058 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom); |
1059 | setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom); |
1060 | |
1061 | // Handle constrained floating-point operations of vector. |
1062 | // The predictor is `hasVSX` because altivec instruction has |
1063 | // no exception but VSX vector instruction has. |
1064 | setOperationAction(ISD::STRICT_FADD, MVT::v4f32, Legal); |
1065 | setOperationAction(ISD::STRICT_FSUB, MVT::v4f32, Legal); |
1066 | setOperationAction(ISD::STRICT_FMUL, MVT::v4f32, Legal); |
1067 | setOperationAction(ISD::STRICT_FDIV, MVT::v4f32, Legal); |
1068 | setOperationAction(ISD::STRICT_FMA, MVT::v4f32, Legal); |
1069 | setOperationAction(ISD::STRICT_FSQRT, MVT::v4f32, Legal); |
1070 | setOperationAction(ISD::STRICT_FMAXNUM, MVT::v4f32, Legal); |
1071 | setOperationAction(ISD::STRICT_FMINNUM, MVT::v4f32, Legal); |
1072 | setOperationAction(ISD::STRICT_FRINT, MVT::v4f32, Legal); |
1073 | setOperationAction(ISD::STRICT_FFLOOR, MVT::v4f32, Legal); |
1074 | setOperationAction(ISD::STRICT_FCEIL, MVT::v4f32, Legal); |
1075 | setOperationAction(ISD::STRICT_FTRUNC, MVT::v4f32, Legal); |
1076 | setOperationAction(ISD::STRICT_FROUND, MVT::v4f32, Legal); |
1077 | |
1078 | setOperationAction(ISD::STRICT_FADD, MVT::v2f64, Legal); |
1079 | setOperationAction(ISD::STRICT_FSUB, MVT::v2f64, Legal); |
1080 | setOperationAction(ISD::STRICT_FMUL, MVT::v2f64, Legal); |
1081 | setOperationAction(ISD::STRICT_FDIV, MVT::v2f64, Legal); |
1082 | setOperationAction(ISD::STRICT_FMA, MVT::v2f64, Legal); |
1083 | setOperationAction(ISD::STRICT_FSQRT, MVT::v2f64, Legal); |
1084 | setOperationAction(ISD::STRICT_FMAXNUM, MVT::v2f64, Legal); |
1085 | setOperationAction(ISD::STRICT_FMINNUM, MVT::v2f64, Legal); |
1086 | setOperationAction(ISD::STRICT_FRINT, MVT::v2f64, Legal); |
1087 | setOperationAction(ISD::STRICT_FFLOOR, MVT::v2f64, Legal); |
1088 | setOperationAction(ISD::STRICT_FCEIL, MVT::v2f64, Legal); |
1089 | setOperationAction(ISD::STRICT_FTRUNC, MVT::v2f64, Legal); |
1090 | setOperationAction(ISD::STRICT_FROUND, MVT::v2f64, Legal); |
1091 | |
1092 | addRegisterClass(MVT::v2i64, &PPC::VSRCRegClass); |
1093 | } |
1094 | |
1095 | if (Subtarget.hasP8Altivec()) { |
1096 | addRegisterClass(MVT::v2i64, &PPC::VRRCRegClass); |
1097 | addRegisterClass(MVT::v1i128, &PPC::VRRCRegClass); |
1098 | } |
1099 | |
1100 | if (Subtarget.hasP9Vector()) { |
1101 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom); |
1102 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom); |
1103 | |
1104 | // 128 bit shifts can be accomplished via 3 instructions for SHL and |
1105 | // SRL, but not for SRA because of the instructions available: |
1106 | // VS{RL} and VS{RL}O. |
1107 | setOperationAction(ISD::SHL, MVT::v1i128, Legal); |
1108 | setOperationAction(ISD::SRL, MVT::v1i128, Legal); |
1109 | setOperationAction(ISD::SRA, MVT::v1i128, Expand); |
1110 | |
1111 | addRegisterClass(MVT::f128, &PPC::VRRCRegClass); |
1112 | setOperationAction(ISD::FADD, MVT::f128, Legal); |
1113 | setOperationAction(ISD::FSUB, MVT::f128, Legal); |
1114 | setOperationAction(ISD::FDIV, MVT::f128, Legal); |
1115 | setOperationAction(ISD::FMUL, MVT::f128, Legal); |
1116 | setOperationAction(ISD::FP_EXTEND, MVT::f128, Legal); |
1117 | // No extending loads to f128 on PPC. |
1118 | for (MVT FPT : MVT::fp_valuetypes()) |
1119 | setLoadExtAction(ISD::EXTLOAD, MVT::f128, FPT, Expand); |
1120 | setOperationAction(ISD::FMA, MVT::f128, Legal); |
1121 | setCondCodeAction(ISD::SETULT, MVT::f128, Expand); |
1122 | setCondCodeAction(ISD::SETUGT, MVT::f128, Expand); |
1123 | setCondCodeAction(ISD::SETUEQ, MVT::f128, Expand); |
1124 | setCondCodeAction(ISD::SETOGE, MVT::f128, Expand); |
1125 | setCondCodeAction(ISD::SETOLE, MVT::f128, Expand); |
1126 | setCondCodeAction(ISD::SETONE, MVT::f128, Expand); |
1127 | |
1128 | setOperationAction(ISD::FTRUNC, MVT::f128, Legal); |
1129 | setOperationAction(ISD::FRINT, MVT::f128, Legal); |
1130 | setOperationAction(ISD::FFLOOR, MVT::f128, Legal); |
1131 | setOperationAction(ISD::FCEIL, MVT::f128, Legal); |
1132 | setOperationAction(ISD::FNEARBYINT, MVT::f128, Legal); |
1133 | setOperationAction(ISD::FROUND, MVT::f128, Legal); |
1134 | |
1135 | setOperationAction(ISD::SELECT, MVT::f128, Expand); |
1136 | setOperationAction(ISD::FP_ROUND, MVT::f64, Legal); |
1137 | setOperationAction(ISD::FP_ROUND, MVT::f32, Legal); |
1138 | setTruncStoreAction(MVT::f128, MVT::f64, Expand); |
1139 | setTruncStoreAction(MVT::f128, MVT::f32, Expand); |
1140 | setOperationAction(ISD::BITCAST, MVT::i128, Custom); |
1141 | // No implementation for these ops for PowerPC. |
1142 | setOperationAction(ISD::FSIN, MVT::f128, Expand); |
1143 | setOperationAction(ISD::FCOS, MVT::f128, Expand); |
1144 | setOperationAction(ISD::FPOW, MVT::f128, Expand); |
1145 | setOperationAction(ISD::FPOWI, MVT::f128, Expand); |
1146 | setOperationAction(ISD::FREM, MVT::f128, Expand); |
1147 | |
1148 | // Handle constrained floating-point operations of fp128 |
1149 | setOperationAction(ISD::STRICT_FADD, MVT::f128, Legal); |
1150 | setOperationAction(ISD::STRICT_FSUB, MVT::f128, Legal); |
1151 | setOperationAction(ISD::STRICT_FMUL, MVT::f128, Legal); |
1152 | setOperationAction(ISD::STRICT_FDIV, MVT::f128, Legal); |
1153 | setOperationAction(ISD::STRICT_FMA, MVT::f128, Legal); |
1154 | setOperationAction(ISD::STRICT_FSQRT, MVT::f128, Legal); |
1155 | setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Legal); |
1156 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Legal); |
1157 | setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal); |
1158 | setOperationAction(ISD::STRICT_FRINT, MVT::f128, Legal); |
1159 | setOperationAction(ISD::STRICT_FNEARBYINT, MVT::f128, Legal); |
1160 | setOperationAction(ISD::STRICT_FFLOOR, MVT::f128, Legal); |
1161 | setOperationAction(ISD::STRICT_FCEIL, MVT::f128, Legal); |
1162 | setOperationAction(ISD::STRICT_FTRUNC, MVT::f128, Legal); |
1163 | setOperationAction(ISD::STRICT_FROUND, MVT::f128, Legal); |
1164 | setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom); |
1165 | setOperationAction(ISD::BSWAP, MVT::v8i16, Legal); |
1166 | setOperationAction(ISD::BSWAP, MVT::v4i32, Legal); |
1167 | setOperationAction(ISD::BSWAP, MVT::v2i64, Legal); |
1168 | setOperationAction(ISD::BSWAP, MVT::v1i128, Legal); |
1169 | } |
1170 | |
1171 | if (Subtarget.hasP9Altivec()) { |
1172 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom); |
1173 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom); |
1174 | |
1175 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Legal); |
1176 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Legal); |
1177 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Legal); |
1178 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Legal); |
1179 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Legal); |
1180 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Legal); |
1181 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Legal); |
1182 | } |
1183 | } |
1184 | |
1185 | if (Subtarget.pairedVectorMemops()) { |
1186 | addRegisterClass(MVT::v256i1, &PPC::VSRpRCRegClass); |
1187 | setOperationAction(ISD::LOAD, MVT::v256i1, Custom); |
1188 | setOperationAction(ISD::STORE, MVT::v256i1, Custom); |
1189 | } |
1190 | if (Subtarget.hasMMA()) { |
1191 | addRegisterClass(MVT::v512i1, &PPC::UACCRCRegClass); |
1192 | setOperationAction(ISD::LOAD, MVT::v512i1, Custom); |
1193 | setOperationAction(ISD::STORE, MVT::v512i1, Custom); |
1194 | setOperationAction(ISD::BUILD_VECTOR, MVT::v512i1, Custom); |
1195 | } |
1196 | |
1197 | if (Subtarget.has64BitSupport()) |
1198 | setOperationAction(ISD::PREFETCH, MVT::Other, Legal); |
1199 | |
1200 | if (Subtarget.isISA3_1()) |
1201 | setOperationAction(ISD::SRA, MVT::v1i128, Legal); |
1202 | |
1203 | setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, isPPC64 ? Legal : Custom); |
1204 | |
1205 | if (!isPPC64) { |
1206 | setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand); |
1207 | setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand); |
1208 | } |
1209 | |
1210 | setBooleanContents(ZeroOrOneBooleanContent); |
1211 | |
1212 | if (Subtarget.hasAltivec()) { |
1213 | // Altivec instructions set fields to all zeros or all ones. |
1214 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
1215 | } |
1216 | |
1217 | if (!isPPC64) { |
1218 | // These libcalls are not available in 32-bit. |
1219 | setLibcallName(RTLIB::SHL_I128, nullptr); |
1220 | setLibcallName(RTLIB::SRL_I128, nullptr); |
1221 | setLibcallName(RTLIB::SRA_I128, nullptr); |
1222 | } |
1223 | |
1224 | if (!isPPC64) |
1225 | setMaxAtomicSizeInBitsSupported(32); |
1226 | |
1227 | setStackPointerRegisterToSaveRestore(isPPC64 ? PPC::X1 : PPC::R1); |
1228 | |
1229 | // We have target-specific dag combine patterns for the following nodes: |
1230 | setTargetDAGCombine(ISD::ADD); |
1231 | setTargetDAGCombine(ISD::SHL); |
1232 | setTargetDAGCombine(ISD::SRA); |
1233 | setTargetDAGCombine(ISD::SRL); |
1234 | setTargetDAGCombine(ISD::MUL); |
1235 | setTargetDAGCombine(ISD::FMA); |
1236 | setTargetDAGCombine(ISD::SINT_TO_FP); |
1237 | setTargetDAGCombine(ISD::BUILD_VECTOR); |
1238 | if (Subtarget.hasFPCVT()) |
1239 | setTargetDAGCombine(ISD::UINT_TO_FP); |
1240 | setTargetDAGCombine(ISD::LOAD); |
1241 | setTargetDAGCombine(ISD::STORE); |
1242 | setTargetDAGCombine(ISD::BR_CC); |
1243 | if (Subtarget.useCRBits()) |
1244 | setTargetDAGCombine(ISD::BRCOND); |
1245 | setTargetDAGCombine(ISD::BSWAP); |
1246 | setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); |
1247 | setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN); |
1248 | setTargetDAGCombine(ISD::INTRINSIC_VOID); |
1249 | |
1250 | setTargetDAGCombine(ISD::SIGN_EXTEND); |
1251 | setTargetDAGCombine(ISD::ZERO_EXTEND); |
1252 | setTargetDAGCombine(ISD::ANY_EXTEND); |
1253 | |
1254 | setTargetDAGCombine(ISD::TRUNCATE); |
1255 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
1256 | |
1257 | |
1258 | if (Subtarget.useCRBits()) { |
1259 | setTargetDAGCombine(ISD::TRUNCATE); |
1260 | setTargetDAGCombine(ISD::SETCC); |
1261 | setTargetDAGCombine(ISD::SELECT_CC); |
1262 | } |
1263 | |
1264 | if (Subtarget.hasP9Altivec()) { |
1265 | setTargetDAGCombine(ISD::ABS); |
1266 | setTargetDAGCombine(ISD::VSELECT); |
1267 | } |
1268 | |
1269 | setLibcallName(RTLIB::LOG_F128, "logf128"); |
1270 | setLibcallName(RTLIB::LOG2_F128, "log2f128"); |
1271 | setLibcallName(RTLIB::LOG10_F128, "log10f128"); |
1272 | setLibcallName(RTLIB::EXP_F128, "expf128"); |
1273 | setLibcallName(RTLIB::EXP2_F128, "exp2f128"); |
1274 | setLibcallName(RTLIB::SIN_F128, "sinf128"); |
1275 | setLibcallName(RTLIB::COS_F128, "cosf128"); |
1276 | setLibcallName(RTLIB::POW_F128, "powf128"); |
1277 | setLibcallName(RTLIB::FMIN_F128, "fminf128"); |
1278 | setLibcallName(RTLIB::FMAX_F128, "fmaxf128"); |
1279 | setLibcallName(RTLIB::POWI_F128, "__powikf2"); |
1280 | setLibcallName(RTLIB::REM_F128, "fmodf128"); |
1281 | |
1282 | // With 32 condition bits, we don't need to sink (and duplicate) compares |
1283 | // aggressively in CodeGenPrep. |
1284 | if (Subtarget.useCRBits()) { |
1285 | setHasMultipleConditionRegisters(); |
1286 | setJumpIsExpensive(); |
1287 | } |
1288 | |
1289 | setMinFunctionAlignment(Align(4)); |
1290 | |
1291 | switch (Subtarget.getCPUDirective()) { |
1292 | default: break; |
1293 | case PPC::DIR_970: |
1294 | case PPC::DIR_A2: |
1295 | case PPC::DIR_E500: |
1296 | case PPC::DIR_E500mc: |
1297 | case PPC::DIR_E5500: |
1298 | case PPC::DIR_PWR4: |
1299 | case PPC::DIR_PWR5: |
1300 | case PPC::DIR_PWR5X: |
1301 | case PPC::DIR_PWR6: |
1302 | case PPC::DIR_PWR6X: |
1303 | case PPC::DIR_PWR7: |
1304 | case PPC::DIR_PWR8: |
1305 | case PPC::DIR_PWR9: |
1306 | case PPC::DIR_PWR10: |
1307 | case PPC::DIR_PWR_FUTURE: |
1308 | setPrefLoopAlignment(Align(16)); |
1309 | setPrefFunctionAlignment(Align(16)); |
1310 | break; |
1311 | } |
1312 | |
1313 | if (Subtarget.enableMachineScheduler()) |
1314 | setSchedulingPreference(Sched::Source); |
1315 | else |
1316 | setSchedulingPreference(Sched::Hybrid); |
1317 | |
1318 | computeRegisterProperties(STI.getRegisterInfo()); |
1319 | |
1320 | // The Freescale cores do better with aggressive inlining of memcpy and |
1321 | // friends. GCC uses same threshold of 128 bytes (= 32 word stores). |
1322 | if (Subtarget.getCPUDirective() == PPC::DIR_E500mc || |
1323 | Subtarget.getCPUDirective() == PPC::DIR_E5500) { |
1324 | MaxStoresPerMemset = 32; |
1325 | MaxStoresPerMemsetOptSize = 16; |
1326 | MaxStoresPerMemcpy = 32; |
1327 | MaxStoresPerMemcpyOptSize = 8; |
1328 | MaxStoresPerMemmove = 32; |
1329 | MaxStoresPerMemmoveOptSize = 8; |
1330 | } else if (Subtarget.getCPUDirective() == PPC::DIR_A2) { |
1331 | // The A2 also benefits from (very) aggressive inlining of memcpy and |
1332 | // friends. The overhead of a the function call, even when warm, can be |
1333 | // over one hundred cycles. |
1334 | MaxStoresPerMemset = 128; |
1335 | MaxStoresPerMemcpy = 128; |
1336 | MaxStoresPerMemmove = 128; |
1337 | MaxLoadsPerMemcmp = 128; |
1338 | } else { |
1339 | MaxLoadsPerMemcmp = 8; |
1340 | MaxLoadsPerMemcmpOptSize = 4; |
1341 | } |
1342 | |
1343 | IsStrictFPEnabled = true; |
1344 | |
1345 | // Let the subtarget (CPU) decide if a predictable select is more expensive |
1346 | // than the corresponding branch. This information is used in CGP to decide |
1347 | // when to convert selects into branches. |
1348 | PredictableSelectIsExpensive = Subtarget.isPredictableSelectIsExpensive(); |
1349 | } |
1350 | |
1351 | /// getMaxByValAlign - Helper for getByValTypeAlignment to determine |
1352 | /// the desired ByVal argument alignment. |
1353 | static void getMaxByValAlign(Type *Ty, Align &MaxAlign, Align MaxMaxAlign) { |
1354 | if (MaxAlign == MaxMaxAlign) |
1355 | return; |
1356 | if (VectorType *VTy = dyn_cast<VectorType>(Ty)) { |
1357 | if (MaxMaxAlign >= 32 && |
1358 | VTy->getPrimitiveSizeInBits().getFixedSize() >= 256) |
1359 | MaxAlign = Align(32); |
1360 | else if (VTy->getPrimitiveSizeInBits().getFixedSize() >= 128 && |
1361 | MaxAlign < 16) |
1362 | MaxAlign = Align(16); |
1363 | } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) { |
1364 | Align EltAlign; |
1365 | getMaxByValAlign(ATy->getElementType(), EltAlign, MaxMaxAlign); |
1366 | if (EltAlign > MaxAlign) |
1367 | MaxAlign = EltAlign; |
1368 | } else if (StructType *STy = dyn_cast<StructType>(Ty)) { |
1369 | for (auto *EltTy : STy->elements()) { |
1370 | Align EltAlign; |
1371 | getMaxByValAlign(EltTy, EltAlign, MaxMaxAlign); |
1372 | if (EltAlign > MaxAlign) |
1373 | MaxAlign = EltAlign; |
1374 | if (MaxAlign == MaxMaxAlign) |
1375 | break; |
1376 | } |
1377 | } |
1378 | } |
1379 | |
1380 | /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate |
1381 | /// function arguments in the caller parameter area. |
1382 | unsigned PPCTargetLowering::getByValTypeAlignment(Type *Ty, |
1383 | const DataLayout &DL) const { |
1384 | // 16byte and wider vectors are passed on 16byte boundary. |
1385 | // The rest is 8 on PPC64 and 4 on PPC32 boundary. |
1386 | Align Alignment = Subtarget.isPPC64() ? Align(8) : Align(4); |
1387 | if (Subtarget.hasAltivec()) |
1388 | getMaxByValAlign(Ty, Alignment, Align(16)); |
1389 | return Alignment.value(); |
1390 | } |
1391 | |
1392 | bool PPCTargetLowering::useSoftFloat() const { |
1393 | return Subtarget.useSoftFloat(); |
1394 | } |
1395 | |
1396 | bool PPCTargetLowering::hasSPE() const { |
1397 | return Subtarget.hasSPE(); |
1398 | } |
1399 | |
1400 | bool PPCTargetLowering::preferIncOfAddToSubOfNot(EVT VT) const { |
1401 | return VT.isScalarInteger(); |
1402 | } |
1403 | |
1404 | const char *PPCTargetLowering::getTargetNodeName(unsigned Opcode) const { |
1405 | switch ((PPCISD::NodeType)Opcode) { |
1406 | case PPCISD::FIRST_NUMBER: break; |
1407 | case PPCISD::FSEL: return "PPCISD::FSEL"; |
1408 | case PPCISD::XSMAXCDP: return "PPCISD::XSMAXCDP"; |
1409 | case PPCISD::XSMINCDP: return "PPCISD::XSMINCDP"; |
1410 | case PPCISD::FCFID: return "PPCISD::FCFID"; |
1411 | case PPCISD::FCFIDU: return "PPCISD::FCFIDU"; |
1412 | case PPCISD::FCFIDS: return "PPCISD::FCFIDS"; |
1413 | case PPCISD::FCFIDUS: return "PPCISD::FCFIDUS"; |
1414 | case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ"; |
1415 | case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ"; |
1416 | case PPCISD::FCTIDUZ: return "PPCISD::FCTIDUZ"; |
1417 | case PPCISD::FCTIWUZ: return "PPCISD::FCTIWUZ"; |
1418 | case PPCISD::FP_TO_UINT_IN_VSR: |
1419 | return "PPCISD::FP_TO_UINT_IN_VSR,"; |
1420 | case PPCISD::FP_TO_SINT_IN_VSR: |
1421 | return "PPCISD::FP_TO_SINT_IN_VSR"; |
1422 | case PPCISD::FRE: return "PPCISD::FRE"; |
1423 | case PPCISD::FRSQRTE: return "PPCISD::FRSQRTE"; |
1424 | case PPCISD::STFIWX: return "PPCISD::STFIWX"; |
1425 | case PPCISD::VPERM: return "PPCISD::VPERM"; |
1426 | case PPCISD::XXSPLT: return "PPCISD::XXSPLT"; |
1427 | case PPCISD::XXSPLTI_SP_TO_DP: |
1428 | return "PPCISD::XXSPLTI_SP_TO_DP"; |
1429 | case PPCISD::XXSPLTI32DX: |
1430 | return "PPCISD::XXSPLTI32DX"; |
1431 | case PPCISD::VECINSERT: return "PPCISD::VECINSERT"; |
1432 | case PPCISD::XXPERMDI: return "PPCISD::XXPERMDI"; |
1433 | case PPCISD::VECSHL: return "PPCISD::VECSHL"; |
1434 | case PPCISD::CMPB: return "PPCISD::CMPB"; |
1435 | case PPCISD::Hi: return "PPCISD::Hi"; |
1436 | case PPCISD::Lo: return "PPCISD::Lo"; |
1437 | case PPCISD::TOC_ENTRY: return "PPCISD::TOC_ENTRY"; |
1438 | case PPCISD::ATOMIC_CMP_SWAP_8: return "PPCISD::ATOMIC_CMP_SWAP_8"; |
1439 | case PPCISD::ATOMIC_CMP_SWAP_16: return "PPCISD::ATOMIC_CMP_SWAP_16"; |
1440 | case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC"; |
1441 | case PPCISD::DYNAREAOFFSET: return "PPCISD::DYNAREAOFFSET"; |
1442 | case PPCISD::PROBED_ALLOCA: return "PPCISD::PROBED_ALLOCA"; |
1443 | case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg"; |
1444 | case PPCISD::SRL: return "PPCISD::SRL"; |
1445 | case PPCISD::SRA: return "PPCISD::SRA"; |
1446 | case PPCISD::SHL: return "PPCISD::SHL"; |
1447 | case PPCISD::SRA_ADDZE: return "PPCISD::SRA_ADDZE"; |
1448 | case PPCISD::CALL: return "PPCISD::CALL"; |
1449 | case PPCISD::CALL_NOP: return "PPCISD::CALL_NOP"; |
1450 | case PPCISD::CALL_NOTOC: return "PPCISD::CALL_NOTOC"; |
1451 | case PPCISD::MTCTR: return "PPCISD::MTCTR"; |
1452 | case PPCISD::BCTRL: return "PPCISD::BCTRL"; |
1453 | case PPCISD::BCTRL_LOAD_TOC: return "PPCISD::BCTRL_LOAD_TOC"; |
1454 | case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG"; |
1455 | case PPCISD::READ_TIME_BASE: return "PPCISD::READ_TIME_BASE"; |
1456 | case PPCISD::EH_SJLJ_SETJMP: return "PPCISD::EH_SJLJ_SETJMP"; |
1457 | case PPCISD::EH_SJLJ_LONGJMP: return "PPCISD::EH_SJLJ_LONGJMP"; |
1458 | case PPCISD::MFOCRF: return "PPCISD::MFOCRF"; |
1459 | case PPCISD::MFVSR: return "PPCISD::MFVSR"; |
1460 | case PPCISD::MTVSRA: return "PPCISD::MTVSRA"; |
1461 | case PPCISD::MTVSRZ: return "PPCISD::MTVSRZ"; |
1462 | case PPCISD::SINT_VEC_TO_FP: return "PPCISD::SINT_VEC_TO_FP"; |
1463 | case PPCISD::UINT_VEC_TO_FP: return "PPCISD::UINT_VEC_TO_FP"; |
1464 | case PPCISD::SCALAR_TO_VECTOR_PERMUTED: |
1465 | return "PPCISD::SCALAR_TO_VECTOR_PERMUTED"; |
1466 | case PPCISD::ANDI_rec_1_EQ_BIT: |
1467 | return "PPCISD::ANDI_rec_1_EQ_BIT"; |
1468 | case PPCISD::ANDI_rec_1_GT_BIT: |
1469 | return "PPCISD::ANDI_rec_1_GT_BIT"; |
1470 | case PPCISD::VCMP: return "PPCISD::VCMP"; |
1471 | case PPCISD::VCMPo: return "PPCISD::VCMPo"; |
1472 | case PPCISD::LBRX: return "PPCISD::LBRX"; |
1473 | case PPCISD::STBRX: return "PPCISD::STBRX"; |
1474 | case PPCISD::LFIWAX: return "PPCISD::LFIWAX"; |
1475 | case PPCISD::LFIWZX: return "PPCISD::LFIWZX"; |
1476 | case PPCISD::LXSIZX: return "PPCISD::LXSIZX"; |
1477 | case PPCISD::STXSIX: return "PPCISD::STXSIX"; |
1478 | case PPCISD::VEXTS: return "PPCISD::VEXTS"; |
1479 | case PPCISD::LXVD2X: return "PPCISD::LXVD2X"; |
1480 | case PPCISD::STXVD2X: return "PPCISD::STXVD2X"; |
1481 | case PPCISD::LOAD_VEC_BE: return "PPCISD::LOAD_VEC_BE"; |
1482 | case PPCISD::STORE_VEC_BE: return "PPCISD::STORE_VEC_BE"; |
1483 | case PPCISD::ST_VSR_SCAL_INT: |
1484 | return "PPCISD::ST_VSR_SCAL_INT"; |
1485 | case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH"; |
1486 | case PPCISD::BDNZ: return "PPCISD::BDNZ"; |
1487 | case PPCISD::BDZ: return "PPCISD::BDZ"; |
1488 | case PPCISD::MFFS: return "PPCISD::MFFS"; |
1489 | case PPCISD::FADDRTZ: return "PPCISD::FADDRTZ"; |
1490 | case PPCISD::TC_RETURN: return "PPCISD::TC_RETURN"; |
1491 | case PPCISD::CR6SET: return "PPCISD::CR6SET"; |
1492 | case PPCISD::CR6UNSET: return "PPCISD::CR6UNSET"; |
1493 | case PPCISD::PPC32_GOT: return "PPCISD::PPC32_GOT"; |
1494 | case PPCISD::PPC32_PICGOT: return "PPCISD::PPC32_PICGOT"; |
1495 | case PPCISD::ADDIS_GOT_TPREL_HA: return "PPCISD::ADDIS_GOT_TPREL_HA"; |
1496 | case PPCISD::LD_GOT_TPREL_L: return "PPCISD::LD_GOT_TPREL_L"; |
1497 | case PPCISD::ADD_TLS: return "PPCISD::ADD_TLS"; |
1498 | case PPCISD::ADDIS_TLSGD_HA: return "PPCISD::ADDIS_TLSGD_HA"; |
1499 | case PPCISD::ADDI_TLSGD_L: return "PPCISD::ADDI_TLSGD_L"; |
1500 | case PPCISD::GET_TLS_ADDR: return "PPCISD::GET_TLS_ADDR"; |
1501 | case PPCISD::ADDI_TLSGD_L_ADDR: return "PPCISD::ADDI_TLSGD_L_ADDR"; |
1502 | case PPCISD::ADDIS_TLSLD_HA: return "PPCISD::ADDIS_TLSLD_HA"; |
1503 | case PPCISD::ADDI_TLSLD_L: return "PPCISD::ADDI_TLSLD_L"; |
1504 | case PPCISD::GET_TLSLD_ADDR: return "PPCISD::GET_TLSLD_ADDR"; |
1505 | case PPCISD::ADDI_TLSLD_L_ADDR: return "PPCISD::ADDI_TLSLD_L_ADDR"; |
1506 | case PPCISD::ADDIS_DTPREL_HA: return "PPCISD::ADDIS_DTPREL_HA"; |
1507 | case PPCISD::ADDI_DTPREL_L: return "PPCISD::ADDI_DTPREL_L"; |
1508 | case PPCISD::PADDI_DTPREL: |
1509 | return "PPCISD::PADDI_DTPREL"; |
1510 | case PPCISD::VADD_SPLAT: return "PPCISD::VADD_SPLAT"; |
1511 | case PPCISD::SC: return "PPCISD::SC"; |
1512 | case PPCISD::CLRBHRB: return "PPCISD::CLRBHRB"; |
1513 | case PPCISD::MFBHRBE: return "PPCISD::MFBHRBE"; |
1514 | case PPCISD::RFEBB: return "PPCISD::RFEBB"; |
1515 | case PPCISD::XXSWAPD: return "PPCISD::XXSWAPD"; |
1516 | case PPCISD::SWAP_NO_CHAIN: return "PPCISD::SWAP_NO_CHAIN"; |
1517 | case PPCISD::VABSD: return "PPCISD::VABSD"; |
1518 | case PPCISD::BUILD_FP128: return "PPCISD::BUILD_FP128"; |
1519 | case PPCISD::BUILD_SPE64: return "PPCISD::BUILD_SPE64"; |
1520 | case PPCISD::EXTRACT_SPE: return "PPCISD::EXTRACT_SPE"; |
1521 | case PPCISD::EXTSWSLI: return "PPCISD::EXTSWSLI"; |
1522 | case PPCISD::LD_VSX_LH: return "PPCISD::LD_VSX_LH"; |
1523 | case PPCISD::FP_EXTEND_HALF: return "PPCISD::FP_EXTEND_HALF"; |
1524 | case PPCISD::MAT_PCREL_ADDR: return "PPCISD::MAT_PCREL_ADDR"; |
1525 | case PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR: |
1526 | return "PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR"; |
1527 | case PPCISD::TLS_LOCAL_EXEC_MAT_ADDR: |
1528 | return "PPCISD::TLS_LOCAL_EXEC_MAT_ADDR"; |
1529 | case PPCISD::ACC_BUILD: return "PPCISD::ACC_BUILD"; |
1530 | case PPCISD::PAIR_BUILD: return "PPCISD::PAIR_BUILD"; |
1531 | case PPCISD::EXTRACT_VSX_REG: return "PPCISD::EXTRACT_VSX_REG"; |
1532 | case PPCISD::XXMFACC: return "PPCISD::XXMFACC"; |
1533 | case PPCISD::LD_SPLAT: return "PPCISD::LD_SPLAT"; |
1534 | case PPCISD::FNMSUB: return "PPCISD::FNMSUB"; |
1535 | case PPCISD::STRICT_FADDRTZ: |
1536 | return "PPCISD::STRICT_FADDRTZ"; |
1537 | case PPCISD::STRICT_FCTIDZ: |
1538 | return "PPCISD::STRICT_FCTIDZ"; |
1539 | case PPCISD::STRICT_FCTIWZ: |
1540 | return "PPCISD::STRICT_FCTIWZ"; |
1541 | case PPCISD::STRICT_FCTIDUZ: |
1542 | return "PPCISD::STRICT_FCTIDUZ"; |
1543 | case PPCISD::STRICT_FCTIWUZ: |
1544 | return "PPCISD::STRICT_FCTIWUZ"; |
1545 | case PPCISD::STRICT_FCFID: |
1546 | return "PPCISD::STRICT_FCFID"; |
1547 | case PPCISD::STRICT_FCFIDU: |
1548 | return "PPCISD::STRICT_FCFIDU"; |
1549 | case PPCISD::STRICT_FCFIDS: |
1550 | return "PPCISD::STRICT_FCFIDS"; |
1551 | case PPCISD::STRICT_FCFIDUS: |
1552 | return "PPCISD::STRICT_FCFIDUS"; |
1553 | case PPCISD::LXVRZX: return "PPCISD::LXVRZX"; |
1554 | } |
1555 | return nullptr; |
1556 | } |
1557 | |
1558 | EVT PPCTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &C, |
1559 | EVT VT) const { |
1560 | if (!VT.isVector()) |
1561 | return Subtarget.useCRBits() ? MVT::i1 : MVT::i32; |
1562 | |
1563 | return VT.changeVectorElementTypeToInteger(); |
1564 | } |
1565 | |
1566 | bool PPCTargetLowering::enableAggressiveFMAFusion(EVT VT) const { |
1567 | assert(VT.isFloatingPoint() && "Non-floating-point FMA?")((VT.isFloatingPoint() && "Non-floating-point FMA?") ? static_cast<void> (0) : __assert_fail ("VT.isFloatingPoint() && \"Non-floating-point FMA?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1567, __PRETTY_FUNCTION__)); |
1568 | return true; |
1569 | } |
1570 | |
1571 | //===----------------------------------------------------------------------===// |
1572 | // Node matching predicates, for use by the tblgen matching code. |
1573 | //===----------------------------------------------------------------------===// |
1574 | |
1575 | /// isFloatingPointZero - Return true if this is 0.0 or -0.0. |
1576 | static bool isFloatingPointZero(SDValue Op) { |
1577 | if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) |
1578 | return CFP->getValueAPF().isZero(); |
1579 | else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { |
1580 | // Maybe this has already been legalized into the constant pool? |
1581 | if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1))) |
1582 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) |
1583 | return CFP->getValueAPF().isZero(); |
1584 | } |
1585 | return false; |
1586 | } |
1587 | |
1588 | /// isConstantOrUndef - Op is either an undef node or a ConstantSDNode. Return |
1589 | /// true if Op is undef or if it matches the specified value. |
1590 | static bool isConstantOrUndef(int Op, int Val) { |
1591 | return Op < 0 || Op == Val; |
1592 | } |
1593 | |
1594 | /// isVPKUHUMShuffleMask - Return true if this is the shuffle mask for a |
1595 | /// VPKUHUM instruction. |
1596 | /// The ShuffleKind distinguishes between big-endian operations with |
1597 | /// two different inputs (0), either-endian operations with two identical |
1598 | /// inputs (1), and little-endian operations with two different inputs (2). |
1599 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1600 | bool PPC::isVPKUHUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1601 | SelectionDAG &DAG) { |
1602 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1603 | if (ShuffleKind == 0) { |
1604 | if (IsLE) |
1605 | return false; |
1606 | for (unsigned i = 0; i != 16; ++i) |
1607 | if (!isConstantOrUndef(N->getMaskElt(i), i*2+1)) |
1608 | return false; |
1609 | } else if (ShuffleKind == 2) { |
1610 | if (!IsLE) |
1611 | return false; |
1612 | for (unsigned i = 0; i != 16; ++i) |
1613 | if (!isConstantOrUndef(N->getMaskElt(i), i*2)) |
1614 | return false; |
1615 | } else if (ShuffleKind == 1) { |
1616 | unsigned j = IsLE ? 0 : 1; |
1617 | for (unsigned i = 0; i != 8; ++i) |
1618 | if (!isConstantOrUndef(N->getMaskElt(i), i*2+j) || |
1619 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j)) |
1620 | return false; |
1621 | } |
1622 | return true; |
1623 | } |
1624 | |
1625 | /// isVPKUWUMShuffleMask - Return true if this is the shuffle mask for a |
1626 | /// VPKUWUM instruction. |
1627 | /// The ShuffleKind distinguishes between big-endian operations with |
1628 | /// two different inputs (0), either-endian operations with two identical |
1629 | /// inputs (1), and little-endian operations with two different inputs (2). |
1630 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1631 | bool PPC::isVPKUWUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1632 | SelectionDAG &DAG) { |
1633 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1634 | if (ShuffleKind == 0) { |
1635 | if (IsLE) |
1636 | return false; |
1637 | for (unsigned i = 0; i != 16; i += 2) |
1638 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+2) || |
1639 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+3)) |
1640 | return false; |
1641 | } else if (ShuffleKind == 2) { |
1642 | if (!IsLE) |
1643 | return false; |
1644 | for (unsigned i = 0; i != 16; i += 2) |
1645 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2) || |
1646 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+1)) |
1647 | return false; |
1648 | } else if (ShuffleKind == 1) { |
1649 | unsigned j = IsLE ? 0 : 2; |
1650 | for (unsigned i = 0; i != 8; i += 2) |
1651 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) || |
1652 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) || |
1653 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) || |
1654 | !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1)) |
1655 | return false; |
1656 | } |
1657 | return true; |
1658 | } |
1659 | |
1660 | /// isVPKUDUMShuffleMask - Return true if this is the shuffle mask for a |
1661 | /// VPKUDUM instruction, AND the VPKUDUM instruction exists for the |
1662 | /// current subtarget. |
1663 | /// |
1664 | /// The ShuffleKind distinguishes between big-endian operations with |
1665 | /// two different inputs (0), either-endian operations with two identical |
1666 | /// inputs (1), and little-endian operations with two different inputs (2). |
1667 | /// For the latter, the input operands are swapped (see PPCInstrAltivec.td). |
1668 | bool PPC::isVPKUDUMShuffleMask(ShuffleVectorSDNode *N, unsigned ShuffleKind, |
1669 | SelectionDAG &DAG) { |
1670 | const PPCSubtarget& Subtarget = |
1671 | static_cast<const PPCSubtarget&>(DAG.getSubtarget()); |
1672 | if (!Subtarget.hasP8Vector()) |
1673 | return false; |
1674 | |
1675 | bool IsLE = DAG.getDataLayout().isLittleEndian(); |
1676 | if (ShuffleKind == 0) { |
1677 | if (IsLE) |
1678 | return false; |
1679 | for (unsigned i = 0; i != 16; i += 4) |
1680 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+4) || |
1681 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+5) || |
1682 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+6) || |
1683 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+7)) |
1684 | return false; |
1685 | } else if (ShuffleKind == 2) { |
1686 | if (!IsLE) |
1687 | return false; |
1688 | for (unsigned i = 0; i != 16; i += 4) |
1689 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2) || |
1690 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+1) || |
1691 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+2) || |
1692 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+3)) |
1693 | return false; |
1694 | } else if (ShuffleKind == 1) { |
1695 | unsigned j = IsLE ? 0 : 4; |
1696 | for (unsigned i = 0; i != 8; i += 4) |
1697 | if (!isConstantOrUndef(N->getMaskElt(i ), i*2+j) || |
1698 | !isConstantOrUndef(N->getMaskElt(i+1), i*2+j+1) || |
1699 | !isConstantOrUndef(N->getMaskElt(i+2), i*2+j+2) || |
1700 | !isConstantOrUndef(N->getMaskElt(i+3), i*2+j+3) || |
1701 | !isConstantOrUndef(N->getMaskElt(i+8), i*2+j) || |
1702 | !isConstantOrUndef(N->getMaskElt(i+9), i*2+j+1) || |
1703 | !isConstantOrUndef(N->getMaskElt(i+10), i*2+j+2) || |
1704 | !isConstantOrUndef(N->getMaskElt(i+11), i*2+j+3)) |
1705 | return false; |
1706 | } |
1707 | return true; |
1708 | } |
1709 | |
1710 | /// isVMerge - Common function, used to match vmrg* shuffles. |
1711 | /// |
1712 | static bool isVMerge(ShuffleVectorSDNode *N, unsigned UnitSize, |
1713 | unsigned LHSStart, unsigned RHSStart) { |
1714 | if (N->getValueType(0) != MVT::v16i8) |
1715 | return false; |
1716 | assert((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) &&(((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && "Unsupported merge size!") ? static_cast<void> (0) : __assert_fail ("(UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && \"Unsupported merge size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1717, __PRETTY_FUNCTION__)) |
1717 | "Unsupported merge size!")(((UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && "Unsupported merge size!") ? static_cast<void> (0) : __assert_fail ("(UnitSize == 1 || UnitSize == 2 || UnitSize == 4) && \"Unsupported merge size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1717, __PRETTY_FUNCTION__)); |
1718 | |
1719 | for (unsigned i = 0; i != 8/UnitSize; ++i) // Step over units |
1720 | for (unsigned j = 0; j != UnitSize; ++j) { // Step over bytes within unit |
1721 | if (!isConstantOrUndef(N->getMaskElt(i*UnitSize*2+j), |
1722 | LHSStart+j+i*UnitSize) || |
1723 | !isConstantOrUndef(N->getMaskElt(i*UnitSize*2+UnitSize+j), |
1724 | RHSStart+j+i*UnitSize)) |
1725 | return false; |
1726 | } |
1727 | return true; |
1728 | } |
1729 | |
1730 | /// isVMRGLShuffleMask - Return true if this is a shuffle mask suitable for |
1731 | /// a VMRGL* instruction with the specified unit size (1,2 or 4 bytes). |
1732 | /// The ShuffleKind distinguishes between big-endian merges with two |
1733 | /// different inputs (0), either-endian merges with two identical inputs (1), |
1734 | /// and little-endian merges with two different inputs (2). For the latter, |
1735 | /// the input operands are swapped (see PPCInstrAltivec.td). |
1736 | bool PPC::isVMRGLShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
1737 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1738 | if (DAG.getDataLayout().isLittleEndian()) { |
1739 | if (ShuffleKind == 1) // unary |
1740 | return isVMerge(N, UnitSize, 0, 0); |
1741 | else if (ShuffleKind == 2) // swapped |
1742 | return isVMerge(N, UnitSize, 0, 16); |
1743 | else |
1744 | return false; |
1745 | } else { |
1746 | if (ShuffleKind == 1) // unary |
1747 | return isVMerge(N, UnitSize, 8, 8); |
1748 | else if (ShuffleKind == 0) // normal |
1749 | return isVMerge(N, UnitSize, 8, 24); |
1750 | else |
1751 | return false; |
1752 | } |
1753 | } |
1754 | |
1755 | /// isVMRGHShuffleMask - Return true if this is a shuffle mask suitable for |
1756 | /// a VMRGH* instruction with the specified unit size (1,2 or 4 bytes). |
1757 | /// The ShuffleKind distinguishes between big-endian merges with two |
1758 | /// different inputs (0), either-endian merges with two identical inputs (1), |
1759 | /// and little-endian merges with two different inputs (2). For the latter, |
1760 | /// the input operands are swapped (see PPCInstrAltivec.td). |
1761 | bool PPC::isVMRGHShuffleMask(ShuffleVectorSDNode *N, unsigned UnitSize, |
1762 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1763 | if (DAG.getDataLayout().isLittleEndian()) { |
1764 | if (ShuffleKind == 1) // unary |
1765 | return isVMerge(N, UnitSize, 8, 8); |
1766 | else if (ShuffleKind == 2) // swapped |
1767 | return isVMerge(N, UnitSize, 8, 24); |
1768 | else |
1769 | return false; |
1770 | } else { |
1771 | if (ShuffleKind == 1) // unary |
1772 | return isVMerge(N, UnitSize, 0, 0); |
1773 | else if (ShuffleKind == 0) // normal |
1774 | return isVMerge(N, UnitSize, 0, 16); |
1775 | else |
1776 | return false; |
1777 | } |
1778 | } |
1779 | |
1780 | /** |
1781 | * Common function used to match vmrgew and vmrgow shuffles |
1782 | * |
1783 | * The indexOffset determines whether to look for even or odd words in |
1784 | * the shuffle mask. This is based on the of the endianness of the target |
1785 | * machine. |
1786 | * - Little Endian: |
1787 | * - Use offset of 0 to check for odd elements |
1788 | * - Use offset of 4 to check for even elements |
1789 | * - Big Endian: |
1790 | * - Use offset of 0 to check for even elements |
1791 | * - Use offset of 4 to check for odd elements |
1792 | * A detailed description of the vector element ordering for little endian and |
1793 | * big endian can be found at |
1794 | * http://www.ibm.com/developerworks/library/l-ibm-xl-c-cpp-compiler/index.html |
1795 | * Targeting your applications - what little endian and big endian IBM XL C/C++ |
1796 | * compiler differences mean to you |
1797 | * |
1798 | * The mask to the shuffle vector instruction specifies the indices of the |
1799 | * elements from the two input vectors to place in the result. The elements are |
1800 | * numbered in array-access order, starting with the first vector. These vectors |
1801 | * are always of type v16i8, thus each vector will contain 16 elements of size |
1802 | * 8. More info on the shuffle vector can be found in the |
1803 | * http://llvm.org/docs/LangRef.html#shufflevector-instruction |
1804 | * Language Reference. |
1805 | * |
1806 | * The RHSStartValue indicates whether the same input vectors are used (unary) |
1807 | * or two different input vectors are used, based on the following: |
1808 | * - If the instruction uses the same vector for both inputs, the range of the |
1809 | * indices will be 0 to 15. In this case, the RHSStart value passed should |
1810 | * be 0. |
1811 | * - If the instruction has two different vectors then the range of the |
1812 | * indices will be 0 to 31. In this case, the RHSStart value passed should |
1813 | * be 16 (indices 0-15 specify elements in the first vector while indices 16 |
1814 | * to 31 specify elements in the second vector). |
1815 | * |
1816 | * \param[in] N The shuffle vector SD Node to analyze |
1817 | * \param[in] IndexOffset Specifies whether to look for even or odd elements |
1818 | * \param[in] RHSStartValue Specifies the starting index for the righthand input |
1819 | * vector to the shuffle_vector instruction |
1820 | * \return true iff this shuffle vector represents an even or odd word merge |
1821 | */ |
1822 | static bool isVMerge(ShuffleVectorSDNode *N, unsigned IndexOffset, |
1823 | unsigned RHSStartValue) { |
1824 | if (N->getValueType(0) != MVT::v16i8) |
1825 | return false; |
1826 | |
1827 | for (unsigned i = 0; i < 2; ++i) |
1828 | for (unsigned j = 0; j < 4; ++j) |
1829 | if (!isConstantOrUndef(N->getMaskElt(i*4+j), |
1830 | i*RHSStartValue+j+IndexOffset) || |
1831 | !isConstantOrUndef(N->getMaskElt(i*4+j+8), |
1832 | i*RHSStartValue+j+IndexOffset+8)) |
1833 | return false; |
1834 | return true; |
1835 | } |
1836 | |
1837 | /** |
1838 | * Determine if the specified shuffle mask is suitable for the vmrgew or |
1839 | * vmrgow instructions. |
1840 | * |
1841 | * \param[in] N The shuffle vector SD Node to analyze |
1842 | * \param[in] CheckEven Check for an even merge (true) or an odd merge (false) |
1843 | * \param[in] ShuffleKind Identify the type of merge: |
1844 | * - 0 = big-endian merge with two different inputs; |
1845 | * - 1 = either-endian merge with two identical inputs; |
1846 | * - 2 = little-endian merge with two different inputs (inputs are swapped for |
1847 | * little-endian merges). |
1848 | * \param[in] DAG The current SelectionDAG |
1849 | * \return true iff this shuffle mask |
1850 | */ |
1851 | bool PPC::isVMRGEOShuffleMask(ShuffleVectorSDNode *N, bool CheckEven, |
1852 | unsigned ShuffleKind, SelectionDAG &DAG) { |
1853 | if (DAG.getDataLayout().isLittleEndian()) { |
1854 | unsigned indexOffset = CheckEven ? 4 : 0; |
1855 | if (ShuffleKind == 1) // Unary |
1856 | return isVMerge(N, indexOffset, 0); |
1857 | else if (ShuffleKind == 2) // swapped |
1858 | return isVMerge(N, indexOffset, 16); |
1859 | else |
1860 | return false; |
1861 | } |
1862 | else { |
1863 | unsigned indexOffset = CheckEven ? 0 : 4; |
1864 | if (ShuffleKind == 1) // Unary |
1865 | return isVMerge(N, indexOffset, 0); |
1866 | else if (ShuffleKind == 0) // Normal |
1867 | return isVMerge(N, indexOffset, 16); |
1868 | else |
1869 | return false; |
1870 | } |
1871 | return false; |
1872 | } |
1873 | |
1874 | /// isVSLDOIShuffleMask - If this is a vsldoi shuffle mask, return the shift |
1875 | /// amount, otherwise return -1. |
1876 | /// The ShuffleKind distinguishes between big-endian operations with two |
1877 | /// different inputs (0), either-endian operations with two identical inputs |
1878 | /// (1), and little-endian operations with two different inputs (2). For the |
1879 | /// latter, the input operands are swapped (see PPCInstrAltivec.td). |
1880 | int PPC::isVSLDOIShuffleMask(SDNode *N, unsigned ShuffleKind, |
1881 | SelectionDAG &DAG) { |
1882 | if (N->getValueType(0) != MVT::v16i8) |
1883 | return -1; |
1884 | |
1885 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
1886 | |
1887 | // Find the first non-undef value in the shuffle mask. |
1888 | unsigned i; |
1889 | for (i = 0; i != 16 && SVOp->getMaskElt(i) < 0; ++i) |
1890 | /*search*/; |
1891 | |
1892 | if (i == 16) return -1; // all undef. |
1893 | |
1894 | // Otherwise, check to see if the rest of the elements are consecutively |
1895 | // numbered from this value. |
1896 | unsigned ShiftAmt = SVOp->getMaskElt(i); |
1897 | if (ShiftAmt < i) return -1; |
1898 | |
1899 | ShiftAmt -= i; |
1900 | bool isLE = DAG.getDataLayout().isLittleEndian(); |
1901 | |
1902 | if ((ShuffleKind == 0 && !isLE) || (ShuffleKind == 2 && isLE)) { |
1903 | // Check the rest of the elements to see if they are consecutive. |
1904 | for (++i; i != 16; ++i) |
1905 | if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i)) |
1906 | return -1; |
1907 | } else if (ShuffleKind == 1) { |
1908 | // Check the rest of the elements to see if they are consecutive. |
1909 | for (++i; i != 16; ++i) |
1910 | if (!isConstantOrUndef(SVOp->getMaskElt(i), (ShiftAmt+i) & 15)) |
1911 | return -1; |
1912 | } else |
1913 | return -1; |
1914 | |
1915 | if (isLE) |
1916 | ShiftAmt = 16 - ShiftAmt; |
1917 | |
1918 | return ShiftAmt; |
1919 | } |
1920 | |
1921 | /// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand |
1922 | /// specifies a splat of a single element that is suitable for input to |
1923 | /// one of the splat operations (VSPLTB/VSPLTH/VSPLTW/XXSPLTW/LXVDSX/etc.). |
1924 | bool PPC::isSplatShuffleMask(ShuffleVectorSDNode *N, unsigned EltSize) { |
1925 | assert(N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) &&((N->getValueType(0) == MVT::v16i8 && isPowerOf2_32 (EltSize) && EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) && EltSize <= 8 && \"Can only handle 1,2,4,8 byte element sizes\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1926, __PRETTY_FUNCTION__)) |
1926 | EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes")((N->getValueType(0) == MVT::v16i8 && isPowerOf2_32 (EltSize) && EltSize <= 8 && "Can only handle 1,2,4,8 byte element sizes" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && isPowerOf2_32(EltSize) && EltSize <= 8 && \"Can only handle 1,2,4,8 byte element sizes\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1926, __PRETTY_FUNCTION__)); |
1927 | |
1928 | // The consecutive indices need to specify an element, not part of two |
1929 | // different elements. So abandon ship early if this isn't the case. |
1930 | if (N->getMaskElt(0) % EltSize != 0) |
1931 | return false; |
1932 | |
1933 | // This is a splat operation if each element of the permute is the same, and |
1934 | // if the value doesn't reference the second vector. |
1935 | unsigned ElementBase = N->getMaskElt(0); |
1936 | |
1937 | // FIXME: Handle UNDEF elements too! |
1938 | if (ElementBase >= 16) |
1939 | return false; |
1940 | |
1941 | // Check that the indices are consecutive, in the case of a multi-byte element |
1942 | // splatted with a v16i8 mask. |
1943 | for (unsigned i = 1; i != EltSize; ++i) |
1944 | if (N->getMaskElt(i) < 0 || N->getMaskElt(i) != (int)(i+ElementBase)) |
1945 | return false; |
1946 | |
1947 | for (unsigned i = EltSize, e = 16; i != e; i += EltSize) { |
1948 | if (N->getMaskElt(i) < 0) continue; |
1949 | for (unsigned j = 0; j != EltSize; ++j) |
1950 | if (N->getMaskElt(i+j) != N->getMaskElt(j)) |
1951 | return false; |
1952 | } |
1953 | return true; |
1954 | } |
1955 | |
1956 | /// Check that the mask is shuffling N byte elements. Within each N byte |
1957 | /// element of the mask, the indices could be either in increasing or |
1958 | /// decreasing order as long as they are consecutive. |
1959 | /// \param[in] N the shuffle vector SD Node to analyze |
1960 | /// \param[in] Width the element width in bytes, could be 2/4/8/16 (HalfWord/ |
1961 | /// Word/DoubleWord/QuadWord). |
1962 | /// \param[in] StepLen the delta indices number among the N byte element, if |
1963 | /// the mask is in increasing/decreasing order then it is 1/-1. |
1964 | /// \return true iff the mask is shuffling N byte elements. |
1965 | static bool isNByteElemShuffleMask(ShuffleVectorSDNode *N, unsigned Width, |
1966 | int StepLen) { |
1967 | assert((Width == 2 || Width == 4 || Width == 8 || Width == 16) &&(((Width == 2 || Width == 4 || Width == 8 || Width == 16) && "Unexpected element width.") ? static_cast<void> (0) : __assert_fail ("(Width == 2 || Width == 4 || Width == 8 || Width == 16) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1968, __PRETTY_FUNCTION__)) |
1968 | "Unexpected element width.")(((Width == 2 || Width == 4 || Width == 8 || Width == 16) && "Unexpected element width.") ? static_cast<void> (0) : __assert_fail ("(Width == 2 || Width == 4 || Width == 8 || Width == 16) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1968, __PRETTY_FUNCTION__)); |
1969 | assert((StepLen == 1 || StepLen == -1) && "Unexpected element width.")(((StepLen == 1 || StepLen == -1) && "Unexpected element width." ) ? static_cast<void> (0) : __assert_fail ("(StepLen == 1 || StepLen == -1) && \"Unexpected element width.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 1969, __PRETTY_FUNCTION__)); |
1970 | |
1971 | unsigned NumOfElem = 16 / Width; |
1972 | unsigned MaskVal[16]; // Width is never greater than 16 |
1973 | for (unsigned i = 0; i < NumOfElem; ++i) { |
1974 | MaskVal[0] = N->getMaskElt(i * Width); |
1975 | if ((StepLen == 1) && (MaskVal[0] % Width)) { |
1976 | return false; |
1977 | } else if ((StepLen == -1) && ((MaskVal[0] + 1) % Width)) { |
1978 | return false; |
1979 | } |
1980 | |
1981 | for (unsigned int j = 1; j < Width; ++j) { |
1982 | MaskVal[j] = N->getMaskElt(i * Width + j); |
1983 | if (MaskVal[j] != MaskVal[j-1] + StepLen) { |
1984 | return false; |
1985 | } |
1986 | } |
1987 | } |
1988 | |
1989 | return true; |
1990 | } |
1991 | |
1992 | bool PPC::isXXINSERTWMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
1993 | unsigned &InsertAtByte, bool &Swap, bool IsLE) { |
1994 | if (!isNByteElemShuffleMask(N, 4, 1)) |
1995 | return false; |
1996 | |
1997 | // Now we look at mask elements 0,4,8,12 |
1998 | unsigned M0 = N->getMaskElt(0) / 4; |
1999 | unsigned M1 = N->getMaskElt(4) / 4; |
2000 | unsigned M2 = N->getMaskElt(8) / 4; |
2001 | unsigned M3 = N->getMaskElt(12) / 4; |
2002 | unsigned LittleEndianShifts[] = { 2, 1, 0, 3 }; |
2003 | unsigned BigEndianShifts[] = { 3, 0, 1, 2 }; |
2004 | |
2005 | // Below, let H and L be arbitrary elements of the shuffle mask |
2006 | // where H is in the range [4,7] and L is in the range [0,3]. |
2007 | // H, 1, 2, 3 or L, 5, 6, 7 |
2008 | if ((M0 > 3 && M1 == 1 && M2 == 2 && M3 == 3) || |
2009 | (M0 < 4 && M1 == 5 && M2 == 6 && M3 == 7)) { |
2010 | ShiftElts = IsLE ? LittleEndianShifts[M0 & 0x3] : BigEndianShifts[M0 & 0x3]; |
2011 | InsertAtByte = IsLE ? 12 : 0; |
2012 | Swap = M0 < 4; |
2013 | return true; |
2014 | } |
2015 | // 0, H, 2, 3 or 4, L, 6, 7 |
2016 | if ((M1 > 3 && M0 == 0 && M2 == 2 && M3 == 3) || |
2017 | (M1 < 4 && M0 == 4 && M2 == 6 && M3 == 7)) { |
2018 | ShiftElts = IsLE ? LittleEndianShifts[M1 & 0x3] : BigEndianShifts[M1 & 0x3]; |
2019 | InsertAtByte = IsLE ? 8 : 4; |
2020 | Swap = M1 < 4; |
2021 | return true; |
2022 | } |
2023 | // 0, 1, H, 3 or 4, 5, L, 7 |
2024 | if ((M2 > 3 && M0 == 0 && M1 == 1 && M3 == 3) || |
2025 | (M2 < 4 && M0 == 4 && M1 == 5 && M3 == 7)) { |
2026 | ShiftElts = IsLE ? LittleEndianShifts[M2 & 0x3] : BigEndianShifts[M2 & 0x3]; |
2027 | InsertAtByte = IsLE ? 4 : 8; |
2028 | Swap = M2 < 4; |
2029 | return true; |
2030 | } |
2031 | // 0, 1, 2, H or 4, 5, 6, L |
2032 | if ((M3 > 3 && M0 == 0 && M1 == 1 && M2 == 2) || |
2033 | (M3 < 4 && M0 == 4 && M1 == 5 && M2 == 6)) { |
2034 | ShiftElts = IsLE ? LittleEndianShifts[M3 & 0x3] : BigEndianShifts[M3 & 0x3]; |
2035 | InsertAtByte = IsLE ? 0 : 12; |
2036 | Swap = M3 < 4; |
2037 | return true; |
2038 | } |
2039 | |
2040 | // If both vector operands for the shuffle are the same vector, the mask will |
2041 | // contain only elements from the first one and the second one will be undef. |
2042 | if (N->getOperand(1).isUndef()) { |
2043 | ShiftElts = 0; |
2044 | Swap = true; |
2045 | unsigned XXINSERTWSrcElem = IsLE ? 2 : 1; |
2046 | if (M0 == XXINSERTWSrcElem && M1 == 1 && M2 == 2 && M3 == 3) { |
2047 | InsertAtByte = IsLE ? 12 : 0; |
2048 | return true; |
2049 | } |
2050 | if (M0 == 0 && M1 == XXINSERTWSrcElem && M2 == 2 && M3 == 3) { |
2051 | InsertAtByte = IsLE ? 8 : 4; |
2052 | return true; |
2053 | } |
2054 | if (M0 == 0 && M1 == 1 && M2 == XXINSERTWSrcElem && M3 == 3) { |
2055 | InsertAtByte = IsLE ? 4 : 8; |
2056 | return true; |
2057 | } |
2058 | if (M0 == 0 && M1 == 1 && M2 == 2 && M3 == XXINSERTWSrcElem) { |
2059 | InsertAtByte = IsLE ? 0 : 12; |
2060 | return true; |
2061 | } |
2062 | } |
2063 | |
2064 | return false; |
2065 | } |
2066 | |
2067 | bool PPC::isXXSLDWIShuffleMask(ShuffleVectorSDNode *N, unsigned &ShiftElts, |
2068 | bool &Swap, bool IsLE) { |
2069 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2069, __PRETTY_FUNCTION__)); |
2070 | // Ensure each byte index of the word is consecutive. |
2071 | if (!isNByteElemShuffleMask(N, 4, 1)) |
2072 | return false; |
2073 | |
2074 | // Now we look at mask elements 0,4,8,12, which are the beginning of words. |
2075 | unsigned M0 = N->getMaskElt(0) / 4; |
2076 | unsigned M1 = N->getMaskElt(4) / 4; |
2077 | unsigned M2 = N->getMaskElt(8) / 4; |
2078 | unsigned M3 = N->getMaskElt(12) / 4; |
2079 | |
2080 | // If both vector operands for the shuffle are the same vector, the mask will |
2081 | // contain only elements from the first one and the second one will be undef. |
2082 | if (N->getOperand(1).isUndef()) { |
2083 | assert(M0 < 4 && "Indexing into an undef vector?")((M0 < 4 && "Indexing into an undef vector?") ? static_cast <void> (0) : __assert_fail ("M0 < 4 && \"Indexing into an undef vector?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2083, __PRETTY_FUNCTION__)); |
2084 | if (M1 != (M0 + 1) % 4 || M2 != (M1 + 1) % 4 || M3 != (M2 + 1) % 4) |
2085 | return false; |
2086 | |
2087 | ShiftElts = IsLE ? (4 - M0) % 4 : M0; |
2088 | Swap = false; |
2089 | return true; |
2090 | } |
2091 | |
2092 | // Ensure each word index of the ShuffleVector Mask is consecutive. |
2093 | if (M1 != (M0 + 1) % 8 || M2 != (M1 + 1) % 8 || M3 != (M2 + 1) % 8) |
2094 | return false; |
2095 | |
2096 | if (IsLE) { |
2097 | if (M0 == 0 || M0 == 7 || M0 == 6 || M0 == 5) { |
2098 | // Input vectors don't need to be swapped if the leading element |
2099 | // of the result is one of the 3 left elements of the second vector |
2100 | // (or if there is no shift to be done at all). |
2101 | Swap = false; |
2102 | ShiftElts = (8 - M0) % 8; |
2103 | } else if (M0 == 4 || M0 == 3 || M0 == 2 || M0 == 1) { |
2104 | // Input vectors need to be swapped if the leading element |
2105 | // of the result is one of the 3 left elements of the first vector |
2106 | // (or if we're shifting by 4 - thereby simply swapping the vectors). |
2107 | Swap = true; |
2108 | ShiftElts = (4 - M0) % 4; |
2109 | } |
2110 | |
2111 | return true; |
2112 | } else { // BE |
2113 | if (M0 == 0 || M0 == 1 || M0 == 2 || M0 == 3) { |
2114 | // Input vectors don't need to be swapped if the leading element |
2115 | // of the result is one of the 4 elements of the first vector. |
2116 | Swap = false; |
2117 | ShiftElts = M0; |
2118 | } else if (M0 == 4 || M0 == 5 || M0 == 6 || M0 == 7) { |
2119 | // Input vectors need to be swapped if the leading element |
2120 | // of the result is one of the 4 elements of the right vector. |
2121 | Swap = true; |
2122 | ShiftElts = M0 - 4; |
2123 | } |
2124 | |
2125 | return true; |
2126 | } |
2127 | } |
2128 | |
2129 | bool static isXXBRShuffleMaskHelper(ShuffleVectorSDNode *N, int Width) { |
2130 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2130, __PRETTY_FUNCTION__)); |
2131 | |
2132 | if (!isNByteElemShuffleMask(N, Width, -1)) |
2133 | return false; |
2134 | |
2135 | for (int i = 0; i < 16; i += Width) |
2136 | if (N->getMaskElt(i) != i + Width - 1) |
2137 | return false; |
2138 | |
2139 | return true; |
2140 | } |
2141 | |
2142 | bool PPC::isXXBRHShuffleMask(ShuffleVectorSDNode *N) { |
2143 | return isXXBRShuffleMaskHelper(N, 2); |
2144 | } |
2145 | |
2146 | bool PPC::isXXBRWShuffleMask(ShuffleVectorSDNode *N) { |
2147 | return isXXBRShuffleMaskHelper(N, 4); |
2148 | } |
2149 | |
2150 | bool PPC::isXXBRDShuffleMask(ShuffleVectorSDNode *N) { |
2151 | return isXXBRShuffleMaskHelper(N, 8); |
2152 | } |
2153 | |
2154 | bool PPC::isXXBRQShuffleMask(ShuffleVectorSDNode *N) { |
2155 | return isXXBRShuffleMaskHelper(N, 16); |
2156 | } |
2157 | |
2158 | /// Can node \p N be lowered to an XXPERMDI instruction? If so, set \p Swap |
2159 | /// if the inputs to the instruction should be swapped and set \p DM to the |
2160 | /// value for the immediate. |
2161 | /// Specifically, set \p Swap to true only if \p N can be lowered to XXPERMDI |
2162 | /// AND element 0 of the result comes from the first input (LE) or second input |
2163 | /// (BE). Set \p DM to the calculated result (0-3) only if \p N can be lowered. |
2164 | /// \return true iff the given mask of shuffle node \p N is a XXPERMDI shuffle |
2165 | /// mask. |
2166 | bool PPC::isXXPERMDIShuffleMask(ShuffleVectorSDNode *N, unsigned &DM, |
2167 | bool &Swap, bool IsLE) { |
2168 | assert(N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8")((N->getValueType(0) == MVT::v16i8 && "Shuffle vector expects v16i8" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::v16i8 && \"Shuffle vector expects v16i8\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2168, __PRETTY_FUNCTION__)); |
2169 | |
2170 | // Ensure each byte index of the double word is consecutive. |
2171 | if (!isNByteElemShuffleMask(N, 8, 1)) |
2172 | return false; |
2173 | |
2174 | unsigned M0 = N->getMaskElt(0) / 8; |
2175 | unsigned M1 = N->getMaskElt(8) / 8; |
2176 | assert(((M0 | M1) < 4) && "A mask element out of bounds?")((((M0 | M1) < 4) && "A mask element out of bounds?" ) ? static_cast<void> (0) : __assert_fail ("((M0 | M1) < 4) && \"A mask element out of bounds?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2176, __PRETTY_FUNCTION__)); |
2177 | |
2178 | // If both vector operands for the shuffle are the same vector, the mask will |
2179 | // contain only elements from the first one and the second one will be undef. |
2180 | if (N->getOperand(1).isUndef()) { |
2181 | if ((M0 | M1) < 2) { |
2182 | DM = IsLE ? (((~M1) & 1) << 1) + ((~M0) & 1) : (M0 << 1) + (M1 & 1); |
2183 | Swap = false; |
2184 | return true; |
2185 | } else |
2186 | return false; |
2187 | } |
2188 | |
2189 | if (IsLE) { |
2190 | if (M0 > 1 && M1 < 2) { |
2191 | Swap = false; |
2192 | } else if (M0 < 2 && M1 > 1) { |
2193 | M0 = (M0 + 2) % 4; |
2194 | M1 = (M1 + 2) % 4; |
2195 | Swap = true; |
2196 | } else |
2197 | return false; |
2198 | |
2199 | // Note: if control flow comes here that means Swap is already set above |
2200 | DM = (((~M1) & 1) << 1) + ((~M0) & 1); |
2201 | return true; |
2202 | } else { // BE |
2203 | if (M0 < 2 && M1 > 1) { |
2204 | Swap = false; |
2205 | } else if (M0 > 1 && M1 < 2) { |
2206 | M0 = (M0 + 2) % 4; |
2207 | M1 = (M1 + 2) % 4; |
2208 | Swap = true; |
2209 | } else |
2210 | return false; |
2211 | |
2212 | // Note: if control flow comes here that means Swap is already set above |
2213 | DM = (M0 << 1) + (M1 & 1); |
2214 | return true; |
2215 | } |
2216 | } |
2217 | |
2218 | |
2219 | /// getSplatIdxForPPCMnemonics - Return the splat index as a value that is |
2220 | /// appropriate for PPC mnemonics (which have a big endian bias - namely |
2221 | /// elements are counted from the left of the vector register). |
2222 | unsigned PPC::getSplatIdxForPPCMnemonics(SDNode *N, unsigned EltSize, |
2223 | SelectionDAG &DAG) { |
2224 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
2225 | assert(isSplatShuffleMask(SVOp, EltSize))((isSplatShuffleMask(SVOp, EltSize)) ? static_cast<void> (0) : __assert_fail ("isSplatShuffleMask(SVOp, EltSize)", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2225, __PRETTY_FUNCTION__)); |
2226 | if (DAG.getDataLayout().isLittleEndian()) |
2227 | return (16 / EltSize) - 1 - (SVOp->getMaskElt(0) / EltSize); |
2228 | else |
2229 | return SVOp->getMaskElt(0) / EltSize; |
2230 | } |
2231 | |
2232 | /// get_VSPLTI_elt - If this is a build_vector of constants which can be formed |
2233 | /// by using a vspltis[bhw] instruction of the specified element size, return |
2234 | /// the constant being splatted. The ByteSize field indicates the number of |
2235 | /// bytes of each element [124] -> [bhw]. |
2236 | SDValue PPC::get_VSPLTI_elt(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) { |
2237 | SDValue OpVal(nullptr, 0); |
2238 | |
2239 | // If ByteSize of the splat is bigger than the element size of the |
2240 | // build_vector, then we have a case where we are checking for a splat where |
2241 | // multiple elements of the buildvector are folded together into a single |
2242 | // logical element of the splat (e.g. "vsplish 1" to splat {0,1}*8). |
2243 | unsigned EltSize = 16/N->getNumOperands(); |
2244 | if (EltSize < ByteSize) { |
2245 | unsigned Multiple = ByteSize/EltSize; // Number of BV entries per spltval. |
2246 | SDValue UniquedVals[4]; |
2247 | assert(Multiple > 1 && Multiple <= 4 && "How can this happen?")((Multiple > 1 && Multiple <= 4 && "How can this happen?" ) ? static_cast<void> (0) : __assert_fail ("Multiple > 1 && Multiple <= 4 && \"How can this happen?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2247, __PRETTY_FUNCTION__)); |
2248 | |
2249 | // See if all of the elements in the buildvector agree across. |
2250 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { |
2251 | if (N->getOperand(i).isUndef()) continue; |
2252 | // If the element isn't a constant, bail fully out. |
2253 | if (!isa<ConstantSDNode>(N->getOperand(i))) return SDValue(); |
2254 | |
2255 | if (!UniquedVals[i&(Multiple-1)].getNode()) |
2256 | UniquedVals[i&(Multiple-1)] = N->getOperand(i); |
2257 | else if (UniquedVals[i&(Multiple-1)] != N->getOperand(i)) |
2258 | return SDValue(); // no match. |
2259 | } |
2260 | |
2261 | // Okay, if we reached this point, UniquedVals[0..Multiple-1] contains |
2262 | // either constant or undef values that are identical for each chunk. See |
2263 | // if these chunks can form into a larger vspltis*. |
2264 | |
2265 | // Check to see if all of the leading entries are either 0 or -1. If |
2266 | // neither, then this won't fit into the immediate field. |
2267 | bool LeadingZero = true; |
2268 | bool LeadingOnes = true; |
2269 | for (unsigned i = 0; i != Multiple-1; ++i) { |
2270 | if (!UniquedVals[i].getNode()) continue; // Must have been undefs. |
2271 | |
2272 | LeadingZero &= isNullConstant(UniquedVals[i]); |
2273 | LeadingOnes &= isAllOnesConstant(UniquedVals[i]); |
2274 | } |
2275 | // Finally, check the least significant entry. |
2276 | if (LeadingZero) { |
2277 | if (!UniquedVals[Multiple-1].getNode()) |
2278 | return DAG.getTargetConstant(0, SDLoc(N), MVT::i32); // 0,0,0,undef |
2279 | int Val = cast<ConstantSDNode>(UniquedVals[Multiple-1])->getZExtValue(); |
2280 | if (Val < 16) // 0,0,0,4 -> vspltisw(4) |
2281 | return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32); |
2282 | } |
2283 | if (LeadingOnes) { |
2284 | if (!UniquedVals[Multiple-1].getNode()) |
2285 | return DAG.getTargetConstant(~0U, SDLoc(N), MVT::i32); // -1,-1,-1,undef |
2286 | int Val =cast<ConstantSDNode>(UniquedVals[Multiple-1])->getSExtValue(); |
2287 | if (Val >= -16) // -1,-1,-1,-2 -> vspltisw(-2) |
2288 | return DAG.getTargetConstant(Val, SDLoc(N), MVT::i32); |
2289 | } |
2290 | |
2291 | return SDValue(); |
2292 | } |
2293 | |
2294 | // Check to see if this buildvec has a single non-undef value in its elements. |
2295 | for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { |
2296 | if (N->getOperand(i).isUndef()) continue; |
2297 | if (!OpVal.getNode()) |
2298 | OpVal = N->getOperand(i); |
2299 | else if (OpVal != N->getOperand(i)) |
2300 | return SDValue(); |
2301 | } |
2302 | |
2303 | if (!OpVal.getNode()) return SDValue(); // All UNDEF: use implicit def. |
2304 | |
2305 | unsigned ValSizeInBytes = EltSize; |
2306 | uint64_t Value = 0; |
2307 | if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) { |
2308 | Value = CN->getZExtValue(); |
2309 | } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) { |
2310 | assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!")((CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!" ) ? static_cast<void> (0) : __assert_fail ("CN->getValueType(0) == MVT::f32 && \"Only one legal FP vector type!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2310, __PRETTY_FUNCTION__)); |
2311 | Value = FloatToBits(CN->getValueAPF().convertToFloat()); |
2312 | } |
2313 | |
2314 | // If the splat value is larger than the element value, then we can never do |
2315 | // this splat. The only case that we could fit the replicated bits into our |
2316 | // immediate field for would be zero, and we prefer to use vxor for it. |
2317 | if (ValSizeInBytes < ByteSize) return SDValue(); |
2318 | |
2319 | // If the element value is larger than the splat value, check if it consists |
2320 | // of a repeated bit pattern of size ByteSize. |
2321 | if (!APInt(ValSizeInBytes * 8, Value).isSplat(ByteSize * 8)) |
2322 | return SDValue(); |
2323 | |
2324 | // Properly sign extend the value. |
2325 | int MaskVal = SignExtend32(Value, ByteSize * 8); |
2326 | |
2327 | // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros. |
2328 | if (MaskVal == 0) return SDValue(); |
2329 | |
2330 | // Finally, if this value fits in a 5 bit sext field, return it |
2331 | if (SignExtend32<5>(MaskVal) == MaskVal) |
2332 | return DAG.getTargetConstant(MaskVal, SDLoc(N), MVT::i32); |
2333 | return SDValue(); |
2334 | } |
2335 | |
2336 | /// isQVALIGNIShuffleMask - If this is a qvaligni shuffle mask, return the shift |
2337 | /// amount, otherwise return -1. |
2338 | int PPC::isQVALIGNIShuffleMask(SDNode *N) { |
2339 | EVT VT = N->getValueType(0); |
2340 | if (VT != MVT::v4f64 && VT != MVT::v4f32 && VT != MVT::v4i1) |
2341 | return -1; |
2342 | |
2343 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N); |
2344 | |
2345 | // Find the first non-undef value in the shuffle mask. |
2346 | unsigned i; |
2347 | for (i = 0; i != 4 && SVOp->getMaskElt(i) < 0; ++i) |
2348 | /*search*/; |
2349 | |
2350 | if (i == 4) return -1; // all undef. |
2351 | |
2352 | // Otherwise, check to see if the rest of the elements are consecutively |
2353 | // numbered from this value. |
2354 | unsigned ShiftAmt = SVOp->getMaskElt(i); |
2355 | if (ShiftAmt < i) return -1; |
2356 | ShiftAmt -= i; |
2357 | |
2358 | // Check the rest of the elements to see if they are consecutive. |
2359 | for (++i; i != 4; ++i) |
2360 | if (!isConstantOrUndef(SVOp->getMaskElt(i), ShiftAmt+i)) |
2361 | return -1; |
2362 | |
2363 | return ShiftAmt; |
2364 | } |
2365 | |
2366 | //===----------------------------------------------------------------------===// |
2367 | // Addressing Mode Selection |
2368 | //===----------------------------------------------------------------------===// |
2369 | |
2370 | /// isIntS16Immediate - This method tests to see if the node is either a 32-bit |
2371 | /// or 64-bit immediate, and if the value can be accurately represented as a |
2372 | /// sign extension from a 16-bit value. If so, this returns true and the |
2373 | /// immediate. |
2374 | bool llvm::isIntS16Immediate(SDNode *N, int16_t &Imm) { |
2375 | if (!isa<ConstantSDNode>(N)) |
2376 | return false; |
2377 | |
2378 | Imm = (int16_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2379 | if (N->getValueType(0) == MVT::i32) |
2380 | return Imm == (int32_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2381 | else |
2382 | return Imm == (int64_t)cast<ConstantSDNode>(N)->getZExtValue(); |
2383 | } |
2384 | bool llvm::isIntS16Immediate(SDValue Op, int16_t &Imm) { |
2385 | return isIntS16Immediate(Op.getNode(), Imm); |
2386 | } |
2387 | |
2388 | |
2389 | /// SelectAddressEVXRegReg - Given the specified address, check to see if it can |
2390 | /// be represented as an indexed [r+r] operation. |
2391 | bool PPCTargetLowering::SelectAddressEVXRegReg(SDValue N, SDValue &Base, |
2392 | SDValue &Index, |
2393 | SelectionDAG &DAG) const { |
2394 | for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end(); |
2395 | UI != E; ++UI) { |
2396 | if (MemSDNode *Memop = dyn_cast<MemSDNode>(*UI)) { |
2397 | if (Memop->getMemoryVT() == MVT::f64) { |
2398 | Base = N.getOperand(0); |
2399 | Index = N.getOperand(1); |
2400 | return true; |
2401 | } |
2402 | } |
2403 | } |
2404 | return false; |
2405 | } |
2406 | |
2407 | /// SelectAddressRegReg - Given the specified addressed, check to see if it |
2408 | /// can be represented as an indexed [r+r] operation. Returns false if it |
2409 | /// can be more efficiently represented as [r+imm]. If \p EncodingAlignment is |
2410 | /// non-zero and N can be represented by a base register plus a signed 16-bit |
2411 | /// displacement, make a more precise judgement by checking (displacement % \p |
2412 | /// EncodingAlignment). |
2413 | bool PPCTargetLowering::SelectAddressRegReg( |
2414 | SDValue N, SDValue &Base, SDValue &Index, SelectionDAG &DAG, |
2415 | MaybeAlign EncodingAlignment) const { |
2416 | // If we have a PC Relative target flag don't select as [reg+reg]. It will be |
2417 | // a [pc+imm]. |
2418 | if (SelectAddressPCRel(N, Base)) |
2419 | return false; |
2420 | |
2421 | int16_t Imm = 0; |
2422 | if (N.getOpcode() == ISD::ADD) { |
2423 | // Is there any SPE load/store (f64), which can't handle 16bit offset? |
2424 | // SPE load/store can only handle 8-bit offsets. |
2425 | if (hasSPE() && SelectAddressEVXRegReg(N, Base, Index, DAG)) |
2426 | return true; |
2427 | if (isIntS16Immediate(N.getOperand(1), Imm) && |
2428 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) |
2429 | return false; // r+i |
2430 | if (N.getOperand(1).getOpcode() == PPCISD::Lo) |
2431 | return false; // r+i |
2432 | |
2433 | Base = N.getOperand(0); |
2434 | Index = N.getOperand(1); |
2435 | return true; |
2436 | } else if (N.getOpcode() == ISD::OR) { |
2437 | if (isIntS16Immediate(N.getOperand(1), Imm) && |
2438 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) |
2439 | return false; // r+i can fold it if we can. |
2440 | |
2441 | // If this is an or of disjoint bitfields, we can codegen this as an add |
2442 | // (for better address arithmetic) if the LHS and RHS of the OR are provably |
2443 | // disjoint. |
2444 | KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0)); |
2445 | |
2446 | if (LHSKnown.Zero.getBoolValue()) { |
2447 | KnownBits RHSKnown = DAG.computeKnownBits(N.getOperand(1)); |
2448 | // If all of the bits are known zero on the LHS or RHS, the add won't |
2449 | // carry. |
2450 | if (~(LHSKnown.Zero | RHSKnown.Zero) == 0) { |
2451 | Base = N.getOperand(0); |
2452 | Index = N.getOperand(1); |
2453 | return true; |
2454 | } |
2455 | } |
2456 | } |
2457 | |
2458 | return false; |
2459 | } |
2460 | |
2461 | // If we happen to be doing an i64 load or store into a stack slot that has |
2462 | // less than a 4-byte alignment, then the frame-index elimination may need to |
2463 | // use an indexed load or store instruction (because the offset may not be a |
2464 | // multiple of 4). The extra register needed to hold the offset comes from the |
2465 | // register scavenger, and it is possible that the scavenger will need to use |
2466 | // an emergency spill slot. As a result, we need to make sure that a spill slot |
2467 | // is allocated when doing an i64 load/store into a less-than-4-byte-aligned |
2468 | // stack slot. |
2469 | static void fixupFuncForFI(SelectionDAG &DAG, int FrameIdx, EVT VT) { |
2470 | // FIXME: This does not handle the LWA case. |
2471 | if (VT != MVT::i64) |
2472 | return; |
2473 | |
2474 | // NOTE: We'll exclude negative FIs here, which come from argument |
2475 | // lowering, because there are no known test cases triggering this problem |
2476 | // using packed structures (or similar). We can remove this exclusion if |
2477 | // we find such a test case. The reason why this is so test-case driven is |
2478 | // because this entire 'fixup' is only to prevent crashes (from the |
2479 | // register scavenger) on not-really-valid inputs. For example, if we have: |
2480 | // %a = alloca i1 |
2481 | // %b = bitcast i1* %a to i64* |
2482 | // store i64* a, i64 b |
2483 | // then the store should really be marked as 'align 1', but is not. If it |
2484 | // were marked as 'align 1' then the indexed form would have been |
2485 | // instruction-selected initially, and the problem this 'fixup' is preventing |
2486 | // won't happen regardless. |
2487 | if (FrameIdx < 0) |
2488 | return; |
2489 | |
2490 | MachineFunction &MF = DAG.getMachineFunction(); |
2491 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
2492 | |
2493 | if (MFI.getObjectAlign(FrameIdx) >= Align(4)) |
2494 | return; |
2495 | |
2496 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
2497 | FuncInfo->setHasNonRISpills(); |
2498 | } |
2499 | |
2500 | /// Returns true if the address N can be represented by a base register plus |
2501 | /// a signed 16-bit displacement [r+imm], and if it is not better |
2502 | /// represented as reg+reg. If \p EncodingAlignment is non-zero, only accept |
2503 | /// displacements that are multiples of that value. |
2504 | bool PPCTargetLowering::SelectAddressRegImm( |
2505 | SDValue N, SDValue &Disp, SDValue &Base, SelectionDAG &DAG, |
2506 | MaybeAlign EncodingAlignment) const { |
2507 | // FIXME dl should come from parent load or store, not from address |
2508 | SDLoc dl(N); |
2509 | |
2510 | // If we have a PC Relative target flag don't select as [reg+imm]. It will be |
2511 | // a [pc+imm]. |
2512 | if (SelectAddressPCRel(N, Base)) |
2513 | return false; |
2514 | |
2515 | // If this can be more profitably realized as r+r, fail. |
2516 | if (SelectAddressRegReg(N, Disp, Base, DAG, EncodingAlignment)) |
2517 | return false; |
2518 | |
2519 | if (N.getOpcode() == ISD::ADD) { |
2520 | int16_t imm = 0; |
2521 | if (isIntS16Immediate(N.getOperand(1), imm) && |
2522 | (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) { |
2523 | Disp = DAG.getTargetConstant(imm, dl, N.getValueType()); |
2524 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) { |
2525 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2526 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2527 | } else { |
2528 | Base = N.getOperand(0); |
2529 | } |
2530 | return true; // [r+i] |
2531 | } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) { |
2532 | // Match LOAD (ADD (X, Lo(G))). |
2533 | assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue()((!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))-> getZExtValue() && "Cannot handle constant offsets yet!" ) ? static_cast<void> (0) : __assert_fail ("!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue() && \"Cannot handle constant offsets yet!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)) |
2534 | && "Cannot handle constant offsets yet!")((!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))-> getZExtValue() && "Cannot handle constant offsets yet!" ) ? static_cast<void> (0) : __assert_fail ("!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getZExtValue() && \"Cannot handle constant offsets yet!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2534, __PRETTY_FUNCTION__)); |
2535 | Disp = N.getOperand(1).getOperand(0); // The global address. |
2536 | assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2539, __PRETTY_FUNCTION__)) |
2537 | Disp.getOpcode() == ISD::TargetGlobalTLSAddress ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2539, __PRETTY_FUNCTION__)) |
2538 | Disp.getOpcode() == ISD::TargetConstantPool ||((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2539, __PRETTY_FUNCTION__)) |
2539 | Disp.getOpcode() == ISD::TargetJumpTable)((Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode () == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD:: TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable ) ? static_cast<void> (0) : __assert_fail ("Disp.getOpcode() == ISD::TargetGlobalAddress || Disp.getOpcode() == ISD::TargetGlobalTLSAddress || Disp.getOpcode() == ISD::TargetConstantPool || Disp.getOpcode() == ISD::TargetJumpTable" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 2539, __PRETTY_FUNCTION__)); |
2540 | Base = N.getOperand(0); |
2541 | return true; // [&g+r] |
2542 | } |
2543 | } else if (N.getOpcode() == ISD::OR) { |
2544 | int16_t imm = 0; |
2545 | if (isIntS16Immediate(N.getOperand(1), imm) && |
2546 | (!EncodingAlignment || isAligned(*EncodingAlignment, imm))) { |
2547 | // If this is an or of disjoint bitfields, we can codegen this as an add |
2548 | // (for better address arithmetic) if the LHS and RHS of the OR are |
2549 | // provably disjoint. |
2550 | KnownBits LHSKnown = DAG.computeKnownBits(N.getOperand(0)); |
2551 | |
2552 | if ((LHSKnown.Zero.getZExtValue()|~(uint64_t)imm) == ~0ULL) { |
2553 | // If all of the bits are known zero on the LHS or RHS, the add won't |
2554 | // carry. |
2555 | if (FrameIndexSDNode *FI = |
2556 | dyn_cast<FrameIndexSDNode>(N.getOperand(0))) { |
2557 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2558 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2559 | } else { |
2560 | Base = N.getOperand(0); |
2561 | } |
2562 | Disp = DAG.getTargetConstant(imm, dl, N.getValueType()); |
2563 | return true; |
2564 | } |
2565 | } |
2566 | } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) { |
2567 | // Loading from a constant address. |
2568 | |
2569 | // If this address fits entirely in a 16-bit sext immediate field, codegen |
2570 | // this as "d, 0" |
2571 | int16_t Imm; |
2572 | if (isIntS16Immediate(CN, Imm) && |
2573 | (!EncodingAlignment || isAligned(*EncodingAlignment, Imm))) { |
2574 | Disp = DAG.getTargetConstant(Imm, dl, CN->getValueType(0)); |
2575 | Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO, |
2576 | CN->getValueType(0)); |
2577 | return true; |
2578 | } |
2579 | |
2580 | // Handle 32-bit sext immediates with LIS + addr mode. |
2581 | if ((CN->getValueType(0) == MVT::i32 || |
2582 | (int64_t)CN->getZExtValue() == (int)CN->getZExtValue()) && |
2583 | (!EncodingAlignment || |
2584 | isAligned(*EncodingAlignment, CN->getZExtValue()))) { |
2585 | int Addr = (int)CN->getZExtValue(); |
2586 | |
2587 | // Otherwise, break this down into an LIS + disp. |
2588 | Disp = DAG.getTargetConstant((short)Addr, dl, MVT::i32); |
2589 | |
2590 | Base = DAG.getTargetConstant((Addr - (signed short)Addr) >> 16, dl, |
2591 | MVT::i32); |
2592 | unsigned Opc = CN->getValueType(0) == MVT::i32 ? PPC::LIS : PPC::LIS8; |
2593 | Base = SDValue(DAG.getMachineNode(Opc, dl, CN->getValueType(0), Base), 0); |
2594 | return true; |
2595 | } |
2596 | } |
2597 | |
2598 | Disp = DAG.getTargetConstant(0, dl, getPointerTy(DAG.getDataLayout())); |
2599 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N)) { |
2600 | Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType()); |
2601 | fixupFuncForFI(DAG, FI->getIndex(), N.getValueType()); |
2602 | } else |
2603 | Base = N; |
2604 | return true; // [r+0] |
2605 | } |
2606 | |
2607 | /// SelectAddressRegRegOnly - Given the specified addressed, force it to be |
2608 | /// represented as an indexed [r+r] operation. |
2609 | bool PPCTargetLowering::SelectAddressRegRegOnly(SDValue N, SDValue &Base, |
2610 | SDValue &Index, |
2611 | SelectionDAG &DAG) const { |
2612 | // Check to see if we can easily represent this as an [r+r] address. This |
2613 | // will fail if it thinks that the address is more profitably represented as |
2614 | // reg+imm, e.g. where imm = 0. |
2615 | if (SelectAddressRegReg(N, Base, Index, DAG)) |
2616 | return true; |
2617 | |
2618 | // If the address is the result of an add, we will utilize the fact that the |
2619 | // address calculation includes an implicit add. However, we can reduce |
2620 | // register pressure if we do not materialize a constant just for use as the |
2621 | // index register. We only get rid of the add if it is not an add of a |
2622 | // value and a 16-bit signed constant and both have a single use. |
2623 | int16_t imm = 0; |
2624 | if (N.getOpcode() == ISD::ADD && |
2625 | (!isIntS16Immediate(N.getOperand(1), imm) || |
2626 | !N.getOperand(1).hasOneUse() || !N.getOperand(0).hasOneUse())) { |
2627 | Base = N.getOperand(0); |
2628 | Index = N.getOperand(1); |
2629 | return true; |
2630 | } |
2631 | |
2632 | // Otherwise, do it the hard way, using R0 as the base register. |
2633 | Base = DAG.getRegister(Subtarget.isPPC64() ? PPC::ZERO8 : PPC::ZERO, |
2634 | N.getValueType()); |
2635 | Index = N; |
2636 | return true; |
2637 | } |
2638 | |
2639 | template <typename Ty> static bool isValidPCRelNode(SDValue N) { |
2640 | Ty *PCRelCand = dyn_cast<Ty>(N); |
2641 | return PCRelCand && (PCRelCand->getTargetFlags() & PPCII::MO_PCREL_FLAG); |
2642 | } |
2643 | |
2644 | /// Returns true if this address is a PC Relative address. |
2645 | /// PC Relative addresses are marked with the flag PPCII::MO_PCREL_FLAG |
2646 | /// or if the node opcode is PPCISD::MAT_PCREL_ADDR. |
2647 | bool PPCTargetLowering::SelectAddressPCRel(SDValue N, SDValue &Base) const { |
2648 | // This is a materialize PC Relative node. Always select this as PC Relative. |
2649 | Base = N; |
2650 | if (N.getOpcode() == PPCISD::MAT_PCREL_ADDR) |
2651 | return true; |
2652 | if (isValidPCRelNode<ConstantPoolSDNode>(N) || |
2653 | isValidPCRelNode<GlobalAddressSDNode>(N) || |
2654 | isValidPCRelNode<JumpTableSDNode>(N) || |
2655 | isValidPCRelNode<BlockAddressSDNode>(N)) |
2656 | return true; |
2657 | return false; |
2658 | } |
2659 | |
2660 | /// Returns true if we should use a direct load into vector instruction |
2661 | /// (such as lxsd or lfd), instead of a load into gpr + direct move sequence. |
2662 | static bool usePartialVectorLoads(SDNode *N, const PPCSubtarget& ST) { |
2663 | |
2664 | // If there are any other uses other than scalar to vector, then we should |
2665 | // keep it as a scalar load -> direct move pattern to prevent multiple |
2666 | // loads. |
2667 | LoadSDNode *LD = dyn_cast<LoadSDNode>(N); |
2668 | if (!LD) |
2669 | return false; |
2670 | |
2671 | EVT MemVT = LD->getMemoryVT(); |
2672 | if (!MemVT.isSimple()) |
2673 | return false; |
2674 | switch(MemVT.getSimpleVT().SimpleTy) { |
2675 | case MVT::i64: |
2676 | break; |
2677 | case MVT::i32: |
2678 | if (!ST.hasP8Vector()) |
2679 | return false; |
2680 | break; |
2681 | case MVT::i16: |
2682 | case MVT::i8: |
2683 | if (!ST.hasP9Vector()) |
2684 | return false; |
2685 | break; |
2686 | default: |
2687 | return false; |
2688 | } |
2689 | |
2690 | SDValue LoadedVal(N, 0); |
2691 | if (!LoadedVal.hasOneUse()) |
2692 | return false; |
2693 | |
2694 | for (SDNode::use_iterator UI = LD->use_begin(), UE = LD->use_end(); |
2695 | UI != UE; ++UI) |
2696 | if (UI.getUse().get().getResNo() == 0 && |
2697 | UI->getOpcode() != ISD::SCALAR_TO_VECTOR && |
2698 | UI->getOpcode() != PPCISD::SCALAR_TO_VECTOR_PERMUTED) |
2699 | return false; |
2700 | |
2701 | return true; |
2702 | } |
2703 | |
2704 | /// getPreIndexedAddressParts - returns true by value, base pointer and |
2705 | /// offset pointer and addressing mode by reference if the node's address |
2706 | /// can be legally represented as pre-indexed load / store address. |
2707 | bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, |
2708 | SDValue &Offset, |
2709 | ISD::MemIndexedMode &AM, |
2710 | SelectionDAG &DAG) const { |
2711 | if (DisablePPCPreinc) return false; |
2712 | |
2713 | bool isLoad = true; |
2714 | SDValue Ptr; |
2715 | EVT VT; |
2716 | unsigned Alignment; |
2717 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { |
2718 | Ptr = LD->getBasePtr(); |
2719 | VT = LD->getMemoryVT(); |
2720 | Alignment = LD->getAlignment(); |
2721 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { |
2722 | Ptr = ST->getBasePtr(); |
2723 | VT = ST->getMemoryVT(); |
2724 | Alignment = ST->getAlignment(); |
2725 | isLoad = false; |
2726 | } else |
2727 | return false; |
2728 | |
2729 | // Do not generate pre-inc forms for specific loads that feed scalar_to_vector |
2730 | // instructions because we can fold these into a more efficient instruction |
2731 | // instead, (such as LXSD). |
2732 | if (isLoad && usePartialVectorLoads(N, Subtarget)) { |
2733 | return false; |
2734 | } |
2735 | |
2736 | // PowerPC doesn't have preinc load/store instructions for vectors |
2737 | if (VT.isVector()) |
2738 | return false; |
2739 | |
2740 | if (SelectAddressRegReg(Ptr, Base, Offset, DAG)) { |
2741 | // Common code will reject creating a pre-inc form if the base pointer |
2742 | // is a frame index, or if N is a store and the base pointer is either |
2743 | // the same as or a predecessor of the value being stored. Check for |
2744 | // those situations here, and try with swapped Base/Offset instead. |
2745 | bool Swap = false; |
2746 | |
2747 | if (isa<FrameIndexSDNode>(Base) || isa<RegisterSDNode>(Base)) |
2748 | Swap = true; |
2749 | else if (!isLoad) { |
2750 | SDValue Val = cast<StoreSDNode>(N)->getValue(); |
2751 | if (Val == Base || Base.getNode()->isPredecessorOf(Val.getNode())) |
2752 | Swap = true; |
2753 | } |
2754 | |
2755 | if (Swap) |
2756 | std::swap(Base, Offset); |
2757 | |
2758 | AM = ISD::PRE_INC; |
2759 | return true; |
2760 | } |
2761 | |
2762 | // LDU/STU can only handle immediates that are a multiple of 4. |
2763 | if (VT != MVT::i64) { |
2764 | if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, None)) |
2765 | return false; |
2766 | } else { |
2767 | // LDU/STU need an address with at least 4-byte alignment. |
2768 | if (Alignment < 4) |
2769 | return false; |
2770 | |
2771 | if (!SelectAddressRegImm(Ptr, Offset, Base, DAG, Align(4))) |
2772 | return false; |
2773 | } |
2774 | |
2775 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { |
2776 | // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of |
2777 | // sext i32 to i64 when addr mode is r+i. |
2778 | if (LD->getValueType(0) == MVT::i64 && LD->getMemoryVT() == MVT::i32 && |
2779 | LD->getExtensionType() == ISD::SEXTLOAD && |
2780 | isa<ConstantSDNode>(Offset)) |
2781 | return false; |
2782 | } |
2783 | |
2784 | AM = ISD::PRE_INC; |
2785 | return true; |
2786 | } |
2787 | |
2788 | //===----------------------------------------------------------------------===// |
2789 | // LowerOperation implementation |
2790 | //===----------------------------------------------------------------------===// |
2791 | |
2792 | /// Return true if we should reference labels using a PICBase, set the HiOpFlags |
2793 | /// and LoOpFlags to the target MO flags. |
2794 | static void getLabelAccessInfo(bool IsPIC, const PPCSubtarget &Subtarget, |
2795 | unsigned &HiOpFlags, unsigned &LoOpFlags, |
2796 | const GlobalValue *GV = nullptr) { |
2797 | HiOpFlags = PPCII::MO_HA; |
2798 | LoOpFlags = PPCII::MO_LO; |
2799 | |
2800 | // Don't use the pic base if not in PIC relocation model. |
2801 | if (IsPIC) { |
2802 | HiOpFlags |= PPCII::MO_PIC_FLAG; |
2803 | LoOpFlags |= PPCII::MO_PIC_FLAG; |
2804 | } |
2805 | } |
2806 | |
2807 | static SDValue LowerLabelRef(SDValue HiPart, SDValue LoPart, bool isPIC, |
2808 | SelectionDAG &DAG) { |
2809 | SDLoc DL(HiPart); |
2810 | EVT PtrVT = HiPart.getValueType(); |
2811 | SDValue Zero = DAG.getConstant(0, DL, PtrVT); |
2812 | |
2813 | SDValue Hi = DAG.getNode(PPCISD::Hi, DL, PtrVT, HiPart, Zero); |
2814 | SDValue Lo = DAG.getNode(PPCISD::Lo, DL, PtrVT, LoPart, Zero); |
2815 | |
2816 | // With PIC, the first instruction is actually "GR+hi(&G)". |
2817 | if (isPIC) |
2818 | Hi = DAG.getNode(ISD::ADD, DL, PtrVT, |
2819 | DAG.getNode(PPCISD::GlobalBaseReg, DL, PtrVT), Hi); |
2820 | |
2821 | // Generate non-pic code that has direct accesses to the constant pool. |
2822 | // The address of the global is just (hi(&g)+lo(&g)). |
2823 | return DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo); |
2824 | } |
2825 | |
2826 | static void setUsesTOCBasePtr(MachineFunction &MF) { |
2827 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
2828 | FuncInfo->setUsesTOCBasePtr(); |
2829 | } |
2830 | |
2831 | static void setUsesTOCBasePtr(SelectionDAG &DAG) { |
2832 | setUsesTOCBasePtr(DAG.getMachineFunction()); |
2833 | } |
2834 | |
2835 | SDValue PPCTargetLowering::getTOCEntry(SelectionDAG &DAG, const SDLoc &dl, |
2836 | SDValue GA) const { |
2837 | const bool Is64Bit = Subtarget.isPPC64(); |
2838 | EVT VT = Is64Bit ? MVT::i64 : MVT::i32; |
2839 | SDValue Reg = Is64Bit ? DAG.getRegister(PPC::X2, VT) |
2840 | : Subtarget.isAIXABI() |
2841 | ? DAG.getRegister(PPC::R2, VT) |
2842 | : DAG.getNode(PPCISD::GlobalBaseReg, dl, VT); |
2843 | SDValue Ops[] = { GA, Reg }; |
2844 | return DAG.getMemIntrinsicNode( |
2845 | PPCISD::TOC_ENTRY, dl, DAG.getVTList(VT, MVT::Other), Ops, VT, |
2846 | MachinePointerInfo::getGOT(DAG.getMachineFunction()), None, |
2847 | MachineMemOperand::MOLoad); |
2848 | } |
2849 | |
2850 | SDValue PPCTargetLowering::LowerConstantPool(SDValue Op, |
2851 | SelectionDAG &DAG) const { |
2852 | EVT PtrVT = Op.getValueType(); |
2853 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); |
2854 | const Constant *C = CP->getConstVal(); |
2855 | |
2856 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2857 | // The actual address of the GlobalValue is stored in the TOC. |
2858 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2859 | if (Subtarget.isUsingPCRelativeCalls()) { |
2860 | SDLoc DL(CP); |
2861 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2862 | SDValue ConstPool = DAG.getTargetConstantPool( |
2863 | C, Ty, CP->getAlign(), CP->getOffset(), PPCII::MO_PCREL_FLAG); |
2864 | return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, ConstPool); |
2865 | } |
2866 | setUsesTOCBasePtr(DAG); |
2867 | SDValue GA = DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0); |
2868 | return getTOCEntry(DAG, SDLoc(CP), GA); |
2869 | } |
2870 | |
2871 | unsigned MOHiFlag, MOLoFlag; |
2872 | bool IsPIC = isPositionIndependent(); |
2873 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
2874 | |
2875 | if (IsPIC && Subtarget.isSVR4ABI()) { |
2876 | SDValue GA = |
2877 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), PPCII::MO_PIC_FLAG); |
2878 | return getTOCEntry(DAG, SDLoc(CP), GA); |
2879 | } |
2880 | |
2881 | SDValue CPIHi = |
2882 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOHiFlag); |
2883 | SDValue CPILo = |
2884 | DAG.getTargetConstantPool(C, PtrVT, CP->getAlign(), 0, MOLoFlag); |
2885 | return LowerLabelRef(CPIHi, CPILo, IsPIC, DAG); |
2886 | } |
2887 | |
2888 | // For 64-bit PowerPC, prefer the more compact relative encodings. |
2889 | // This trades 32 bits per jump table entry for one or two instructions |
2890 | // on the jump site. |
2891 | unsigned PPCTargetLowering::getJumpTableEncoding() const { |
2892 | if (isJumpTableRelative()) |
2893 | return MachineJumpTableInfo::EK_LabelDifference32; |
2894 | |
2895 | return TargetLowering::getJumpTableEncoding(); |
2896 | } |
2897 | |
2898 | bool PPCTargetLowering::isJumpTableRelative() const { |
2899 | if (UseAbsoluteJumpTables) |
2900 | return false; |
2901 | if (Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2902 | return true; |
2903 | return TargetLowering::isJumpTableRelative(); |
2904 | } |
2905 | |
2906 | SDValue PPCTargetLowering::getPICJumpTableRelocBase(SDValue Table, |
2907 | SelectionDAG &DAG) const { |
2908 | if (!Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2909 | return TargetLowering::getPICJumpTableRelocBase(Table, DAG); |
2910 | |
2911 | switch (getTargetMachine().getCodeModel()) { |
2912 | case CodeModel::Small: |
2913 | case CodeModel::Medium: |
2914 | return TargetLowering::getPICJumpTableRelocBase(Table, DAG); |
2915 | default: |
2916 | return DAG.getNode(PPCISD::GlobalBaseReg, SDLoc(), |
2917 | getPointerTy(DAG.getDataLayout())); |
2918 | } |
2919 | } |
2920 | |
2921 | const MCExpr * |
2922 | PPCTargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, |
2923 | unsigned JTI, |
2924 | MCContext &Ctx) const { |
2925 | if (!Subtarget.isPPC64() || Subtarget.isAIXABI()) |
2926 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); |
2927 | |
2928 | switch (getTargetMachine().getCodeModel()) { |
2929 | case CodeModel::Small: |
2930 | case CodeModel::Medium: |
2931 | return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx); |
2932 | default: |
2933 | return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx); |
2934 | } |
2935 | } |
2936 | |
2937 | SDValue PPCTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { |
2938 | EVT PtrVT = Op.getValueType(); |
2939 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Op); |
2940 | |
2941 | // isUsingPCRelativeCalls() returns true when PCRelative is enabled |
2942 | if (Subtarget.isUsingPCRelativeCalls()) { |
2943 | SDLoc DL(JT); |
2944 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2945 | SDValue GA = |
2946 | DAG.getTargetJumpTable(JT->getIndex(), Ty, PPCII::MO_PCREL_FLAG); |
2947 | SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
2948 | return MatAddr; |
2949 | } |
2950 | |
2951 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2952 | // The actual address of the GlobalValue is stored in the TOC. |
2953 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2954 | setUsesTOCBasePtr(DAG); |
2955 | SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT); |
2956 | return getTOCEntry(DAG, SDLoc(JT), GA); |
2957 | } |
2958 | |
2959 | unsigned MOHiFlag, MOLoFlag; |
2960 | bool IsPIC = isPositionIndependent(); |
2961 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
2962 | |
2963 | if (IsPIC && Subtarget.isSVR4ABI()) { |
2964 | SDValue GA = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, |
2965 | PPCII::MO_PIC_FLAG); |
2966 | return getTOCEntry(DAG, SDLoc(GA), GA); |
2967 | } |
2968 | |
2969 | SDValue JTIHi = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOHiFlag); |
2970 | SDValue JTILo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, MOLoFlag); |
2971 | return LowerLabelRef(JTIHi, JTILo, IsPIC, DAG); |
2972 | } |
2973 | |
2974 | SDValue PPCTargetLowering::LowerBlockAddress(SDValue Op, |
2975 | SelectionDAG &DAG) const { |
2976 | EVT PtrVT = Op.getValueType(); |
2977 | BlockAddressSDNode *BASDN = cast<BlockAddressSDNode>(Op); |
2978 | const BlockAddress *BA = BASDN->getBlockAddress(); |
2979 | |
2980 | // isUsingPCRelativeCalls() returns true when PCRelative is enabled |
2981 | if (Subtarget.isUsingPCRelativeCalls()) { |
2982 | SDLoc DL(BASDN); |
2983 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
2984 | SDValue GA = DAG.getTargetBlockAddress(BA, Ty, BASDN->getOffset(), |
2985 | PPCII::MO_PCREL_FLAG); |
2986 | SDValue MatAddr = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
2987 | return MatAddr; |
2988 | } |
2989 | |
2990 | // 64-bit SVR4 ABI and AIX ABI code are always position-independent. |
2991 | // The actual BlockAddress is stored in the TOC. |
2992 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
2993 | setUsesTOCBasePtr(DAG); |
2994 | SDValue GA = DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset()); |
2995 | return getTOCEntry(DAG, SDLoc(BASDN), GA); |
2996 | } |
2997 | |
2998 | // 32-bit position-independent ELF stores the BlockAddress in the .got. |
2999 | if (Subtarget.is32BitELFABI() && isPositionIndependent()) |
3000 | return getTOCEntry( |
3001 | DAG, SDLoc(BASDN), |
3002 | DAG.getTargetBlockAddress(BA, PtrVT, BASDN->getOffset())); |
3003 | |
3004 | unsigned MOHiFlag, MOLoFlag; |
3005 | bool IsPIC = isPositionIndependent(); |
3006 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag); |
3007 | SDValue TgtBAHi = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOHiFlag); |
3008 | SDValue TgtBALo = DAG.getTargetBlockAddress(BA, PtrVT, 0, MOLoFlag); |
3009 | return LowerLabelRef(TgtBAHi, TgtBALo, IsPIC, DAG); |
3010 | } |
3011 | |
3012 | SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
3013 | SelectionDAG &DAG) const { |
3014 | // FIXME: TLS addresses currently use medium model code sequences, |
3015 | // which is the most useful form. Eventually support for small and |
3016 | // large models could be added if users need it, at the cost of |
3017 | // additional complexity. |
3018 | if (Subtarget.isUsingPCRelativeCalls() && !EnablePPCPCRelTLS) |
3019 | report_fatal_error("Thread local storage is not supported with pc-relative" |
3020 | " addressing - please compile with -mno-pcrel"); |
3021 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); |
3022 | if (DAG.getTarget().useEmulatedTLS()) |
3023 | return LowerToTLSEmulatedModel(GA, DAG); |
3024 | |
3025 | SDLoc dl(GA); |
3026 | const GlobalValue *GV = GA->getGlobal(); |
3027 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3028 | bool is64bit = Subtarget.isPPC64(); |
3029 | const Module *M = DAG.getMachineFunction().getFunction().getParent(); |
3030 | PICLevel::Level picLevel = M->getPICLevel(); |
3031 | |
3032 | const TargetMachine &TM = getTargetMachine(); |
3033 | TLSModel::Model Model = TM.getTLSModel(GV); |
3034 | |
3035 | if (Model == TLSModel::LocalExec) { |
3036 | if (Subtarget.isUsingPCRelativeCalls()) { |
3037 | SDValue TLSReg = DAG.getRegister(PPC::X13, MVT::i64); |
3038 | SDValue TGA = DAG.getTargetGlobalAddress( |
3039 | GV, dl, PtrVT, 0, (PPCII::MO_PCREL_FLAG | PPCII::MO_TPREL_FLAG)); |
3040 | SDValue MatAddr = |
3041 | DAG.getNode(PPCISD::TLS_LOCAL_EXEC_MAT_ADDR, dl, PtrVT, TGA); |
3042 | return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TLSReg, MatAddr); |
3043 | } |
3044 | |
3045 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3046 | PPCII::MO_TPREL_HA); |
3047 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3048 | PPCII::MO_TPREL_LO); |
3049 | SDValue TLSReg = is64bit ? DAG.getRegister(PPC::X13, MVT::i64) |
3050 | : DAG.getRegister(PPC::R2, MVT::i32); |
3051 | |
3052 | SDValue Hi = DAG.getNode(PPCISD::Hi, dl, PtrVT, TGAHi, TLSReg); |
3053 | return DAG.getNode(PPCISD::Lo, dl, PtrVT, TGALo, Hi); |
3054 | } |
3055 | |
3056 | if (Model == TLSModel::InitialExec) { |
3057 | bool IsPCRel = Subtarget.isUsingPCRelativeCalls(); |
3058 | SDValue TGA = DAG.getTargetGlobalAddress( |
3059 | GV, dl, PtrVT, 0, IsPCRel ? PPCII::MO_GOT_TPREL_PCREL_FLAG : 0); |
3060 | SDValue TGATLS = DAG.getTargetGlobalAddress( |
3061 | GV, dl, PtrVT, 0, |
3062 | IsPCRel ? (PPCII::MO_TLS | PPCII::MO_PCREL_FLAG) : PPCII::MO_TLS); |
3063 | SDValue TPOffset; |
3064 | if (IsPCRel) { |
3065 | SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3066 | TPOffset = DAG.getLoad(MVT::i64, dl, DAG.getEntryNode(), MatPCRel, |
3067 | MachinePointerInfo()); |
3068 | } else { |
3069 | SDValue GOTPtr; |
3070 | if (is64bit) { |
3071 | setUsesTOCBasePtr(DAG); |
3072 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3073 | GOTPtr = |
3074 | DAG.getNode(PPCISD::ADDIS_GOT_TPREL_HA, dl, PtrVT, GOTReg, TGA); |
3075 | } else { |
3076 | if (!TM.isPositionIndependent()) |
3077 | GOTPtr = DAG.getNode(PPCISD::PPC32_GOT, dl, PtrVT); |
3078 | else if (picLevel == PICLevel::SmallPIC) |
3079 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3080 | else |
3081 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3082 | } |
3083 | TPOffset = DAG.getNode(PPCISD::LD_GOT_TPREL_L, dl, PtrVT, TGA, GOTPtr); |
3084 | } |
3085 | return DAG.getNode(PPCISD::ADD_TLS, dl, PtrVT, TPOffset, TGATLS); |
3086 | } |
3087 | |
3088 | if (Model == TLSModel::GeneralDynamic) { |
3089 | if (Subtarget.isUsingPCRelativeCalls()) { |
3090 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3091 | PPCII::MO_GOT_TLSGD_PCREL_FLAG); |
3092 | return DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3093 | } |
3094 | |
3095 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0); |
3096 | SDValue GOTPtr; |
3097 | if (is64bit) { |
3098 | setUsesTOCBasePtr(DAG); |
3099 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3100 | GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSGD_HA, dl, PtrVT, |
3101 | GOTReg, TGA); |
3102 | } else { |
3103 | if (picLevel == PICLevel::SmallPIC) |
3104 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3105 | else |
3106 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3107 | } |
3108 | return DAG.getNode(PPCISD::ADDI_TLSGD_L_ADDR, dl, PtrVT, |
3109 | GOTPtr, TGA, TGA); |
3110 | } |
3111 | |
3112 | if (Model == TLSModel::LocalDynamic) { |
3113 | if (Subtarget.isUsingPCRelativeCalls()) { |
3114 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, |
3115 | PPCII::MO_GOT_TLSLD_PCREL_FLAG); |
3116 | SDValue MatPCRel = |
3117 | DAG.getNode(PPCISD::TLS_DYNAMIC_MAT_PCREL_ADDR, dl, PtrVT, TGA); |
3118 | return DAG.getNode(PPCISD::PADDI_DTPREL, dl, PtrVT, MatPCRel, TGA); |
3119 | } |
3120 | |
3121 | SDValue TGA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, 0); |
3122 | SDValue GOTPtr; |
3123 | if (is64bit) { |
3124 | setUsesTOCBasePtr(DAG); |
3125 | SDValue GOTReg = DAG.getRegister(PPC::X2, MVT::i64); |
3126 | GOTPtr = DAG.getNode(PPCISD::ADDIS_TLSLD_HA, dl, PtrVT, |
3127 | GOTReg, TGA); |
3128 | } else { |
3129 | if (picLevel == PICLevel::SmallPIC) |
3130 | GOTPtr = DAG.getNode(PPCISD::GlobalBaseReg, dl, PtrVT); |
3131 | else |
3132 | GOTPtr = DAG.getNode(PPCISD::PPC32_PICGOT, dl, PtrVT); |
3133 | } |
3134 | SDValue TLSAddr = DAG.getNode(PPCISD::ADDI_TLSLD_L_ADDR, dl, |
3135 | PtrVT, GOTPtr, TGA, TGA); |
3136 | SDValue DtvOffsetHi = DAG.getNode(PPCISD::ADDIS_DTPREL_HA, dl, |
3137 | PtrVT, TLSAddr, TGA); |
3138 | return DAG.getNode(PPCISD::ADDI_DTPREL_L, dl, PtrVT, DtvOffsetHi, TGA); |
3139 | } |
3140 | |
3141 | llvm_unreachable("Unknown TLS model!")::llvm::llvm_unreachable_internal("Unknown TLS model!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3141); |
3142 | } |
3143 | |
3144 | SDValue PPCTargetLowering::LowerGlobalAddress(SDValue Op, |
3145 | SelectionDAG &DAG) const { |
3146 | EVT PtrVT = Op.getValueType(); |
3147 | GlobalAddressSDNode *GSDN = cast<GlobalAddressSDNode>(Op); |
3148 | SDLoc DL(GSDN); |
3149 | const GlobalValue *GV = GSDN->getGlobal(); |
3150 | |
3151 | // 64-bit SVR4 ABI & AIX ABI code is always position-independent. |
3152 | // The actual address of the GlobalValue is stored in the TOC. |
3153 | if (Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) { |
3154 | if (Subtarget.isUsingPCRelativeCalls()) { |
3155 | EVT Ty = getPointerTy(DAG.getDataLayout()); |
3156 | if (isAccessedAsGotIndirect(Op)) { |
3157 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(), |
3158 | PPCII::MO_PCREL_FLAG | |
3159 | PPCII::MO_GOT_FLAG); |
3160 | SDValue MatPCRel = DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
3161 | SDValue Load = DAG.getLoad(MVT::i64, DL, DAG.getEntryNode(), MatPCRel, |
3162 | MachinePointerInfo()); |
3163 | return Load; |
3164 | } else { |
3165 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, Ty, GSDN->getOffset(), |
3166 | PPCII::MO_PCREL_FLAG); |
3167 | return DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, Ty, GA); |
3168 | } |
3169 | } |
3170 | setUsesTOCBasePtr(DAG); |
3171 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset()); |
3172 | return getTOCEntry(DAG, DL, GA); |
3173 | } |
3174 | |
3175 | unsigned MOHiFlag, MOLoFlag; |
3176 | bool IsPIC = isPositionIndependent(); |
3177 | getLabelAccessInfo(IsPIC, Subtarget, MOHiFlag, MOLoFlag, GV); |
3178 | |
3179 | if (IsPIC && Subtarget.isSVR4ABI()) { |
3180 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, |
3181 | GSDN->getOffset(), |
3182 | PPCII::MO_PIC_FLAG); |
3183 | return getTOCEntry(DAG, DL, GA); |
3184 | } |
3185 | |
3186 | SDValue GAHi = |
3187 | DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOHiFlag); |
3188 | SDValue GALo = |
3189 | DAG.getTargetGlobalAddress(GV, DL, PtrVT, GSDN->getOffset(), MOLoFlag); |
3190 | |
3191 | return LowerLabelRef(GAHi, GALo, IsPIC, DAG); |
3192 | } |
3193 | |
3194 | SDValue PPCTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
3195 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); |
3196 | SDLoc dl(Op); |
3197 | |
3198 | if (Op.getValueType() == MVT::v2i64) { |
3199 | // When the operands themselves are v2i64 values, we need to do something |
3200 | // special because VSX has no underlying comparison operations for these. |
3201 | if (Op.getOperand(0).getValueType() == MVT::v2i64) { |
3202 | // Equality can be handled by casting to the legal type for Altivec |
3203 | // comparisons, everything else needs to be expanded. |
3204 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
3205 | return DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, |
3206 | DAG.getSetCC(dl, MVT::v4i32, |
3207 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(0)), |
3208 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op.getOperand(1)), |
3209 | CC)); |
3210 | } |
3211 | |
3212 | return SDValue(); |
3213 | } |
3214 | |
3215 | // We handle most of these in the usual way. |
3216 | return Op; |
3217 | } |
3218 | |
3219 | // If we're comparing for equality to zero, expose the fact that this is |
3220 | // implemented as a ctlz/srl pair on ppc, so that the dag combiner can |
3221 | // fold the new nodes. |
3222 | if (SDValue V = lowerCmpEqZeroToCtlzSrl(Op, DAG)) |
3223 | return V; |
3224 | |
3225 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { |
3226 | // Leave comparisons against 0 and -1 alone for now, since they're usually |
3227 | // optimized. FIXME: revisit this when we can custom lower all setcc |
3228 | // optimizations. |
3229 | if (C->isAllOnesValue() || C->isNullValue()) |
3230 | return SDValue(); |
3231 | } |
3232 | |
3233 | // If we have an integer seteq/setne, turn it into a compare against zero |
3234 | // by xor'ing the rhs with the lhs, which is faster than setting a |
3235 | // condition register, reading it back out, and masking the correct bit. The |
3236 | // normal approach here uses sub to do this instead of xor. Using xor exposes |
3237 | // the result to other bit-twiddling opportunities. |
3238 | EVT LHSVT = Op.getOperand(0).getValueType(); |
3239 | if (LHSVT.isInteger() && (CC == ISD::SETEQ || CC == ISD::SETNE)) { |
3240 | EVT VT = Op.getValueType(); |
3241 | SDValue Sub = DAG.getNode(ISD::XOR, dl, LHSVT, Op.getOperand(0), |
3242 | Op.getOperand(1)); |
3243 | return DAG.getSetCC(dl, VT, Sub, DAG.getConstant(0, dl, LHSVT), CC); |
3244 | } |
3245 | return SDValue(); |
3246 | } |
3247 | |
3248 | SDValue PPCTargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const { |
3249 | SDNode *Node = Op.getNode(); |
3250 | EVT VT = Node->getValueType(0); |
3251 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3252 | SDValue InChain = Node->getOperand(0); |
3253 | SDValue VAListPtr = Node->getOperand(1); |
3254 | const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue(); |
3255 | SDLoc dl(Node); |
3256 | |
3257 | assert(!Subtarget.isPPC64() && "LowerVAARG is PPC32 only")((!Subtarget.isPPC64() && "LowerVAARG is PPC32 only") ? static_cast<void> (0) : __assert_fail ("!Subtarget.isPPC64() && \"LowerVAARG is PPC32 only\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3257, __PRETTY_FUNCTION__)); |
3258 | |
3259 | // gpr_index |
3260 | SDValue GprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain, |
3261 | VAListPtr, MachinePointerInfo(SV), MVT::i8); |
3262 | InChain = GprIndex.getValue(1); |
3263 | |
3264 | if (VT == MVT::i64) { |
3265 | // Check if GprIndex is even |
3266 | SDValue GprAnd = DAG.getNode(ISD::AND, dl, MVT::i32, GprIndex, |
3267 | DAG.getConstant(1, dl, MVT::i32)); |
3268 | SDValue CC64 = DAG.getSetCC(dl, MVT::i32, GprAnd, |
3269 | DAG.getConstant(0, dl, MVT::i32), ISD::SETNE); |
3270 | SDValue GprIndexPlusOne = DAG.getNode(ISD::ADD, dl, MVT::i32, GprIndex, |
3271 | DAG.getConstant(1, dl, MVT::i32)); |
3272 | // Align GprIndex to be even if it isn't |
3273 | GprIndex = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC64, GprIndexPlusOne, |
3274 | GprIndex); |
3275 | } |
3276 | |
3277 | // fpr index is 1 byte after gpr |
3278 | SDValue FprPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3279 | DAG.getConstant(1, dl, MVT::i32)); |
3280 | |
3281 | // fpr |
3282 | SDValue FprIndex = DAG.getExtLoad(ISD::ZEXTLOAD, dl, MVT::i32, InChain, |
3283 | FprPtr, MachinePointerInfo(SV), MVT::i8); |
3284 | InChain = FprIndex.getValue(1); |
3285 | |
3286 | SDValue RegSaveAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3287 | DAG.getConstant(8, dl, MVT::i32)); |
3288 | |
3289 | SDValue OverflowAreaPtr = DAG.getNode(ISD::ADD, dl, PtrVT, VAListPtr, |
3290 | DAG.getConstant(4, dl, MVT::i32)); |
3291 | |
3292 | // areas |
3293 | SDValue OverflowArea = |
3294 | DAG.getLoad(MVT::i32, dl, InChain, OverflowAreaPtr, MachinePointerInfo()); |
3295 | InChain = OverflowArea.getValue(1); |
3296 | |
3297 | SDValue RegSaveArea = |
3298 | DAG.getLoad(MVT::i32, dl, InChain, RegSaveAreaPtr, MachinePointerInfo()); |
3299 | InChain = RegSaveArea.getValue(1); |
3300 | |
3301 | // select overflow_area if index > 8 |
3302 | SDValue CC = DAG.getSetCC(dl, MVT::i32, VT.isInteger() ? GprIndex : FprIndex, |
3303 | DAG.getConstant(8, dl, MVT::i32), ISD::SETLT); |
3304 | |
3305 | // adjustment constant gpr_index * 4/8 |
3306 | SDValue RegConstant = DAG.getNode(ISD::MUL, dl, MVT::i32, |
3307 | VT.isInteger() ? GprIndex : FprIndex, |
3308 | DAG.getConstant(VT.isInteger() ? 4 : 8, dl, |
3309 | MVT::i32)); |
3310 | |
3311 | // OurReg = RegSaveArea + RegConstant |
3312 | SDValue OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, RegSaveArea, |
3313 | RegConstant); |
3314 | |
3315 | // Floating types are 32 bytes into RegSaveArea |
3316 | if (VT.isFloatingPoint()) |
3317 | OurReg = DAG.getNode(ISD::ADD, dl, PtrVT, OurReg, |
3318 | DAG.getConstant(32, dl, MVT::i32)); |
3319 | |
3320 | // increase {f,g}pr_index by 1 (or 2 if VT is i64) |
3321 | SDValue IndexPlus1 = DAG.getNode(ISD::ADD, dl, MVT::i32, |
3322 | VT.isInteger() ? GprIndex : FprIndex, |
3323 | DAG.getConstant(VT == MVT::i64 ? 2 : 1, dl, |
3324 | MVT::i32)); |
3325 | |
3326 | InChain = DAG.getTruncStore(InChain, dl, IndexPlus1, |
3327 | VT.isInteger() ? VAListPtr : FprPtr, |
3328 | MachinePointerInfo(SV), MVT::i8); |
3329 | |
3330 | // determine if we should load from reg_save_area or overflow_area |
3331 | SDValue Result = DAG.getNode(ISD::SELECT, dl, PtrVT, CC, OurReg, OverflowArea); |
3332 | |
3333 | // increase overflow_area by 4/8 if gpr/fpr > 8 |
3334 | SDValue OverflowAreaPlusN = DAG.getNode(ISD::ADD, dl, PtrVT, OverflowArea, |
3335 | DAG.getConstant(VT.isInteger() ? 4 : 8, |
3336 | dl, MVT::i32)); |
3337 | |
3338 | OverflowArea = DAG.getNode(ISD::SELECT, dl, MVT::i32, CC, OverflowArea, |
3339 | OverflowAreaPlusN); |
3340 | |
3341 | InChain = DAG.getTruncStore(InChain, dl, OverflowArea, OverflowAreaPtr, |
3342 | MachinePointerInfo(), MVT::i32); |
3343 | |
3344 | return DAG.getLoad(VT, dl, InChain, Result, MachinePointerInfo()); |
3345 | } |
3346 | |
3347 | SDValue PPCTargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const { |
3348 | assert(!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only")((!Subtarget.isPPC64() && "LowerVACOPY is PPC32 only" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.isPPC64() && \"LowerVACOPY is PPC32 only\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3348, __PRETTY_FUNCTION__)); |
3349 | |
3350 | // We have to copy the entire va_list struct: |
3351 | // 2*sizeof(char) + 2 Byte alignment + 2*sizeof(char*) = 12 Byte |
3352 | return DAG.getMemcpy(Op.getOperand(0), Op, Op.getOperand(1), Op.getOperand(2), |
3353 | DAG.getConstant(12, SDLoc(Op), MVT::i32), Align(8), |
3354 | false, true, false, MachinePointerInfo(), |
3355 | MachinePointerInfo()); |
3356 | } |
3357 | |
3358 | SDValue PPCTargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op, |
3359 | SelectionDAG &DAG) const { |
3360 | if (Subtarget.isAIXABI()) |
3361 | report_fatal_error("ADJUST_TRAMPOLINE operation is not supported on AIX."); |
3362 | |
3363 | return Op.getOperand(0); |
3364 | } |
3365 | |
3366 | SDValue PPCTargetLowering::LowerINIT_TRAMPOLINE(SDValue Op, |
3367 | SelectionDAG &DAG) const { |
3368 | if (Subtarget.isAIXABI()) |
3369 | report_fatal_error("INIT_TRAMPOLINE operation is not supported on AIX."); |
3370 | |
3371 | SDValue Chain = Op.getOperand(0); |
3372 | SDValue Trmp = Op.getOperand(1); // trampoline |
3373 | SDValue FPtr = Op.getOperand(2); // nested function |
3374 | SDValue Nest = Op.getOperand(3); // 'nest' parameter value |
3375 | SDLoc dl(Op); |
3376 | |
3377 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
3378 | bool isPPC64 = (PtrVT == MVT::i64); |
3379 | Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext()); |
3380 | |
3381 | TargetLowering::ArgListTy Args; |
3382 | TargetLowering::ArgListEntry Entry; |
3383 | |
3384 | Entry.Ty = IntPtrTy; |
3385 | Entry.Node = Trmp; Args.push_back(Entry); |
3386 | |
3387 | // TrampSize == (isPPC64 ? 48 : 40); |
3388 | Entry.Node = DAG.getConstant(isPPC64 ? 48 : 40, dl, |
3389 | isPPC64 ? MVT::i64 : MVT::i32); |
3390 | Args.push_back(Entry); |
3391 | |
3392 | Entry.Node = FPtr; Args.push_back(Entry); |
3393 | Entry.Node = Nest; Args.push_back(Entry); |
3394 | |
3395 | // Lower to a call to __trampoline_setup(Trmp, TrampSize, FPtr, ctx_reg) |
3396 | TargetLowering::CallLoweringInfo CLI(DAG); |
3397 | CLI.setDebugLoc(dl).setChain(Chain).setLibCallee( |
3398 | CallingConv::C, Type::getVoidTy(*DAG.getContext()), |
3399 | DAG.getExternalSymbol("__trampoline_setup", PtrVT), std::move(Args)); |
3400 | |
3401 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
3402 | return CallResult.second; |
3403 | } |
3404 | |
3405 | SDValue PPCTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
3406 | MachineFunction &MF = DAG.getMachineFunction(); |
3407 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3408 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3409 | |
3410 | SDLoc dl(Op); |
3411 | |
3412 | if (Subtarget.isPPC64() || Subtarget.isAIXABI()) { |
3413 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
3414 | // memory location argument. |
3415 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
3416 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
3417 | return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), |
3418 | MachinePointerInfo(SV)); |
3419 | } |
3420 | |
3421 | // For the 32-bit SVR4 ABI we follow the layout of the va_list struct. |
3422 | // We suppose the given va_list is already allocated. |
3423 | // |
3424 | // typedef struct { |
3425 | // char gpr; /* index into the array of 8 GPRs |
3426 | // * stored in the register save area |
3427 | // * gpr=0 corresponds to r3, |
3428 | // * gpr=1 to r4, etc. |
3429 | // */ |
3430 | // char fpr; /* index into the array of 8 FPRs |
3431 | // * stored in the register save area |
3432 | // * fpr=0 corresponds to f1, |
3433 | // * fpr=1 to f2, etc. |
3434 | // */ |
3435 | // char *overflow_arg_area; |
3436 | // /* location on stack that holds |
3437 | // * the next overflow argument |
3438 | // */ |
3439 | // char *reg_save_area; |
3440 | // /* where r3:r10 and f1:f8 (if saved) |
3441 | // * are stored |
3442 | // */ |
3443 | // } va_list[1]; |
3444 | |
3445 | SDValue ArgGPR = DAG.getConstant(FuncInfo->getVarArgsNumGPR(), dl, MVT::i32); |
3446 | SDValue ArgFPR = DAG.getConstant(FuncInfo->getVarArgsNumFPR(), dl, MVT::i32); |
3447 | SDValue StackOffsetFI = DAG.getFrameIndex(FuncInfo->getVarArgsStackOffset(), |
3448 | PtrVT); |
3449 | SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), |
3450 | PtrVT); |
3451 | |
3452 | uint64_t FrameOffset = PtrVT.getSizeInBits()/8; |
3453 | SDValue ConstFrameOffset = DAG.getConstant(FrameOffset, dl, PtrVT); |
3454 | |
3455 | uint64_t StackOffset = PtrVT.getSizeInBits()/8 - 1; |
3456 | SDValue ConstStackOffset = DAG.getConstant(StackOffset, dl, PtrVT); |
3457 | |
3458 | uint64_t FPROffset = 1; |
3459 | SDValue ConstFPROffset = DAG.getConstant(FPROffset, dl, PtrVT); |
3460 | |
3461 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
3462 | |
3463 | // Store first byte : number of int regs |
3464 | SDValue firstStore = |
3465 | DAG.getTruncStore(Op.getOperand(0), dl, ArgGPR, Op.getOperand(1), |
3466 | MachinePointerInfo(SV), MVT::i8); |
3467 | uint64_t nextOffset = FPROffset; |
3468 | SDValue nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, Op.getOperand(1), |
3469 | ConstFPROffset); |
3470 | |
3471 | // Store second byte : number of float regs |
3472 | SDValue secondStore = |
3473 | DAG.getTruncStore(firstStore, dl, ArgFPR, nextPtr, |
3474 | MachinePointerInfo(SV, nextOffset), MVT::i8); |
3475 | nextOffset += StackOffset; |
3476 | nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstStackOffset); |
3477 | |
3478 | // Store second word : arguments given on stack |
3479 | SDValue thirdStore = DAG.getStore(secondStore, dl, StackOffsetFI, nextPtr, |
3480 | MachinePointerInfo(SV, nextOffset)); |
3481 | nextOffset += FrameOffset; |
3482 | nextPtr = DAG.getNode(ISD::ADD, dl, PtrVT, nextPtr, ConstFrameOffset); |
3483 | |
3484 | // Store third word : arguments given in registers |
3485 | return DAG.getStore(thirdStore, dl, FR, nextPtr, |
3486 | MachinePointerInfo(SV, nextOffset)); |
3487 | } |
3488 | |
3489 | /// FPR - The set of FP registers that should be allocated for arguments |
3490 | /// on Darwin and AIX. |
3491 | static const MCPhysReg FPR[] = {PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, |
3492 | PPC::F6, PPC::F7, PPC::F8, PPC::F9, PPC::F10, |
3493 | PPC::F11, PPC::F12, PPC::F13}; |
3494 | |
3495 | /// CalculateStackSlotSize - Calculates the size reserved for this argument on |
3496 | /// the stack. |
3497 | static unsigned CalculateStackSlotSize(EVT ArgVT, ISD::ArgFlagsTy Flags, |
3498 | unsigned PtrByteSize) { |
3499 | unsigned ArgSize = ArgVT.getStoreSize(); |
3500 | if (Flags.isByVal()) |
3501 | ArgSize = Flags.getByValSize(); |
3502 | |
3503 | // Round up to multiples of the pointer size, except for array members, |
3504 | // which are always packed. |
3505 | if (!Flags.isInConsecutiveRegs()) |
3506 | ArgSize = ((ArgSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
3507 | |
3508 | return ArgSize; |
3509 | } |
3510 | |
3511 | /// CalculateStackSlotAlignment - Calculates the alignment of this argument |
3512 | /// on the stack. |
3513 | static Align CalculateStackSlotAlignment(EVT ArgVT, EVT OrigVT, |
3514 | ISD::ArgFlagsTy Flags, |
3515 | unsigned PtrByteSize) { |
3516 | Align Alignment(PtrByteSize); |
3517 | |
3518 | // Altivec parameters are padded to a 16 byte boundary. |
3519 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
3520 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
3521 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 || |
3522 | ArgVT == MVT::v1i128 || ArgVT == MVT::f128) |
3523 | Alignment = Align(16); |
3524 | |
3525 | // ByVal parameters are aligned as requested. |
3526 | if (Flags.isByVal()) { |
3527 | auto BVAlign = Flags.getNonZeroByValAlign(); |
3528 | if (BVAlign > PtrByteSize) { |
3529 | if (BVAlign.value() % PtrByteSize != 0) |
3530 | llvm_unreachable(::llvm::llvm_unreachable_internal("ByVal alignment is not a multiple of the pointer size" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3531) |
3531 | "ByVal alignment is not a multiple of the pointer size")::llvm::llvm_unreachable_internal("ByVal alignment is not a multiple of the pointer size" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3531); |
3532 | |
3533 | Alignment = BVAlign; |
3534 | } |
3535 | } |
3536 | |
3537 | // Array members are always packed to their original alignment. |
3538 | if (Flags.isInConsecutiveRegs()) { |
3539 | // If the array member was split into multiple registers, the first |
3540 | // needs to be aligned to the size of the full type. (Except for |
3541 | // ppcf128, which is only aligned as its f64 components.) |
3542 | if (Flags.isSplit() && OrigVT != MVT::ppcf128) |
3543 | Alignment = Align(OrigVT.getStoreSize()); |
3544 | else |
3545 | Alignment = Align(ArgVT.getStoreSize()); |
3546 | } |
3547 | |
3548 | return Alignment; |
3549 | } |
3550 | |
3551 | /// CalculateStackSlotUsed - Return whether this argument will use its |
3552 | /// stack slot (instead of being passed in registers). ArgOffset, |
3553 | /// AvailableFPRs, and AvailableVRs must hold the current argument |
3554 | /// position, and will be updated to account for this argument. |
3555 | static bool CalculateStackSlotUsed(EVT ArgVT, EVT OrigVT, ISD::ArgFlagsTy Flags, |
3556 | unsigned PtrByteSize, unsigned LinkageSize, |
3557 | unsigned ParamAreaSize, unsigned &ArgOffset, |
3558 | unsigned &AvailableFPRs, |
3559 | unsigned &AvailableVRs) { |
3560 | bool UseMemory = false; |
3561 | |
3562 | // Respect alignment of argument on the stack. |
3563 | Align Alignment = |
3564 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
3565 | ArgOffset = alignTo(ArgOffset, Alignment); |
3566 | // If there's no space left in the argument save area, we must |
3567 | // use memory (this check also catches zero-sized arguments). |
3568 | if (ArgOffset >= LinkageSize + ParamAreaSize) |
3569 | UseMemory = true; |
3570 | |
3571 | // Allocate argument on the stack. |
3572 | ArgOffset += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
3573 | if (Flags.isInConsecutiveRegsLast()) |
3574 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
3575 | // If we overran the argument save area, we must use memory |
3576 | // (this check catches arguments passed partially in memory) |
3577 | if (ArgOffset > LinkageSize + ParamAreaSize) |
3578 | UseMemory = true; |
3579 | |
3580 | // However, if the argument is actually passed in an FPR or a VR, |
3581 | // we don't use memory after all. |
3582 | if (!Flags.isByVal()) { |
3583 | if (ArgVT == MVT::f32 || ArgVT == MVT::f64) |
3584 | if (AvailableFPRs > 0) { |
3585 | --AvailableFPRs; |
3586 | return false; |
3587 | } |
3588 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
3589 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
3590 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64 || |
3591 | ArgVT == MVT::v1i128 || ArgVT == MVT::f128) |
3592 | if (AvailableVRs > 0) { |
3593 | --AvailableVRs; |
3594 | return false; |
3595 | } |
3596 | } |
3597 | |
3598 | return UseMemory; |
3599 | } |
3600 | |
3601 | /// EnsureStackAlignment - Round stack frame size up from NumBytes to |
3602 | /// ensure minimum alignment required for target. |
3603 | static unsigned EnsureStackAlignment(const PPCFrameLowering *Lowering, |
3604 | unsigned NumBytes) { |
3605 | return alignTo(NumBytes, Lowering->getStackAlign()); |
3606 | } |
3607 | |
3608 | SDValue PPCTargetLowering::LowerFormalArguments( |
3609 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3610 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3611 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3612 | if (Subtarget.isAIXABI()) |
3613 | return LowerFormalArguments_AIX(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3614 | InVals); |
3615 | if (Subtarget.is64BitELFABI()) |
3616 | return LowerFormalArguments_64SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3617 | InVals); |
3618 | if (Subtarget.is32BitELFABI()) |
3619 | return LowerFormalArguments_32SVR4(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3620 | InVals); |
3621 | |
3622 | return LowerFormalArguments_Darwin(Chain, CallConv, isVarArg, Ins, dl, DAG, |
3623 | InVals); |
3624 | } |
3625 | |
3626 | SDValue PPCTargetLowering::LowerFormalArguments_32SVR4( |
3627 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3628 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3629 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3630 | |
3631 | // 32-bit SVR4 ABI Stack Frame Layout: |
3632 | // +-----------------------------------+ |
3633 | // +--> | Back chain | |
3634 | // | +-----------------------------------+ |
3635 | // | | Floating-point register save area | |
3636 | // | +-----------------------------------+ |
3637 | // | | General register save area | |
3638 | // | +-----------------------------------+ |
3639 | // | | CR save word | |
3640 | // | +-----------------------------------+ |
3641 | // | | VRSAVE save word | |
3642 | // | +-----------------------------------+ |
3643 | // | | Alignment padding | |
3644 | // | +-----------------------------------+ |
3645 | // | | Vector register save area | |
3646 | // | +-----------------------------------+ |
3647 | // | | Local variable space | |
3648 | // | +-----------------------------------+ |
3649 | // | | Parameter list area | |
3650 | // | +-----------------------------------+ |
3651 | // | | LR save word | |
3652 | // | +-----------------------------------+ |
3653 | // SP--> +--- | Back chain | |
3654 | // +-----------------------------------+ |
3655 | // |
3656 | // Specifications: |
3657 | // System V Application Binary Interface PowerPC Processor Supplement |
3658 | // AltiVec Technology Programming Interface Manual |
3659 | |
3660 | MachineFunction &MF = DAG.getMachineFunction(); |
3661 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3662 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3663 | |
3664 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3665 | // Potential tail calls could cause overwriting of argument stack slots. |
3666 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
3667 | (CallConv == CallingConv::Fast)); |
3668 | const Align PtrAlign(4); |
3669 | |
3670 | // Assign locations to all of the incoming arguments. |
3671 | SmallVector<CCValAssign, 16> ArgLocs; |
3672 | PPCCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
3673 | *DAG.getContext()); |
3674 | |
3675 | // Reserve space for the linkage area on the stack. |
3676 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
3677 | CCInfo.AllocateStack(LinkageSize, PtrAlign); |
3678 | if (useSoftFloat()) |
3679 | CCInfo.PreAnalyzeFormalArguments(Ins); |
3680 | |
3681 | CCInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4); |
3682 | CCInfo.clearWasPPCF128(); |
3683 | |
3684 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
3685 | CCValAssign &VA = ArgLocs[i]; |
3686 | |
3687 | // Arguments stored in registers. |
3688 | if (VA.isRegLoc()) { |
3689 | const TargetRegisterClass *RC; |
3690 | EVT ValVT = VA.getValVT(); |
3691 | |
3692 | switch (ValVT.getSimpleVT().SimpleTy) { |
3693 | default: |
3694 | llvm_unreachable("ValVT not supported by formal arguments Lowering")::llvm::llvm_unreachable_internal("ValVT not supported by formal arguments Lowering" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3694); |
3695 | case MVT::i1: |
3696 | case MVT::i32: |
3697 | RC = &PPC::GPRCRegClass; |
3698 | break; |
3699 | case MVT::f32: |
3700 | if (Subtarget.hasP8Vector()) |
3701 | RC = &PPC::VSSRCRegClass; |
3702 | else if (Subtarget.hasSPE()) |
3703 | RC = &PPC::GPRCRegClass; |
3704 | else |
3705 | RC = &PPC::F4RCRegClass; |
3706 | break; |
3707 | case MVT::f64: |
3708 | if (Subtarget.hasVSX()) |
3709 | RC = &PPC::VSFRCRegClass; |
3710 | else if (Subtarget.hasSPE()) |
3711 | // SPE passes doubles in GPR pairs. |
3712 | RC = &PPC::GPRCRegClass; |
3713 | else |
3714 | RC = &PPC::F8RCRegClass; |
3715 | break; |
3716 | case MVT::v16i8: |
3717 | case MVT::v8i16: |
3718 | case MVT::v4i32: |
3719 | RC = &PPC::VRRCRegClass; |
3720 | break; |
3721 | case MVT::v4f32: |
3722 | RC = &PPC::VRRCRegClass; |
3723 | break; |
3724 | case MVT::v2f64: |
3725 | case MVT::v2i64: |
3726 | RC = &PPC::VRRCRegClass; |
3727 | break; |
3728 | } |
3729 | |
3730 | SDValue ArgValue; |
3731 | // Transform the arguments stored in physical registers into |
3732 | // virtual ones. |
3733 | if (VA.getLocVT() == MVT::f64 && Subtarget.hasSPE()) { |
3734 | assert(i + 1 < e && "No second half of double precision argument")((i + 1 < e && "No second half of double precision argument" ) ? static_cast<void> (0) : __assert_fail ("i + 1 < e && \"No second half of double precision argument\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3734, __PRETTY_FUNCTION__)); |
3735 | unsigned RegLo = MF.addLiveIn(VA.getLocReg(), RC); |
3736 | unsigned RegHi = MF.addLiveIn(ArgLocs[++i].getLocReg(), RC); |
3737 | SDValue ArgValueLo = DAG.getCopyFromReg(Chain, dl, RegLo, MVT::i32); |
3738 | SDValue ArgValueHi = DAG.getCopyFromReg(Chain, dl, RegHi, MVT::i32); |
3739 | if (!Subtarget.isLittleEndian()) |
3740 | std::swap (ArgValueLo, ArgValueHi); |
3741 | ArgValue = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, ArgValueLo, |
3742 | ArgValueHi); |
3743 | } else { |
3744 | unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); |
3745 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, |
3746 | ValVT == MVT::i1 ? MVT::i32 : ValVT); |
3747 | if (ValVT == MVT::i1) |
3748 | ArgValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgValue); |
3749 | } |
3750 | |
3751 | InVals.push_back(ArgValue); |
3752 | } else { |
3753 | // Argument stored in memory. |
3754 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3754, __PRETTY_FUNCTION__)); |
3755 | |
3756 | // Get the extended size of the argument type in stack |
3757 | unsigned ArgSize = VA.getLocVT().getStoreSize(); |
3758 | // Get the actual size of the argument type |
3759 | unsigned ObjSize = VA.getValVT().getStoreSize(); |
3760 | unsigned ArgOffset = VA.getLocMemOffset(); |
3761 | // Stack objects in PPC32 are right justified. |
3762 | ArgOffset += ArgSize - ObjSize; |
3763 | int FI = MFI.CreateFixedObject(ArgSize, ArgOffset, isImmutable); |
3764 | |
3765 | // Create load nodes to retrieve arguments from the stack. |
3766 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
3767 | InVals.push_back( |
3768 | DAG.getLoad(VA.getValVT(), dl, Chain, FIN, MachinePointerInfo())); |
3769 | } |
3770 | } |
3771 | |
3772 | // Assign locations to all of the incoming aggregate by value arguments. |
3773 | // Aggregates passed by value are stored in the local variable space of the |
3774 | // caller's stack frame, right above the parameter list area. |
3775 | SmallVector<CCValAssign, 16> ByValArgLocs; |
3776 | CCState CCByValInfo(CallConv, isVarArg, DAG.getMachineFunction(), |
3777 | ByValArgLocs, *DAG.getContext()); |
3778 | |
3779 | // Reserve stack space for the allocations in CCInfo. |
3780 | CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign); |
3781 | |
3782 | CCByValInfo.AnalyzeFormalArguments(Ins, CC_PPC32_SVR4_ByVal); |
3783 | |
3784 | // Area that is at least reserved in the caller of this function. |
3785 | unsigned MinReservedArea = CCByValInfo.getNextStackOffset(); |
3786 | MinReservedArea = std::max(MinReservedArea, LinkageSize); |
3787 | |
3788 | // Set the size that is at least reserved in caller of this function. Tail |
3789 | // call optimized function's reserved stack space needs to be aligned so that |
3790 | // taking the difference between two stack areas will result in an aligned |
3791 | // stack. |
3792 | MinReservedArea = |
3793 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
3794 | FuncInfo->setMinReservedArea(MinReservedArea); |
3795 | |
3796 | SmallVector<SDValue, 8> MemOps; |
3797 | |
3798 | // If the function takes variable number of arguments, make a frame index for |
3799 | // the start of the first vararg value... for expansion of llvm.va_start. |
3800 | if (isVarArg) { |
3801 | static const MCPhysReg GPArgRegs[] = { |
3802 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
3803 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
3804 | }; |
3805 | const unsigned NumGPArgRegs = array_lengthof(GPArgRegs); |
3806 | |
3807 | static const MCPhysReg FPArgRegs[] = { |
3808 | PPC::F1, PPC::F2, PPC::F3, PPC::F4, PPC::F5, PPC::F6, PPC::F7, |
3809 | PPC::F8 |
3810 | }; |
3811 | unsigned NumFPArgRegs = array_lengthof(FPArgRegs); |
3812 | |
3813 | if (useSoftFloat() || hasSPE()) |
3814 | NumFPArgRegs = 0; |
3815 | |
3816 | FuncInfo->setVarArgsNumGPR(CCInfo.getFirstUnallocated(GPArgRegs)); |
3817 | FuncInfo->setVarArgsNumFPR(CCInfo.getFirstUnallocated(FPArgRegs)); |
3818 | |
3819 | // Make room for NumGPArgRegs and NumFPArgRegs. |
3820 | int Depth = NumGPArgRegs * PtrVT.getSizeInBits()/8 + |
3821 | NumFPArgRegs * MVT(MVT::f64).getSizeInBits()/8; |
3822 | |
3823 | FuncInfo->setVarArgsStackOffset( |
3824 | MFI.CreateFixedObject(PtrVT.getSizeInBits()/8, |
3825 | CCInfo.getNextStackOffset(), true)); |
3826 | |
3827 | FuncInfo->setVarArgsFrameIndex( |
3828 | MFI.CreateStackObject(Depth, Align(8), false)); |
3829 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
3830 | |
3831 | // The fixed integer arguments of a variadic function are stored to the |
3832 | // VarArgsFrameIndex on the stack so that they may be loaded by |
3833 | // dereferencing the result of va_next. |
3834 | for (unsigned GPRIndex = 0; GPRIndex != NumGPArgRegs; ++GPRIndex) { |
3835 | // Get an existing live-in vreg, or add a new one. |
3836 | unsigned VReg = MF.getRegInfo().getLiveInVirtReg(GPArgRegs[GPRIndex]); |
3837 | if (!VReg) |
3838 | VReg = MF.addLiveIn(GPArgRegs[GPRIndex], &PPC::GPRCRegClass); |
3839 | |
3840 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
3841 | SDValue Store = |
3842 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
3843 | MemOps.push_back(Store); |
3844 | // Increment the address by four for the next argument to store |
3845 | SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT); |
3846 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
3847 | } |
3848 | |
3849 | // FIXME 32-bit SVR4: We only need to save FP argument registers if CR bit 6 |
3850 | // is set. |
3851 | // The double arguments are stored to the VarArgsFrameIndex |
3852 | // on the stack. |
3853 | for (unsigned FPRIndex = 0; FPRIndex != NumFPArgRegs; ++FPRIndex) { |
3854 | // Get an existing live-in vreg, or add a new one. |
3855 | unsigned VReg = MF.getRegInfo().getLiveInVirtReg(FPArgRegs[FPRIndex]); |
3856 | if (!VReg) |
3857 | VReg = MF.addLiveIn(FPArgRegs[FPRIndex], &PPC::F8RCRegClass); |
3858 | |
3859 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::f64); |
3860 | SDValue Store = |
3861 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
3862 | MemOps.push_back(Store); |
3863 | // Increment the address by eight for the next argument to store |
3864 | SDValue PtrOff = DAG.getConstant(MVT(MVT::f64).getSizeInBits()/8, dl, |
3865 | PtrVT); |
3866 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
3867 | } |
3868 | } |
3869 | |
3870 | if (!MemOps.empty()) |
3871 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
3872 | |
3873 | return Chain; |
3874 | } |
3875 | |
3876 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
3877 | // value to MVT::i64 and then truncate to the correct register size. |
3878 | SDValue PPCTargetLowering::extendArgForPPC64(ISD::ArgFlagsTy Flags, |
3879 | EVT ObjectVT, SelectionDAG &DAG, |
3880 | SDValue ArgVal, |
3881 | const SDLoc &dl) const { |
3882 | if (Flags.isSExt()) |
3883 | ArgVal = DAG.getNode(ISD::AssertSext, dl, MVT::i64, ArgVal, |
3884 | DAG.getValueType(ObjectVT)); |
3885 | else if (Flags.isZExt()) |
3886 | ArgVal = DAG.getNode(ISD::AssertZext, dl, MVT::i64, ArgVal, |
3887 | DAG.getValueType(ObjectVT)); |
3888 | |
3889 | return DAG.getNode(ISD::TRUNCATE, dl, ObjectVT, ArgVal); |
3890 | } |
3891 | |
3892 | SDValue PPCTargetLowering::LowerFormalArguments_64SVR4( |
3893 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
3894 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
3895 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3896 | // TODO: add description of PPC stack frame format, or at least some docs. |
3897 | // |
3898 | bool isELFv2ABI = Subtarget.isELFv2ABI(); |
3899 | bool isLittleEndian = Subtarget.isLittleEndian(); |
3900 | MachineFunction &MF = DAG.getMachineFunction(); |
3901 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3902 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
3903 | |
3904 | assert(!(CallConv == CallingConv::Fast && isVarArg) &&((!(CallConv == CallingConv::Fast && isVarArg) && "fastcc not supported on varargs functions") ? static_cast< void> (0) : __assert_fail ("!(CallConv == CallingConv::Fast && isVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3905, __PRETTY_FUNCTION__)) |
3905 | "fastcc not supported on varargs functions")((!(CallConv == CallingConv::Fast && isVarArg) && "fastcc not supported on varargs functions") ? static_cast< void> (0) : __assert_fail ("!(CallConv == CallingConv::Fast && isVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3905, __PRETTY_FUNCTION__)); |
3906 | |
3907 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
3908 | // Potential tail calls could cause overwriting of argument stack slots. |
3909 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
3910 | (CallConv == CallingConv::Fast)); |
3911 | unsigned PtrByteSize = 8; |
3912 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
3913 | |
3914 | static const MCPhysReg GPR[] = { |
3915 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
3916 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
3917 | }; |
3918 | static const MCPhysReg VR[] = { |
3919 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
3920 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
3921 | }; |
3922 | |
3923 | const unsigned Num_GPR_Regs = array_lengthof(GPR); |
3924 | const unsigned Num_FPR_Regs = useSoftFloat() ? 0 : 13; |
3925 | const unsigned Num_VR_Regs = array_lengthof(VR); |
3926 | |
3927 | // Do a first pass over the arguments to determine whether the ABI |
3928 | // guarantees that our caller has allocated the parameter save area |
3929 | // on its stack frame. In the ELFv1 ABI, this is always the case; |
3930 | // in the ELFv2 ABI, it is true if this is a vararg function or if |
3931 | // any parameter is located in a stack slot. |
3932 | |
3933 | bool HasParameterArea = !isELFv2ABI || isVarArg; |
3934 | unsigned ParamAreaSize = Num_GPR_Regs * PtrByteSize; |
3935 | unsigned NumBytes = LinkageSize; |
3936 | unsigned AvailableFPRs = Num_FPR_Regs; |
3937 | unsigned AvailableVRs = Num_VR_Regs; |
3938 | for (unsigned i = 0, e = Ins.size(); i != e; ++i) { |
3939 | if (Ins[i].Flags.isNest()) |
3940 | continue; |
3941 | |
3942 | if (CalculateStackSlotUsed(Ins[i].VT, Ins[i].ArgVT, Ins[i].Flags, |
3943 | PtrByteSize, LinkageSize, ParamAreaSize, |
3944 | NumBytes, AvailableFPRs, AvailableVRs)) |
3945 | HasParameterArea = true; |
3946 | } |
3947 | |
3948 | // Add DAG nodes to load the arguments or copy them out of registers. On |
3949 | // entry to a function on PPC, the arguments start after the linkage area, |
3950 | // although the first ones are often in registers. |
3951 | |
3952 | unsigned ArgOffset = LinkageSize; |
3953 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
3954 | SmallVector<SDValue, 8> MemOps; |
3955 | Function::const_arg_iterator FuncArg = MF.getFunction().arg_begin(); |
3956 | unsigned CurArgIdx = 0; |
3957 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) { |
3958 | SDValue ArgVal; |
3959 | bool needsLoad = false; |
3960 | EVT ObjectVT = Ins[ArgNo].VT; |
3961 | EVT OrigVT = Ins[ArgNo].ArgVT; |
3962 | unsigned ObjSize = ObjectVT.getStoreSize(); |
3963 | unsigned ArgSize = ObjSize; |
3964 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
3965 | if (Ins[ArgNo].isOrigArg()) { |
3966 | std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx); |
3967 | CurArgIdx = Ins[ArgNo].getOrigArgIndex(); |
3968 | } |
3969 | // We re-align the argument offset for each argument, except when using the |
3970 | // fast calling convention, when we need to make sure we do that only when |
3971 | // we'll actually use a stack slot. |
3972 | unsigned CurArgOffset; |
3973 | Align Alignment; |
3974 | auto ComputeArgOffset = [&]() { |
3975 | /* Respect alignment of argument on the stack. */ |
3976 | Alignment = |
3977 | CalculateStackSlotAlignment(ObjectVT, OrigVT, Flags, PtrByteSize); |
3978 | ArgOffset = alignTo(ArgOffset, Alignment); |
3979 | CurArgOffset = ArgOffset; |
3980 | }; |
3981 | |
3982 | if (CallConv != CallingConv::Fast) { |
3983 | ComputeArgOffset(); |
3984 | |
3985 | /* Compute GPR index associated with argument offset. */ |
3986 | GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
3987 | GPR_idx = std::min(GPR_idx, Num_GPR_Regs); |
3988 | } |
3989 | |
3990 | // FIXME the codegen can be much improved in some cases. |
3991 | // We do not have to keep everything in memory. |
3992 | if (Flags.isByVal()) { |
3993 | assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit")((Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit" ) ? static_cast<void> (0) : __assert_fail ("Ins[ArgNo].isOrigArg() && \"Byval arguments cannot be implicit\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 3993, __PRETTY_FUNCTION__)); |
3994 | |
3995 | if (CallConv == CallingConv::Fast) |
3996 | ComputeArgOffset(); |
3997 | |
3998 | // ObjSize is the true size, ArgSize rounded up to multiple of registers. |
3999 | ObjSize = Flags.getByValSize(); |
4000 | ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4001 | // Empty aggregate parameters do not take up registers. Examples: |
4002 | // struct { } a; |
4003 | // union { } b; |
4004 | // int c[0]; |
4005 | // etc. However, we have to provide a place-holder in InVals, so |
4006 | // pretend we have an 8-byte item at the current address for that |
4007 | // purpose. |
4008 | if (!ObjSize) { |
4009 | int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true); |
4010 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4011 | InVals.push_back(FIN); |
4012 | continue; |
4013 | } |
4014 | |
4015 | // Create a stack object covering all stack doublewords occupied |
4016 | // by the argument. If the argument is (fully or partially) on |
4017 | // the stack, or if the argument is fully in registers but the |
4018 | // caller has allocated the parameter save anyway, we can refer |
4019 | // directly to the caller's stack frame. Otherwise, create a |
4020 | // local copy in our own frame. |
4021 | int FI; |
4022 | if (HasParameterArea || |
4023 | ArgSize + ArgOffset > LinkageSize + Num_GPR_Regs * PtrByteSize) |
4024 | FI = MFI.CreateFixedObject(ArgSize, ArgOffset, false, true); |
4025 | else |
4026 | FI = MFI.CreateStackObject(ArgSize, Alignment, false); |
4027 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4028 | |
4029 | // Handle aggregates smaller than 8 bytes. |
4030 | if (ObjSize < PtrByteSize) { |
4031 | // The value of the object is its address, which differs from the |
4032 | // address of the enclosing doubleword on big-endian systems. |
4033 | SDValue Arg = FIN; |
4034 | if (!isLittleEndian) { |
4035 | SDValue ArgOff = DAG.getConstant(PtrByteSize - ObjSize, dl, PtrVT); |
4036 | Arg = DAG.getNode(ISD::ADD, dl, ArgOff.getValueType(), Arg, ArgOff); |
4037 | } |
4038 | InVals.push_back(Arg); |
4039 | |
4040 | if (GPR_idx != Num_GPR_Regs) { |
4041 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4042 | FuncInfo->addLiveInAttr(VReg, Flags); |
4043 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4044 | SDValue Store; |
4045 | |
4046 | if (ObjSize==1 || ObjSize==2 || ObjSize==4) { |
4047 | EVT ObjType = (ObjSize == 1 ? MVT::i8 : |
4048 | (ObjSize == 2 ? MVT::i16 : MVT::i32)); |
4049 | Store = DAG.getTruncStore(Val.getValue(1), dl, Val, Arg, |
4050 | MachinePointerInfo(&*FuncArg), ObjType); |
4051 | } else { |
4052 | // For sizes that don't fit a truncating store (3, 5, 6, 7), |
4053 | // store the whole register as-is to the parameter save area |
4054 | // slot. |
4055 | Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, |
4056 | MachinePointerInfo(&*FuncArg)); |
4057 | } |
4058 | |
4059 | MemOps.push_back(Store); |
4060 | } |
4061 | // Whether we copied from a register or not, advance the offset |
4062 | // into the parameter save area by a full doubleword. |
4063 | ArgOffset += PtrByteSize; |
4064 | continue; |
4065 | } |
4066 | |
4067 | // The value of the object is its address, which is the address of |
4068 | // its first stack doubleword. |
4069 | InVals.push_back(FIN); |
4070 | |
4071 | // Store whatever pieces of the object are in registers to memory. |
4072 | for (unsigned j = 0; j < ArgSize; j += PtrByteSize) { |
4073 | if (GPR_idx == Num_GPR_Regs) |
4074 | break; |
4075 | |
4076 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4077 | FuncInfo->addLiveInAttr(VReg, Flags); |
4078 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4079 | SDValue Addr = FIN; |
4080 | if (j) { |
4081 | SDValue Off = DAG.getConstant(j, dl, PtrVT); |
4082 | Addr = DAG.getNode(ISD::ADD, dl, Off.getValueType(), Addr, Off); |
4083 | } |
4084 | SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, Addr, |
4085 | MachinePointerInfo(&*FuncArg, j)); |
4086 | MemOps.push_back(Store); |
4087 | ++GPR_idx; |
4088 | } |
4089 | ArgOffset += ArgSize; |
4090 | continue; |
4091 | } |
4092 | |
4093 | switch (ObjectVT.getSimpleVT().SimpleTy) { |
4094 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4094); |
4095 | case MVT::i1: |
4096 | case MVT::i32: |
4097 | case MVT::i64: |
4098 | if (Flags.isNest()) { |
4099 | // The 'nest' parameter, if any, is passed in R11. |
4100 | unsigned VReg = MF.addLiveIn(PPC::X11, &PPC::G8RCRegClass); |
4101 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4102 | |
4103 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4104 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4105 | |
4106 | break; |
4107 | } |
4108 | |
4109 | // These can be scalar arguments or elements of an integer array type |
4110 | // passed directly. Clang may use those instead of "byval" aggregate |
4111 | // types to avoid forcing arguments to memory unnecessarily. |
4112 | if (GPR_idx != Num_GPR_Regs) { |
4113 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4114 | FuncInfo->addLiveInAttr(VReg, Flags); |
4115 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4116 | |
4117 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4118 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
4119 | // value to MVT::i64 and then truncate to the correct register size. |
4120 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4121 | } else { |
4122 | if (CallConv == CallingConv::Fast) |
4123 | ComputeArgOffset(); |
4124 | |
4125 | needsLoad = true; |
4126 | ArgSize = PtrByteSize; |
4127 | } |
4128 | if (CallConv != CallingConv::Fast || needsLoad) |
4129 | ArgOffset += 8; |
4130 | break; |
4131 | |
4132 | case MVT::f32: |
4133 | case MVT::f64: |
4134 | // These can be scalar arguments or elements of a float array type |
4135 | // passed directly. The latter are used to implement ELFv2 homogenous |
4136 | // float aggregates. |
4137 | if (FPR_idx != Num_FPR_Regs) { |
4138 | unsigned VReg; |
4139 | |
4140 | if (ObjectVT == MVT::f32) |
4141 | VReg = MF.addLiveIn(FPR[FPR_idx], |
4142 | Subtarget.hasP8Vector() |
4143 | ? &PPC::VSSRCRegClass |
4144 | : &PPC::F4RCRegClass); |
4145 | else |
4146 | VReg = MF.addLiveIn(FPR[FPR_idx], Subtarget.hasVSX() |
4147 | ? &PPC::VSFRCRegClass |
4148 | : &PPC::F8RCRegClass); |
4149 | |
4150 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4151 | ++FPR_idx; |
4152 | } else if (GPR_idx != Num_GPR_Regs && CallConv != CallingConv::Fast) { |
4153 | // FIXME: We may want to re-enable this for CallingConv::Fast on the P8 |
4154 | // once we support fp <-> gpr moves. |
4155 | |
4156 | // This can only ever happen in the presence of f32 array types, |
4157 | // since otherwise we never run out of FPRs before running out |
4158 | // of GPRs. |
4159 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx++], &PPC::G8RCRegClass); |
4160 | FuncInfo->addLiveInAttr(VReg, Flags); |
4161 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4162 | |
4163 | if (ObjectVT == MVT::f32) { |
4164 | if ((ArgOffset % PtrByteSize) == (isLittleEndian ? 4 : 0)) |
4165 | ArgVal = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgVal, |
4166 | DAG.getConstant(32, dl, MVT::i32)); |
4167 | ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, ArgVal); |
4168 | } |
4169 | |
4170 | ArgVal = DAG.getNode(ISD::BITCAST, dl, ObjectVT, ArgVal); |
4171 | } else { |
4172 | if (CallConv == CallingConv::Fast) |
4173 | ComputeArgOffset(); |
4174 | |
4175 | needsLoad = true; |
4176 | } |
4177 | |
4178 | // When passing an array of floats, the array occupies consecutive |
4179 | // space in the argument area; only round up to the next doubleword |
4180 | // at the end of the array. Otherwise, each float takes 8 bytes. |
4181 | if (CallConv != CallingConv::Fast || needsLoad) { |
4182 | ArgSize = Flags.isInConsecutiveRegs() ? ObjSize : PtrByteSize; |
4183 | ArgOffset += ArgSize; |
4184 | if (Flags.isInConsecutiveRegsLast()) |
4185 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4186 | } |
4187 | break; |
4188 | case MVT::v4f32: |
4189 | case MVT::v4i32: |
4190 | case MVT::v8i16: |
4191 | case MVT::v16i8: |
4192 | case MVT::v2f64: |
4193 | case MVT::v2i64: |
4194 | case MVT::v1i128: |
4195 | case MVT::f128: |
4196 | // These can be scalar arguments or elements of a vector array type |
4197 | // passed directly. The latter are used to implement ELFv2 homogenous |
4198 | // vector aggregates. |
4199 | if (VR_idx != Num_VR_Regs) { |
4200 | unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass); |
4201 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4202 | ++VR_idx; |
4203 | } else { |
4204 | if (CallConv == CallingConv::Fast) |
4205 | ComputeArgOffset(); |
4206 | needsLoad = true; |
4207 | } |
4208 | if (CallConv != CallingConv::Fast || needsLoad) |
4209 | ArgOffset += 16; |
4210 | break; |
4211 | } |
4212 | |
4213 | // We need to load the argument to a virtual register if we determined |
4214 | // above that we ran out of physical registers of the appropriate type. |
4215 | if (needsLoad) { |
4216 | if (ObjSize < ArgSize && !isLittleEndian) |
4217 | CurArgOffset += ArgSize - ObjSize; |
4218 | int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, isImmutable); |
4219 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4220 | ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo()); |
4221 | } |
4222 | |
4223 | InVals.push_back(ArgVal); |
4224 | } |
4225 | |
4226 | // Area that is at least reserved in the caller of this function. |
4227 | unsigned MinReservedArea; |
4228 | if (HasParameterArea) |
4229 | MinReservedArea = std::max(ArgOffset, LinkageSize + 8 * PtrByteSize); |
4230 | else |
4231 | MinReservedArea = LinkageSize; |
4232 | |
4233 | // Set the size that is at least reserved in caller of this function. Tail |
4234 | // call optimized functions' reserved stack space needs to be aligned so that |
4235 | // taking the difference between two stack areas will result in an aligned |
4236 | // stack. |
4237 | MinReservedArea = |
4238 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
4239 | FuncInfo->setMinReservedArea(MinReservedArea); |
4240 | |
4241 | // If the function takes variable number of arguments, make a frame index for |
4242 | // the start of the first vararg value... for expansion of llvm.va_start. |
4243 | // On ELFv2ABI spec, it writes: |
4244 | // C programs that are intended to be *portable* across different compilers |
4245 | // and architectures must use the header file <stdarg.h> to deal with variable |
4246 | // argument lists. |
4247 | if (isVarArg && MFI.hasVAStart()) { |
4248 | int Depth = ArgOffset; |
4249 | |
4250 | FuncInfo->setVarArgsFrameIndex( |
4251 | MFI.CreateFixedObject(PtrByteSize, Depth, true)); |
4252 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
4253 | |
4254 | // If this function is vararg, store any remaining integer argument regs |
4255 | // to their spots on the stack so that they may be loaded by dereferencing |
4256 | // the result of va_next. |
4257 | for (GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
4258 | GPR_idx < Num_GPR_Regs; ++GPR_idx) { |
4259 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4260 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4261 | SDValue Store = |
4262 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
4263 | MemOps.push_back(Store); |
4264 | // Increment the address by four for the next argument to store |
4265 | SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT); |
4266 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
4267 | } |
4268 | } |
4269 | |
4270 | if (!MemOps.empty()) |
4271 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
4272 | |
4273 | return Chain; |
4274 | } |
4275 | |
4276 | SDValue PPCTargetLowering::LowerFormalArguments_Darwin( |
4277 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
4278 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
4279 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
4280 | // TODO: add description of PPC stack frame format, or at least some docs. |
4281 | // |
4282 | MachineFunction &MF = DAG.getMachineFunction(); |
4283 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4284 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
4285 | |
4286 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
4287 | bool isPPC64 = PtrVT == MVT::i64; |
4288 | // Potential tail calls could cause overwriting of argument stack slots. |
4289 | bool isImmutable = !(getTargetMachine().Options.GuaranteedTailCallOpt && |
4290 | (CallConv == CallingConv::Fast)); |
4291 | unsigned PtrByteSize = isPPC64 ? 8 : 4; |
4292 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
4293 | unsigned ArgOffset = LinkageSize; |
4294 | // Area that is at least reserved in caller of this function. |
4295 | unsigned MinReservedArea = ArgOffset; |
4296 | |
4297 | static const MCPhysReg GPR_32[] = { // 32-bit registers. |
4298 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
4299 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
4300 | }; |
4301 | static const MCPhysReg GPR_64[] = { // 64-bit registers. |
4302 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
4303 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
4304 | }; |
4305 | static const MCPhysReg VR[] = { |
4306 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
4307 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
4308 | }; |
4309 | |
4310 | const unsigned Num_GPR_Regs = array_lengthof(GPR_32); |
4311 | const unsigned Num_FPR_Regs = useSoftFloat() ? 0 : 13; |
4312 | const unsigned Num_VR_Regs = array_lengthof( VR); |
4313 | |
4314 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
4315 | |
4316 | const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32; |
4317 | |
4318 | // In 32-bit non-varargs functions, the stack space for vectors is after the |
4319 | // stack space for non-vectors. We do not use this space unless we have |
4320 | // too many vectors to fit in registers, something that only occurs in |
4321 | // constructed examples:), but we have to walk the arglist to figure |
4322 | // that out...for the pathological case, compute VecArgOffset as the |
4323 | // start of the vector parameter area. Computing VecArgOffset is the |
4324 | // entire point of the following loop. |
4325 | unsigned VecArgOffset = ArgOffset; |
4326 | if (!isVarArg && !isPPC64) { |
4327 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; |
4328 | ++ArgNo) { |
4329 | EVT ObjectVT = Ins[ArgNo].VT; |
4330 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
4331 | |
4332 | if (Flags.isByVal()) { |
4333 | // ObjSize is the true size, ArgSize rounded up to multiple of regs. |
4334 | unsigned ObjSize = Flags.getByValSize(); |
4335 | unsigned ArgSize = |
4336 | ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4337 | VecArgOffset += ArgSize; |
4338 | continue; |
4339 | } |
4340 | |
4341 | switch(ObjectVT.getSimpleVT().SimpleTy) { |
4342 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4342); |
4343 | case MVT::i1: |
4344 | case MVT::i32: |
4345 | case MVT::f32: |
4346 | VecArgOffset += 4; |
4347 | break; |
4348 | case MVT::i64: // PPC64 |
4349 | case MVT::f64: |
4350 | // FIXME: We are guaranteed to be !isPPC64 at this point. |
4351 | // Does MVT::i64 apply? |
4352 | VecArgOffset += 8; |
4353 | break; |
4354 | case MVT::v4f32: |
4355 | case MVT::v4i32: |
4356 | case MVT::v8i16: |
4357 | case MVT::v16i8: |
4358 | // Nothing to do, we're only looking at Nonvector args here. |
4359 | break; |
4360 | } |
4361 | } |
4362 | } |
4363 | // We've found where the vector parameter area in memory is. Skip the |
4364 | // first 12 parameters; these don't use that memory. |
4365 | VecArgOffset = ((VecArgOffset+15)/16)*16; |
4366 | VecArgOffset += 12*16; |
4367 | |
4368 | // Add DAG nodes to load the arguments or copy them out of registers. On |
4369 | // entry to a function on PPC, the arguments start after the linkage area, |
4370 | // although the first ones are often in registers. |
4371 | |
4372 | SmallVector<SDValue, 8> MemOps; |
4373 | unsigned nAltivecParamsAtEnd = 0; |
4374 | Function::const_arg_iterator FuncArg = MF.getFunction().arg_begin(); |
4375 | unsigned CurArgIdx = 0; |
4376 | for (unsigned ArgNo = 0, e = Ins.size(); ArgNo != e; ++ArgNo) { |
4377 | SDValue ArgVal; |
4378 | bool needsLoad = false; |
4379 | EVT ObjectVT = Ins[ArgNo].VT; |
4380 | unsigned ObjSize = ObjectVT.getSizeInBits()/8; |
4381 | unsigned ArgSize = ObjSize; |
4382 | ISD::ArgFlagsTy Flags = Ins[ArgNo].Flags; |
4383 | if (Ins[ArgNo].isOrigArg()) { |
4384 | std::advance(FuncArg, Ins[ArgNo].getOrigArgIndex() - CurArgIdx); |
4385 | CurArgIdx = Ins[ArgNo].getOrigArgIndex(); |
4386 | } |
4387 | unsigned CurArgOffset = ArgOffset; |
4388 | |
4389 | // Varargs or 64 bit Altivec parameters are padded to a 16 byte boundary. |
4390 | if (ObjectVT==MVT::v4f32 || ObjectVT==MVT::v4i32 || |
4391 | ObjectVT==MVT::v8i16 || ObjectVT==MVT::v16i8) { |
4392 | if (isVarArg || isPPC64) { |
4393 | MinReservedArea = ((MinReservedArea+15)/16)*16; |
4394 | MinReservedArea += CalculateStackSlotSize(ObjectVT, |
4395 | Flags, |
4396 | PtrByteSize); |
4397 | } else nAltivecParamsAtEnd++; |
4398 | } else |
4399 | // Calculate min reserved area. |
4400 | MinReservedArea += CalculateStackSlotSize(Ins[ArgNo].VT, |
4401 | Flags, |
4402 | PtrByteSize); |
4403 | |
4404 | // FIXME the codegen can be much improved in some cases. |
4405 | // We do not have to keep everything in memory. |
4406 | if (Flags.isByVal()) { |
4407 | assert(Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit")((Ins[ArgNo].isOrigArg() && "Byval arguments cannot be implicit" ) ? static_cast<void> (0) : __assert_fail ("Ins[ArgNo].isOrigArg() && \"Byval arguments cannot be implicit\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4407, __PRETTY_FUNCTION__)); |
4408 | |
4409 | // ObjSize is the true size, ArgSize rounded up to multiple of registers. |
4410 | ObjSize = Flags.getByValSize(); |
4411 | ArgSize = ((ObjSize + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
4412 | // Objects of size 1 and 2 are right justified, everything else is |
4413 | // left justified. This means the memory address is adjusted forwards. |
4414 | if (ObjSize==1 || ObjSize==2) { |
4415 | CurArgOffset = CurArgOffset + (4 - ObjSize); |
4416 | } |
4417 | // The value of the object is its address. |
4418 | int FI = MFI.CreateFixedObject(ObjSize, CurArgOffset, false, true); |
4419 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4420 | InVals.push_back(FIN); |
4421 | if (ObjSize==1 || ObjSize==2) { |
4422 | if (GPR_idx != Num_GPR_Regs) { |
4423 | unsigned VReg; |
4424 | if (isPPC64) |
4425 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4426 | else |
4427 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4428 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4429 | EVT ObjType = ObjSize == 1 ? MVT::i8 : MVT::i16; |
4430 | SDValue Store = |
4431 | DAG.getTruncStore(Val.getValue(1), dl, Val, FIN, |
4432 | MachinePointerInfo(&*FuncArg), ObjType); |
4433 | MemOps.push_back(Store); |
4434 | ++GPR_idx; |
4435 | } |
4436 | |
4437 | ArgOffset += PtrByteSize; |
4438 | |
4439 | continue; |
4440 | } |
4441 | for (unsigned j = 0; j < ArgSize; j += PtrByteSize) { |
4442 | // Store whatever pieces of the object are in registers |
4443 | // to memory. ArgOffset will be the address of the beginning |
4444 | // of the object. |
4445 | if (GPR_idx != Num_GPR_Regs) { |
4446 | unsigned VReg; |
4447 | if (isPPC64) |
4448 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4449 | else |
4450 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4451 | int FI = MFI.CreateFixedObject(PtrByteSize, ArgOffset, true); |
4452 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4453 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4454 | SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, |
4455 | MachinePointerInfo(&*FuncArg, j)); |
4456 | MemOps.push_back(Store); |
4457 | ++GPR_idx; |
4458 | ArgOffset += PtrByteSize; |
4459 | } else { |
4460 | ArgOffset += ArgSize - (ArgOffset-CurArgOffset); |
4461 | break; |
4462 | } |
4463 | } |
4464 | continue; |
4465 | } |
4466 | |
4467 | switch (ObjectVT.getSimpleVT().SimpleTy) { |
4468 | default: llvm_unreachable("Unhandled argument type!")::llvm::llvm_unreachable_internal("Unhandled argument type!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4468); |
4469 | case MVT::i1: |
4470 | case MVT::i32: |
4471 | if (!isPPC64) { |
4472 | if (GPR_idx != Num_GPR_Regs) { |
4473 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4474 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); |
4475 | |
4476 | if (ObjectVT == MVT::i1) |
4477 | ArgVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, ArgVal); |
4478 | |
4479 | ++GPR_idx; |
4480 | } else { |
4481 | needsLoad = true; |
4482 | ArgSize = PtrByteSize; |
4483 | } |
4484 | // All int arguments reserve stack space in the Darwin ABI. |
4485 | ArgOffset += PtrByteSize; |
4486 | break; |
4487 | } |
4488 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
4489 | case MVT::i64: // PPC64 |
4490 | if (GPR_idx != Num_GPR_Regs) { |
4491 | unsigned VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4492 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64); |
4493 | |
4494 | if (ObjectVT == MVT::i32 || ObjectVT == MVT::i1) |
4495 | // PPC64 passes i8, i16, and i32 values in i64 registers. Promote |
4496 | // value to MVT::i64 and then truncate to the correct register size. |
4497 | ArgVal = extendArgForPPC64(Flags, ObjectVT, DAG, ArgVal, dl); |
4498 | |
4499 | ++GPR_idx; |
4500 | } else { |
4501 | needsLoad = true; |
4502 | ArgSize = PtrByteSize; |
4503 | } |
4504 | // All int arguments reserve stack space in the Darwin ABI. |
4505 | ArgOffset += 8; |
4506 | break; |
4507 | |
4508 | case MVT::f32: |
4509 | case MVT::f64: |
4510 | // Every 4 bytes of argument space consumes one of the GPRs available for |
4511 | // argument passing. |
4512 | if (GPR_idx != Num_GPR_Regs) { |
4513 | ++GPR_idx; |
4514 | if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64) |
4515 | ++GPR_idx; |
4516 | } |
4517 | if (FPR_idx != Num_FPR_Regs) { |
4518 | unsigned VReg; |
4519 | |
4520 | if (ObjectVT == MVT::f32) |
4521 | VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F4RCRegClass); |
4522 | else |
4523 | VReg = MF.addLiveIn(FPR[FPR_idx], &PPC::F8RCRegClass); |
4524 | |
4525 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4526 | ++FPR_idx; |
4527 | } else { |
4528 | needsLoad = true; |
4529 | } |
4530 | |
4531 | // All FP arguments reserve stack space in the Darwin ABI. |
4532 | ArgOffset += isPPC64 ? 8 : ObjSize; |
4533 | break; |
4534 | case MVT::v4f32: |
4535 | case MVT::v4i32: |
4536 | case MVT::v8i16: |
4537 | case MVT::v16i8: |
4538 | // Note that vector arguments in registers don't reserve stack space, |
4539 | // except in varargs functions. |
4540 | if (VR_idx != Num_VR_Regs) { |
4541 | unsigned VReg = MF.addLiveIn(VR[VR_idx], &PPC::VRRCRegClass); |
4542 | ArgVal = DAG.getCopyFromReg(Chain, dl, VReg, ObjectVT); |
4543 | if (isVarArg) { |
4544 | while ((ArgOffset % 16) != 0) { |
4545 | ArgOffset += PtrByteSize; |
4546 | if (GPR_idx != Num_GPR_Regs) |
4547 | GPR_idx++; |
4548 | } |
4549 | ArgOffset += 16; |
4550 | GPR_idx = std::min(GPR_idx+4, Num_GPR_Regs); // FIXME correct for ppc64? |
4551 | } |
4552 | ++VR_idx; |
4553 | } else { |
4554 | if (!isVarArg && !isPPC64) { |
4555 | // Vectors go after all the nonvectors. |
4556 | CurArgOffset = VecArgOffset; |
4557 | VecArgOffset += 16; |
4558 | } else { |
4559 | // Vectors are aligned. |
4560 | ArgOffset = ((ArgOffset+15)/16)*16; |
4561 | CurArgOffset = ArgOffset; |
4562 | ArgOffset += 16; |
4563 | } |
4564 | needsLoad = true; |
4565 | } |
4566 | break; |
4567 | } |
4568 | |
4569 | // We need to load the argument to a virtual register if we determined above |
4570 | // that we ran out of physical registers of the appropriate type. |
4571 | if (needsLoad) { |
4572 | int FI = MFI.CreateFixedObject(ObjSize, |
4573 | CurArgOffset + (ArgSize - ObjSize), |
4574 | isImmutable); |
4575 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
4576 | ArgVal = DAG.getLoad(ObjectVT, dl, Chain, FIN, MachinePointerInfo()); |
4577 | } |
4578 | |
4579 | InVals.push_back(ArgVal); |
4580 | } |
4581 | |
4582 | // Allow for Altivec parameters at the end, if needed. |
4583 | if (nAltivecParamsAtEnd) { |
4584 | MinReservedArea = ((MinReservedArea+15)/16)*16; |
4585 | MinReservedArea += 16*nAltivecParamsAtEnd; |
4586 | } |
4587 | |
4588 | // Area that is at least reserved in the caller of this function. |
4589 | MinReservedArea = std::max(MinReservedArea, LinkageSize + 8 * PtrByteSize); |
4590 | |
4591 | // Set the size that is at least reserved in caller of this function. Tail |
4592 | // call optimized functions' reserved stack space needs to be aligned so that |
4593 | // taking the difference between two stack areas will result in an aligned |
4594 | // stack. |
4595 | MinReservedArea = |
4596 | EnsureStackAlignment(Subtarget.getFrameLowering(), MinReservedArea); |
4597 | FuncInfo->setMinReservedArea(MinReservedArea); |
4598 | |
4599 | // If the function takes variable number of arguments, make a frame index for |
4600 | // the start of the first vararg value... for expansion of llvm.va_start. |
4601 | if (isVarArg) { |
4602 | int Depth = ArgOffset; |
4603 | |
4604 | FuncInfo->setVarArgsFrameIndex( |
4605 | MFI.CreateFixedObject(PtrVT.getSizeInBits()/8, |
4606 | Depth, true)); |
4607 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
4608 | |
4609 | // If this function is vararg, store any remaining integer argument regs |
4610 | // to their spots on the stack so that they may be loaded by dereferencing |
4611 | // the result of va_next. |
4612 | for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) { |
4613 | unsigned VReg; |
4614 | |
4615 | if (isPPC64) |
4616 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::G8RCRegClass); |
4617 | else |
4618 | VReg = MF.addLiveIn(GPR[GPR_idx], &PPC::GPRCRegClass); |
4619 | |
4620 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
4621 | SDValue Store = |
4622 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
4623 | MemOps.push_back(Store); |
4624 | // Increment the address by four for the next argument to store |
4625 | SDValue PtrOff = DAG.getConstant(PtrVT.getSizeInBits()/8, dl, PtrVT); |
4626 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
4627 | } |
4628 | } |
4629 | |
4630 | if (!MemOps.empty()) |
4631 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
4632 | |
4633 | return Chain; |
4634 | } |
4635 | |
4636 | /// CalculateTailCallSPDiff - Get the amount the stack pointer has to be |
4637 | /// adjusted to accommodate the arguments for the tailcall. |
4638 | static int CalculateTailCallSPDiff(SelectionDAG& DAG, bool isTailCall, |
4639 | unsigned ParamSize) { |
4640 | |
4641 | if (!isTailCall) return 0; |
4642 | |
4643 | PPCFunctionInfo *FI = DAG.getMachineFunction().getInfo<PPCFunctionInfo>(); |
4644 | unsigned CallerMinReservedArea = FI->getMinReservedArea(); |
4645 | int SPDiff = (int)CallerMinReservedArea - (int)ParamSize; |
4646 | // Remember only if the new adjustment is bigger. |
4647 | if (SPDiff < FI->getTailCallSPDelta()) |
4648 | FI->setTailCallSPDelta(SPDiff); |
4649 | |
4650 | return SPDiff; |
4651 | } |
4652 | |
4653 | static bool isFunctionGlobalAddress(SDValue Callee); |
4654 | |
4655 | static bool callsShareTOCBase(const Function *Caller, SDValue Callee, |
4656 | const TargetMachine &TM) { |
4657 | // It does not make sense to call callsShareTOCBase() with a caller that |
4658 | // is PC Relative since PC Relative callers do not have a TOC. |
4659 | #ifndef NDEBUG |
4660 | const PPCSubtarget *STICaller = &TM.getSubtarget<PPCSubtarget>(*Caller); |
4661 | assert(!STICaller->isUsingPCRelativeCalls() &&((!STICaller->isUsingPCRelativeCalls() && "PC Relative callers do not have a TOC and cannot share a TOC Base" ) ? static_cast<void> (0) : __assert_fail ("!STICaller->isUsingPCRelativeCalls() && \"PC Relative callers do not have a TOC and cannot share a TOC Base\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4662, __PRETTY_FUNCTION__)) |
4662 | "PC Relative callers do not have a TOC and cannot share a TOC Base")((!STICaller->isUsingPCRelativeCalls() && "PC Relative callers do not have a TOC and cannot share a TOC Base" ) ? static_cast<void> (0) : __assert_fail ("!STICaller->isUsingPCRelativeCalls() && \"PC Relative callers do not have a TOC and cannot share a TOC Base\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4662, __PRETTY_FUNCTION__)); |
4663 | #endif |
4664 | |
4665 | // Callee is either a GlobalAddress or an ExternalSymbol. ExternalSymbols |
4666 | // don't have enough information to determine if the caller and callee share |
4667 | // the same TOC base, so we have to pessimistically assume they don't for |
4668 | // correctness. |
4669 | GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
4670 | if (!G) |
4671 | return false; |
4672 | |
4673 | const GlobalValue *GV = G->getGlobal(); |
4674 | |
4675 | // If the callee is preemptable, then the static linker will use a plt-stub |
4676 | // which saves the toc to the stack, and needs a nop after the call |
4677 | // instruction to convert to a toc-restore. |
4678 | if (!TM.shouldAssumeDSOLocal(*Caller->getParent(), GV)) |
4679 | return false; |
4680 | |
4681 | // Functions with PC Relative enabled may clobber the TOC in the same DSO. |
4682 | // We may need a TOC restore in the situation where the caller requires a |
4683 | // valid TOC but the callee is PC Relative and does not. |
4684 | const Function *F = dyn_cast<Function>(GV); |
4685 | const GlobalAlias *Alias = dyn_cast<GlobalAlias>(GV); |
4686 | |
4687 | // If we have an Alias we can try to get the function from there. |
4688 | if (Alias) { |
4689 | const GlobalObject *GlobalObj = Alias->getBaseObject(); |
4690 | F = dyn_cast<Function>(GlobalObj); |
4691 | } |
4692 | |
4693 | // If we still have no valid function pointer we do not have enough |
4694 | // information to determine if the callee uses PC Relative calls so we must |
4695 | // assume that it does. |
4696 | if (!F) |
4697 | return false; |
4698 | |
4699 | // If the callee uses PC Relative we cannot guarantee that the callee won't |
4700 | // clobber the TOC of the caller and so we must assume that the two |
4701 | // functions do not share a TOC base. |
4702 | const PPCSubtarget *STICallee = &TM.getSubtarget<PPCSubtarget>(*F); |
4703 | if (STICallee->isUsingPCRelativeCalls()) |
4704 | return false; |
4705 | |
4706 | // The medium and large code models are expected to provide a sufficiently |
4707 | // large TOC to provide all data addressing needs of a module with a |
4708 | // single TOC. |
4709 | if (CodeModel::Medium == TM.getCodeModel() || |
4710 | CodeModel::Large == TM.getCodeModel()) |
4711 | return true; |
4712 | |
4713 | // Otherwise we need to ensure callee and caller are in the same section, |
4714 | // since the linker may allocate multiple TOCs, and we don't know which |
4715 | // sections will belong to the same TOC base. |
4716 | if (!GV->isStrongDefinitionForLinker()) |
4717 | return false; |
4718 | |
4719 | // Any explicitly-specified sections and section prefixes must also match. |
4720 | // Also, if we're using -ffunction-sections, then each function is always in |
4721 | // a different section (the same is true for COMDAT functions). |
4722 | if (TM.getFunctionSections() || GV->hasComdat() || Caller->hasComdat() || |
4723 | GV->getSection() != Caller->getSection()) |
4724 | return false; |
4725 | if (const auto *F = dyn_cast<Function>(GV)) { |
4726 | if (F->getSectionPrefix() != Caller->getSectionPrefix()) |
4727 | return false; |
4728 | } |
4729 | |
4730 | return true; |
4731 | } |
4732 | |
4733 | static bool |
4734 | needStackSlotPassParameters(const PPCSubtarget &Subtarget, |
4735 | const SmallVectorImpl<ISD::OutputArg> &Outs) { |
4736 | assert(Subtarget.is64BitELFABI())((Subtarget.is64BitELFABI()) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64BitELFABI()", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 4736, __PRETTY_FUNCTION__)); |
4737 | |
4738 | const unsigned PtrByteSize = 8; |
4739 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
4740 | |
4741 | static const MCPhysReg GPR[] = { |
4742 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
4743 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
4744 | }; |
4745 | static const MCPhysReg VR[] = { |
4746 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
4747 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
4748 | }; |
4749 | |
4750 | const unsigned NumGPRs = array_lengthof(GPR); |
4751 | const unsigned NumFPRs = 13; |
4752 | const unsigned NumVRs = array_lengthof(VR); |
4753 | const unsigned ParamAreaSize = NumGPRs * PtrByteSize; |
4754 | |
4755 | unsigned NumBytes = LinkageSize; |
4756 | unsigned AvailableFPRs = NumFPRs; |
4757 | unsigned AvailableVRs = NumVRs; |
4758 | |
4759 | for (const ISD::OutputArg& Param : Outs) { |
4760 | if (Param.Flags.isNest()) continue; |
4761 | |
4762 | if (CalculateStackSlotUsed(Param.VT, Param.ArgVT, Param.Flags, PtrByteSize, |
4763 | LinkageSize, ParamAreaSize, NumBytes, |
4764 | AvailableFPRs, AvailableVRs)) |
4765 | return true; |
4766 | } |
4767 | return false; |
4768 | } |
4769 | |
4770 | static bool hasSameArgumentList(const Function *CallerFn, const CallBase &CB) { |
4771 | if (CB.arg_size() != CallerFn->arg_size()) |
4772 | return false; |
4773 | |
4774 | auto CalleeArgIter = CB.arg_begin(); |
4775 | auto CalleeArgEnd = CB.arg_end(); |
4776 | Function::const_arg_iterator CallerArgIter = CallerFn->arg_begin(); |
4777 | |
4778 | for (; CalleeArgIter != CalleeArgEnd; ++CalleeArgIter, ++CallerArgIter) { |
4779 | const Value* CalleeArg = *CalleeArgIter; |
4780 | const Value* CallerArg = &(*CallerArgIter); |
4781 | if (CalleeArg == CallerArg) |
4782 | continue; |
4783 | |
4784 | // e.g. @caller([4 x i64] %a, [4 x i64] %b) { |
4785 | // tail call @callee([4 x i64] undef, [4 x i64] %b) |
4786 | // } |
4787 | // 1st argument of callee is undef and has the same type as caller. |
4788 | if (CalleeArg->getType() == CallerArg->getType() && |
4789 | isa<UndefValue>(CalleeArg)) |
4790 | continue; |
4791 | |
4792 | return false; |
4793 | } |
4794 | |
4795 | return true; |
4796 | } |
4797 | |
4798 | // Returns true if TCO is possible between the callers and callees |
4799 | // calling conventions. |
4800 | static bool |
4801 | areCallingConvEligibleForTCO_64SVR4(CallingConv::ID CallerCC, |
4802 | CallingConv::ID CalleeCC) { |
4803 | // Tail calls are possible with fastcc and ccc. |
4804 | auto isTailCallableCC = [] (CallingConv::ID CC){ |
4805 | return CC == CallingConv::C || CC == CallingConv::Fast; |
4806 | }; |
4807 | if (!isTailCallableCC(CallerCC) || !isTailCallableCC(CalleeCC)) |
4808 | return false; |
4809 | |
4810 | // We can safely tail call both fastcc and ccc callees from a c calling |
4811 | // convention caller. If the caller is fastcc, we may have less stack space |
4812 | // than a non-fastcc caller with the same signature so disable tail-calls in |
4813 | // that case. |
4814 | return CallerCC == CallingConv::C || CallerCC == CalleeCC; |
4815 | } |
4816 | |
4817 | bool PPCTargetLowering::IsEligibleForTailCallOptimization_64SVR4( |
4818 | SDValue Callee, CallingConv::ID CalleeCC, const CallBase *CB, bool isVarArg, |
4819 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
4820 | const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const { |
4821 | bool TailCallOpt = getTargetMachine().Options.GuaranteedTailCallOpt; |
4822 | |
4823 | if (DisableSCO && !TailCallOpt) return false; |
4824 | |
4825 | // Variadic argument functions are not supported. |
4826 | if (isVarArg) return false; |
4827 | |
4828 | auto &Caller = DAG.getMachineFunction().getFunction(); |
4829 | // Check that the calling conventions are compatible for tco. |
4830 | if (!areCallingConvEligibleForTCO_64SVR4(Caller.getCallingConv(), CalleeCC)) |
4831 | return false; |
4832 | |
4833 | // Caller contains any byval parameter is not supported. |
4834 | if (any_of(Ins, [](const ISD::InputArg &IA) { return IA.Flags.isByVal(); })) |
4835 | return false; |
4836 | |
4837 | // Callee contains any byval parameter is not supported, too. |
4838 | // Note: This is a quick work around, because in some cases, e.g. |
4839 | // caller's stack size > callee's stack size, we are still able to apply |
4840 | // sibling call optimization. For example, gcc is able to do SCO for caller1 |
4841 | // in the following example, but not for caller2. |
4842 | // struct test { |
4843 | // long int a; |
4844 | // char ary[56]; |
4845 | // } gTest; |
4846 | // __attribute__((noinline)) int callee(struct test v, struct test *b) { |
4847 | // b->a = v.a; |
4848 | // return 0; |
4849 | // } |
4850 | // void caller1(struct test a, struct test c, struct test *b) { |
4851 | // callee(gTest, b); } |
4852 | // void caller2(struct test *b) { callee(gTest, b); } |
4853 | if (any_of(Outs, [](const ISD::OutputArg& OA) { return OA.Flags.isByVal(); })) |
4854 | return false; |
4855 | |
4856 | // If callee and caller use different calling conventions, we cannot pass |
4857 | // parameters on stack since offsets for the parameter area may be different. |
4858 | if (Caller.getCallingConv() != CalleeCC && |
4859 | needStackSlotPassParameters(Subtarget, Outs)) |
4860 | return false; |
4861 | |
4862 | // All variants of 64-bit ELF ABIs without PC-Relative addressing require that |
4863 | // the caller and callee share the same TOC for TCO/SCO. If the caller and |
4864 | // callee potentially have different TOC bases then we cannot tail call since |
4865 | // we need to restore the TOC pointer after the call. |
4866 | // ref: https://bugzilla.mozilla.org/show_bug.cgi?id=973977 |
4867 | // We cannot guarantee this for indirect calls or calls to external functions. |
4868 | // When PC-Relative addressing is used, the concept of the TOC is no longer |
4869 | // applicable so this check is not required. |
4870 | // Check first for indirect calls. |
4871 | if (!Subtarget.isUsingPCRelativeCalls() && |
4872 | !isFunctionGlobalAddress(Callee) && !isa<ExternalSymbolSDNode>(Callee)) |
4873 | return false; |
4874 | |
4875 | // Check if we share the TOC base. |
4876 | if (!Subtarget.isUsingPCRelativeCalls() && |
4877 | !callsShareTOCBase(&Caller, Callee, getTargetMachine())) |
4878 | return false; |
4879 | |
4880 | // TCO allows altering callee ABI, so we don't have to check further. |
4881 | if (CalleeCC == CallingConv::Fast && TailCallOpt) |
4882 | return true; |
4883 | |
4884 | if (DisableSCO) return false; |
4885 | |
4886 | // If callee use the same argument list that caller is using, then we can |
4887 | // apply SCO on this case. If it is not, then we need to check if callee needs |
4888 | // stack for passing arguments. |
4889 | // PC Relative tail calls may not have a CallBase. |
4890 | // If there is no CallBase we cannot verify if we have the same argument |
4891 | // list so assume that we don't have the same argument list. |
4892 | if (CB && !hasSameArgumentList(&Caller, *CB) && |
4893 | needStackSlotPassParameters(Subtarget, Outs)) |
4894 | return false; |
4895 | else if (!CB && needStackSlotPassParameters(Subtarget, Outs)) |
4896 | return false; |
4897 | |
4898 | return true; |
4899 | } |
4900 | |
4901 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
4902 | /// for tail call optimization. Targets which want to do tail call |
4903 | /// optimization should implement this function. |
4904 | bool |
4905 | PPCTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee, |
4906 | CallingConv::ID CalleeCC, |
4907 | bool isVarArg, |
4908 | const SmallVectorImpl<ISD::InputArg> &Ins, |
4909 | SelectionDAG& DAG) const { |
4910 | if (!getTargetMachine().Options.GuaranteedTailCallOpt) |
4911 | return false; |
4912 | |
4913 | // Variable argument functions are not supported. |
4914 | if (isVarArg) |
4915 | return false; |
4916 | |
4917 | MachineFunction &MF = DAG.getMachineFunction(); |
4918 | CallingConv::ID CallerCC = MF.getFunction().getCallingConv(); |
4919 | if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC) { |
4920 | // Functions containing by val parameters are not supported. |
4921 | for (unsigned i = 0; i != Ins.size(); i++) { |
4922 | ISD::ArgFlagsTy Flags = Ins[i].Flags; |
4923 | if (Flags.isByVal()) return false; |
4924 | } |
4925 | |
4926 | // Non-PIC/GOT tail calls are supported. |
4927 | if (getTargetMachine().getRelocationModel() != Reloc::PIC_) |
4928 | return true; |
4929 | |
4930 | // At the moment we can only do local tail calls (in same module, hidden |
4931 | // or protected) if we are generating PIC. |
4932 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) |
4933 | return G->getGlobal()->hasHiddenVisibility() |
4934 | || G->getGlobal()->hasProtectedVisibility(); |
4935 | } |
4936 | |
4937 | return false; |
4938 | } |
4939 | |
4940 | /// isCallCompatibleAddress - Return the immediate to use if the specified |
4941 | /// 32-bit value is representable in the immediate field of a BxA instruction. |
4942 | static SDNode *isBLACompatibleAddress(SDValue Op, SelectionDAG &DAG) { |
4943 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); |
4944 | if (!C) return nullptr; |
4945 | |
4946 | int Addr = C->getZExtValue(); |
4947 | if ((Addr & 3) != 0 || // Low 2 bits are implicitly zero. |
4948 | SignExtend32<26>(Addr) != Addr) |
4949 | return nullptr; // Top 6 bits have to be sext of immediate. |
4950 | |
4951 | return DAG |
4952 | .getConstant( |
4953 | (int)C->getZExtValue() >> 2, SDLoc(Op), |
4954 | DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout())) |
4955 | .getNode(); |
4956 | } |
4957 | |
4958 | namespace { |
4959 | |
4960 | struct TailCallArgumentInfo { |
4961 | SDValue Arg; |
4962 | SDValue FrameIdxOp; |
4963 | int FrameIdx = 0; |
4964 | |
4965 | TailCallArgumentInfo() = default; |
4966 | }; |
4967 | |
4968 | } // end anonymous namespace |
4969 | |
4970 | /// StoreTailCallArgumentsToStackSlot - Stores arguments to their stack slot. |
4971 | static void StoreTailCallArgumentsToStackSlot( |
4972 | SelectionDAG &DAG, SDValue Chain, |
4973 | const SmallVectorImpl<TailCallArgumentInfo> &TailCallArgs, |
4974 | SmallVectorImpl<SDValue> &MemOpChains, const SDLoc &dl) { |
4975 | for (unsigned i = 0, e = TailCallArgs.size(); i != e; ++i) { |
4976 | SDValue Arg = TailCallArgs[i].Arg; |
4977 | SDValue FIN = TailCallArgs[i].FrameIdxOp; |
4978 | int FI = TailCallArgs[i].FrameIdx; |
4979 | // Store relative to framepointer. |
4980 | MemOpChains.push_back(DAG.getStore( |
4981 | Chain, dl, Arg, FIN, |
4982 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI))); |
4983 | } |
4984 | } |
4985 | |
4986 | /// EmitTailCallStoreFPAndRetAddr - Move the frame pointer and return address to |
4987 | /// the appropriate stack slot for the tail call optimized function call. |
4988 | static SDValue EmitTailCallStoreFPAndRetAddr(SelectionDAG &DAG, SDValue Chain, |
4989 | SDValue OldRetAddr, SDValue OldFP, |
4990 | int SPDiff, const SDLoc &dl) { |
4991 | if (SPDiff) { |
4992 | // Calculate the new stack slot for the return address. |
4993 | MachineFunction &MF = DAG.getMachineFunction(); |
4994 | const PPCSubtarget &Subtarget = MF.getSubtarget<PPCSubtarget>(); |
4995 | const PPCFrameLowering *FL = Subtarget.getFrameLowering(); |
4996 | bool isPPC64 = Subtarget.isPPC64(); |
4997 | int SlotSize = isPPC64 ? 8 : 4; |
4998 | int NewRetAddrLoc = SPDiff + FL->getReturnSaveOffset(); |
4999 | int NewRetAddr = MF.getFrameInfo().CreateFixedObject(SlotSize, |
5000 | NewRetAddrLoc, true); |
5001 | EVT VT = isPPC64 ? MVT::i64 : MVT::i32; |
5002 | SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewRetAddr, VT); |
5003 | Chain = DAG.getStore(Chain, dl, OldRetAddr, NewRetAddrFrIdx, |
5004 | MachinePointerInfo::getFixedStack(MF, NewRetAddr)); |
5005 | } |
5006 | return Chain; |
5007 | } |
5008 | |
5009 | /// CalculateTailCallArgDest - Remember Argument for later processing. Calculate |
5010 | /// the position of the argument. |
5011 | static void |
5012 | CalculateTailCallArgDest(SelectionDAG &DAG, MachineFunction &MF, bool isPPC64, |
5013 | SDValue Arg, int SPDiff, unsigned ArgOffset, |
5014 | SmallVectorImpl<TailCallArgumentInfo>& TailCallArguments) { |
5015 | int Offset = ArgOffset + SPDiff; |
5016 | uint32_t OpSize = (Arg.getValueSizeInBits() + 7) / 8; |
5017 | int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true); |
5018 | EVT VT = isPPC64 ? MVT::i64 : MVT::i32; |
5019 | SDValue FIN = DAG.getFrameIndex(FI, VT); |
5020 | TailCallArgumentInfo Info; |
5021 | Info.Arg = Arg; |
5022 | Info.FrameIdxOp = FIN; |
5023 | Info.FrameIdx = FI; |
5024 | TailCallArguments.push_back(Info); |
5025 | } |
5026 | |
5027 | /// EmitTCFPAndRetAddrLoad - Emit load from frame pointer and return address |
5028 | /// stack slot. Returns the chain as result and the loaded frame pointers in |
5029 | /// LROpOut/FPOpout. Used when tail calling. |
5030 | SDValue PPCTargetLowering::EmitTailCallLoadFPAndRetAddr( |
5031 | SelectionDAG &DAG, int SPDiff, SDValue Chain, SDValue &LROpOut, |
5032 | SDValue &FPOpOut, const SDLoc &dl) const { |
5033 | if (SPDiff) { |
5034 | // Load the LR and FP stack slot for later adjusting. |
5035 | EVT VT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
5036 | LROpOut = getReturnAddrFrameIndex(DAG); |
5037 | LROpOut = DAG.getLoad(VT, dl, Chain, LROpOut, MachinePointerInfo()); |
5038 | Chain = SDValue(LROpOut.getNode(), 1); |
5039 | } |
5040 | return Chain; |
5041 | } |
5042 | |
5043 | /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified |
5044 | /// by "Src" to address "Dst" of size "Size". Alignment information is |
5045 | /// specified by the specific parameter attribute. The copy will be passed as |
5046 | /// a byval function parameter. |
5047 | /// Sometimes what we are copying is the end of a larger object, the part that |
5048 | /// does not fit in registers. |
5049 | static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, |
5050 | SDValue Chain, ISD::ArgFlagsTy Flags, |
5051 | SelectionDAG &DAG, const SDLoc &dl) { |
5052 | SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32); |
5053 | return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, |
5054 | Flags.getNonZeroByValAlign(), false, false, false, |
5055 | MachinePointerInfo(), MachinePointerInfo()); |
5056 | } |
5057 | |
5058 | /// LowerMemOpCallTo - Store the argument to the stack or remember it in case of |
5059 | /// tail calls. |
5060 | static void LowerMemOpCallTo( |
5061 | SelectionDAG &DAG, MachineFunction &MF, SDValue Chain, SDValue Arg, |
5062 | SDValue PtrOff, int SPDiff, unsigned ArgOffset, bool isPPC64, |
5063 | bool isTailCall, bool isVector, SmallVectorImpl<SDValue> &MemOpChains, |
5064 | SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments, const SDLoc &dl) { |
5065 | EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); |
5066 | if (!isTailCall) { |
5067 | if (isVector) { |
5068 | SDValue StackPtr; |
5069 | if (isPPC64) |
5070 | StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
5071 | else |
5072 | StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
5073 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, |
5074 | DAG.getConstant(ArgOffset, dl, PtrVT)); |
5075 | } |
5076 | MemOpChains.push_back( |
5077 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
5078 | // Calculate and remember argument location. |
5079 | } else CalculateTailCallArgDest(DAG, MF, isPPC64, Arg, SPDiff, ArgOffset, |
5080 | TailCallArguments); |
5081 | } |
5082 | |
5083 | static void |
5084 | PrepareTailCall(SelectionDAG &DAG, SDValue &InFlag, SDValue &Chain, |
5085 | const SDLoc &dl, int SPDiff, unsigned NumBytes, SDValue LROp, |
5086 | SDValue FPOp, |
5087 | SmallVectorImpl<TailCallArgumentInfo> &TailCallArguments) { |
5088 | // Emit a sequence of copyto/copyfrom virtual registers for arguments that |
5089 | // might overwrite each other in case of tail call optimization. |
5090 | SmallVector<SDValue, 8> MemOpChains2; |
5091 | // Do not flag preceding copytoreg stuff together with the following stuff. |
5092 | InFlag = SDValue(); |
5093 | StoreTailCallArgumentsToStackSlot(DAG, Chain, TailCallArguments, |
5094 | MemOpChains2, dl); |
5095 | if (!MemOpChains2.empty()) |
5096 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2); |
5097 | |
5098 | // Store the return address to the appropriate stack slot. |
5099 | Chain = EmitTailCallStoreFPAndRetAddr(DAG, Chain, LROp, FPOp, SPDiff, dl); |
5100 | |
5101 | // Emit callseq_end just before tailcall node. |
5102 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true), |
5103 | DAG.getIntPtrConstant(0, dl, true), InFlag, dl); |
5104 | InFlag = Chain.getValue(1); |
5105 | } |
5106 | |
5107 | // Is this global address that of a function that can be called by name? (as |
5108 | // opposed to something that must hold a descriptor for an indirect call). |
5109 | static bool isFunctionGlobalAddress(SDValue Callee) { |
5110 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { |
5111 | if (Callee.getOpcode() == ISD::GlobalTLSAddress || |
5112 | Callee.getOpcode() == ISD::TargetGlobalTLSAddress) |
5113 | return false; |
5114 | |
5115 | return G->getGlobal()->getValueType()->isFunctionTy(); |
5116 | } |
5117 | |
5118 | return false; |
5119 | } |
5120 | |
5121 | SDValue PPCTargetLowering::LowerCallResult( |
5122 | SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, |
5123 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5124 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
5125 | SmallVector<CCValAssign, 16> RVLocs; |
5126 | CCState CCRetInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
5127 | *DAG.getContext()); |
5128 | |
5129 | CCRetInfo.AnalyzeCallResult( |
5130 | Ins, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
5131 | ? RetCC_PPC_Cold |
5132 | : RetCC_PPC); |
5133 | |
5134 | // Copy all of the result registers out of their specified physreg. |
5135 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
5136 | CCValAssign &VA = RVLocs[i]; |
5137 | assert(VA.isRegLoc() && "Can only return in registers!")((VA.isRegLoc() && "Can only return in registers!") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5137, __PRETTY_FUNCTION__)); |
5138 | |
5139 | SDValue Val; |
5140 | |
5141 | if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) { |
5142 | SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, |
5143 | InFlag); |
5144 | Chain = Lo.getValue(1); |
5145 | InFlag = Lo.getValue(2); |
5146 | VA = RVLocs[++i]; // skip ahead to next loc |
5147 | SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, |
5148 | InFlag); |
5149 | Chain = Hi.getValue(1); |
5150 | InFlag = Hi.getValue(2); |
5151 | if (!Subtarget.isLittleEndian()) |
5152 | std::swap (Lo, Hi); |
5153 | Val = DAG.getNode(PPCISD::BUILD_SPE64, dl, MVT::f64, Lo, Hi); |
5154 | } else { |
5155 | Val = DAG.getCopyFromReg(Chain, dl, |
5156 | VA.getLocReg(), VA.getLocVT(), InFlag); |
5157 | Chain = Val.getValue(1); |
5158 | InFlag = Val.getValue(2); |
5159 | } |
5160 | |
5161 | switch (VA.getLocInfo()) { |
5162 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5162); |
5163 | case CCValAssign::Full: break; |
5164 | case CCValAssign::AExt: |
5165 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5166 | break; |
5167 | case CCValAssign::ZExt: |
5168 | Val = DAG.getNode(ISD::AssertZext, dl, VA.getLocVT(), Val, |
5169 | DAG.getValueType(VA.getValVT())); |
5170 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5171 | break; |
5172 | case CCValAssign::SExt: |
5173 | Val = DAG.getNode(ISD::AssertSext, dl, VA.getLocVT(), Val, |
5174 | DAG.getValueType(VA.getValVT())); |
5175 | Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val); |
5176 | break; |
5177 | } |
5178 | |
5179 | InVals.push_back(Val); |
5180 | } |
5181 | |
5182 | return Chain; |
5183 | } |
5184 | |
5185 | static bool isIndirectCall(const SDValue &Callee, SelectionDAG &DAG, |
5186 | const PPCSubtarget &Subtarget, bool isPatchPoint) { |
5187 | // PatchPoint calls are not indirect. |
5188 | if (isPatchPoint) |
5189 | return false; |
5190 | |
5191 | if (isFunctionGlobalAddress(Callee) || dyn_cast<ExternalSymbolSDNode>(Callee)) |
5192 | return false; |
5193 | |
5194 | // Darwin, and 32-bit ELF can use a BLA. The descriptor based ABIs can not |
5195 | // becuase the immediate function pointer points to a descriptor instead of |
5196 | // a function entry point. The ELFv2 ABI cannot use a BLA because the function |
5197 | // pointer immediate points to the global entry point, while the BLA would |
5198 | // need to jump to the local entry point (see rL211174). |
5199 | if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI() && |
5200 | isBLACompatibleAddress(Callee, DAG)) |
5201 | return false; |
5202 | |
5203 | return true; |
5204 | } |
5205 | |
5206 | // AIX and 64-bit ELF ABIs w/o PCRel require a TOC save/restore around calls. |
5207 | static inline bool isTOCSaveRestoreRequired(const PPCSubtarget &Subtarget) { |
5208 | return Subtarget.isAIXABI() || |
5209 | (Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls()); |
5210 | } |
5211 | |
5212 | static unsigned getCallOpcode(PPCTargetLowering::CallFlags CFlags, |
5213 | const Function &Caller, |
5214 | const SDValue &Callee, |
5215 | const PPCSubtarget &Subtarget, |
5216 | const TargetMachine &TM) { |
5217 | if (CFlags.IsTailCall) |
5218 | return PPCISD::TC_RETURN; |
5219 | |
5220 | // This is a call through a function pointer. |
5221 | if (CFlags.IsIndirect) { |
5222 | // AIX and the 64-bit ELF ABIs need to maintain the TOC pointer accross |
5223 | // indirect calls. The save of the caller's TOC pointer to the stack will be |
5224 | // inserted into the DAG as part of call lowering. The restore of the TOC |
5225 | // pointer is modeled by using a pseudo instruction for the call opcode that |
5226 | // represents the 2 instruction sequence of an indirect branch and link, |
5227 | // immediately followed by a load of the TOC pointer from the the stack save |
5228 | // slot into gpr2. For 64-bit ELFv2 ABI with PCRel, do not restore the TOC |
5229 | // as it is not saved or used. |
5230 | return isTOCSaveRestoreRequired(Subtarget) ? PPCISD::BCTRL_LOAD_TOC |
5231 | : PPCISD::BCTRL; |
5232 | } |
5233 | |
5234 | if (Subtarget.isUsingPCRelativeCalls()) { |
5235 | assert(Subtarget.is64BitELFABI() && "PC Relative is only on ELF ABI.")((Subtarget.is64BitELFABI() && "PC Relative is only on ELF ABI." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.is64BitELFABI() && \"PC Relative is only on ELF ABI.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5235, __PRETTY_FUNCTION__)); |
5236 | return PPCISD::CALL_NOTOC; |
5237 | } |
5238 | |
5239 | // The ABIs that maintain a TOC pointer accross calls need to have a nop |
5240 | // immediately following the call instruction if the caller and callee may |
5241 | // have different TOC bases. At link time if the linker determines the calls |
5242 | // may not share a TOC base, the call is redirected to a trampoline inserted |
5243 | // by the linker. The trampoline will (among other things) save the callers |
5244 | // TOC pointer at an ABI designated offset in the linkage area and the linker |
5245 | // will rewrite the nop to be a load of the TOC pointer from the linkage area |
5246 | // into gpr2. |
5247 | if (Subtarget.isAIXABI() || Subtarget.is64BitELFABI()) |
5248 | return callsShareTOCBase(&Caller, Callee, TM) ? PPCISD::CALL |
5249 | : PPCISD::CALL_NOP; |
5250 | |
5251 | return PPCISD::CALL; |
5252 | } |
5253 | |
5254 | static SDValue transformCallee(const SDValue &Callee, SelectionDAG &DAG, |
5255 | const SDLoc &dl, const PPCSubtarget &Subtarget) { |
5256 | if (!Subtarget.usesFunctionDescriptors() && !Subtarget.isELFv2ABI()) |
5257 | if (SDNode *Dest = isBLACompatibleAddress(Callee, DAG)) |
5258 | return SDValue(Dest, 0); |
5259 | |
5260 | // Returns true if the callee is local, and false otherwise. |
5261 | auto isLocalCallee = [&]() { |
5262 | const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee); |
5263 | const Module *Mod = DAG.getMachineFunction().getFunction().getParent(); |
5264 | const GlobalValue *GV = G ? G->getGlobal() : nullptr; |
5265 | |
5266 | return DAG.getTarget().shouldAssumeDSOLocal(*Mod, GV) && |
5267 | !dyn_cast_or_null<GlobalIFunc>(GV); |
5268 | }; |
5269 | |
5270 | // The PLT is only used in 32-bit ELF PIC mode. Attempting to use the PLT in |
5271 | // a static relocation model causes some versions of GNU LD (2.17.50, at |
5272 | // least) to force BSS-PLT, instead of secure-PLT, even if all objects are |
5273 | // built with secure-PLT. |
5274 | bool UsePlt = |
5275 | Subtarget.is32BitELFABI() && !isLocalCallee() && |
5276 | Subtarget.getTargetMachine().getRelocationModel() == Reloc::PIC_; |
5277 | |
5278 | const auto getAIXFuncEntryPointSymbolSDNode = [&](const GlobalValue *GV) { |
5279 | const TargetMachine &TM = Subtarget.getTargetMachine(); |
5280 | const TargetLoweringObjectFile *TLOF = TM.getObjFileLowering(); |
5281 | MCSymbolXCOFF *S = |
5282 | cast<MCSymbolXCOFF>(TLOF->getFunctionEntryPointSymbol(GV, TM)); |
5283 | |
5284 | MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout()); |
5285 | return DAG.getMCSymbol(S, PtrVT); |
5286 | }; |
5287 | |
5288 | if (isFunctionGlobalAddress(Callee)) { |
5289 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Callee)->getGlobal(); |
5290 | |
5291 | if (Subtarget.isAIXABI()) { |
5292 | assert(!isa<GlobalIFunc>(GV) && "IFunc is not supported on AIX.")((!isa<GlobalIFunc>(GV) && "IFunc is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("!isa<GlobalIFunc>(GV) && \"IFunc is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5292, __PRETTY_FUNCTION__)); |
5293 | return getAIXFuncEntryPointSymbolSDNode(GV); |
5294 | } |
5295 | return DAG.getTargetGlobalAddress(GV, dl, Callee.getValueType(), 0, |
5296 | UsePlt ? PPCII::MO_PLT : 0); |
5297 | } |
5298 | |
5299 | if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { |
5300 | const char *SymName = S->getSymbol(); |
5301 | if (Subtarget.isAIXABI()) { |
5302 | // If there exists a user-declared function whose name is the same as the |
5303 | // ExternalSymbol's, then we pick up the user-declared version. |
5304 | const Module *Mod = DAG.getMachineFunction().getFunction().getParent(); |
5305 | if (const Function *F = |
5306 | dyn_cast_or_null<Function>(Mod->getNamedValue(SymName))) |
5307 | return getAIXFuncEntryPointSymbolSDNode(F); |
5308 | |
5309 | // On AIX, direct function calls reference the symbol for the function's |
5310 | // entry point, which is named by prepending a "." before the function's |
5311 | // C-linkage name. A Qualname is returned here because an external |
5312 | // function entry point is a csect with XTY_ER property. |
5313 | const auto getExternalFunctionEntryPointSymbol = [&](StringRef SymName) { |
5314 | auto &Context = DAG.getMachineFunction().getMMI().getContext(); |
5315 | MCSectionXCOFF *Sec = Context.getXCOFFSection( |
5316 | (Twine(".") + Twine(SymName)).str(), XCOFF::XMC_PR, XCOFF::XTY_ER, |
5317 | SectionKind::getMetadata()); |
5318 | return Sec->getQualNameSymbol(); |
5319 | }; |
5320 | |
5321 | SymName = getExternalFunctionEntryPointSymbol(SymName)->getName().data(); |
5322 | } |
5323 | return DAG.getTargetExternalSymbol(SymName, Callee.getValueType(), |
5324 | UsePlt ? PPCII::MO_PLT : 0); |
5325 | } |
5326 | |
5327 | // No transformation needed. |
5328 | assert(Callee.getNode() && "What no callee?")((Callee.getNode() && "What no callee?") ? static_cast <void> (0) : __assert_fail ("Callee.getNode() && \"What no callee?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5328, __PRETTY_FUNCTION__)); |
5329 | return Callee; |
5330 | } |
5331 | |
5332 | static SDValue getOutputChainFromCallSeq(SDValue CallSeqStart) { |
5333 | assert(CallSeqStart.getOpcode() == ISD::CALLSEQ_START &&((CallSeqStart.getOpcode() == ISD::CALLSEQ_START && "Expected a CALLSEQ_STARTSDNode." ) ? static_cast<void> (0) : __assert_fail ("CallSeqStart.getOpcode() == ISD::CALLSEQ_START && \"Expected a CALLSEQ_STARTSDNode.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5334, __PRETTY_FUNCTION__)) |
5334 | "Expected a CALLSEQ_STARTSDNode.")((CallSeqStart.getOpcode() == ISD::CALLSEQ_START && "Expected a CALLSEQ_STARTSDNode." ) ? static_cast<void> (0) : __assert_fail ("CallSeqStart.getOpcode() == ISD::CALLSEQ_START && \"Expected a CALLSEQ_STARTSDNode.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5334, __PRETTY_FUNCTION__)); |
5335 | |
5336 | // The last operand is the chain, except when the node has glue. If the node |
5337 | // has glue, then the last operand is the glue, and the chain is the second |
5338 | // last operand. |
5339 | SDValue LastValue = CallSeqStart.getValue(CallSeqStart->getNumValues() - 1); |
5340 | if (LastValue.getValueType() != MVT::Glue) |
5341 | return LastValue; |
5342 | |
5343 | return CallSeqStart.getValue(CallSeqStart->getNumValues() - 2); |
5344 | } |
5345 | |
5346 | // Creates the node that moves a functions address into the count register |
5347 | // to prepare for an indirect call instruction. |
5348 | static void prepareIndirectCall(SelectionDAG &DAG, SDValue &Callee, |
5349 | SDValue &Glue, SDValue &Chain, |
5350 | const SDLoc &dl) { |
5351 | SDValue MTCTROps[] = {Chain, Callee, Glue}; |
5352 | EVT ReturnTypes[] = {MVT::Other, MVT::Glue}; |
5353 | Chain = DAG.getNode(PPCISD::MTCTR, dl, makeArrayRef(ReturnTypes, 2), |
5354 | makeArrayRef(MTCTROps, Glue.getNode() ? 3 : 2)); |
5355 | // The glue is the second value produced. |
5356 | Glue = Chain.getValue(1); |
5357 | } |
5358 | |
5359 | static void prepareDescriptorIndirectCall(SelectionDAG &DAG, SDValue &Callee, |
5360 | SDValue &Glue, SDValue &Chain, |
5361 | SDValue CallSeqStart, |
5362 | const CallBase *CB, const SDLoc &dl, |
5363 | bool hasNest, |
5364 | const PPCSubtarget &Subtarget) { |
5365 | // Function pointers in the 64-bit SVR4 ABI do not point to the function |
5366 | // entry point, but to the function descriptor (the function entry point |
5367 | // address is part of the function descriptor though). |
5368 | // The function descriptor is a three doubleword structure with the |
5369 | // following fields: function entry point, TOC base address and |
5370 | // environment pointer. |
5371 | // Thus for a call through a function pointer, the following actions need |
5372 | // to be performed: |
5373 | // 1. Save the TOC of the caller in the TOC save area of its stack |
5374 | // frame (this is done in LowerCall_Darwin() or LowerCall_64SVR4()). |
5375 | // 2. Load the address of the function entry point from the function |
5376 | // descriptor. |
5377 | // 3. Load the TOC of the callee from the function descriptor into r2. |
5378 | // 4. Load the environment pointer from the function descriptor into |
5379 | // r11. |
5380 | // 5. Branch to the function entry point address. |
5381 | // 6. On return of the callee, the TOC of the caller needs to be |
5382 | // restored (this is done in FinishCall()). |
5383 | // |
5384 | // The loads are scheduled at the beginning of the call sequence, and the |
5385 | // register copies are flagged together to ensure that no other |
5386 | // operations can be scheduled in between. E.g. without flagging the |
5387 | // copies together, a TOC access in the caller could be scheduled between |
5388 | // the assignment of the callee TOC and the branch to the callee, which leads |
5389 | // to incorrect code. |
5390 | |
5391 | // Start by loading the function address from the descriptor. |
5392 | SDValue LDChain = getOutputChainFromCallSeq(CallSeqStart); |
5393 | auto MMOFlags = Subtarget.hasInvariantFunctionDescriptors() |
5394 | ? (MachineMemOperand::MODereferenceable | |
5395 | MachineMemOperand::MOInvariant) |
5396 | : MachineMemOperand::MONone; |
5397 | |
5398 | MachinePointerInfo MPI(CB ? CB->getCalledOperand() : nullptr); |
5399 | |
5400 | // Registers used in building the DAG. |
5401 | const MCRegister EnvPtrReg = Subtarget.getEnvironmentPointerRegister(); |
5402 | const MCRegister TOCReg = Subtarget.getTOCPointerRegister(); |
5403 | |
5404 | // Offsets of descriptor members. |
5405 | const unsigned TOCAnchorOffset = Subtarget.descriptorTOCAnchorOffset(); |
5406 | const unsigned EnvPtrOffset = Subtarget.descriptorEnvironmentPointerOffset(); |
5407 | |
5408 | const MVT RegVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
5409 | const unsigned Alignment = Subtarget.isPPC64() ? 8 : 4; |
5410 | |
5411 | // One load for the functions entry point address. |
5412 | SDValue LoadFuncPtr = DAG.getLoad(RegVT, dl, LDChain, Callee, MPI, |
5413 | Alignment, MMOFlags); |
5414 | |
5415 | // One for loading the TOC anchor for the module that contains the called |
5416 | // function. |
5417 | SDValue TOCOff = DAG.getIntPtrConstant(TOCAnchorOffset, dl); |
5418 | SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, Callee, TOCOff); |
5419 | SDValue TOCPtr = |
5420 | DAG.getLoad(RegVT, dl, LDChain, AddTOC, |
5421 | MPI.getWithOffset(TOCAnchorOffset), Alignment, MMOFlags); |
5422 | |
5423 | // One for loading the environment pointer. |
5424 | SDValue PtrOff = DAG.getIntPtrConstant(EnvPtrOffset, dl); |
5425 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, RegVT, Callee, PtrOff); |
5426 | SDValue LoadEnvPtr = |
5427 | DAG.getLoad(RegVT, dl, LDChain, AddPtr, |
5428 | MPI.getWithOffset(EnvPtrOffset), Alignment, MMOFlags); |
5429 | |
5430 | |
5431 | // Then copy the newly loaded TOC anchor to the TOC pointer. |
5432 | SDValue TOCVal = DAG.getCopyToReg(Chain, dl, TOCReg, TOCPtr, Glue); |
5433 | Chain = TOCVal.getValue(0); |
5434 | Glue = TOCVal.getValue(1); |
5435 | |
5436 | // If the function call has an explicit 'nest' parameter, it takes the |
5437 | // place of the environment pointer. |
5438 | assert((!hasNest || !Subtarget.isAIXABI()) &&(((!hasNest || !Subtarget.isAIXABI()) && "Nest parameter is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("(!hasNest || !Subtarget.isAIXABI()) && \"Nest parameter is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5439, __PRETTY_FUNCTION__)) |
5439 | "Nest parameter is not supported on AIX.")(((!hasNest || !Subtarget.isAIXABI()) && "Nest parameter is not supported on AIX." ) ? static_cast<void> (0) : __assert_fail ("(!hasNest || !Subtarget.isAIXABI()) && \"Nest parameter is not supported on AIX.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5439, __PRETTY_FUNCTION__)); |
5440 | if (!hasNest) { |
5441 | SDValue EnvVal = DAG.getCopyToReg(Chain, dl, EnvPtrReg, LoadEnvPtr, Glue); |
5442 | Chain = EnvVal.getValue(0); |
5443 | Glue = EnvVal.getValue(1); |
5444 | } |
5445 | |
5446 | // The rest of the indirect call sequence is the same as the non-descriptor |
5447 | // DAG. |
5448 | prepareIndirectCall(DAG, LoadFuncPtr, Glue, Chain, dl); |
5449 | } |
5450 | |
5451 | static void |
5452 | buildCallOperands(SmallVectorImpl<SDValue> &Ops, |
5453 | PPCTargetLowering::CallFlags CFlags, const SDLoc &dl, |
5454 | SelectionDAG &DAG, |
5455 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, |
5456 | SDValue Glue, SDValue Chain, SDValue &Callee, int SPDiff, |
5457 | const PPCSubtarget &Subtarget) { |
5458 | const bool IsPPC64 = Subtarget.isPPC64(); |
5459 | // MVT for a general purpose register. |
5460 | const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32; |
5461 | |
5462 | // First operand is always the chain. |
5463 | Ops.push_back(Chain); |
5464 | |
5465 | // If it's a direct call pass the callee as the second operand. |
5466 | if (!CFlags.IsIndirect) |
5467 | Ops.push_back(Callee); |
5468 | else { |
5469 | assert(!CFlags.IsPatchPoint && "Patch point calls are not indirect.")((!CFlags.IsPatchPoint && "Patch point calls are not indirect." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsPatchPoint && \"Patch point calls are not indirect.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5469, __PRETTY_FUNCTION__)); |
5470 | |
5471 | // For the TOC based ABIs, we have saved the TOC pointer to the linkage area |
5472 | // on the stack (this would have been done in `LowerCall_64SVR4` or |
5473 | // `LowerCall_AIX`). The call instruction is a pseudo instruction that |
5474 | // represents both the indirect branch and a load that restores the TOC |
5475 | // pointer from the linkage area. The operand for the TOC restore is an add |
5476 | // of the TOC save offset to the stack pointer. This must be the second |
5477 | // operand: after the chain input but before any other variadic arguments. |
5478 | // For 64-bit ELFv2 ABI with PCRel, do not restore the TOC as it is not |
5479 | // saved or used. |
5480 | if (isTOCSaveRestoreRequired(Subtarget)) { |
5481 | const MCRegister StackPtrReg = Subtarget.getStackPointerRegister(); |
5482 | |
5483 | SDValue StackPtr = DAG.getRegister(StackPtrReg, RegVT); |
5484 | unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset(); |
5485 | SDValue TOCOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
5486 | SDValue AddTOC = DAG.getNode(ISD::ADD, dl, RegVT, StackPtr, TOCOff); |
5487 | Ops.push_back(AddTOC); |
5488 | } |
5489 | |
5490 | // Add the register used for the environment pointer. |
5491 | if (Subtarget.usesFunctionDescriptors() && !CFlags.HasNest) |
5492 | Ops.push_back(DAG.getRegister(Subtarget.getEnvironmentPointerRegister(), |
5493 | RegVT)); |
5494 | |
5495 | |
5496 | // Add CTR register as callee so a bctr can be emitted later. |
5497 | if (CFlags.IsTailCall) |
5498 | Ops.push_back(DAG.getRegister(IsPPC64 ? PPC::CTR8 : PPC::CTR, RegVT)); |
5499 | } |
5500 | |
5501 | // If this is a tail call add stack pointer delta. |
5502 | if (CFlags.IsTailCall) |
5503 | Ops.push_back(DAG.getConstant(SPDiff, dl, MVT::i32)); |
5504 | |
5505 | // Add argument registers to the end of the list so that they are known live |
5506 | // into the call. |
5507 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) |
5508 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, |
5509 | RegsToPass[i].second.getValueType())); |
5510 | |
5511 | // We cannot add R2/X2 as an operand here for PATCHPOINT, because there is |
5512 | // no way to mark dependencies as implicit here. |
5513 | // We will add the R2/X2 dependency in EmitInstrWithCustomInserter. |
5514 | if ((Subtarget.is64BitELFABI() || Subtarget.isAIXABI()) && |
5515 | !CFlags.IsPatchPoint && !Subtarget.isUsingPCRelativeCalls()) |
5516 | Ops.push_back(DAG.getRegister(Subtarget.getTOCPointerRegister(), RegVT)); |
5517 | |
5518 | // Add implicit use of CR bit 6 for 32-bit SVR4 vararg calls |
5519 | if (CFlags.IsVarArg && Subtarget.is32BitELFABI()) |
5520 | Ops.push_back(DAG.getRegister(PPC::CR1EQ, MVT::i32)); |
5521 | |
5522 | // Add a register mask operand representing the call-preserved registers. |
5523 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
5524 | const uint32_t *Mask = |
5525 | TRI->getCallPreservedMask(DAG.getMachineFunction(), CFlags.CallConv); |
5526 | assert(Mask && "Missing call preserved mask for calling convention")((Mask && "Missing call preserved mask for calling convention" ) ? static_cast<void> (0) : __assert_fail ("Mask && \"Missing call preserved mask for calling convention\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5526, __PRETTY_FUNCTION__)); |
5527 | Ops.push_back(DAG.getRegisterMask(Mask)); |
5528 | |
5529 | // If the glue is valid, it is the last operand. |
5530 | if (Glue.getNode()) |
5531 | Ops.push_back(Glue); |
5532 | } |
5533 | |
5534 | SDValue PPCTargetLowering::FinishCall( |
5535 | CallFlags CFlags, const SDLoc &dl, SelectionDAG &DAG, |
5536 | SmallVector<std::pair<unsigned, SDValue>, 8> &RegsToPass, SDValue Glue, |
5537 | SDValue Chain, SDValue CallSeqStart, SDValue &Callee, int SPDiff, |
5538 | unsigned NumBytes, const SmallVectorImpl<ISD::InputArg> &Ins, |
5539 | SmallVectorImpl<SDValue> &InVals, const CallBase *CB) const { |
5540 | |
5541 | if ((Subtarget.is64BitELFABI() && !Subtarget.isUsingPCRelativeCalls()) || |
5542 | Subtarget.isAIXABI()) |
5543 | setUsesTOCBasePtr(DAG); |
5544 | |
5545 | unsigned CallOpc = |
5546 | getCallOpcode(CFlags, DAG.getMachineFunction().getFunction(), Callee, |
5547 | Subtarget, DAG.getTarget()); |
5548 | |
5549 | if (!CFlags.IsIndirect) |
5550 | Callee = transformCallee(Callee, DAG, dl, Subtarget); |
5551 | else if (Subtarget.usesFunctionDescriptors()) |
5552 | prepareDescriptorIndirectCall(DAG, Callee, Glue, Chain, CallSeqStart, CB, |
5553 | dl, CFlags.HasNest, Subtarget); |
5554 | else |
5555 | prepareIndirectCall(DAG, Callee, Glue, Chain, dl); |
5556 | |
5557 | // Build the operand list for the call instruction. |
5558 | SmallVector<SDValue, 8> Ops; |
5559 | buildCallOperands(Ops, CFlags, dl, DAG, RegsToPass, Glue, Chain, Callee, |
5560 | SPDiff, Subtarget); |
5561 | |
5562 | // Emit tail call. |
5563 | if (CFlags.IsTailCall) { |
5564 | // Indirect tail call when using PC Relative calls do not have the same |
5565 | // constraints. |
5566 | assert(((Callee.getOpcode() == ISD::Register &&((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5567 | cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5568 | Callee.getOpcode() == ISD::TargetExternalSymbol ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5569 | Callee.getOpcode() == ISD::TargetGlobalAddress ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5570 | isa<ConstantSDNode>(Callee) ||((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5571 | (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) &&((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5572 | "Expecting a global address, external symbol, absolute value, "((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5573 | "register or an indirect tail call when PC Relative calls are "((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)) |
5574 | "used.")((((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode >(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && "Expecting a global address, external symbol, absolute value, " "register or an indirect tail call when PC Relative calls are " "used.") ? static_cast<void> (0) : __assert_fail ("((Callee.getOpcode() == ISD::Register && cast<RegisterSDNode>(Callee)->getReg() == PPC::CTR) || Callee.getOpcode() == ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress || isa<ConstantSDNode>(Callee) || (CFlags.IsIndirect && Subtarget.isUsingPCRelativeCalls())) && \"Expecting a global address, external symbol, absolute value, \" \"register or an indirect tail call when PC Relative calls are \" \"used.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5574, __PRETTY_FUNCTION__)); |
5575 | // PC Relative calls also use TC_RETURN as the way to mark tail calls. |
5576 | assert(CallOpc == PPCISD::TC_RETURN &&((CallOpc == PPCISD::TC_RETURN && "Unexpected call opcode for a tail call." ) ? static_cast<void> (0) : __assert_fail ("CallOpc == PPCISD::TC_RETURN && \"Unexpected call opcode for a tail call.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5577, __PRETTY_FUNCTION__)) |
5577 | "Unexpected call opcode for a tail call.")((CallOpc == PPCISD::TC_RETURN && "Unexpected call opcode for a tail call." ) ? static_cast<void> (0) : __assert_fail ("CallOpc == PPCISD::TC_RETURN && \"Unexpected call opcode for a tail call.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5577, __PRETTY_FUNCTION__)); |
5578 | DAG.getMachineFunction().getFrameInfo().setHasTailCall(); |
5579 | return DAG.getNode(CallOpc, dl, MVT::Other, Ops); |
5580 | } |
5581 | |
5582 | std::array<EVT, 2> ReturnTypes = {{MVT::Other, MVT::Glue}}; |
5583 | Chain = DAG.getNode(CallOpc, dl, ReturnTypes, Ops); |
5584 | DAG.addNoMergeSiteInfo(Chain.getNode(), CFlags.NoMerge); |
5585 | Glue = Chain.getValue(1); |
5586 | |
5587 | // When performing tail call optimization the callee pops its arguments off |
5588 | // the stack. Account for this here so these bytes can be pushed back on in |
5589 | // PPCFrameLowering::eliminateCallFramePseudoInstr. |
5590 | int BytesCalleePops = (CFlags.CallConv == CallingConv::Fast && |
5591 | getTargetMachine().Options.GuaranteedTailCallOpt) |
5592 | ? NumBytes |
5593 | : 0; |
5594 | |
5595 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true), |
5596 | DAG.getIntPtrConstant(BytesCalleePops, dl, true), |
5597 | Glue, dl); |
5598 | Glue = Chain.getValue(1); |
5599 | |
5600 | return LowerCallResult(Chain, Glue, CFlags.CallConv, CFlags.IsVarArg, Ins, dl, |
5601 | DAG, InVals); |
5602 | } |
5603 | |
5604 | SDValue |
5605 | PPCTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
5606 | SmallVectorImpl<SDValue> &InVals) const { |
5607 | SelectionDAG &DAG = CLI.DAG; |
5608 | SDLoc &dl = CLI.DL; |
5609 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
5610 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
5611 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
5612 | SDValue Chain = CLI.Chain; |
5613 | SDValue Callee = CLI.Callee; |
5614 | bool &isTailCall = CLI.IsTailCall; |
5615 | CallingConv::ID CallConv = CLI.CallConv; |
5616 | bool isVarArg = CLI.IsVarArg; |
5617 | bool isPatchPoint = CLI.IsPatchPoint; |
5618 | const CallBase *CB = CLI.CB; |
5619 | |
5620 | if (isTailCall) { |
5621 | if (Subtarget.useLongCalls() && !(CB && CB->isMustTailCall())) |
5622 | isTailCall = false; |
5623 | else if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) |
5624 | isTailCall = IsEligibleForTailCallOptimization_64SVR4( |
5625 | Callee, CallConv, CB, isVarArg, Outs, Ins, DAG); |
5626 | else |
5627 | isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg, |
5628 | Ins, DAG); |
5629 | if (isTailCall) { |
5630 | ++NumTailCalls; |
5631 | if (!getTargetMachine().Options.GuaranteedTailCallOpt) |
5632 | ++NumSiblingCalls; |
5633 | |
5634 | // PC Relative calls no longer guarantee that the callee is a Global |
5635 | // Address Node. The callee could be an indirect tail call in which |
5636 | // case the SDValue for the callee could be a load (to load the address |
5637 | // of a function pointer) or it may be a register copy (to move the |
5638 | // address of the callee from a function parameter into a virtual |
5639 | // register). It may also be an ExternalSymbolSDNode (ex memcopy). |
5640 | assert((Subtarget.isUsingPCRelativeCalls() ||(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5642, __PRETTY_FUNCTION__)) |
5641 | isa<GlobalAddressSDNode>(Callee)) &&(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5642, __PRETTY_FUNCTION__)) |
5642 | "Callee should be an llvm::Function object.")(((Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode >(Callee)) && "Callee should be an llvm::Function object." ) ? static_cast<void> (0) : __assert_fail ("(Subtarget.isUsingPCRelativeCalls() || isa<GlobalAddressSDNode>(Callee)) && \"Callee should be an llvm::Function object.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5642, __PRETTY_FUNCTION__)); |
5643 | |
5644 | LLVM_DEBUG(dbgs() << "TCO caller: " << DAG.getMachineFunction().getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "TCO caller: " << DAG .getMachineFunction().getName() << "\nTCO callee: "; } } while (false) |
5645 | << "\nTCO callee: ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "TCO caller: " << DAG .getMachineFunction().getName() << "\nTCO callee: "; } } while (false); |
5646 | LLVM_DEBUG(Callee.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { Callee.dump(); } } while (false); |
5647 | } |
5648 | } |
5649 | |
5650 | if (!isTailCall && CB && CB->isMustTailCall()) |
5651 | report_fatal_error("failed to perform tail call elimination on a call " |
5652 | "site marked musttail"); |
5653 | |
5654 | // When long calls (i.e. indirect calls) are always used, calls are always |
5655 | // made via function pointer. If we have a function name, first translate it |
5656 | // into a pointer. |
5657 | if (Subtarget.useLongCalls() && isa<GlobalAddressSDNode>(Callee) && |
5658 | !isTailCall) |
5659 | Callee = LowerGlobalAddress(Callee, DAG); |
5660 | |
5661 | CallFlags CFlags( |
5662 | CallConv, isTailCall, isVarArg, isPatchPoint, |
5663 | isIndirectCall(Callee, DAG, Subtarget, isPatchPoint), |
5664 | // hasNest |
5665 | Subtarget.is64BitELFABI() && |
5666 | any_of(Outs, [](ISD::OutputArg Arg) { return Arg.Flags.isNest(); }), |
5667 | CLI.NoMerge); |
5668 | |
5669 | if (Subtarget.isSVR4ABI() && Subtarget.isPPC64()) |
5670 | return LowerCall_64SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5671 | InVals, CB); |
5672 | |
5673 | if (Subtarget.isSVR4ABI()) |
5674 | return LowerCall_32SVR4(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5675 | InVals, CB); |
5676 | |
5677 | if (Subtarget.isAIXABI()) |
5678 | return LowerCall_AIX(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5679 | InVals, CB); |
5680 | |
5681 | return LowerCall_Darwin(Chain, Callee, CFlags, Outs, OutVals, Ins, dl, DAG, |
5682 | InVals, CB); |
5683 | } |
5684 | |
5685 | SDValue PPCTargetLowering::LowerCall_32SVR4( |
5686 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
5687 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
5688 | const SmallVectorImpl<SDValue> &OutVals, |
5689 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5690 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
5691 | const CallBase *CB) const { |
5692 | // See PPCTargetLowering::LowerFormalArguments_32SVR4() for a description |
5693 | // of the 32-bit SVR4 ABI stack frame layout. |
5694 | |
5695 | const CallingConv::ID CallConv = CFlags.CallConv; |
5696 | const bool IsVarArg = CFlags.IsVarArg; |
5697 | const bool IsTailCall = CFlags.IsTailCall; |
5698 | |
5699 | assert((CallConv == CallingConv::C ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5701, __PRETTY_FUNCTION__)) |
5700 | CallConv == CallingConv::Cold ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5701, __PRETTY_FUNCTION__)) |
5701 | CallConv == CallingConv::Fast) && "Unknown calling convention!")(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unknown calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unknown calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5701, __PRETTY_FUNCTION__)); |
5702 | |
5703 | const Align PtrAlign(4); |
5704 | |
5705 | MachineFunction &MF = DAG.getMachineFunction(); |
5706 | |
5707 | // Mark this function as potentially containing a function that contains a |
5708 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
5709 | // and restoring the callers stack pointer in this functions epilog. This is |
5710 | // done because by tail calling the called function might overwrite the value |
5711 | // in this function's (MF) stack pointer stack slot 0(SP). |
5712 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
5713 | CallConv == CallingConv::Fast) |
5714 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
5715 | |
5716 | // Count how many bytes are to be pushed on the stack, including the linkage |
5717 | // area, parameter list area and the part of the local variable space which |
5718 | // contains copies of aggregates which are passed by value. |
5719 | |
5720 | // Assign locations to all of the outgoing arguments. |
5721 | SmallVector<CCValAssign, 16> ArgLocs; |
5722 | PPCCCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
5723 | |
5724 | // Reserve space for the linkage area on the stack. |
5725 | CCInfo.AllocateStack(Subtarget.getFrameLowering()->getLinkageSize(), |
5726 | PtrAlign); |
5727 | if (useSoftFloat()) |
5728 | CCInfo.PreAnalyzeCallOperands(Outs); |
5729 | |
5730 | if (IsVarArg) { |
5731 | // Handle fixed and variable vector arguments differently. |
5732 | // Fixed vector arguments go into registers as long as registers are |
5733 | // available. Variable vector arguments always go into memory. |
5734 | unsigned NumArgs = Outs.size(); |
5735 | |
5736 | for (unsigned i = 0; i != NumArgs; ++i) { |
5737 | MVT ArgVT = Outs[i].VT; |
5738 | ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; |
5739 | bool Result; |
5740 | |
5741 | if (Outs[i].IsFixed) { |
5742 | Result = CC_PPC32_SVR4(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, |
5743 | CCInfo); |
5744 | } else { |
5745 | Result = CC_PPC32_SVR4_VarArg(i, ArgVT, ArgVT, CCValAssign::Full, |
5746 | ArgFlags, CCInfo); |
5747 | } |
5748 | |
5749 | if (Result) { |
5750 | #ifndef NDEBUG |
5751 | errs() << "Call operand #" << i << " has unhandled type " |
5752 | << EVT(ArgVT).getEVTString() << "\n"; |
5753 | #endif |
5754 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5754); |
5755 | } |
5756 | } |
5757 | } else { |
5758 | // All arguments are treated the same. |
5759 | CCInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4); |
5760 | } |
5761 | CCInfo.clearWasPPCF128(); |
5762 | |
5763 | // Assign locations to all of the outgoing aggregate by value arguments. |
5764 | SmallVector<CCValAssign, 16> ByValArgLocs; |
5765 | CCState CCByValInfo(CallConv, IsVarArg, MF, ByValArgLocs, *DAG.getContext()); |
5766 | |
5767 | // Reserve stack space for the allocations in CCInfo. |
5768 | CCByValInfo.AllocateStack(CCInfo.getNextStackOffset(), PtrAlign); |
5769 | |
5770 | CCByValInfo.AnalyzeCallOperands(Outs, CC_PPC32_SVR4_ByVal); |
5771 | |
5772 | // Size of the linkage area, parameter list area and the part of the local |
5773 | // space variable where copies of aggregates which are passed by value are |
5774 | // stored. |
5775 | unsigned NumBytes = CCByValInfo.getNextStackOffset(); |
5776 | |
5777 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
5778 | // call optimization. |
5779 | int SPDiff = CalculateTailCallSPDiff(DAG, IsTailCall, NumBytes); |
5780 | |
5781 | // Adjust the stack pointer for the new arguments... |
5782 | // These operations are automatically eliminated by the prolog/epilog pass |
5783 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
5784 | SDValue CallSeqStart = Chain; |
5785 | |
5786 | // Load the return address and frame pointer so it can be moved somewhere else |
5787 | // later. |
5788 | SDValue LROp, FPOp; |
5789 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
5790 | |
5791 | // Set up a copy of the stack pointer for use loading and storing any |
5792 | // arguments that may not fit in the registers available for argument |
5793 | // passing. |
5794 | SDValue StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
5795 | |
5796 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
5797 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
5798 | SmallVector<SDValue, 8> MemOpChains; |
5799 | |
5800 | bool seenFloatArg = false; |
5801 | // Walk the register/memloc assignments, inserting copies/loads. |
5802 | // i - Tracks the index into the list of registers allocated for the call |
5803 | // RealArgIdx - Tracks the index into the list of actual function arguments |
5804 | // j - Tracks the index into the list of byval arguments |
5805 | for (unsigned i = 0, RealArgIdx = 0, j = 0, e = ArgLocs.size(); |
5806 | i != e; |
5807 | ++i, ++RealArgIdx) { |
5808 | CCValAssign &VA = ArgLocs[i]; |
5809 | SDValue Arg = OutVals[RealArgIdx]; |
5810 | ISD::ArgFlagsTy Flags = Outs[RealArgIdx].Flags; |
5811 | |
5812 | if (Flags.isByVal()) { |
5813 | // Argument is an aggregate which is passed by value, thus we need to |
5814 | // create a copy of it in the local variable space of the current stack |
5815 | // frame (which is the stack frame of the caller) and pass the address of |
5816 | // this copy to the callee. |
5817 | assert((j < ByValArgLocs.size()) && "Index out of bounds!")(((j < ByValArgLocs.size()) && "Index out of bounds!" ) ? static_cast<void> (0) : __assert_fail ("(j < ByValArgLocs.size()) && \"Index out of bounds!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5817, __PRETTY_FUNCTION__)); |
5818 | CCValAssign &ByValVA = ByValArgLocs[j++]; |
5819 | assert((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!")(((VA.getValNo() == ByValVA.getValNo()) && "ValNo mismatch!" ) ? static_cast<void> (0) : __assert_fail ("(VA.getValNo() == ByValVA.getValNo()) && \"ValNo mismatch!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5819, __PRETTY_FUNCTION__)); |
5820 | |
5821 | // Memory reserved in the local variable space of the callers stack frame. |
5822 | unsigned LocMemOffset = ByValVA.getLocMemOffset(); |
5823 | |
5824 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); |
5825 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()), |
5826 | StackPtr, PtrOff); |
5827 | |
5828 | // Create a copy of the argument in the local area of the current |
5829 | // stack frame. |
5830 | SDValue MemcpyCall = |
5831 | CreateCopyOfByValArgument(Arg, PtrOff, |
5832 | CallSeqStart.getNode()->getOperand(0), |
5833 | Flags, DAG, dl); |
5834 | |
5835 | // This must go outside the CALLSEQ_START..END. |
5836 | SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, NumBytes, 0, |
5837 | SDLoc(MemcpyCall)); |
5838 | DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), |
5839 | NewCallSeqStart.getNode()); |
5840 | Chain = CallSeqStart = NewCallSeqStart; |
5841 | |
5842 | // Pass the address of the aggregate copy on the stack either in a |
5843 | // physical register or in the parameter list area of the current stack |
5844 | // frame to the callee. |
5845 | Arg = PtrOff; |
5846 | } |
5847 | |
5848 | // When useCRBits() is true, there can be i1 arguments. |
5849 | // It is because getRegisterType(MVT::i1) => MVT::i1, |
5850 | // and for other integer types getRegisterType() => MVT::i32. |
5851 | // Extend i1 and ensure callee will get i32. |
5852 | if (Arg.getValueType() == MVT::i1) |
5853 | Arg = DAG.getNode(Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, |
5854 | dl, MVT::i32, Arg); |
5855 | |
5856 | if (VA.isRegLoc()) { |
5857 | seenFloatArg |= VA.getLocVT().isFloatingPoint(); |
5858 | // Put argument in a physical register. |
5859 | if (Subtarget.hasSPE() && Arg.getValueType() == MVT::f64) { |
5860 | bool IsLE = Subtarget.isLittleEndian(); |
5861 | SDValue SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
5862 | DAG.getIntPtrConstant(IsLE ? 0 : 1, dl)); |
5863 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), SVal.getValue(0))); |
5864 | SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
5865 | DAG.getIntPtrConstant(IsLE ? 1 : 0, dl)); |
5866 | RegsToPass.push_back(std::make_pair(ArgLocs[++i].getLocReg(), |
5867 | SVal.getValue(0))); |
5868 | } else |
5869 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
5870 | } else { |
5871 | // Put argument in the parameter list area of the current stack frame. |
5872 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5872, __PRETTY_FUNCTION__)); |
5873 | unsigned LocMemOffset = VA.getLocMemOffset(); |
5874 | |
5875 | if (!IsTailCall) { |
5876 | SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl); |
5877 | PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(MF.getDataLayout()), |
5878 | StackPtr, PtrOff); |
5879 | |
5880 | MemOpChains.push_back( |
5881 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
5882 | } else { |
5883 | // Calculate and remember argument location. |
5884 | CalculateTailCallArgDest(DAG, MF, false, Arg, SPDiff, LocMemOffset, |
5885 | TailCallArguments); |
5886 | } |
5887 | } |
5888 | } |
5889 | |
5890 | if (!MemOpChains.empty()) |
5891 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
5892 | |
5893 | // Build a sequence of copy-to-reg nodes chained together with token chain |
5894 | // and flag operands which copy the outgoing args into the appropriate regs. |
5895 | SDValue InFlag; |
5896 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
5897 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
5898 | RegsToPass[i].second, InFlag); |
5899 | InFlag = Chain.getValue(1); |
5900 | } |
5901 | |
5902 | // Set CR bit 6 to true if this is a vararg call with floating args passed in |
5903 | // registers. |
5904 | if (IsVarArg) { |
5905 | SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue); |
5906 | SDValue Ops[] = { Chain, InFlag }; |
5907 | |
5908 | Chain = DAG.getNode(seenFloatArg ? PPCISD::CR6SET : PPCISD::CR6UNSET, |
5909 | dl, VTs, makeArrayRef(Ops, InFlag.getNode() ? 2 : 1)); |
5910 | |
5911 | InFlag = Chain.getValue(1); |
5912 | } |
5913 | |
5914 | if (IsTailCall) |
5915 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
5916 | TailCallArguments); |
5917 | |
5918 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
5919 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
5920 | } |
5921 | |
5922 | // Copy an argument into memory, being careful to do this outside the |
5923 | // call sequence for the call to which the argument belongs. |
5924 | SDValue PPCTargetLowering::createMemcpyOutsideCallSeq( |
5925 | SDValue Arg, SDValue PtrOff, SDValue CallSeqStart, ISD::ArgFlagsTy Flags, |
5926 | SelectionDAG &DAG, const SDLoc &dl) const { |
5927 | SDValue MemcpyCall = CreateCopyOfByValArgument(Arg, PtrOff, |
5928 | CallSeqStart.getNode()->getOperand(0), |
5929 | Flags, DAG, dl); |
5930 | // The MEMCPY must go outside the CALLSEQ_START..END. |
5931 | int64_t FrameSize = CallSeqStart.getConstantOperandVal(1); |
5932 | SDValue NewCallSeqStart = DAG.getCALLSEQ_START(MemcpyCall, FrameSize, 0, |
5933 | SDLoc(MemcpyCall)); |
5934 | DAG.ReplaceAllUsesWith(CallSeqStart.getNode(), |
5935 | NewCallSeqStart.getNode()); |
5936 | return NewCallSeqStart; |
5937 | } |
5938 | |
5939 | SDValue PPCTargetLowering::LowerCall_64SVR4( |
5940 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
5941 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
5942 | const SmallVectorImpl<SDValue> &OutVals, |
5943 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
5944 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
5945 | const CallBase *CB) const { |
5946 | bool isELFv2ABI = Subtarget.isELFv2ABI(); |
5947 | bool isLittleEndian = Subtarget.isLittleEndian(); |
5948 | unsigned NumOps = Outs.size(); |
5949 | bool IsSibCall = false; |
5950 | bool IsFastCall = CFlags.CallConv == CallingConv::Fast; |
5951 | |
5952 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
5953 | unsigned PtrByteSize = 8; |
5954 | |
5955 | MachineFunction &MF = DAG.getMachineFunction(); |
5956 | |
5957 | if (CFlags.IsTailCall && !getTargetMachine().Options.GuaranteedTailCallOpt) |
5958 | IsSibCall = true; |
5959 | |
5960 | // Mark this function as potentially containing a function that contains a |
5961 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
5962 | // and restoring the callers stack pointer in this functions epilog. This is |
5963 | // done because by tail calling the called function might overwrite the value |
5964 | // in this function's (MF) stack pointer stack slot 0(SP). |
5965 | if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall) |
5966 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
5967 | |
5968 | assert(!(IsFastCall && CFlags.IsVarArg) &&((!(IsFastCall && CFlags.IsVarArg) && "fastcc not supported on varargs functions" ) ? static_cast<void> (0) : __assert_fail ("!(IsFastCall && CFlags.IsVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5969, __PRETTY_FUNCTION__)) |
5969 | "fastcc not supported on varargs functions")((!(IsFastCall && CFlags.IsVarArg) && "fastcc not supported on varargs functions" ) ? static_cast<void> (0) : __assert_fail ("!(IsFastCall && CFlags.IsVarArg) && \"fastcc not supported on varargs functions\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 5969, __PRETTY_FUNCTION__)); |
5970 | |
5971 | // Count how many bytes are to be pushed on the stack, including the linkage |
5972 | // area, and parameter passing area. On ELFv1, the linkage area is 48 bytes |
5973 | // reserved space for [SP][CR][LR][2 x unused][TOC]; on ELFv2, the linkage |
5974 | // area is 32 bytes reserved space for [SP][CR][LR][TOC]. |
5975 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
5976 | unsigned NumBytes = LinkageSize; |
5977 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
5978 | |
5979 | static const MCPhysReg GPR[] = { |
5980 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
5981 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
5982 | }; |
5983 | static const MCPhysReg VR[] = { |
5984 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
5985 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
5986 | }; |
5987 | |
5988 | const unsigned NumGPRs = array_lengthof(GPR); |
5989 | const unsigned NumFPRs = useSoftFloat() ? 0 : 13; |
5990 | const unsigned NumVRs = array_lengthof(VR); |
5991 | |
5992 | // On ELFv2, we can avoid allocating the parameter area if all the arguments |
5993 | // can be passed to the callee in registers. |
5994 | // For the fast calling convention, there is another check below. |
5995 | // Note: We should keep consistent with LowerFormalArguments_64SVR4() |
5996 | bool HasParameterArea = !isELFv2ABI || CFlags.IsVarArg || IsFastCall; |
5997 | if (!HasParameterArea) { |
5998 | unsigned ParamAreaSize = NumGPRs * PtrByteSize; |
5999 | unsigned AvailableFPRs = NumFPRs; |
6000 | unsigned AvailableVRs = NumVRs; |
6001 | unsigned NumBytesTmp = NumBytes; |
6002 | for (unsigned i = 0; i != NumOps; ++i) { |
6003 | if (Outs[i].Flags.isNest()) continue; |
6004 | if (CalculateStackSlotUsed(Outs[i].VT, Outs[i].ArgVT, Outs[i].Flags, |
6005 | PtrByteSize, LinkageSize, ParamAreaSize, |
6006 | NumBytesTmp, AvailableFPRs, AvailableVRs)) |
6007 | HasParameterArea = true; |
6008 | } |
6009 | } |
6010 | |
6011 | // When using the fast calling convention, we don't provide backing for |
6012 | // arguments that will be in registers. |
6013 | unsigned NumGPRsUsed = 0, NumFPRsUsed = 0, NumVRsUsed = 0; |
6014 | |
6015 | // Avoid allocating parameter area for fastcc functions if all the arguments |
6016 | // can be passed in the registers. |
6017 | if (IsFastCall) |
6018 | HasParameterArea = false; |
6019 | |
6020 | // Add up all the space actually used. |
6021 | for (unsigned i = 0; i != NumOps; ++i) { |
6022 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6023 | EVT ArgVT = Outs[i].VT; |
6024 | EVT OrigVT = Outs[i].ArgVT; |
6025 | |
6026 | if (Flags.isNest()) |
6027 | continue; |
6028 | |
6029 | if (IsFastCall) { |
6030 | if (Flags.isByVal()) { |
6031 | NumGPRsUsed += (Flags.getByValSize()+7)/8; |
6032 | if (NumGPRsUsed > NumGPRs) |
6033 | HasParameterArea = true; |
6034 | } else { |
6035 | switch (ArgVT.getSimpleVT().SimpleTy) { |
6036 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6036); |
6037 | case MVT::i1: |
6038 | case MVT::i32: |
6039 | case MVT::i64: |
6040 | if (++NumGPRsUsed <= NumGPRs) |
6041 | continue; |
6042 | break; |
6043 | case MVT::v4i32: |
6044 | case MVT::v8i16: |
6045 | case MVT::v16i8: |
6046 | case MVT::v2f64: |
6047 | case MVT::v2i64: |
6048 | case MVT::v1i128: |
6049 | case MVT::f128: |
6050 | if (++NumVRsUsed <= NumVRs) |
6051 | continue; |
6052 | break; |
6053 | case MVT::v4f32: |
6054 | if (++NumVRsUsed <= NumVRs) |
6055 | continue; |
6056 | break; |
6057 | case MVT::f32: |
6058 | case MVT::f64: |
6059 | if (++NumFPRsUsed <= NumFPRs) |
6060 | continue; |
6061 | break; |
6062 | } |
6063 | HasParameterArea = true; |
6064 | } |
6065 | } |
6066 | |
6067 | /* Respect alignment of argument on the stack. */ |
6068 | auto Alignement = |
6069 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
6070 | NumBytes = alignTo(NumBytes, Alignement); |
6071 | |
6072 | NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
6073 | if (Flags.isInConsecutiveRegsLast()) |
6074 | NumBytes = ((NumBytes + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
6075 | } |
6076 | |
6077 | unsigned NumBytesActuallyUsed = NumBytes; |
6078 | |
6079 | // In the old ELFv1 ABI, |
6080 | // the prolog code of the callee may store up to 8 GPR argument registers to |
6081 | // the stack, allowing va_start to index over them in memory if its varargs. |
6082 | // Because we cannot tell if this is needed on the caller side, we have to |
6083 | // conservatively assume that it is needed. As such, make sure we have at |
6084 | // least enough stack space for the caller to store the 8 GPRs. |
6085 | // In the ELFv2 ABI, we allocate the parameter area iff a callee |
6086 | // really requires memory operands, e.g. a vararg function. |
6087 | if (HasParameterArea) |
6088 | NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize); |
6089 | else |
6090 | NumBytes = LinkageSize; |
6091 | |
6092 | // Tail call needs the stack to be aligned. |
6093 | if (getTargetMachine().Options.GuaranteedTailCallOpt && IsFastCall) |
6094 | NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes); |
6095 | |
6096 | int SPDiff = 0; |
6097 | |
6098 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
6099 | // call optimization. |
6100 | if (!IsSibCall) |
6101 | SPDiff = CalculateTailCallSPDiff(DAG, CFlags.IsTailCall, NumBytes); |
6102 | |
6103 | // To protect arguments on the stack from being clobbered in a tail call, |
6104 | // force all the loads to happen before doing any other lowering. |
6105 | if (CFlags.IsTailCall) |
6106 | Chain = DAG.getStackArgumentTokenFactor(Chain); |
6107 | |
6108 | // Adjust the stack pointer for the new arguments... |
6109 | // These operations are automatically eliminated by the prolog/epilog pass |
6110 | if (!IsSibCall) |
6111 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
6112 | SDValue CallSeqStart = Chain; |
6113 | |
6114 | // Load the return address and frame pointer so it can be move somewhere else |
6115 | // later. |
6116 | SDValue LROp, FPOp; |
6117 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
6118 | |
6119 | // Set up a copy of the stack pointer for use loading and storing any |
6120 | // arguments that may not fit in the registers available for argument |
6121 | // passing. |
6122 | SDValue StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
6123 | |
6124 | // Figure out which arguments are going to go in registers, and which in |
6125 | // memory. Also, if this is a vararg function, floating point operations |
6126 | // must be stored to our stack, and loaded into integer regs as well, if |
6127 | // any integer regs are available for argument passing. |
6128 | unsigned ArgOffset = LinkageSize; |
6129 | |
6130 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
6131 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
6132 | |
6133 | SmallVector<SDValue, 8> MemOpChains; |
6134 | for (unsigned i = 0; i != NumOps; ++i) { |
6135 | SDValue Arg = OutVals[i]; |
6136 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6137 | EVT ArgVT = Outs[i].VT; |
6138 | EVT OrigVT = Outs[i].ArgVT; |
6139 | |
6140 | // PtrOff will be used to store the current argument to the stack if a |
6141 | // register cannot be found for it. |
6142 | SDValue PtrOff; |
6143 | |
6144 | // We re-align the argument offset for each argument, except when using the |
6145 | // fast calling convention, when we need to make sure we do that only when |
6146 | // we'll actually use a stack slot. |
6147 | auto ComputePtrOff = [&]() { |
6148 | /* Respect alignment of argument on the stack. */ |
6149 | auto Alignment = |
6150 | CalculateStackSlotAlignment(ArgVT, OrigVT, Flags, PtrByteSize); |
6151 | ArgOffset = alignTo(ArgOffset, Alignment); |
6152 | |
6153 | PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType()); |
6154 | |
6155 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6156 | }; |
6157 | |
6158 | if (!IsFastCall) { |
6159 | ComputePtrOff(); |
6160 | |
6161 | /* Compute GPR index associated with argument offset. */ |
6162 | GPR_idx = (ArgOffset - LinkageSize) / PtrByteSize; |
6163 | GPR_idx = std::min(GPR_idx, NumGPRs); |
6164 | } |
6165 | |
6166 | // Promote integers to 64-bit values. |
6167 | if (Arg.getValueType() == MVT::i32 || Arg.getValueType() == MVT::i1) { |
6168 | // FIXME: Should this use ANY_EXTEND if neither sext nor zext? |
6169 | unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
6170 | Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg); |
6171 | } |
6172 | |
6173 | // FIXME memcpy is used way more than necessary. Correctness first. |
6174 | // Note: "by value" is code for passing a structure by value, not |
6175 | // basic types. |
6176 | if (Flags.isByVal()) { |
6177 | // Note: Size includes alignment padding, so |
6178 | // struct x { short a; char b; } |
6179 | // will have Size = 4. With #pragma pack(1), it will have Size = 3. |
6180 | // These are the proper values we need for right-justifying the |
6181 | // aggregate in a parameter register. |
6182 | unsigned Size = Flags.getByValSize(); |
6183 | |
6184 | // An empty aggregate parameter takes up no storage and no |
6185 | // registers. |
6186 | if (Size == 0) |
6187 | continue; |
6188 | |
6189 | if (IsFastCall) |
6190 | ComputePtrOff(); |
6191 | |
6192 | // All aggregates smaller than 8 bytes must be passed right-justified. |
6193 | if (Size==1 || Size==2 || Size==4) { |
6194 | EVT VT = (Size==1) ? MVT::i8 : ((Size==2) ? MVT::i16 : MVT::i32); |
6195 | if (GPR_idx != NumGPRs) { |
6196 | SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg, |
6197 | MachinePointerInfo(), VT); |
6198 | MemOpChains.push_back(Load.getValue(1)); |
6199 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6200 | |
6201 | ArgOffset += PtrByteSize; |
6202 | continue; |
6203 | } |
6204 | } |
6205 | |
6206 | if (GPR_idx == NumGPRs && Size < 8) { |
6207 | SDValue AddPtr = PtrOff; |
6208 | if (!isLittleEndian) { |
6209 | SDValue Const = DAG.getConstant(PtrByteSize - Size, dl, |
6210 | PtrOff.getValueType()); |
6211 | AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6212 | } |
6213 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6214 | CallSeqStart, |
6215 | Flags, DAG, dl); |
6216 | ArgOffset += PtrByteSize; |
6217 | continue; |
6218 | } |
6219 | // Copy entire object into memory. There are cases where gcc-generated |
6220 | // code assumes it is there, even if it could be put entirely into |
6221 | // registers. (This is not what the doc says.) |
6222 | |
6223 | // FIXME: The above statement is likely due to a misunderstanding of the |
6224 | // documents. All arguments must be copied into the parameter area BY |
6225 | // THE CALLEE in the event that the callee takes the address of any |
6226 | // formal argument. That has not yet been implemented. However, it is |
6227 | // reasonable to use the stack area as a staging area for the register |
6228 | // load. |
6229 | |
6230 | // Skip this for small aggregates, as we will use the same slot for a |
6231 | // right-justified copy, below. |
6232 | if (Size >= 8) |
6233 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff, |
6234 | CallSeqStart, |
6235 | Flags, DAG, dl); |
6236 | |
6237 | // When a register is available, pass a small aggregate right-justified. |
6238 | if (Size < 8 && GPR_idx != NumGPRs) { |
6239 | // The easiest way to get this right-justified in a register |
6240 | // is to copy the structure into the rightmost portion of a |
6241 | // local variable slot, then load the whole slot into the |
6242 | // register. |
6243 | // FIXME: The memcpy seems to produce pretty awful code for |
6244 | // small aggregates, particularly for packed ones. |
6245 | // FIXME: It would be preferable to use the slot in the |
6246 | // parameter save area instead of a new local variable. |
6247 | SDValue AddPtr = PtrOff; |
6248 | if (!isLittleEndian) { |
6249 | SDValue Const = DAG.getConstant(8 - Size, dl, PtrOff.getValueType()); |
6250 | AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6251 | } |
6252 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6253 | CallSeqStart, |
6254 | Flags, DAG, dl); |
6255 | |
6256 | // Load the slot into the register. |
6257 | SDValue Load = |
6258 | DAG.getLoad(PtrVT, dl, Chain, PtrOff, MachinePointerInfo()); |
6259 | MemOpChains.push_back(Load.getValue(1)); |
6260 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6261 | |
6262 | // Done with this argument. |
6263 | ArgOffset += PtrByteSize; |
6264 | continue; |
6265 | } |
6266 | |
6267 | // For aggregates larger than PtrByteSize, copy the pieces of the |
6268 | // object that fit into registers from the parameter save area. |
6269 | for (unsigned j=0; j<Size; j+=PtrByteSize) { |
6270 | SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType()); |
6271 | SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); |
6272 | if (GPR_idx != NumGPRs) { |
6273 | SDValue Load = |
6274 | DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo()); |
6275 | MemOpChains.push_back(Load.getValue(1)); |
6276 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6277 | ArgOffset += PtrByteSize; |
6278 | } else { |
6279 | ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize; |
6280 | break; |
6281 | } |
6282 | } |
6283 | continue; |
6284 | } |
6285 | |
6286 | switch (Arg.getSimpleValueType().SimpleTy) { |
6287 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6287); |
6288 | case MVT::i1: |
6289 | case MVT::i32: |
6290 | case MVT::i64: |
6291 | if (Flags.isNest()) { |
6292 | // The 'nest' parameter, if any, is passed in R11. |
6293 | RegsToPass.push_back(std::make_pair(PPC::X11, Arg)); |
6294 | break; |
6295 | } |
6296 | |
6297 | // These can be scalar arguments or elements of an integer array type |
6298 | // passed directly. Clang may use those instead of "byval" aggregate |
6299 | // types to avoid forcing arguments to memory unnecessarily. |
6300 | if (GPR_idx != NumGPRs) { |
6301 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg)); |
6302 | } else { |
6303 | if (IsFastCall) |
6304 | ComputePtrOff(); |
6305 | |
6306 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6307, __PRETTY_FUNCTION__)) |
6307 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6307, __PRETTY_FUNCTION__)); |
6308 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6309 | true, CFlags.IsTailCall, false, MemOpChains, |
6310 | TailCallArguments, dl); |
6311 | if (IsFastCall) |
6312 | ArgOffset += PtrByteSize; |
6313 | } |
6314 | if (!IsFastCall) |
6315 | ArgOffset += PtrByteSize; |
6316 | break; |
6317 | case MVT::f32: |
6318 | case MVT::f64: { |
6319 | // These can be scalar arguments or elements of a float array type |
6320 | // passed directly. The latter are used to implement ELFv2 homogenous |
6321 | // float aggregates. |
6322 | |
6323 | // Named arguments go into FPRs first, and once they overflow, the |
6324 | // remaining arguments go into GPRs and then the parameter save area. |
6325 | // Unnamed arguments for vararg functions always go to GPRs and |
6326 | // then the parameter save area. For now, put all arguments to vararg |
6327 | // routines always in both locations (FPR *and* GPR or stack slot). |
6328 | bool NeedGPROrStack = CFlags.IsVarArg || FPR_idx == NumFPRs; |
6329 | bool NeededLoad = false; |
6330 | |
6331 | // First load the argument into the next available FPR. |
6332 | if (FPR_idx != NumFPRs) |
6333 | RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg)); |
6334 | |
6335 | // Next, load the argument into GPR or stack slot if needed. |
6336 | if (!NeedGPROrStack) |
6337 | ; |
6338 | else if (GPR_idx != NumGPRs && !IsFastCall) { |
6339 | // FIXME: We may want to re-enable this for CallingConv::Fast on the P8 |
6340 | // once we support fp <-> gpr moves. |
6341 | |
6342 | // In the non-vararg case, this can only ever happen in the |
6343 | // presence of f32 array types, since otherwise we never run |
6344 | // out of FPRs before running out of GPRs. |
6345 | SDValue ArgVal; |
6346 | |
6347 | // Double values are always passed in a single GPR. |
6348 | if (Arg.getValueType() != MVT::f32) { |
6349 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg); |
6350 | |
6351 | // Non-array float values are extended and passed in a GPR. |
6352 | } else if (!Flags.isInConsecutiveRegs()) { |
6353 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6354 | ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal); |
6355 | |
6356 | // If we have an array of floats, we collect every odd element |
6357 | // together with its predecessor into one GPR. |
6358 | } else if (ArgOffset % PtrByteSize != 0) { |
6359 | SDValue Lo, Hi; |
6360 | Lo = DAG.getNode(ISD::BITCAST, dl, MVT::i32, OutVals[i - 1]); |
6361 | Hi = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6362 | if (!isLittleEndian) |
6363 | std::swap(Lo, Hi); |
6364 | ArgVal = DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); |
6365 | |
6366 | // The final element, if even, goes into the first half of a GPR. |
6367 | } else if (Flags.isInConsecutiveRegsLast()) { |
6368 | ArgVal = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Arg); |
6369 | ArgVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i64, ArgVal); |
6370 | if (!isLittleEndian) |
6371 | ArgVal = DAG.getNode(ISD::SHL, dl, MVT::i64, ArgVal, |
6372 | DAG.getConstant(32, dl, MVT::i32)); |
6373 | |
6374 | // Non-final even elements are skipped; they will be handled |
6375 | // together the with subsequent argument on the next go-around. |
6376 | } else |
6377 | ArgVal = SDValue(); |
6378 | |
6379 | if (ArgVal.getNode()) |
6380 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], ArgVal)); |
6381 | } else { |
6382 | if (IsFastCall) |
6383 | ComputePtrOff(); |
6384 | |
6385 | // Single-precision floating-point values are mapped to the |
6386 | // second (rightmost) word of the stack doubleword. |
6387 | if (Arg.getValueType() == MVT::f32 && |
6388 | !isLittleEndian && !Flags.isInConsecutiveRegs()) { |
6389 | SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType()); |
6390 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour); |
6391 | } |
6392 | |
6393 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6394, __PRETTY_FUNCTION__)) |
6394 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6394, __PRETTY_FUNCTION__)); |
6395 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6396 | true, CFlags.IsTailCall, false, MemOpChains, |
6397 | TailCallArguments, dl); |
6398 | |
6399 | NeededLoad = true; |
6400 | } |
6401 | // When passing an array of floats, the array occupies consecutive |
6402 | // space in the argument area; only round up to the next doubleword |
6403 | // at the end of the array. Otherwise, each float takes 8 bytes. |
6404 | if (!IsFastCall || NeededLoad) { |
6405 | ArgOffset += (Arg.getValueType() == MVT::f32 && |
6406 | Flags.isInConsecutiveRegs()) ? 4 : 8; |
6407 | if (Flags.isInConsecutiveRegsLast()) |
6408 | ArgOffset = ((ArgOffset + PtrByteSize - 1)/PtrByteSize) * PtrByteSize; |
6409 | } |
6410 | break; |
6411 | } |
6412 | case MVT::v4f32: |
6413 | case MVT::v4i32: |
6414 | case MVT::v8i16: |
6415 | case MVT::v16i8: |
6416 | case MVT::v2f64: |
6417 | case MVT::v2i64: |
6418 | case MVT::v1i128: |
6419 | case MVT::f128: |
6420 | // These can be scalar arguments or elements of a vector array type |
6421 | // passed directly. The latter are used to implement ELFv2 homogenous |
6422 | // vector aggregates. |
6423 | |
6424 | // For a varargs call, named arguments go into VRs or on the stack as |
6425 | // usual; unnamed arguments always go to the stack or the corresponding |
6426 | // GPRs when within range. For now, we always put the value in both |
6427 | // locations (or even all three). |
6428 | if (CFlags.IsVarArg) { |
6429 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist if we have a varargs call." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist if we have a varargs call.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6430, __PRETTY_FUNCTION__)) |
6430 | "Parameter area must exist if we have a varargs call.")((HasParameterArea && "Parameter area must exist if we have a varargs call." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist if we have a varargs call.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6430, __PRETTY_FUNCTION__)); |
6431 | // We could elide this store in the case where the object fits |
6432 | // entirely in R registers. Maybe later. |
6433 | SDValue Store = |
6434 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6435 | MemOpChains.push_back(Store); |
6436 | if (VR_idx != NumVRs) { |
6437 | SDValue Load = |
6438 | DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo()); |
6439 | MemOpChains.push_back(Load.getValue(1)); |
6440 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load)); |
6441 | } |
6442 | ArgOffset += 16; |
6443 | for (unsigned i=0; i<16; i+=PtrByteSize) { |
6444 | if (GPR_idx == NumGPRs) |
6445 | break; |
6446 | SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, |
6447 | DAG.getConstant(i, dl, PtrVT)); |
6448 | SDValue Load = |
6449 | DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo()); |
6450 | MemOpChains.push_back(Load.getValue(1)); |
6451 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6452 | } |
6453 | break; |
6454 | } |
6455 | |
6456 | // Non-varargs Altivec params go into VRs or on the stack. |
6457 | if (VR_idx != NumVRs) { |
6458 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg)); |
6459 | } else { |
6460 | if (IsFastCall) |
6461 | ComputePtrOff(); |
6462 | |
6463 | assert(HasParameterArea &&((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6464, __PRETTY_FUNCTION__)) |
6464 | "Parameter area must exist to pass an argument in memory.")((HasParameterArea && "Parameter area must exist to pass an argument in memory." ) ? static_cast<void> (0) : __assert_fail ("HasParameterArea && \"Parameter area must exist to pass an argument in memory.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6464, __PRETTY_FUNCTION__)); |
6465 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6466 | true, CFlags.IsTailCall, true, MemOpChains, |
6467 | TailCallArguments, dl); |
6468 | if (IsFastCall) |
6469 | ArgOffset += 16; |
6470 | } |
6471 | |
6472 | if (!IsFastCall) |
6473 | ArgOffset += 16; |
6474 | break; |
6475 | } |
6476 | } |
6477 | |
6478 | assert((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) &&(((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && "mismatch in size of parameter area") ? static_cast<void> (0) : __assert_fail ("(!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && \"mismatch in size of parameter area\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6479, __PRETTY_FUNCTION__)) |
6479 | "mismatch in size of parameter area")(((!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && "mismatch in size of parameter area") ? static_cast<void> (0) : __assert_fail ("(!HasParameterArea || NumBytesActuallyUsed == ArgOffset) && \"mismatch in size of parameter area\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6479, __PRETTY_FUNCTION__)); |
6480 | (void)NumBytesActuallyUsed; |
6481 | |
6482 | if (!MemOpChains.empty()) |
6483 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
6484 | |
6485 | // Check if this is an indirect call (MTCTR/BCTRL). |
6486 | // See prepareDescriptorIndirectCall and buildCallOperands for more |
6487 | // information about calls through function pointers in the 64-bit SVR4 ABI. |
6488 | if (CFlags.IsIndirect) { |
6489 | // For 64-bit ELFv2 ABI with PCRel, do not save the TOC of the |
6490 | // caller in the TOC save area. |
6491 | if (isTOCSaveRestoreRequired(Subtarget)) { |
6492 | assert(!CFlags.IsTailCall && "Indirect tails calls not supported")((!CFlags.IsTailCall && "Indirect tails calls not supported" ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tails calls not supported\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6492, __PRETTY_FUNCTION__)); |
6493 | // Load r2 into a virtual register and store it to the TOC save area. |
6494 | setUsesTOCBasePtr(DAG); |
6495 | SDValue Val = DAG.getCopyFromReg(Chain, dl, PPC::X2, MVT::i64); |
6496 | // TOC save area offset. |
6497 | unsigned TOCSaveOffset = Subtarget.getFrameLowering()->getTOCSaveOffset(); |
6498 | SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
6499 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6500 | Chain = DAG.getStore(Val.getValue(1), dl, Val, AddPtr, |
6501 | MachinePointerInfo::getStack( |
6502 | DAG.getMachineFunction(), TOCSaveOffset)); |
6503 | } |
6504 | // In the ELFv2 ABI, R12 must contain the address of an indirect callee. |
6505 | // This does not mean the MTCTR instruction must use R12; it's easier |
6506 | // to model this as an extra parameter, so do that. |
6507 | if (isELFv2ABI && !CFlags.IsPatchPoint) |
6508 | RegsToPass.push_back(std::make_pair((unsigned)PPC::X12, Callee)); |
6509 | } |
6510 | |
6511 | // Build a sequence of copy-to-reg nodes chained together with token chain |
6512 | // and flag operands which copy the outgoing args into the appropriate regs. |
6513 | SDValue InFlag; |
6514 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
6515 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
6516 | RegsToPass[i].second, InFlag); |
6517 | InFlag = Chain.getValue(1); |
6518 | } |
6519 | |
6520 | if (CFlags.IsTailCall && !IsSibCall) |
6521 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
6522 | TailCallArguments); |
6523 | |
6524 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
6525 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
6526 | } |
6527 | |
6528 | SDValue PPCTargetLowering::LowerCall_Darwin( |
6529 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
6530 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
6531 | const SmallVectorImpl<SDValue> &OutVals, |
6532 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
6533 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
6534 | const CallBase *CB) const { |
6535 | unsigned NumOps = Outs.size(); |
6536 | |
6537 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
6538 | bool isPPC64 = PtrVT == MVT::i64; |
6539 | unsigned PtrByteSize = isPPC64 ? 8 : 4; |
6540 | |
6541 | MachineFunction &MF = DAG.getMachineFunction(); |
6542 | |
6543 | // Mark this function as potentially containing a function that contains a |
6544 | // tail call. As a consequence the frame pointer will be used for dynamicalloc |
6545 | // and restoring the callers stack pointer in this functions epilog. This is |
6546 | // done because by tail calling the called function might overwrite the value |
6547 | // in this function's (MF) stack pointer stack slot 0(SP). |
6548 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
6549 | CFlags.CallConv == CallingConv::Fast) |
6550 | MF.getInfo<PPCFunctionInfo>()->setHasFastCall(); |
6551 | |
6552 | // Count how many bytes are to be pushed on the stack, including the linkage |
6553 | // area, and parameter passing area. We start with 24/48 bytes, which is |
6554 | // prereserved space for [SP][CR][LR][3 x unused]. |
6555 | unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
6556 | unsigned NumBytes = LinkageSize; |
6557 | |
6558 | // Add up all the space actually used. |
6559 | // In 32-bit non-varargs calls, Altivec parameters all go at the end; usually |
6560 | // they all go in registers, but we must reserve stack space for them for |
6561 | // possible use by the caller. In varargs or 64-bit calls, parameters are |
6562 | // assigned stack space in order, with padding so Altivec parameters are |
6563 | // 16-byte aligned. |
6564 | unsigned nAltivecParamsAtEnd = 0; |
6565 | for (unsigned i = 0; i != NumOps; ++i) { |
6566 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6567 | EVT ArgVT = Outs[i].VT; |
6568 | // Varargs Altivec parameters are padded to a 16 byte boundary. |
6569 | if (ArgVT == MVT::v4f32 || ArgVT == MVT::v4i32 || |
6570 | ArgVT == MVT::v8i16 || ArgVT == MVT::v16i8 || |
6571 | ArgVT == MVT::v2f64 || ArgVT == MVT::v2i64) { |
6572 | if (!CFlags.IsVarArg && !isPPC64) { |
6573 | // Non-varargs Altivec parameters go after all the non-Altivec |
6574 | // parameters; handle those later so we know how much padding we need. |
6575 | nAltivecParamsAtEnd++; |
6576 | continue; |
6577 | } |
6578 | // Varargs and 64-bit Altivec parameters are padded to 16 byte boundary. |
6579 | NumBytes = ((NumBytes+15)/16)*16; |
6580 | } |
6581 | NumBytes += CalculateStackSlotSize(ArgVT, Flags, PtrByteSize); |
6582 | } |
6583 | |
6584 | // Allow for Altivec parameters at the end, if needed. |
6585 | if (nAltivecParamsAtEnd) { |
6586 | NumBytes = ((NumBytes+15)/16)*16; |
6587 | NumBytes += 16*nAltivecParamsAtEnd; |
6588 | } |
6589 | |
6590 | // The prolog code of the callee may store up to 8 GPR argument registers to |
6591 | // the stack, allowing va_start to index over them in memory if its varargs. |
6592 | // Because we cannot tell if this is needed on the caller side, we have to |
6593 | // conservatively assume that it is needed. As such, make sure we have at |
6594 | // least enough stack space for the caller to store the 8 GPRs. |
6595 | NumBytes = std::max(NumBytes, LinkageSize + 8 * PtrByteSize); |
6596 | |
6597 | // Tail call needs the stack to be aligned. |
6598 | if (getTargetMachine().Options.GuaranteedTailCallOpt && |
6599 | CFlags.CallConv == CallingConv::Fast) |
6600 | NumBytes = EnsureStackAlignment(Subtarget.getFrameLowering(), NumBytes); |
6601 | |
6602 | // Calculate by how many bytes the stack has to be adjusted in case of tail |
6603 | // call optimization. |
6604 | int SPDiff = CalculateTailCallSPDiff(DAG, CFlags.IsTailCall, NumBytes); |
6605 | |
6606 | // To protect arguments on the stack from being clobbered in a tail call, |
6607 | // force all the loads to happen before doing any other lowering. |
6608 | if (CFlags.IsTailCall) |
6609 | Chain = DAG.getStackArgumentTokenFactor(Chain); |
6610 | |
6611 | // Adjust the stack pointer for the new arguments... |
6612 | // These operations are automatically eliminated by the prolog/epilog pass |
6613 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
6614 | SDValue CallSeqStart = Chain; |
6615 | |
6616 | // Load the return address and frame pointer so it can be move somewhere else |
6617 | // later. |
6618 | SDValue LROp, FPOp; |
6619 | Chain = EmitTailCallLoadFPAndRetAddr(DAG, SPDiff, Chain, LROp, FPOp, dl); |
6620 | |
6621 | // Set up a copy of the stack pointer for use loading and storing any |
6622 | // arguments that may not fit in the registers available for argument |
6623 | // passing. |
6624 | SDValue StackPtr; |
6625 | if (isPPC64) |
6626 | StackPtr = DAG.getRegister(PPC::X1, MVT::i64); |
6627 | else |
6628 | StackPtr = DAG.getRegister(PPC::R1, MVT::i32); |
6629 | |
6630 | // Figure out which arguments are going to go in registers, and which in |
6631 | // memory. Also, if this is a vararg function, floating point operations |
6632 | // must be stored to our stack, and loaded into integer regs as well, if |
6633 | // any integer regs are available for argument passing. |
6634 | unsigned ArgOffset = LinkageSize; |
6635 | unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0; |
6636 | |
6637 | static const MCPhysReg GPR_32[] = { // 32-bit registers. |
6638 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
6639 | PPC::R7, PPC::R8, PPC::R9, PPC::R10, |
6640 | }; |
6641 | static const MCPhysReg GPR_64[] = { // 64-bit registers. |
6642 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
6643 | PPC::X7, PPC::X8, PPC::X9, PPC::X10, |
6644 | }; |
6645 | static const MCPhysReg VR[] = { |
6646 | PPC::V2, PPC::V3, PPC::V4, PPC::V5, PPC::V6, PPC::V7, PPC::V8, |
6647 | PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13 |
6648 | }; |
6649 | const unsigned NumGPRs = array_lengthof(GPR_32); |
6650 | const unsigned NumFPRs = 13; |
6651 | const unsigned NumVRs = array_lengthof(VR); |
6652 | |
6653 | const MCPhysReg *GPR = isPPC64 ? GPR_64 : GPR_32; |
6654 | |
6655 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
6656 | SmallVector<TailCallArgumentInfo, 8> TailCallArguments; |
6657 | |
6658 | SmallVector<SDValue, 8> MemOpChains; |
6659 | for (unsigned i = 0; i != NumOps; ++i) { |
6660 | SDValue Arg = OutVals[i]; |
6661 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
6662 | |
6663 | // PtrOff will be used to store the current argument to the stack if a |
6664 | // register cannot be found for it. |
6665 | SDValue PtrOff; |
6666 | |
6667 | PtrOff = DAG.getConstant(ArgOffset, dl, StackPtr.getValueType()); |
6668 | |
6669 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
6670 | |
6671 | // On PPC64, promote integers to 64-bit values. |
6672 | if (isPPC64 && Arg.getValueType() == MVT::i32) { |
6673 | // FIXME: Should this use ANY_EXTEND if neither sext nor zext? |
6674 | unsigned ExtOp = Flags.isSExt() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
6675 | Arg = DAG.getNode(ExtOp, dl, MVT::i64, Arg); |
6676 | } |
6677 | |
6678 | // FIXME memcpy is used way more than necessary. Correctness first. |
6679 | // Note: "by value" is code for passing a structure by value, not |
6680 | // basic types. |
6681 | if (Flags.isByVal()) { |
6682 | unsigned Size = Flags.getByValSize(); |
6683 | // Very small objects are passed right-justified. Everything else is |
6684 | // passed left-justified. |
6685 | if (Size==1 || Size==2) { |
6686 | EVT VT = (Size==1) ? MVT::i8 : MVT::i16; |
6687 | if (GPR_idx != NumGPRs) { |
6688 | SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, PtrVT, Chain, Arg, |
6689 | MachinePointerInfo(), VT); |
6690 | MemOpChains.push_back(Load.getValue(1)); |
6691 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6692 | |
6693 | ArgOffset += PtrByteSize; |
6694 | } else { |
6695 | SDValue Const = DAG.getConstant(PtrByteSize - Size, dl, |
6696 | PtrOff.getValueType()); |
6697 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, Const); |
6698 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, AddPtr, |
6699 | CallSeqStart, |
6700 | Flags, DAG, dl); |
6701 | ArgOffset += PtrByteSize; |
6702 | } |
6703 | continue; |
6704 | } |
6705 | // Copy entire object into memory. There are cases where gcc-generated |
6706 | // code assumes it is there, even if it could be put entirely into |
6707 | // registers. (This is not what the doc says.) |
6708 | Chain = CallSeqStart = createMemcpyOutsideCallSeq(Arg, PtrOff, |
6709 | CallSeqStart, |
6710 | Flags, DAG, dl); |
6711 | |
6712 | // For small aggregates (Darwin only) and aggregates >= PtrByteSize, |
6713 | // copy the pieces of the object that fit into registers from the |
6714 | // parameter save area. |
6715 | for (unsigned j=0; j<Size; j+=PtrByteSize) { |
6716 | SDValue Const = DAG.getConstant(j, dl, PtrOff.getValueType()); |
6717 | SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const); |
6718 | if (GPR_idx != NumGPRs) { |
6719 | SDValue Load = |
6720 | DAG.getLoad(PtrVT, dl, Chain, AddArg, MachinePointerInfo()); |
6721 | MemOpChains.push_back(Load.getValue(1)); |
6722 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6723 | ArgOffset += PtrByteSize; |
6724 | } else { |
6725 | ArgOffset += ((Size - j + PtrByteSize-1)/PtrByteSize)*PtrByteSize; |
6726 | break; |
6727 | } |
6728 | } |
6729 | continue; |
6730 | } |
6731 | |
6732 | switch (Arg.getSimpleValueType().SimpleTy) { |
6733 | default: llvm_unreachable("Unexpected ValueType for argument!")::llvm::llvm_unreachable_internal("Unexpected ValueType for argument!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6733); |
6734 | case MVT::i1: |
6735 | case MVT::i32: |
6736 | case MVT::i64: |
6737 | if (GPR_idx != NumGPRs) { |
6738 | if (Arg.getValueType() == MVT::i1) |
6739 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, PtrVT, Arg); |
6740 | |
6741 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg)); |
6742 | } else { |
6743 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6744 | isPPC64, CFlags.IsTailCall, false, MemOpChains, |
6745 | TailCallArguments, dl); |
6746 | } |
6747 | ArgOffset += PtrByteSize; |
6748 | break; |
6749 | case MVT::f32: |
6750 | case MVT::f64: |
6751 | if (FPR_idx != NumFPRs) { |
6752 | RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg)); |
6753 | |
6754 | if (CFlags.IsVarArg) { |
6755 | SDValue Store = |
6756 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6757 | MemOpChains.push_back(Store); |
6758 | |
6759 | // Float varargs are always shadowed in available integer registers |
6760 | if (GPR_idx != NumGPRs) { |
6761 | SDValue Load = |
6762 | DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo()); |
6763 | MemOpChains.push_back(Load.getValue(1)); |
6764 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6765 | } |
6766 | if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){ |
6767 | SDValue ConstFour = DAG.getConstant(4, dl, PtrOff.getValueType()); |
6768 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, ConstFour); |
6769 | SDValue Load = |
6770 | DAG.getLoad(PtrVT, dl, Store, PtrOff, MachinePointerInfo()); |
6771 | MemOpChains.push_back(Load.getValue(1)); |
6772 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6773 | } |
6774 | } else { |
6775 | // If we have any FPRs remaining, we may also have GPRs remaining. |
6776 | // Args passed in FPRs consume either 1 (f32) or 2 (f64) available |
6777 | // GPRs. |
6778 | if (GPR_idx != NumGPRs) |
6779 | ++GPR_idx; |
6780 | if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && |
6781 | !isPPC64) // PPC64 has 64-bit GPR's obviously :) |
6782 | ++GPR_idx; |
6783 | } |
6784 | } else |
6785 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6786 | isPPC64, CFlags.IsTailCall, false, MemOpChains, |
6787 | TailCallArguments, dl); |
6788 | if (isPPC64) |
6789 | ArgOffset += 8; |
6790 | else |
6791 | ArgOffset += Arg.getValueType() == MVT::f32 ? 4 : 8; |
6792 | break; |
6793 | case MVT::v4f32: |
6794 | case MVT::v4i32: |
6795 | case MVT::v8i16: |
6796 | case MVT::v16i8: |
6797 | if (CFlags.IsVarArg) { |
6798 | // These go aligned on the stack, or in the corresponding R registers |
6799 | // when within range. The Darwin PPC ABI doc claims they also go in |
6800 | // V registers; in fact gcc does this only for arguments that are |
6801 | // prototyped, not for those that match the ... We do it for all |
6802 | // arguments, seems to work. |
6803 | while (ArgOffset % 16 !=0) { |
6804 | ArgOffset += PtrByteSize; |
6805 | if (GPR_idx != NumGPRs) |
6806 | GPR_idx++; |
6807 | } |
6808 | // We could elide this store in the case where the object fits |
6809 | // entirely in R registers. Maybe later. |
6810 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, |
6811 | DAG.getConstant(ArgOffset, dl, PtrVT)); |
6812 | SDValue Store = |
6813 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo()); |
6814 | MemOpChains.push_back(Store); |
6815 | if (VR_idx != NumVRs) { |
6816 | SDValue Load = |
6817 | DAG.getLoad(MVT::v4f32, dl, Store, PtrOff, MachinePointerInfo()); |
6818 | MemOpChains.push_back(Load.getValue(1)); |
6819 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Load)); |
6820 | } |
6821 | ArgOffset += 16; |
6822 | for (unsigned i=0; i<16; i+=PtrByteSize) { |
6823 | if (GPR_idx == NumGPRs) |
6824 | break; |
6825 | SDValue Ix = DAG.getNode(ISD::ADD, dl, PtrVT, PtrOff, |
6826 | DAG.getConstant(i, dl, PtrVT)); |
6827 | SDValue Load = |
6828 | DAG.getLoad(PtrVT, dl, Store, Ix, MachinePointerInfo()); |
6829 | MemOpChains.push_back(Load.getValue(1)); |
6830 | RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load)); |
6831 | } |
6832 | break; |
6833 | } |
6834 | |
6835 | // Non-varargs Altivec params generally go in registers, but have |
6836 | // stack space allocated at the end. |
6837 | if (VR_idx != NumVRs) { |
6838 | // Doesn't have GPR space allocated. |
6839 | RegsToPass.push_back(std::make_pair(VR[VR_idx++], Arg)); |
6840 | } else if (nAltivecParamsAtEnd==0) { |
6841 | // We are emitting Altivec params in order. |
6842 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6843 | isPPC64, CFlags.IsTailCall, true, MemOpChains, |
6844 | TailCallArguments, dl); |
6845 | ArgOffset += 16; |
6846 | } |
6847 | break; |
6848 | } |
6849 | } |
6850 | // If all Altivec parameters fit in registers, as they usually do, |
6851 | // they get stack space following the non-Altivec parameters. We |
6852 | // don't track this here because nobody below needs it. |
6853 | // If there are more Altivec parameters than fit in registers emit |
6854 | // the stores here. |
6855 | if (!CFlags.IsVarArg && nAltivecParamsAtEnd > NumVRs) { |
6856 | unsigned j = 0; |
6857 | // Offset is aligned; skip 1st 12 params which go in V registers. |
6858 | ArgOffset = ((ArgOffset+15)/16)*16; |
6859 | ArgOffset += 12*16; |
6860 | for (unsigned i = 0; i != NumOps; ++i) { |
6861 | SDValue Arg = OutVals[i]; |
6862 | EVT ArgType = Outs[i].VT; |
6863 | if (ArgType==MVT::v4f32 || ArgType==MVT::v4i32 || |
6864 | ArgType==MVT::v8i16 || ArgType==MVT::v16i8) { |
6865 | if (++j > NumVRs) { |
6866 | SDValue PtrOff; |
6867 | // We are emitting Altivec params in order. |
6868 | LowerMemOpCallTo(DAG, MF, Chain, Arg, PtrOff, SPDiff, ArgOffset, |
6869 | isPPC64, CFlags.IsTailCall, true, MemOpChains, |
6870 | TailCallArguments, dl); |
6871 | ArgOffset += 16; |
6872 | } |
6873 | } |
6874 | } |
6875 | } |
6876 | |
6877 | if (!MemOpChains.empty()) |
6878 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
6879 | |
6880 | // On Darwin, R12 must contain the address of an indirect callee. This does |
6881 | // not mean the MTCTR instruction must use R12; it's easier to model this as |
6882 | // an extra parameter, so do that. |
6883 | if (CFlags.IsIndirect) { |
6884 | assert(!CFlags.IsTailCall && "Indirect tail-calls not supported.")((!CFlags.IsTailCall && "Indirect tail-calls not supported." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tail-calls not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6884, __PRETTY_FUNCTION__)); |
6885 | RegsToPass.push_back(std::make_pair((unsigned)(isPPC64 ? PPC::X12 : |
6886 | PPC::R12), Callee)); |
6887 | } |
6888 | |
6889 | // Build a sequence of copy-to-reg nodes chained together with token chain |
6890 | // and flag operands which copy the outgoing args into the appropriate regs. |
6891 | SDValue InFlag; |
6892 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
6893 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
6894 | RegsToPass[i].second, InFlag); |
6895 | InFlag = Chain.getValue(1); |
6896 | } |
6897 | |
6898 | if (CFlags.IsTailCall) |
6899 | PrepareTailCall(DAG, InFlag, Chain, dl, SPDiff, NumBytes, LROp, FPOp, |
6900 | TailCallArguments); |
6901 | |
6902 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
6903 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
6904 | } |
6905 | |
6906 | static bool CC_AIX(unsigned ValNo, MVT ValVT, MVT LocVT, |
6907 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
6908 | CCState &State) { |
6909 | |
6910 | const PPCSubtarget &Subtarget = static_cast<const PPCSubtarget &>( |
6911 | State.getMachineFunction().getSubtarget()); |
6912 | const bool IsPPC64 = Subtarget.isPPC64(); |
6913 | const Align PtrAlign = IsPPC64 ? Align(8) : Align(4); |
6914 | const MVT RegVT = IsPPC64 ? MVT::i64 : MVT::i32; |
6915 | |
6916 | assert((!ValVT.isInteger() ||(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6918, __PRETTY_FUNCTION__)) |
6917 | (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) &&(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6918, __PRETTY_FUNCTION__)) |
6918 | "Integer argument exceeds register size: should have been legalized")(((!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT .getFixedSizeInBits())) && "Integer argument exceeds register size: should have been legalized" ) ? static_cast<void> (0) : __assert_fail ("(!ValVT.isInteger() || (ValVT.getFixedSizeInBits() <= RegVT.getFixedSizeInBits())) && \"Integer argument exceeds register size: should have been legalized\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6918, __PRETTY_FUNCTION__)); |
6919 | |
6920 | if (ValVT == MVT::f128) |
6921 | report_fatal_error("f128 is unimplemented on AIX."); |
6922 | |
6923 | if (ArgFlags.isNest()) |
6924 | report_fatal_error("Nest arguments are unimplemented."); |
6925 | |
6926 | if (ValVT.isVector() || LocVT.isVector()) |
6927 | report_fatal_error("Vector arguments are unimplemented on AIX."); |
6928 | |
6929 | static const MCPhysReg GPR_32[] = {// 32-bit registers. |
6930 | PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
6931 | PPC::R7, PPC::R8, PPC::R9, PPC::R10}; |
6932 | static const MCPhysReg GPR_64[] = {// 64-bit registers. |
6933 | PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
6934 | PPC::X7, PPC::X8, PPC::X9, PPC::X10}; |
6935 | |
6936 | if (ArgFlags.isByVal()) { |
6937 | if (ArgFlags.getNonZeroByValAlign() > PtrAlign) |
6938 | report_fatal_error("Pass-by-value arguments with alignment greater than " |
6939 | "register width are not supported."); |
6940 | |
6941 | const unsigned ByValSize = ArgFlags.getByValSize(); |
6942 | |
6943 | // An empty aggregate parameter takes up no storage and no registers, |
6944 | // but needs a MemLoc for a stack slot for the formal arguments side. |
6945 | if (ByValSize == 0) { |
6946 | State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6947 | State.getNextStackOffset(), RegVT, |
6948 | LocInfo)); |
6949 | return false; |
6950 | } |
6951 | |
6952 | const unsigned StackSize = alignTo(ByValSize, PtrAlign); |
6953 | unsigned Offset = State.AllocateStack(StackSize, PtrAlign); |
6954 | for (const unsigned E = Offset + StackSize; Offset < E; |
6955 | Offset += PtrAlign.value()) { |
6956 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) |
6957 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
6958 | else { |
6959 | State.addLoc(CCValAssign::getMem(ValNo, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6960 | Offset, MVT::INVALID_SIMPLE_VALUE_TYPE, |
6961 | LocInfo)); |
6962 | break; |
6963 | } |
6964 | } |
6965 | return false; |
6966 | } |
6967 | |
6968 | // Arguments always reserve parameter save area. |
6969 | switch (ValVT.SimpleTy) { |
6970 | default: |
6971 | report_fatal_error("Unhandled value type for argument."); |
6972 | case MVT::i64: |
6973 | // i64 arguments should have been split to i32 for PPC32. |
6974 | assert(IsPPC64 && "PPC32 should have split i64 values.")((IsPPC64 && "PPC32 should have split i64 values.") ? static_cast<void> (0) : __assert_fail ("IsPPC64 && \"PPC32 should have split i64 values.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 6974, __PRETTY_FUNCTION__)); |
6975 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
6976 | case MVT::i1: |
6977 | case MVT::i32: { |
6978 | const unsigned Offset = State.AllocateStack(PtrAlign.value(), PtrAlign); |
6979 | // AIX integer arguments are always passed in register width. |
6980 | if (ValVT.getFixedSizeInBits() < RegVT.getFixedSizeInBits()) |
6981 | LocInfo = ArgFlags.isSExt() ? CCValAssign::LocInfo::SExt |
6982 | : CCValAssign::LocInfo::ZExt; |
6983 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) |
6984 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
6985 | else |
6986 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, RegVT, LocInfo)); |
6987 | |
6988 | return false; |
6989 | } |
6990 | case MVT::f32: |
6991 | case MVT::f64: { |
6992 | // Parameter save area (PSA) is reserved even if the float passes in fpr. |
6993 | const unsigned StoreSize = LocVT.getStoreSize(); |
6994 | // Floats are always 4-byte aligned in the PSA on AIX. |
6995 | // This includes f64 in 64-bit mode for ABI compatibility. |
6996 | const unsigned Offset = |
6997 | State.AllocateStack(IsPPC64 ? 8 : StoreSize, Align(4)); |
6998 | unsigned FReg = State.AllocateReg(FPR); |
6999 | if (FReg) |
7000 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, FReg, LocVT, LocInfo)); |
7001 | |
7002 | // Reserve and initialize GPRs or initialize the PSA as required. |
7003 | for (unsigned I = 0; I < StoreSize; I += PtrAlign.value()) { |
7004 | if (unsigned Reg = State.AllocateReg(IsPPC64 ? GPR_64 : GPR_32)) { |
7005 | assert(FReg && "An FPR should be available when a GPR is reserved.")((FReg && "An FPR should be available when a GPR is reserved." ) ? static_cast<void> (0) : __assert_fail ("FReg && \"An FPR should be available when a GPR is reserved.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7005, __PRETTY_FUNCTION__)); |
7006 | if (State.isVarArg()) { |
7007 | // Successfully reserved GPRs are only initialized for vararg calls. |
7008 | // Custom handling is required for: |
7009 | // f64 in PPC32 needs to be split into 2 GPRs. |
7010 | // f32 in PPC64 needs to occupy only lower 32 bits of 64-bit GPR. |
7011 | State.addLoc( |
7012 | CCValAssign::getCustomReg(ValNo, ValVT, Reg, RegVT, LocInfo)); |
7013 | } |
7014 | } else { |
7015 | // If there are insufficient GPRs, the PSA needs to be initialized. |
7016 | // Initialization occurs even if an FPR was initialized for |
7017 | // compatibility with the AIX XL compiler. The full memory for the |
7018 | // argument will be initialized even if a prior word is saved in GPR. |
7019 | // A custom memLoc is used when the argument also passes in FPR so |
7020 | // that the callee handling can skip over it easily. |
7021 | State.addLoc( |
7022 | FReg ? CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, |
7023 | LocInfo) |
7024 | : CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
7025 | break; |
7026 | } |
7027 | } |
7028 | |
7029 | return false; |
7030 | } |
7031 | } |
7032 | return true; |
7033 | } |
7034 | |
7035 | static const TargetRegisterClass *getRegClassForSVT(MVT::SimpleValueType SVT, |
7036 | bool IsPPC64) { |
7037 | assert((IsPPC64 || SVT != MVT::i64) &&(((IsPPC64 || SVT != MVT::i64) && "i64 should have been split for 32-bit codegen." ) ? static_cast<void> (0) : __assert_fail ("(IsPPC64 || SVT != MVT::i64) && \"i64 should have been split for 32-bit codegen.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7038, __PRETTY_FUNCTION__)) |
7038 | "i64 should have been split for 32-bit codegen.")(((IsPPC64 || SVT != MVT::i64) && "i64 should have been split for 32-bit codegen." ) ? static_cast<void> (0) : __assert_fail ("(IsPPC64 || SVT != MVT::i64) && \"i64 should have been split for 32-bit codegen.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7038, __PRETTY_FUNCTION__)); |
7039 | |
7040 | switch (SVT) { |
7041 | default: |
7042 | report_fatal_error("Unexpected value type for formal argument"); |
7043 | case MVT::i1: |
7044 | case MVT::i32: |
7045 | case MVT::i64: |
7046 | return IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
7047 | case MVT::f32: |
7048 | return &PPC::F4RCRegClass; |
7049 | case MVT::f64: |
7050 | return &PPC::F8RCRegClass; |
7051 | } |
7052 | } |
7053 | |
7054 | static SDValue truncateScalarIntegerArg(ISD::ArgFlagsTy Flags, EVT ValVT, |
7055 | SelectionDAG &DAG, SDValue ArgValue, |
7056 | MVT LocVT, const SDLoc &dl) { |
7057 | assert(ValVT.isScalarInteger() && LocVT.isScalarInteger())((ValVT.isScalarInteger() && LocVT.isScalarInteger()) ? static_cast<void> (0) : __assert_fail ("ValVT.isScalarInteger() && LocVT.isScalarInteger()" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7057, __PRETTY_FUNCTION__)); |
7058 | assert(ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits())((ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits()) ? static_cast<void> (0) : __assert_fail ("ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits()" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7058, __PRETTY_FUNCTION__)); |
7059 | |
7060 | if (Flags.isSExt()) |
7061 | ArgValue = DAG.getNode(ISD::AssertSext, dl, LocVT, ArgValue, |
7062 | DAG.getValueType(ValVT)); |
7063 | else if (Flags.isZExt()) |
7064 | ArgValue = DAG.getNode(ISD::AssertZext, dl, LocVT, ArgValue, |
7065 | DAG.getValueType(ValVT)); |
7066 | |
7067 | return DAG.getNode(ISD::TRUNCATE, dl, ValVT, ArgValue); |
7068 | } |
7069 | |
7070 | static unsigned mapArgRegToOffsetAIX(unsigned Reg, const PPCFrameLowering *FL) { |
7071 | const unsigned LASize = FL->getLinkageSize(); |
7072 | |
7073 | if (PPC::GPRCRegClass.contains(Reg)) { |
7074 | assert(Reg >= PPC::R3 && Reg <= PPC::R10 &&((Reg >= PPC::R3 && Reg <= PPC::R10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::R3 && Reg <= PPC::R10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7075, __PRETTY_FUNCTION__)) |
7075 | "Reg must be a valid argument register!")((Reg >= PPC::R3 && Reg <= PPC::R10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::R3 && Reg <= PPC::R10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7075, __PRETTY_FUNCTION__)); |
7076 | return LASize + 4 * (Reg - PPC::R3); |
7077 | } |
7078 | |
7079 | if (PPC::G8RCRegClass.contains(Reg)) { |
7080 | assert(Reg >= PPC::X3 && Reg <= PPC::X10 &&((Reg >= PPC::X3 && Reg <= PPC::X10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::X3 && Reg <= PPC::X10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7081, __PRETTY_FUNCTION__)) |
7081 | "Reg must be a valid argument register!")((Reg >= PPC::X3 && Reg <= PPC::X10 && "Reg must be a valid argument register!" ) ? static_cast<void> (0) : __assert_fail ("Reg >= PPC::X3 && Reg <= PPC::X10 && \"Reg must be a valid argument register!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7081, __PRETTY_FUNCTION__)); |
7082 | return LASize + 8 * (Reg - PPC::X3); |
7083 | } |
7084 | |
7085 | llvm_unreachable("Only general purpose registers expected.")::llvm::llvm_unreachable_internal("Only general purpose registers expected." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7085); |
7086 | } |
7087 | |
7088 | // AIX ABI Stack Frame Layout: |
7089 | // |
7090 | // Low Memory +--------------------------------------------+ |
7091 | // SP +---> | Back chain | ---+ |
7092 | // | +--------------------------------------------+ | |
7093 | // | | Saved Condition Register | | |
7094 | // | +--------------------------------------------+ | |
7095 | // | | Saved Linkage Register | | |
7096 | // | +--------------------------------------------+ | Linkage Area |
7097 | // | | Reserved for compilers | | |
7098 | // | +--------------------------------------------+ | |
7099 | // | | Reserved for binders | | |
7100 | // | +--------------------------------------------+ | |
7101 | // | | Saved TOC pointer | ---+ |
7102 | // | +--------------------------------------------+ |
7103 | // | | Parameter save area | |
7104 | // | +--------------------------------------------+ |
7105 | // | | Alloca space | |
7106 | // | +--------------------------------------------+ |
7107 | // | | Local variable space | |
7108 | // | +--------------------------------------------+ |
7109 | // | | Float/int conversion temporary | |
7110 | // | +--------------------------------------------+ |
7111 | // | | Save area for AltiVec registers | |
7112 | // | +--------------------------------------------+ |
7113 | // | | AltiVec alignment padding | |
7114 | // | +--------------------------------------------+ |
7115 | // | | Save area for VRSAVE register | |
7116 | // | +--------------------------------------------+ |
7117 | // | | Save area for General Purpose registers | |
7118 | // | +--------------------------------------------+ |
7119 | // | | Save area for Floating Point registers | |
7120 | // | +--------------------------------------------+ |
7121 | // +---- | Back chain | |
7122 | // High Memory +--------------------------------------------+ |
7123 | // |
7124 | // Specifications: |
7125 | // AIX 7.2 Assembler Language Reference |
7126 | // Subroutine linkage convention |
7127 | |
7128 | SDValue PPCTargetLowering::LowerFormalArguments_AIX( |
7129 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
7130 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
7131 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
7132 | |
7133 | assert((CallConv == CallingConv::C || CallConv == CallingConv::Cold ||(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7135, __PRETTY_FUNCTION__)) |
7134 | CallConv == CallingConv::Fast) &&(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7135, __PRETTY_FUNCTION__)) |
7135 | "Unexpected calling convention!")(((CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CallConv == CallingConv::C || CallConv == CallingConv::Cold || CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7135, __PRETTY_FUNCTION__)); |
7136 | |
7137 | if (getTargetMachine().Options.GuaranteedTailCallOpt) |
7138 | report_fatal_error("Tail call support is unimplemented on AIX."); |
7139 | |
7140 | if (useSoftFloat()) |
7141 | report_fatal_error("Soft float support is unimplemented on AIX."); |
7142 | |
7143 | const PPCSubtarget &Subtarget = |
7144 | static_cast<const PPCSubtarget &>(DAG.getSubtarget()); |
7145 | |
7146 | const bool IsPPC64 = Subtarget.isPPC64(); |
7147 | const unsigned PtrByteSize = IsPPC64 ? 8 : 4; |
7148 | |
7149 | // Assign locations to all of the incoming arguments. |
7150 | SmallVector<CCValAssign, 16> ArgLocs; |
7151 | MachineFunction &MF = DAG.getMachineFunction(); |
7152 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
7153 | CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext()); |
7154 | |
7155 | const EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7156 | // Reserve space for the linkage area on the stack. |
7157 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
7158 | CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize)); |
7159 | CCInfo.AnalyzeFormalArguments(Ins, CC_AIX); |
7160 | |
7161 | SmallVector<SDValue, 8> MemOps; |
7162 | |
7163 | for (size_t I = 0, End = ArgLocs.size(); I != End; /* No increment here */) { |
7164 | CCValAssign &VA = ArgLocs[I++]; |
7165 | MVT LocVT = VA.getLocVT(); |
7166 | ISD::ArgFlagsTy Flags = Ins[VA.getValNo()].Flags; |
7167 | |
7168 | // For compatibility with the AIX XL compiler, the float args in the |
7169 | // parameter save area are initialized even if the argument is available |
7170 | // in register. The caller is required to initialize both the register |
7171 | // and memory, however, the callee can choose to expect it in either. |
7172 | // The memloc is dismissed here because the argument is retrieved from |
7173 | // the register. |
7174 | if (VA.isMemLoc() && VA.needsCustom()) |
7175 | continue; |
7176 | |
7177 | if (Flags.isByVal() && VA.isMemLoc()) { |
7178 | const unsigned Size = |
7179 | alignTo(Flags.getByValSize() ? Flags.getByValSize() : PtrByteSize, |
7180 | PtrByteSize); |
7181 | const int FI = MF.getFrameInfo().CreateFixedObject( |
7182 | Size, VA.getLocMemOffset(), /* IsImmutable */ false, |
7183 | /* IsAliased */ true); |
7184 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7185 | InVals.push_back(FIN); |
7186 | |
7187 | continue; |
7188 | } |
7189 | |
7190 | if (Flags.isByVal()) { |
7191 | assert(VA.isRegLoc() && "MemLocs should already be handled.")((VA.isRegLoc() && "MemLocs should already be handled." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"MemLocs should already be handled.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7191, __PRETTY_FUNCTION__)); |
7192 | |
7193 | const MCPhysReg ArgReg = VA.getLocReg(); |
7194 | const PPCFrameLowering *FL = Subtarget.getFrameLowering(); |
7195 | |
7196 | if (Flags.getNonZeroByValAlign() > PtrByteSize) |
7197 | report_fatal_error("Over aligned byvals not supported yet."); |
7198 | |
7199 | const unsigned StackSize = alignTo(Flags.getByValSize(), PtrByteSize); |
7200 | const int FI = MF.getFrameInfo().CreateFixedObject( |
7201 | StackSize, mapArgRegToOffsetAIX(ArgReg, FL), /* IsImmutable */ false, |
7202 | /* IsAliased */ true); |
7203 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7204 | InVals.push_back(FIN); |
7205 | |
7206 | // Add live ins for all the RegLocs for the same ByVal. |
7207 | const TargetRegisterClass *RegClass = |
7208 | IsPPC64 ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
7209 | |
7210 | auto HandleRegLoc = [&, RegClass, LocVT](const MCPhysReg PhysReg, |
7211 | unsigned Offset) { |
7212 | const unsigned VReg = MF.addLiveIn(PhysReg, RegClass); |
7213 | // Since the callers side has left justified the aggregate in the |
7214 | // register, we can simply store the entire register into the stack |
7215 | // slot. |
7216 | SDValue CopyFrom = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); |
7217 | // The store to the fixedstack object is needed becuase accessing a |
7218 | // field of the ByVal will use a gep and load. Ideally we will optimize |
7219 | // to extracting the value from the register directly, and elide the |
7220 | // stores when the arguments address is not taken, but that will need to |
7221 | // be future work. |
7222 | SDValue Store = DAG.getStore( |
7223 | CopyFrom.getValue(1), dl, CopyFrom, |
7224 | DAG.getObjectPtrOffset(dl, FIN, TypeSize::Fixed(Offset)), |
7225 | MachinePointerInfo::getFixedStack(MF, FI, Offset)); |
7226 | |
7227 | MemOps.push_back(Store); |
7228 | }; |
7229 | |
7230 | unsigned Offset = 0; |
7231 | HandleRegLoc(VA.getLocReg(), Offset); |
7232 | Offset += PtrByteSize; |
7233 | for (; Offset != StackSize && ArgLocs[I].isRegLoc(); |
7234 | Offset += PtrByteSize) { |
7235 | assert(ArgLocs[I].getValNo() == VA.getValNo() &&((ArgLocs[I].getValNo() == VA.getValNo() && "RegLocs should be for ByVal argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"RegLocs should be for ByVal argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7236, __PRETTY_FUNCTION__)) |
7236 | "RegLocs should be for ByVal argument.")((ArgLocs[I].getValNo() == VA.getValNo() && "RegLocs should be for ByVal argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"RegLocs should be for ByVal argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7236, __PRETTY_FUNCTION__)); |
7237 | |
7238 | const CCValAssign RL = ArgLocs[I++]; |
7239 | HandleRegLoc(RL.getLocReg(), Offset); |
7240 | } |
7241 | |
7242 | if (Offset != StackSize) { |
7243 | assert(ArgLocs[I].getValNo() == VA.getValNo() &&((ArgLocs[I].getValNo() == VA.getValNo() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7244, __PRETTY_FUNCTION__)) |
7244 | "Expected MemLoc for remaining bytes.")((ArgLocs[I].getValNo() == VA.getValNo() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == VA.getValNo() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7244, __PRETTY_FUNCTION__)); |
7245 | assert(ArgLocs[I].isMemLoc() && "Expected MemLoc for remaining bytes.")((ArgLocs[I].isMemLoc() && "Expected MemLoc for remaining bytes." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].isMemLoc() && \"Expected MemLoc for remaining bytes.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7245, __PRETTY_FUNCTION__)); |
7246 | // Consume the MemLoc.The InVal has already been emitted, so nothing |
7247 | // more needs to be done. |
7248 | ++I; |
7249 | } |
7250 | |
7251 | continue; |
7252 | } |
7253 | |
7254 | EVT ValVT = VA.getValVT(); |
7255 | if (VA.isRegLoc() && !VA.needsCustom()) { |
7256 | MVT::SimpleValueType SVT = ValVT.getSimpleVT().SimpleTy; |
7257 | unsigned VReg = |
7258 | MF.addLiveIn(VA.getLocReg(), getRegClassForSVT(SVT, IsPPC64)); |
7259 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl, VReg, LocVT); |
7260 | if (ValVT.isScalarInteger() && |
7261 | (ValVT.getFixedSizeInBits() < LocVT.getFixedSizeInBits())) { |
7262 | ArgValue = |
7263 | truncateScalarIntegerArg(Flags, ValVT, DAG, ArgValue, LocVT, dl); |
7264 | } |
7265 | InVals.push_back(ArgValue); |
7266 | continue; |
7267 | } |
7268 | if (VA.isMemLoc()) { |
7269 | const unsigned LocSize = LocVT.getStoreSize(); |
7270 | const unsigned ValSize = ValVT.getStoreSize(); |
7271 | assert((ValSize <= LocSize) &&(((ValSize <= LocSize) && "Object size is larger than size of MemLoc" ) ? static_cast<void> (0) : __assert_fail ("(ValSize <= LocSize) && \"Object size is larger than size of MemLoc\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7272, __PRETTY_FUNCTION__)) |
7272 | "Object size is larger than size of MemLoc")(((ValSize <= LocSize) && "Object size is larger than size of MemLoc" ) ? static_cast<void> (0) : __assert_fail ("(ValSize <= LocSize) && \"Object size is larger than size of MemLoc\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7272, __PRETTY_FUNCTION__)); |
7273 | int CurArgOffset = VA.getLocMemOffset(); |
7274 | // Objects are right-justified because AIX is big-endian. |
7275 | if (LocSize > ValSize) |
7276 | CurArgOffset += LocSize - ValSize; |
7277 | // Potential tail calls could cause overwriting of argument stack slots. |
7278 | const bool IsImmutable = |
7279 | !(getTargetMachine().Options.GuaranteedTailCallOpt && |
7280 | (CallConv == CallingConv::Fast)); |
7281 | int FI = MFI.CreateFixedObject(ValSize, CurArgOffset, IsImmutable); |
7282 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); |
7283 | SDValue ArgValue = |
7284 | DAG.getLoad(ValVT, dl, Chain, FIN, MachinePointerInfo()); |
7285 | InVals.push_back(ArgValue); |
7286 | continue; |
7287 | } |
7288 | } |
7289 | |
7290 | // On AIX a minimum of 8 words is saved to the parameter save area. |
7291 | const unsigned MinParameterSaveArea = 8 * PtrByteSize; |
7292 | // Area that is at least reserved in the caller of this function. |
7293 | unsigned CallerReservedArea = |
7294 | std::max(CCInfo.getNextStackOffset(), LinkageSize + MinParameterSaveArea); |
7295 | |
7296 | // Set the size that is at least reserved in caller of this function. Tail |
7297 | // call optimized function's reserved stack space needs to be aligned so |
7298 | // that taking the difference between two stack areas will result in an |
7299 | // aligned stack. |
7300 | CallerReservedArea = |
7301 | EnsureStackAlignment(Subtarget.getFrameLowering(), CallerReservedArea); |
7302 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
7303 | FuncInfo->setMinReservedArea(CallerReservedArea); |
7304 | |
7305 | if (isVarArg) { |
7306 | FuncInfo->setVarArgsFrameIndex( |
7307 | MFI.CreateFixedObject(PtrByteSize, CCInfo.getNextStackOffset(), true)); |
7308 | SDValue FIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT); |
7309 | |
7310 | static const MCPhysReg GPR_32[] = {PPC::R3, PPC::R4, PPC::R5, PPC::R6, |
7311 | PPC::R7, PPC::R8, PPC::R9, PPC::R10}; |
7312 | |
7313 | static const MCPhysReg GPR_64[] = {PPC::X3, PPC::X4, PPC::X5, PPC::X6, |
7314 | PPC::X7, PPC::X8, PPC::X9, PPC::X10}; |
7315 | const unsigned NumGPArgRegs = array_lengthof(IsPPC64 ? GPR_64 : GPR_32); |
7316 | |
7317 | // The fixed integer arguments of a variadic function are stored to the |
7318 | // VarArgsFrameIndex on the stack so that they may be loaded by |
7319 | // dereferencing the result of va_next. |
7320 | for (unsigned GPRIndex = |
7321 | (CCInfo.getNextStackOffset() - LinkageSize) / PtrByteSize; |
7322 | GPRIndex < NumGPArgRegs; ++GPRIndex) { |
7323 | |
7324 | const unsigned VReg = |
7325 | IsPPC64 ? MF.addLiveIn(GPR_64[GPRIndex], &PPC::G8RCRegClass) |
7326 | : MF.addLiveIn(GPR_32[GPRIndex], &PPC::GPRCRegClass); |
7327 | |
7328 | SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, PtrVT); |
7329 | SDValue Store = |
7330 | DAG.getStore(Val.getValue(1), dl, Val, FIN, MachinePointerInfo()); |
7331 | MemOps.push_back(Store); |
7332 | // Increment the address for the next argument to store. |
7333 | SDValue PtrOff = DAG.getConstant(PtrByteSize, dl, PtrVT); |
7334 | FIN = DAG.getNode(ISD::ADD, dl, PtrOff.getValueType(), FIN, PtrOff); |
7335 | } |
7336 | } |
7337 | |
7338 | if (!MemOps.empty()) |
7339 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
7340 | |
7341 | return Chain; |
7342 | } |
7343 | |
7344 | SDValue PPCTargetLowering::LowerCall_AIX( |
7345 | SDValue Chain, SDValue Callee, CallFlags CFlags, |
7346 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7347 | const SmallVectorImpl<SDValue> &OutVals, |
7348 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
7349 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
7350 | const CallBase *CB) const { |
7351 | // See PPCTargetLowering::LowerFormalArguments_AIX() for a description of the |
7352 | // AIX ABI stack frame layout. |
7353 | |
7354 | assert((CFlags.CallConv == CallingConv::C ||(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7357, __PRETTY_FUNCTION__)) |
7355 | CFlags.CallConv == CallingConv::Cold ||(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7357, __PRETTY_FUNCTION__)) |
7356 | CFlags.CallConv == CallingConv::Fast) &&(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7357, __PRETTY_FUNCTION__)) |
7357 | "Unexpected calling convention!")(((CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv ::Cold || CFlags.CallConv == CallingConv::Fast) && "Unexpected calling convention!" ) ? static_cast<void> (0) : __assert_fail ("(CFlags.CallConv == CallingConv::C || CFlags.CallConv == CallingConv::Cold || CFlags.CallConv == CallingConv::Fast) && \"Unexpected calling convention!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7357, __PRETTY_FUNCTION__)); |
7358 | |
7359 | if (CFlags.IsPatchPoint) |
7360 | report_fatal_error("This call type is unimplemented on AIX."); |
7361 | |
7362 | const PPCSubtarget& Subtarget = |
7363 | static_cast<const PPCSubtarget&>(DAG.getSubtarget()); |
7364 | if (Subtarget.hasAltivec()) |
7365 | report_fatal_error("Altivec support is unimplemented on AIX."); |
7366 | |
7367 | MachineFunction &MF = DAG.getMachineFunction(); |
7368 | SmallVector<CCValAssign, 16> ArgLocs; |
7369 | CCState CCInfo(CFlags.CallConv, CFlags.IsVarArg, MF, ArgLocs, |
7370 | *DAG.getContext()); |
7371 | |
7372 | // Reserve space for the linkage save area (LSA) on the stack. |
7373 | // In both PPC32 and PPC64 there are 6 reserved slots in the LSA: |
7374 | // [SP][CR][LR][2 x reserved][TOC]. |
7375 | // The LSA is 24 bytes (6x4) in PPC32 and 48 bytes (6x8) in PPC64. |
7376 | const unsigned LinkageSize = Subtarget.getFrameLowering()->getLinkageSize(); |
7377 | const bool IsPPC64 = Subtarget.isPPC64(); |
7378 | const EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7379 | const unsigned PtrByteSize = IsPPC64 ? 8 : 4; |
7380 | CCInfo.AllocateStack(LinkageSize, Align(PtrByteSize)); |
7381 | CCInfo.AnalyzeCallOperands(Outs, CC_AIX); |
7382 | |
7383 | // The prolog code of the callee may store up to 8 GPR argument registers to |
7384 | // the stack, allowing va_start to index over them in memory if the callee |
7385 | // is variadic. |
7386 | // Because we cannot tell if this is needed on the caller side, we have to |
7387 | // conservatively assume that it is needed. As such, make sure we have at |
7388 | // least enough stack space for the caller to store the 8 GPRs. |
7389 | const unsigned MinParameterSaveAreaSize = 8 * PtrByteSize; |
7390 | const unsigned NumBytes = std::max(LinkageSize + MinParameterSaveAreaSize, |
7391 | CCInfo.getNextStackOffset()); |
7392 | |
7393 | // Adjust the stack pointer for the new arguments... |
7394 | // These operations are automatically eliminated by the prolog/epilog pass. |
7395 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
7396 | SDValue CallSeqStart = Chain; |
7397 | |
7398 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
7399 | SmallVector<SDValue, 8> MemOpChains; |
7400 | |
7401 | // Set up a copy of the stack pointer for loading and storing any |
7402 | // arguments that may not fit in the registers available for argument |
7403 | // passing. |
7404 | const SDValue StackPtr = IsPPC64 ? DAG.getRegister(PPC::X1, MVT::i64) |
7405 | : DAG.getRegister(PPC::R1, MVT::i32); |
7406 | |
7407 | for (unsigned I = 0, E = ArgLocs.size(); I != E;) { |
7408 | const unsigned ValNo = ArgLocs[I].getValNo(); |
7409 | SDValue Arg = OutVals[ValNo]; |
7410 | ISD::ArgFlagsTy Flags = Outs[ValNo].Flags; |
7411 | |
7412 | if (Flags.isByVal()) { |
7413 | const unsigned ByValSize = Flags.getByValSize(); |
7414 | |
7415 | // Nothing to do for zero-sized ByVals on the caller side. |
7416 | if (!ByValSize) { |
7417 | ++I; |
7418 | continue; |
7419 | } |
7420 | |
7421 | auto GetLoad = [&](EVT VT, unsigned LoadOffset) { |
7422 | return DAG.getExtLoad( |
7423 | ISD::ZEXTLOAD, dl, PtrVT, Chain, |
7424 | (LoadOffset != 0) |
7425 | ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset)) |
7426 | : Arg, |
7427 | MachinePointerInfo(), VT); |
7428 | }; |
7429 | |
7430 | unsigned LoadOffset = 0; |
7431 | |
7432 | // Initialize registers, which are fully occupied by the by-val argument. |
7433 | while (LoadOffset + PtrByteSize <= ByValSize && ArgLocs[I].isRegLoc()) { |
7434 | SDValue Load = GetLoad(PtrVT, LoadOffset); |
7435 | MemOpChains.push_back(Load.getValue(1)); |
7436 | LoadOffset += PtrByteSize; |
7437 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7438 | assert(ByValVA.getValNo() == ValNo &&((ByValVA.getValNo() == ValNo && "Unexpected location for pass-by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ByValVA.getValNo() == ValNo && \"Unexpected location for pass-by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7439, __PRETTY_FUNCTION__)) |
7439 | "Unexpected location for pass-by-value argument.")((ByValVA.getValNo() == ValNo && "Unexpected location for pass-by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ByValVA.getValNo() == ValNo && \"Unexpected location for pass-by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7439, __PRETTY_FUNCTION__)); |
7440 | RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), Load)); |
7441 | } |
7442 | |
7443 | if (LoadOffset == ByValSize) |
7444 | continue; |
7445 | |
7446 | // There must be one more loc to handle the remainder. |
7447 | assert(ArgLocs[I].getValNo() == ValNo &&((ArgLocs[I].getValNo() == ValNo && "Expected additional location for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == ValNo && \"Expected additional location for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7448, __PRETTY_FUNCTION__)) |
7448 | "Expected additional location for by-value argument.")((ArgLocs[I].getValNo() == ValNo && "Expected additional location for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ArgLocs[I].getValNo() == ValNo && \"Expected additional location for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7448, __PRETTY_FUNCTION__)); |
7449 | |
7450 | if (ArgLocs[I].isMemLoc()) { |
7451 | assert(LoadOffset < ByValSize && "Unexpected memloc for by-val arg.")((LoadOffset < ByValSize && "Unexpected memloc for by-val arg." ) ? static_cast<void> (0) : __assert_fail ("LoadOffset < ByValSize && \"Unexpected memloc for by-val arg.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7451, __PRETTY_FUNCTION__)); |
7452 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7453 | ISD::ArgFlagsTy MemcpyFlags = Flags; |
7454 | // Only memcpy the bytes that don't pass in register. |
7455 | MemcpyFlags.setByValSize(ByValSize - LoadOffset); |
7456 | Chain = CallSeqStart = createMemcpyOutsideCallSeq( |
7457 | (LoadOffset != 0) |
7458 | ? DAG.getObjectPtrOffset(dl, Arg, TypeSize::Fixed(LoadOffset)) |
7459 | : Arg, |
7460 | DAG.getObjectPtrOffset(dl, StackPtr, |
7461 | TypeSize::Fixed(ByValVA.getLocMemOffset())), |
7462 | CallSeqStart, MemcpyFlags, DAG, dl); |
7463 | continue; |
7464 | } |
7465 | |
7466 | // Initialize the final register residue. |
7467 | // Any residue that occupies the final by-val arg register must be |
7468 | // left-justified on AIX. Loads must be a power-of-2 size and cannot be |
7469 | // larger than the ByValSize. For example: a 7 byte by-val arg requires 4, |
7470 | // 2 and 1 byte loads. |
7471 | const unsigned ResidueBytes = ByValSize % PtrByteSize; |
7472 | assert(ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize &&((ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && "Unexpected register residue for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && \"Unexpected register residue for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7473, __PRETTY_FUNCTION__)) |
7473 | "Unexpected register residue for by-value argument.")((ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && "Unexpected register residue for by-value argument." ) ? static_cast<void> (0) : __assert_fail ("ResidueBytes != 0 && LoadOffset + PtrByteSize > ByValSize && \"Unexpected register residue for by-value argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7473, __PRETTY_FUNCTION__)); |
7474 | SDValue ResidueVal; |
7475 | for (unsigned Bytes = 0; Bytes != ResidueBytes;) { |
7476 | const unsigned N = PowerOf2Floor(ResidueBytes - Bytes); |
7477 | const MVT VT = |
7478 | N == 1 ? MVT::i8 |
7479 | : ((N == 2) ? MVT::i16 : (N == 4 ? MVT::i32 : MVT::i64)); |
7480 | SDValue Load = GetLoad(VT, LoadOffset); |
7481 | MemOpChains.push_back(Load.getValue(1)); |
7482 | LoadOffset += N; |
7483 | Bytes += N; |
7484 | |
7485 | // By-val arguments are passed left-justfied in register. |
7486 | // Every load here needs to be shifted, otherwise a full register load |
7487 | // should have been used. |
7488 | assert(PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) &&((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7490, __PRETTY_FUNCTION__)) |
7489 | "Unexpected load emitted during handling of pass-by-value "((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7490, __PRETTY_FUNCTION__)) |
7490 | "argument.")((PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && "Unexpected load emitted during handling of pass-by-value " "argument." ) ? static_cast<void> (0) : __assert_fail ("PtrVT.getSimpleVT().getSizeInBits() > (Bytes * 8) && \"Unexpected load emitted during handling of pass-by-value \" \"argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7490, __PRETTY_FUNCTION__)); |
7491 | unsigned NumSHLBits = PtrVT.getSimpleVT().getSizeInBits() - (Bytes * 8); |
7492 | EVT ShiftAmountTy = |
7493 | getShiftAmountTy(Load->getValueType(0), DAG.getDataLayout()); |
7494 | SDValue SHLAmt = DAG.getConstant(NumSHLBits, dl, ShiftAmountTy); |
7495 | SDValue ShiftedLoad = |
7496 | DAG.getNode(ISD::SHL, dl, Load.getValueType(), Load, SHLAmt); |
7497 | ResidueVal = ResidueVal ? DAG.getNode(ISD::OR, dl, PtrVT, ResidueVal, |
7498 | ShiftedLoad) |
7499 | : ShiftedLoad; |
7500 | } |
7501 | |
7502 | const CCValAssign &ByValVA = ArgLocs[I++]; |
7503 | RegsToPass.push_back(std::make_pair(ByValVA.getLocReg(), ResidueVal)); |
7504 | continue; |
7505 | } |
7506 | |
7507 | CCValAssign &VA = ArgLocs[I++]; |
7508 | const MVT LocVT = VA.getLocVT(); |
7509 | const MVT ValVT = VA.getValVT(); |
7510 | |
7511 | switch (VA.getLocInfo()) { |
7512 | default: |
7513 | report_fatal_error("Unexpected argument extension type."); |
7514 | case CCValAssign::Full: |
7515 | break; |
7516 | case CCValAssign::ZExt: |
7517 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); |
7518 | break; |
7519 | case CCValAssign::SExt: |
7520 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); |
7521 | break; |
7522 | } |
7523 | |
7524 | if (VA.isRegLoc() && !VA.needsCustom()) { |
7525 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
7526 | continue; |
7527 | } |
7528 | |
7529 | if (VA.isMemLoc()) { |
7530 | SDValue PtrOff = |
7531 | DAG.getConstant(VA.getLocMemOffset(), dl, StackPtr.getValueType()); |
7532 | PtrOff = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
7533 | MemOpChains.push_back( |
7534 | DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo())); |
7535 | |
7536 | continue; |
7537 | } |
7538 | |
7539 | // Custom handling is used for GPR initializations for vararg float |
7540 | // arguments. |
7541 | assert(VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg &&((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7543, __PRETTY_FUNCTION__)) |
7542 | ValVT.isFloatingPoint() && LocVT.isInteger() &&((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7543, __PRETTY_FUNCTION__)) |
7543 | "Unexpected register handling for calling convention.")((VA.isRegLoc() && VA.needsCustom() && CFlags .IsVarArg && ValVT.isFloatingPoint() && LocVT .isInteger() && "Unexpected register handling for calling convention." ) ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && VA.needsCustom() && CFlags.IsVarArg && ValVT.isFloatingPoint() && LocVT.isInteger() && \"Unexpected register handling for calling convention.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7543, __PRETTY_FUNCTION__)); |
7544 | |
7545 | SDValue ArgAsInt = |
7546 | DAG.getBitcast(MVT::getIntegerVT(ValVT.getSizeInBits()), Arg); |
7547 | |
7548 | if (Arg.getValueType().getStoreSize() == LocVT.getStoreSize()) |
7549 | // f32 in 32-bit GPR |
7550 | // f64 in 64-bit GPR |
7551 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgAsInt)); |
7552 | else if (Arg.getValueType().getFixedSizeInBits() < |
7553 | LocVT.getFixedSizeInBits()) |
7554 | // f32 in 64-bit GPR. |
7555 | RegsToPass.push_back(std::make_pair( |
7556 | VA.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, LocVT))); |
7557 | else { |
7558 | // f64 in two 32-bit GPRs |
7559 | // The 2 GPRs are marked custom and expected to be adjacent in ArgLocs. |
7560 | assert(Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 &&((Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && "Unexpected custom register for argument!" ) ? static_cast<void> (0) : __assert_fail ("Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && \"Unexpected custom register for argument!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7561, __PRETTY_FUNCTION__)) |
7561 | "Unexpected custom register for argument!")((Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && "Unexpected custom register for argument!" ) ? static_cast<void> (0) : __assert_fail ("Arg.getValueType() == MVT::f64 && CFlags.IsVarArg && !IsPPC64 && \"Unexpected custom register for argument!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7561, __PRETTY_FUNCTION__)); |
7562 | CCValAssign &GPR1 = VA; |
7563 | SDValue MSWAsI64 = DAG.getNode(ISD::SRL, dl, MVT::i64, ArgAsInt, |
7564 | DAG.getConstant(32, dl, MVT::i8)); |
7565 | RegsToPass.push_back(std::make_pair( |
7566 | GPR1.getLocReg(), DAG.getZExtOrTrunc(MSWAsI64, dl, MVT::i32))); |
7567 | |
7568 | if (I != E) { |
7569 | // If only 1 GPR was available, there will only be one custom GPR and |
7570 | // the argument will also pass in memory. |
7571 | CCValAssign &PeekArg = ArgLocs[I]; |
7572 | if (PeekArg.isRegLoc() && PeekArg.getValNo() == PeekArg.getValNo()) { |
7573 | assert(PeekArg.needsCustom() && "A second custom GPR is expected.")((PeekArg.needsCustom() && "A second custom GPR is expected." ) ? static_cast<void> (0) : __assert_fail ("PeekArg.needsCustom() && \"A second custom GPR is expected.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7573, __PRETTY_FUNCTION__)); |
7574 | CCValAssign &GPR2 = ArgLocs[I++]; |
7575 | RegsToPass.push_back(std::make_pair( |
7576 | GPR2.getLocReg(), DAG.getZExtOrTrunc(ArgAsInt, dl, MVT::i32))); |
7577 | } |
7578 | } |
7579 | } |
7580 | } |
7581 | |
7582 | if (!MemOpChains.empty()) |
7583 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
7584 | |
7585 | // For indirect calls, we need to save the TOC base to the stack for |
7586 | // restoration after the call. |
7587 | if (CFlags.IsIndirect) { |
7588 | assert(!CFlags.IsTailCall && "Indirect tail-calls not supported.")((!CFlags.IsTailCall && "Indirect tail-calls not supported." ) ? static_cast<void> (0) : __assert_fail ("!CFlags.IsTailCall && \"Indirect tail-calls not supported.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7588, __PRETTY_FUNCTION__)); |
7589 | const MCRegister TOCBaseReg = Subtarget.getTOCPointerRegister(); |
7590 | const MCRegister StackPtrReg = Subtarget.getStackPointerRegister(); |
7591 | const MVT PtrVT = Subtarget.isPPC64() ? MVT::i64 : MVT::i32; |
7592 | const unsigned TOCSaveOffset = |
7593 | Subtarget.getFrameLowering()->getTOCSaveOffset(); |
7594 | |
7595 | setUsesTOCBasePtr(DAG); |
7596 | SDValue Val = DAG.getCopyFromReg(Chain, dl, TOCBaseReg, PtrVT); |
7597 | SDValue PtrOff = DAG.getIntPtrConstant(TOCSaveOffset, dl); |
7598 | SDValue StackPtr = DAG.getRegister(StackPtrReg, PtrVT); |
7599 | SDValue AddPtr = DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, PtrOff); |
7600 | Chain = DAG.getStore( |
7601 | Val.getValue(1), dl, Val, AddPtr, |
7602 | MachinePointerInfo::getStack(DAG.getMachineFunction(), TOCSaveOffset)); |
7603 | } |
7604 | |
7605 | // Build a sequence of copy-to-reg nodes chained together with token chain |
7606 | // and flag operands which copy the outgoing args into the appropriate regs. |
7607 | SDValue InFlag; |
7608 | for (auto Reg : RegsToPass) { |
7609 | Chain = DAG.getCopyToReg(Chain, dl, Reg.first, Reg.second, InFlag); |
7610 | InFlag = Chain.getValue(1); |
7611 | } |
7612 | |
7613 | const int SPDiff = 0; |
7614 | return FinishCall(CFlags, dl, DAG, RegsToPass, InFlag, Chain, CallSeqStart, |
7615 | Callee, SPDiff, NumBytes, Ins, InVals, CB); |
7616 | } |
7617 | |
7618 | bool |
7619 | PPCTargetLowering::CanLowerReturn(CallingConv::ID CallConv, |
7620 | MachineFunction &MF, bool isVarArg, |
7621 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7622 | LLVMContext &Context) const { |
7623 | SmallVector<CCValAssign, 16> RVLocs; |
7624 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
7625 | return CCInfo.CheckReturn( |
7626 | Outs, (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
7627 | ? RetCC_PPC_Cold |
7628 | : RetCC_PPC); |
7629 | } |
7630 | |
7631 | SDValue |
7632 | PPCTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
7633 | bool isVarArg, |
7634 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
7635 | const SmallVectorImpl<SDValue> &OutVals, |
7636 | const SDLoc &dl, SelectionDAG &DAG) const { |
7637 | SmallVector<CCValAssign, 16> RVLocs; |
7638 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
7639 | *DAG.getContext()); |
7640 | CCInfo.AnalyzeReturn(Outs, |
7641 | (Subtarget.isSVR4ABI() && CallConv == CallingConv::Cold) |
7642 | ? RetCC_PPC_Cold |
7643 | : RetCC_PPC); |
7644 | |
7645 | SDValue Flag; |
7646 | SmallVector<SDValue, 4> RetOps(1, Chain); |
7647 | |
7648 | // Copy the result values into the output registers. |
7649 | for (unsigned i = 0, RealResIdx = 0; i != RVLocs.size(); ++i, ++RealResIdx) { |
7650 | CCValAssign &VA = RVLocs[i]; |
7651 | assert(VA.isRegLoc() && "Can only return in registers!")((VA.isRegLoc() && "Can only return in registers!") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7651, __PRETTY_FUNCTION__)); |
7652 | |
7653 | SDValue Arg = OutVals[RealResIdx]; |
7654 | |
7655 | if (Subtarget.isAIXABI() && |
7656 | (VA.getLocVT().isVector() || VA.getValVT().isVector())) |
7657 | report_fatal_error("Returning vector types not yet supported on AIX."); |
7658 | |
7659 | switch (VA.getLocInfo()) { |
7660 | default: llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7660); |
7661 | case CCValAssign::Full: break; |
7662 | case CCValAssign::AExt: |
7663 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); |
7664 | break; |
7665 | case CCValAssign::ZExt: |
7666 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); |
7667 | break; |
7668 | case CCValAssign::SExt: |
7669 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); |
7670 | break; |
7671 | } |
7672 | if (Subtarget.hasSPE() && VA.getLocVT() == MVT::f64) { |
7673 | bool isLittleEndian = Subtarget.isLittleEndian(); |
7674 | // Legalize ret f64 -> ret 2 x i32. |
7675 | SDValue SVal = |
7676 | DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
7677 | DAG.getIntPtrConstant(isLittleEndian ? 0 : 1, dl)); |
7678 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag); |
7679 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
7680 | SVal = DAG.getNode(PPCISD::EXTRACT_SPE, dl, MVT::i32, Arg, |
7681 | DAG.getIntPtrConstant(isLittleEndian ? 1 : 0, dl)); |
7682 | Flag = Chain.getValue(1); |
7683 | VA = RVLocs[++i]; // skip ahead to next loc |
7684 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), SVal, Flag); |
7685 | } else |
7686 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); |
7687 | Flag = Chain.getValue(1); |
7688 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
7689 | } |
7690 | |
7691 | RetOps[0] = Chain; // Update chain. |
7692 | |
7693 | // Add the flag if we have it. |
7694 | if (Flag.getNode()) |
7695 | RetOps.push_back(Flag); |
7696 | |
7697 | return DAG.getNode(PPCISD::RET_FLAG, dl, MVT::Other, RetOps); |
7698 | } |
7699 | |
7700 | SDValue |
7701 | PPCTargetLowering::LowerGET_DYNAMIC_AREA_OFFSET(SDValue Op, |
7702 | SelectionDAG &DAG) const { |
7703 | SDLoc dl(Op); |
7704 | |
7705 | // Get the correct type for integers. |
7706 | EVT IntVT = Op.getValueType(); |
7707 | |
7708 | // Get the inputs. |
7709 | SDValue Chain = Op.getOperand(0); |
7710 | SDValue FPSIdx = getFramePointerFrameIndex(DAG); |
7711 | // Build a DYNAREAOFFSET node. |
7712 | SDValue Ops[2] = {Chain, FPSIdx}; |
7713 | SDVTList VTs = DAG.getVTList(IntVT); |
7714 | return DAG.getNode(PPCISD::DYNAREAOFFSET, dl, VTs, Ops); |
7715 | } |
7716 | |
7717 | SDValue PPCTargetLowering::LowerSTACKRESTORE(SDValue Op, |
7718 | SelectionDAG &DAG) const { |
7719 | // When we pop the dynamic allocation we need to restore the SP link. |
7720 | SDLoc dl(Op); |
7721 | |
7722 | // Get the correct type for pointers. |
7723 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7724 | |
7725 | // Construct the stack pointer operand. |
7726 | bool isPPC64 = Subtarget.isPPC64(); |
7727 | unsigned SP = isPPC64 ? PPC::X1 : PPC::R1; |
7728 | SDValue StackPtr = DAG.getRegister(SP, PtrVT); |
7729 | |
7730 | // Get the operands for the STACKRESTORE. |
7731 | SDValue Chain = Op.getOperand(0); |
7732 | SDValue SaveSP = Op.getOperand(1); |
7733 | |
7734 | // Load the old link SP. |
7735 | SDValue LoadLinkSP = |
7736 | DAG.getLoad(PtrVT, dl, Chain, StackPtr, MachinePointerInfo()); |
7737 | |
7738 | // Restore the stack pointer. |
7739 | Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), dl, SP, SaveSP); |
7740 | |
7741 | // Store the old link SP. |
7742 | return DAG.getStore(Chain, dl, LoadLinkSP, StackPtr, MachinePointerInfo()); |
7743 | } |
7744 | |
7745 | SDValue PPCTargetLowering::getReturnAddrFrameIndex(SelectionDAG &DAG) const { |
7746 | MachineFunction &MF = DAG.getMachineFunction(); |
7747 | bool isPPC64 = Subtarget.isPPC64(); |
7748 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7749 | |
7750 | // Get current frame pointer save index. The users of this index will be |
7751 | // primarily DYNALLOC instructions. |
7752 | PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); |
7753 | int RASI = FI->getReturnAddrSaveIndex(); |
7754 | |
7755 | // If the frame pointer save index hasn't been defined yet. |
7756 | if (!RASI) { |
7757 | // Find out what the fix offset of the frame pointer save area. |
7758 | int LROffset = Subtarget.getFrameLowering()->getReturnSaveOffset(); |
7759 | // Allocate the frame index for frame pointer save area. |
7760 | RASI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, LROffset, false); |
7761 | // Save the result. |
7762 | FI->setReturnAddrSaveIndex(RASI); |
7763 | } |
7764 | return DAG.getFrameIndex(RASI, PtrVT); |
7765 | } |
7766 | |
7767 | SDValue |
7768 | PPCTargetLowering::getFramePointerFrameIndex(SelectionDAG & DAG) const { |
7769 | MachineFunction &MF = DAG.getMachineFunction(); |
7770 | bool isPPC64 = Subtarget.isPPC64(); |
7771 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
7772 | |
7773 | // Get current frame pointer save index. The users of this index will be |
7774 | // primarily DYNALLOC instructions. |
7775 | PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); |
7776 | int FPSI = FI->getFramePointerSaveIndex(); |
7777 | |
7778 | // If the frame pointer save index hasn't been defined yet. |
7779 | if (!FPSI) { |
7780 | // Find out what the fix offset of the frame pointer save area. |
7781 | int FPOffset = Subtarget.getFrameLowering()->getFramePointerSaveOffset(); |
7782 | // Allocate the frame index for frame pointer save area. |
7783 | FPSI = MF.getFrameInfo().CreateFixedObject(isPPC64? 8 : 4, FPOffset, true); |
7784 | // Save the result. |
7785 | FI->setFramePointerSaveIndex(FPSI); |
7786 | } |
7787 | return DAG.getFrameIndex(FPSI, PtrVT); |
7788 | } |
7789 | |
7790 | SDValue PPCTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, |
7791 | SelectionDAG &DAG) const { |
7792 | MachineFunction &MF = DAG.getMachineFunction(); |
7793 | // Get the inputs. |
7794 | SDValue Chain = Op.getOperand(0); |
7795 | SDValue Size = Op.getOperand(1); |
7796 | SDLoc dl(Op); |
7797 | |
7798 | // Get the correct type for pointers. |
7799 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7800 | // Negate the size. |
7801 | SDValue NegSize = DAG.getNode(ISD::SUB, dl, PtrVT, |
7802 | DAG.getConstant(0, dl, PtrVT), Size); |
7803 | // Construct a node for the frame pointer save index. |
7804 | SDValue FPSIdx = getFramePointerFrameIndex(DAG); |
7805 | SDValue Ops[3] = { Chain, NegSize, FPSIdx }; |
7806 | SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other); |
7807 | if (hasInlineStackProbe(MF)) |
7808 | return DAG.getNode(PPCISD::PROBED_ALLOCA, dl, VTs, Ops); |
7809 | return DAG.getNode(PPCISD::DYNALLOC, dl, VTs, Ops); |
7810 | } |
7811 | |
7812 | SDValue PPCTargetLowering::LowerEH_DWARF_CFA(SDValue Op, |
7813 | SelectionDAG &DAG) const { |
7814 | MachineFunction &MF = DAG.getMachineFunction(); |
7815 | |
7816 | bool isPPC64 = Subtarget.isPPC64(); |
7817 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
7818 | |
7819 | int FI = MF.getFrameInfo().CreateFixedObject(isPPC64 ? 8 : 4, 0, false); |
7820 | return DAG.getFrameIndex(FI, PtrVT); |
7821 | } |
7822 | |
7823 | SDValue PPCTargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op, |
7824 | SelectionDAG &DAG) const { |
7825 | SDLoc DL(Op); |
7826 | return DAG.getNode(PPCISD::EH_SJLJ_SETJMP, DL, |
7827 | DAG.getVTList(MVT::i32, MVT::Other), |
7828 | Op.getOperand(0), Op.getOperand(1)); |
7829 | } |
7830 | |
7831 | SDValue PPCTargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op, |
7832 | SelectionDAG &DAG) const { |
7833 | SDLoc DL(Op); |
7834 | return DAG.getNode(PPCISD::EH_SJLJ_LONGJMP, DL, MVT::Other, |
7835 | Op.getOperand(0), Op.getOperand(1)); |
7836 | } |
7837 | |
7838 | SDValue PPCTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
7839 | if (Op.getValueType().isVector()) |
7840 | return LowerVectorLoad(Op, DAG); |
7841 | |
7842 | assert(Op.getValueType() == MVT::i1 &&((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 loads" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 loads\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7843, __PRETTY_FUNCTION__)) |
7843 | "Custom lowering only for i1 loads")((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 loads" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 loads\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7843, __PRETTY_FUNCTION__)); |
7844 | |
7845 | // First, load 8 bits into 32 bits, then truncate to 1 bit. |
7846 | |
7847 | SDLoc dl(Op); |
7848 | LoadSDNode *LD = cast<LoadSDNode>(Op); |
7849 | |
7850 | SDValue Chain = LD->getChain(); |
7851 | SDValue BasePtr = LD->getBasePtr(); |
7852 | MachineMemOperand *MMO = LD->getMemOperand(); |
7853 | |
7854 | SDValue NewLD = |
7855 | DAG.getExtLoad(ISD::EXTLOAD, dl, getPointerTy(DAG.getDataLayout()), Chain, |
7856 | BasePtr, MVT::i8, MMO); |
7857 | SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewLD); |
7858 | |
7859 | SDValue Ops[] = { Result, SDValue(NewLD.getNode(), 1) }; |
7860 | return DAG.getMergeValues(Ops, dl); |
7861 | } |
7862 | |
7863 | SDValue PPCTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
7864 | if (Op.getOperand(1).getValueType().isVector()) |
7865 | return LowerVectorStore(Op, DAG); |
7866 | |
7867 | assert(Op.getOperand(1).getValueType() == MVT::i1 &&((Op.getOperand(1).getValueType() == MVT::i1 && "Custom lowering only for i1 stores" ) ? static_cast<void> (0) : __assert_fail ("Op.getOperand(1).getValueType() == MVT::i1 && \"Custom lowering only for i1 stores\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7868, __PRETTY_FUNCTION__)) |
7868 | "Custom lowering only for i1 stores")((Op.getOperand(1).getValueType() == MVT::i1 && "Custom lowering only for i1 stores" ) ? static_cast<void> (0) : __assert_fail ("Op.getOperand(1).getValueType() == MVT::i1 && \"Custom lowering only for i1 stores\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7868, __PRETTY_FUNCTION__)); |
7869 | |
7870 | // First, zero extend to 32 bits, then use a truncating store to 8 bits. |
7871 | |
7872 | SDLoc dl(Op); |
7873 | StoreSDNode *ST = cast<StoreSDNode>(Op); |
7874 | |
7875 | SDValue Chain = ST->getChain(); |
7876 | SDValue BasePtr = ST->getBasePtr(); |
7877 | SDValue Value = ST->getValue(); |
7878 | MachineMemOperand *MMO = ST->getMemOperand(); |
7879 | |
7880 | Value = DAG.getNode(ISD::ZERO_EXTEND, dl, getPointerTy(DAG.getDataLayout()), |
7881 | Value); |
7882 | return DAG.getTruncStore(Chain, dl, Value, BasePtr, MVT::i8, MMO); |
7883 | } |
7884 | |
7885 | // FIXME: Remove this once the ANDI glue bug is fixed: |
7886 | SDValue PPCTargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const { |
7887 | assert(Op.getValueType() == MVT::i1 &&((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 results" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 results\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7888, __PRETTY_FUNCTION__)) |
7888 | "Custom lowering only for i1 results")((Op.getValueType() == MVT::i1 && "Custom lowering only for i1 results" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::i1 && \"Custom lowering only for i1 results\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7888, __PRETTY_FUNCTION__)); |
7889 | |
7890 | SDLoc DL(Op); |
7891 | return DAG.getNode(PPCISD::ANDI_rec_1_GT_BIT, DL, MVT::i1, Op.getOperand(0)); |
7892 | } |
7893 | |
7894 | SDValue PPCTargetLowering::LowerTRUNCATEVector(SDValue Op, |
7895 | SelectionDAG &DAG) const { |
7896 | |
7897 | // Implements a vector truncate that fits in a vector register as a shuffle. |
7898 | // We want to legalize vector truncates down to where the source fits in |
7899 | // a vector register (and target is therefore smaller than vector register |
7900 | // size). At that point legalization will try to custom lower the sub-legal |
7901 | // result and get here - where we can contain the truncate as a single target |
7902 | // operation. |
7903 | |
7904 | // For example a trunc <2 x i16> to <2 x i8> could be visualized as follows: |
7905 | // <MSB1|LSB1, MSB2|LSB2> to <LSB1, LSB2> |
7906 | // |
7907 | // We will implement it for big-endian ordering as this (where x denotes |
7908 | // undefined): |
7909 | // < MSB1|LSB1, MSB2|LSB2, uu, uu, uu, uu, uu, uu> to |
7910 | // < LSB1, LSB2, u, u, u, u, u, u, u, u, u, u, u, u, u, u> |
7911 | // |
7912 | // The same operation in little-endian ordering will be: |
7913 | // <uu, uu, uu, uu, uu, uu, LSB2|MSB2, LSB1|MSB1> to |
7914 | // <u, u, u, u, u, u, u, u, u, u, u, u, u, u, LSB2, LSB1> |
7915 | |
7916 | EVT TrgVT = Op.getValueType(); |
7917 | assert(TrgVT.isVector() && "Vector type expected.")((TrgVT.isVector() && "Vector type expected.") ? static_cast <void> (0) : __assert_fail ("TrgVT.isVector() && \"Vector type expected.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 7917, __PRETTY_FUNCTION__)); |
7918 | unsigned TrgNumElts = TrgVT.getVectorNumElements(); |
7919 | EVT EltVT = TrgVT.getVectorElementType(); |
7920 | if (!isOperationCustom(Op.getOpcode(), TrgVT) || |
7921 | TrgVT.getSizeInBits() > 128 || !isPowerOf2_32(TrgNumElts) || |
7922 | !isPowerOf2_32(EltVT.getSizeInBits())) |
7923 | return SDValue(); |
7924 | |
7925 | SDValue N1 = Op.getOperand(0); |
7926 | EVT SrcVT = N1.getValueType(); |
7927 | unsigned SrcSize = SrcVT.getSizeInBits(); |
7928 | if (SrcSize > 256 || |
7929 | !isPowerOf2_32(SrcVT.getVectorNumElements()) || |
7930 | !isPowerOf2_32(SrcVT.getVectorElementType().getSizeInBits())) |
7931 | return SDValue(); |
7932 | if (SrcSize == 256 && SrcVT.getVectorNumElements() < 2) |
7933 | return SDValue(); |
7934 | |
7935 | unsigned WideNumElts = 128 / EltVT.getSizeInBits(); |
7936 | EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts); |
7937 | |
7938 | SDLoc DL(Op); |
7939 | SDValue Op1, Op2; |
7940 | if (SrcSize == 256) { |
7941 | EVT VecIdxTy = getVectorIdxTy(DAG.getDataLayout()); |
7942 | EVT SplitVT = |
7943 | N1.getValueType().getHalfNumVectorElementsVT(*DAG.getContext()); |
7944 | unsigned SplitNumElts = SplitVT.getVectorNumElements(); |
7945 | Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1, |
7946 | DAG.getConstant(0, DL, VecIdxTy)); |
7947 | Op2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, N1, |
7948 | DAG.getConstant(SplitNumElts, DL, VecIdxTy)); |
7949 | } |
7950 | else { |
7951 | Op1 = SrcSize == 128 ? N1 : widenVec(DAG, N1, DL); |
7952 | Op2 = DAG.getUNDEF(WideVT); |
7953 | } |
7954 | |
7955 | // First list the elements we want to keep. |
7956 | unsigned SizeMult = SrcSize / TrgVT.getSizeInBits(); |
7957 | SmallVector<int, 16> ShuffV; |
7958 | if (Subtarget.isLittleEndian()) |
7959 | for (unsigned i = 0; i < TrgNumElts; ++i) |
7960 | ShuffV.push_back(i * SizeMult); |
7961 | else |
7962 | for (unsigned i = 1; i <= TrgNumElts; ++i) |
7963 | ShuffV.push_back(i * SizeMult - 1); |
7964 | |
7965 | // Populate the remaining elements with undefs. |
7966 | for (unsigned i = TrgNumElts; i < WideNumElts; ++i) |
7967 | // ShuffV.push_back(i + WideNumElts); |
7968 | ShuffV.push_back(WideNumElts + 1); |
7969 | |
7970 | Op1 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op1); |
7971 | Op2 = DAG.getNode(ISD::BITCAST, DL, WideVT, Op2); |
7972 | return DAG.getVectorShuffle(WideVT, DL, Op1, Op2, ShuffV); |
7973 | } |
7974 | |
7975 | /// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when |
7976 | /// possible. |
7977 | SDValue PPCTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { |
7978 | // Not FP, or using SPE? Not a fsel. |
7979 | if (!Op.getOperand(0).getValueType().isFloatingPoint() || |
7980 | !Op.getOperand(2).getValueType().isFloatingPoint() || Subtarget.hasSPE()) |
7981 | return Op; |
7982 | |
7983 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); |
7984 | |
7985 | EVT ResVT = Op.getValueType(); |
7986 | EVT CmpVT = Op.getOperand(0).getValueType(); |
7987 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
7988 | SDValue TV = Op.getOperand(2), FV = Op.getOperand(3); |
7989 | SDLoc dl(Op); |
7990 | SDNodeFlags Flags = Op.getNode()->getFlags(); |
7991 | |
7992 | // We have xsmaxcdp/xsmincdp which are OK to emit even in the |
7993 | // presence of infinities. |
7994 | if (Subtarget.hasP9Vector() && LHS == TV && RHS == FV) { |
7995 | switch (CC) { |
7996 | default: |
7997 | break; |
7998 | case ISD::SETOGT: |
7999 | case ISD::SETGT: |
8000 | return DAG.getNode(PPCISD::XSMAXCDP, dl, Op.getValueType(), LHS, RHS); |
8001 | case ISD::SETOLT: |
8002 | case ISD::SETLT: |
8003 | return DAG.getNode(PPCISD::XSMINCDP, dl, Op.getValueType(), LHS, RHS); |
8004 | } |
8005 | } |
8006 | |
8007 | // We might be able to do better than this under some circumstances, but in |
8008 | // general, fsel-based lowering of select is a finite-math-only optimization. |
8009 | // For more information, see section F.3 of the 2.06 ISA specification. |
8010 | // With ISA 3.0 |
8011 | if ((!DAG.getTarget().Options.NoInfsFPMath && !Flags.hasNoInfs()) || |
8012 | (!DAG.getTarget().Options.NoNaNsFPMath && !Flags.hasNoNaNs())) |
8013 | return Op; |
8014 | |
8015 | // If the RHS of the comparison is a 0.0, we don't need to do the |
8016 | // subtraction at all. |
8017 | SDValue Sel1; |
8018 | if (isFloatingPointZero(RHS)) |
8019 | switch (CC) { |
8020 | default: break; // SETUO etc aren't handled by fsel. |
8021 | case ISD::SETNE: |
8022 | std::swap(TV, FV); |
8023 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8024 | case ISD::SETEQ: |
8025 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8026 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8027 | Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV); |
8028 | if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits |
8029 | Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1); |
8030 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8031 | DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), Sel1, FV); |
8032 | case ISD::SETULT: |
8033 | case ISD::SETLT: |
8034 | std::swap(TV, FV); // fsel is natively setge, swap operands for setlt |
8035 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8036 | case ISD::SETOGE: |
8037 | case ISD::SETGE: |
8038 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8039 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8040 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, LHS, TV, FV); |
8041 | case ISD::SETUGT: |
8042 | case ISD::SETGT: |
8043 | std::swap(TV, FV); // fsel is natively setge, swap operands for setlt |
8044 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8045 | case ISD::SETOLE: |
8046 | case ISD::SETLE: |
8047 | if (LHS.getValueType() == MVT::f32) // Comparison is always 64-bits |
8048 | LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, LHS); |
8049 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8050 | DAG.getNode(ISD::FNEG, dl, MVT::f64, LHS), TV, FV); |
8051 | } |
8052 | |
8053 | SDValue Cmp; |
8054 | switch (CC) { |
8055 | default: break; // SETUO etc aren't handled by fsel. |
8056 | case ISD::SETNE: |
8057 | std::swap(TV, FV); |
8058 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
8059 | case ISD::SETEQ: |
8060 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8061 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8062 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8063 | Sel1 = DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8064 | if (Sel1.getValueType() == MVT::f32) // Comparison is always 64-bits |
8065 | Sel1 = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Sel1); |
8066 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, |
8067 | DAG.getNode(ISD::FNEG, dl, MVT::f64, Cmp), Sel1, FV); |
8068 | case ISD::SETULT: |
8069 | case ISD::SETLT: |
8070 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8071 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8072 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8073 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV); |
8074 | case ISD::SETOGE: |
8075 | case ISD::SETGE: |
8076 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, LHS, RHS, Flags); |
8077 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8078 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8079 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8080 | case ISD::SETUGT: |
8081 | case ISD::SETGT: |
8082 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags); |
8083 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8084 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8085 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, FV, TV); |
8086 | case ISD::SETOLE: |
8087 | case ISD::SETLE: |
8088 | Cmp = DAG.getNode(ISD::FSUB, dl, CmpVT, RHS, LHS, Flags); |
8089 | if (Cmp.getValueType() == MVT::f32) // Comparison is always 64-bits |
8090 | Cmp = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Cmp); |
8091 | return DAG.getNode(PPCISD::FSEL, dl, ResVT, Cmp, TV, FV); |
8092 | } |
8093 | return Op; |
8094 | } |
8095 | |
8096 | static unsigned getPPCStrictOpcode(unsigned Opc) { |
8097 | switch (Opc) { |
8098 | default: |
8099 | llvm_unreachable("No strict version of this opcode!")::llvm::llvm_unreachable_internal("No strict version of this opcode!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8099); |
8100 | case PPCISD::FCTIDZ: |
8101 | return PPCISD::STRICT_FCTIDZ; |
8102 | case PPCISD::FCTIWZ: |
8103 | return PPCISD::STRICT_FCTIWZ; |
8104 | case PPCISD::FCTIDUZ: |
8105 | return PPCISD::STRICT_FCTIDUZ; |
8106 | case PPCISD::FCTIWUZ: |
8107 | return PPCISD::STRICT_FCTIWUZ; |
8108 | case PPCISD::FCFID: |
8109 | return PPCISD::STRICT_FCFID; |
8110 | case PPCISD::FCFIDU: |
8111 | return PPCISD::STRICT_FCFIDU; |
8112 | case PPCISD::FCFIDS: |
8113 | return PPCISD::STRICT_FCFIDS; |
8114 | case PPCISD::FCFIDUS: |
8115 | return PPCISD::STRICT_FCFIDUS; |
8116 | } |
8117 | } |
8118 | |
8119 | static SDValue convertFPToInt(SDValue Op, SelectionDAG &DAG, |
8120 | const PPCSubtarget &Subtarget) { |
8121 | SDLoc dl(Op); |
8122 | bool IsStrict = Op->isStrictFPOpcode(); |
8123 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8124 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8125 | |
8126 | // TODO: Any other flags to propagate? |
8127 | SDNodeFlags Flags; |
8128 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8129 | |
8130 | // For strict nodes, source is the second operand. |
8131 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8132 | SDValue Chain = IsStrict ? Op.getOperand(0) : SDValue(); |
8133 | assert(Src.getValueType().isFloatingPoint())((Src.getValueType().isFloatingPoint()) ? static_cast<void > (0) : __assert_fail ("Src.getValueType().isFloatingPoint()" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8133, __PRETTY_FUNCTION__)); |
8134 | if (Src.getValueType() == MVT::f32) { |
8135 | if (IsStrict) { |
8136 | Src = |
8137 | DAG.getNode(ISD::STRICT_FP_EXTEND, dl, |
8138 | DAG.getVTList(MVT::f64, MVT::Other), {Chain, Src}, Flags); |
8139 | Chain = Src.getValue(1); |
8140 | } else |
8141 | Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src); |
8142 | } |
8143 | SDValue Conv; |
8144 | unsigned Opc = ISD::DELETED_NODE; |
8145 | switch (Op.getSimpleValueType().SimpleTy) { |
8146 | default: llvm_unreachable("Unhandled FP_TO_INT type in custom expander!")::llvm::llvm_unreachable_internal("Unhandled FP_TO_INT type in custom expander!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8146); |
8147 | case MVT::i32: |
8148 | Opc = IsSigned ? PPCISD::FCTIWZ |
8149 | : (Subtarget.hasFPCVT() ? PPCISD::FCTIWUZ : PPCISD::FCTIDZ); |
8150 | break; |
8151 | case MVT::i64: |
8152 | assert((IsSigned || Subtarget.hasFPCVT()) &&(((IsSigned || Subtarget.hasFPCVT()) && "i64 FP_TO_UINT is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"i64 FP_TO_UINT is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8153, __PRETTY_FUNCTION__)) |
8153 | "i64 FP_TO_UINT is supported only with FPCVT")(((IsSigned || Subtarget.hasFPCVT()) && "i64 FP_TO_UINT is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"i64 FP_TO_UINT is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8153, __PRETTY_FUNCTION__)); |
8154 | Opc = IsSigned ? PPCISD::FCTIDZ : PPCISD::FCTIDUZ; |
8155 | } |
8156 | if (IsStrict) { |
8157 | Opc = getPPCStrictOpcode(Opc); |
8158 | Conv = DAG.getNode(Opc, dl, DAG.getVTList(MVT::f64, MVT::Other), |
8159 | {Chain, Src}, Flags); |
8160 | } else { |
8161 | Conv = DAG.getNode(Opc, dl, MVT::f64, Src); |
8162 | } |
8163 | return Conv; |
8164 | } |
8165 | |
8166 | void PPCTargetLowering::LowerFP_TO_INTForReuse(SDValue Op, ReuseLoadInfo &RLI, |
8167 | SelectionDAG &DAG, |
8168 | const SDLoc &dl) const { |
8169 | SDValue Tmp = convertFPToInt(Op, DAG, Subtarget); |
8170 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8171 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8172 | bool IsStrict = Op->isStrictFPOpcode(); |
8173 | |
8174 | // Convert the FP value to an int value through memory. |
8175 | bool i32Stack = Op.getValueType() == MVT::i32 && Subtarget.hasSTFIWX() && |
8176 | (IsSigned || Subtarget.hasFPCVT()); |
8177 | SDValue FIPtr = DAG.CreateStackTemporary(i32Stack ? MVT::i32 : MVT::f64); |
8178 | int FI = cast<FrameIndexSDNode>(FIPtr)->getIndex(); |
8179 | MachinePointerInfo MPI = |
8180 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); |
8181 | |
8182 | // Emit a store to the stack slot. |
8183 | SDValue Chain = IsStrict ? Tmp.getValue(1) : DAG.getEntryNode(); |
8184 | Align Alignment(DAG.getEVTAlign(Tmp.getValueType())); |
8185 | if (i32Stack) { |
8186 | MachineFunction &MF = DAG.getMachineFunction(); |
8187 | Alignment = Align(4); |
8188 | MachineMemOperand *MMO = |
8189 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Alignment); |
8190 | SDValue Ops[] = { Chain, Tmp, FIPtr }; |
8191 | Chain = DAG.getMemIntrinsicNode(PPCISD::STFIWX, dl, |
8192 | DAG.getVTList(MVT::Other), Ops, MVT::i32, MMO); |
8193 | } else |
8194 | Chain = DAG.getStore(Chain, dl, Tmp, FIPtr, MPI, Alignment); |
8195 | |
8196 | // Result is a load from the stack slot. If loading 4 bytes, make sure to |
8197 | // add in a bias on big endian. |
8198 | if (Op.getValueType() == MVT::i32 && !i32Stack) { |
8199 | FIPtr = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, |
8200 | DAG.getConstant(4, dl, FIPtr.getValueType())); |
8201 | MPI = MPI.getWithOffset(Subtarget.isLittleEndian() ? 0 : 4); |
8202 | } |
8203 | |
8204 | RLI.Chain = Chain; |
8205 | RLI.Ptr = FIPtr; |
8206 | RLI.MPI = MPI; |
8207 | RLI.Alignment = Alignment; |
8208 | } |
8209 | |
8210 | /// Custom lowers floating point to integer conversions to use |
8211 | /// the direct move instructions available in ISA 2.07 to avoid the |
8212 | /// need for load/store combinations. |
8213 | SDValue PPCTargetLowering::LowerFP_TO_INTDirectMove(SDValue Op, |
8214 | SelectionDAG &DAG, |
8215 | const SDLoc &dl) const { |
8216 | SDValue Conv = convertFPToInt(Op, DAG, Subtarget); |
8217 | SDValue Mov = DAG.getNode(PPCISD::MFVSR, dl, Op.getValueType(), Conv); |
8218 | if (Op->isStrictFPOpcode()) |
8219 | return DAG.getMergeValues({Mov, Conv.getValue(1)}, dl); |
8220 | else |
8221 | return Mov; |
8222 | } |
8223 | |
8224 | SDValue PPCTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG, |
8225 | const SDLoc &dl) const { |
8226 | bool IsStrict = Op->isStrictFPOpcode(); |
8227 | bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT || |
8228 | Op.getOpcode() == ISD::STRICT_FP_TO_SINT; |
8229 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8230 | EVT SrcVT = Src.getValueType(); |
8231 | EVT DstVT = Op.getValueType(); |
8232 | |
8233 | // FP to INT conversions are legal for f128. |
8234 | if (SrcVT == MVT::f128) |
8235 | return Op; |
8236 | |
8237 | // Expand ppcf128 to i32 by hand for the benefit of llvm-gcc bootstrap on |
8238 | // PPC (the libcall is not available). |
8239 | if (SrcVT == MVT::ppcf128) { |
8240 | if (DstVT == MVT::i32) { |
8241 | // TODO: Conservatively pass only nofpexcept flag here. Need to check and |
8242 | // set other fast-math flags to FP operations in both strict and |
8243 | // non-strict cases. (FP_TO_SINT, FSUB) |
8244 | SDNodeFlags Flags; |
8245 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8246 | |
8247 | if (IsSigned) { |
8248 | SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src, |
8249 | DAG.getIntPtrConstant(0, dl)); |
8250 | SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::f64, Src, |
8251 | DAG.getIntPtrConstant(1, dl)); |
8252 | |
8253 | // Add the two halves of the long double in round-to-zero mode, and use |
8254 | // a smaller FP_TO_SINT. |
8255 | if (IsStrict) { |
8256 | SDValue Res = DAG.getNode(PPCISD::STRICT_FADDRTZ, dl, |
8257 | DAG.getVTList(MVT::f64, MVT::Other), |
8258 | {Op.getOperand(0), Lo, Hi}, Flags); |
8259 | return DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, |
8260 | DAG.getVTList(MVT::i32, MVT::Other), |
8261 | {Res.getValue(1), Res}, Flags); |
8262 | } else { |
8263 | SDValue Res = DAG.getNode(PPCISD::FADDRTZ, dl, MVT::f64, Lo, Hi); |
8264 | return DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Res); |
8265 | } |
8266 | } else { |
8267 | const uint64_t TwoE31[] = {0x41e0000000000000LL, 0}; |
8268 | APFloat APF = APFloat(APFloat::PPCDoubleDouble(), APInt(128, TwoE31)); |
8269 | SDValue Cst = DAG.getConstantFP(APF, dl, SrcVT); |
8270 | SDValue SignMask = DAG.getConstant(0x80000000, dl, DstVT); |
8271 | if (IsStrict) { |
8272 | // Sel = Src < 0x80000000 |
8273 | // FltOfs = select Sel, 0.0, 0x80000000 |
8274 | // IntOfs = select Sel, 0, 0x80000000 |
8275 | // Result = fp_to_sint(Src - FltOfs) ^ IntOfs |
8276 | SDValue Chain = Op.getOperand(0); |
8277 | EVT SetCCVT = |
8278 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT); |
8279 | EVT DstSetCCVT = |
8280 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), DstVT); |
8281 | SDValue Sel = DAG.getSetCC(dl, SetCCVT, Src, Cst, ISD::SETLT, |
8282 | Chain, true); |
8283 | Chain = Sel.getValue(1); |
8284 | |
8285 | SDValue FltOfs = DAG.getSelect( |
8286 | dl, SrcVT, Sel, DAG.getConstantFP(0.0, dl, SrcVT), Cst); |
8287 | Sel = DAG.getBoolExtOrTrunc(Sel, dl, DstSetCCVT, DstVT); |
8288 | |
8289 | SDValue Val = DAG.getNode(ISD::STRICT_FSUB, dl, |
8290 | DAG.getVTList(SrcVT, MVT::Other), |
8291 | {Chain, Src, FltOfs}, Flags); |
8292 | Chain = Val.getValue(1); |
8293 | SDValue SInt = DAG.getNode(ISD::STRICT_FP_TO_SINT, dl, |
8294 | DAG.getVTList(DstVT, MVT::Other), |
8295 | {Chain, Val}, Flags); |
8296 | Chain = SInt.getValue(1); |
8297 | SDValue IntOfs = DAG.getSelect( |
8298 | dl, DstVT, Sel, DAG.getConstant(0, dl, DstVT), SignMask); |
8299 | SDValue Result = DAG.getNode(ISD::XOR, dl, DstVT, SInt, IntOfs); |
8300 | return DAG.getMergeValues({Result, Chain}, dl); |
8301 | } else { |
8302 | // X>=2^31 ? (int)(X-2^31)+0x80000000 : (int)X |
8303 | // FIXME: generated code sucks. |
8304 | SDValue True = DAG.getNode(ISD::FSUB, dl, MVT::ppcf128, Src, Cst); |
8305 | True = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, True); |
8306 | True = DAG.getNode(ISD::ADD, dl, MVT::i32, True, SignMask); |
8307 | SDValue False = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Src); |
8308 | return DAG.getSelectCC(dl, Src, Cst, True, False, ISD::SETGE); |
8309 | } |
8310 | } |
8311 | } |
8312 | |
8313 | return SDValue(); |
8314 | } |
8315 | |
8316 | if (Subtarget.hasDirectMove() && Subtarget.isPPC64()) |
8317 | return LowerFP_TO_INTDirectMove(Op, DAG, dl); |
8318 | |
8319 | ReuseLoadInfo RLI; |
8320 | LowerFP_TO_INTForReuse(Op, RLI, DAG, dl); |
8321 | |
8322 | return DAG.getLoad(Op.getValueType(), dl, RLI.Chain, RLI.Ptr, RLI.MPI, |
8323 | RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges); |
8324 | } |
8325 | |
8326 | // We're trying to insert a regular store, S, and then a load, L. If the |
8327 | // incoming value, O, is a load, we might just be able to have our load use the |
8328 | // address used by O. However, we don't know if anything else will store to |
8329 | // that address before we can load from it. To prevent this situation, we need |
8330 | // to insert our load, L, into the chain as a peer of O. To do this, we give L |
8331 | // the same chain operand as O, we create a token factor from the chain results |
8332 | // of O and L, and we replace all uses of O's chain result with that token |
8333 | // factor (see spliceIntoChain below for this last part). |
8334 | bool PPCTargetLowering::canReuseLoadAddress(SDValue Op, EVT MemVT, |
8335 | ReuseLoadInfo &RLI, |
8336 | SelectionDAG &DAG, |
8337 | ISD::LoadExtType ET) const { |
8338 | // Conservatively skip reusing for constrained FP nodes. |
8339 | if (Op->isStrictFPOpcode()) |
8340 | return false; |
8341 | |
8342 | SDLoc dl(Op); |
8343 | bool ValidFPToUint = Op.getOpcode() == ISD::FP_TO_UINT && |
8344 | (Subtarget.hasFPCVT() || Op.getValueType() == MVT::i32); |
8345 | if (ET == ISD::NON_EXTLOAD && |
8346 | (ValidFPToUint || Op.getOpcode() == ISD::FP_TO_SINT) && |
8347 | isOperationLegalOrCustom(Op.getOpcode(), |
8348 | Op.getOperand(0).getValueType())) { |
8349 | |
8350 | LowerFP_TO_INTForReuse(Op, RLI, DAG, dl); |
8351 | return true; |
8352 | } |
8353 | |
8354 | LoadSDNode *LD = dyn_cast<LoadSDNode>(Op); |
8355 | if (!LD || LD->getExtensionType() != ET || LD->isVolatile() || |
8356 | LD->isNonTemporal()) |
8357 | return false; |
8358 | if (LD->getMemoryVT() != MemVT) |
8359 | return false; |
8360 | |
8361 | RLI.Ptr = LD->getBasePtr(); |
8362 | if (LD->isIndexed() && !LD->getOffset().isUndef()) { |
8363 | assert(LD->getAddressingMode() == ISD::PRE_INC &&((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8364, __PRETTY_FUNCTION__)) |
8364 | "Non-pre-inc AM on PPC?")((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8364, __PRETTY_FUNCTION__)); |
8365 | RLI.Ptr = DAG.getNode(ISD::ADD, dl, RLI.Ptr.getValueType(), RLI.Ptr, |
8366 | LD->getOffset()); |
8367 | } |
8368 | |
8369 | RLI.Chain = LD->getChain(); |
8370 | RLI.MPI = LD->getPointerInfo(); |
8371 | RLI.IsDereferenceable = LD->isDereferenceable(); |
8372 | RLI.IsInvariant = LD->isInvariant(); |
8373 | RLI.Alignment = LD->getAlign(); |
8374 | RLI.AAInfo = LD->getAAInfo(); |
8375 | RLI.Ranges = LD->getRanges(); |
8376 | |
8377 | RLI.ResChain = SDValue(LD, LD->isIndexed() ? 2 : 1); |
8378 | return true; |
8379 | } |
8380 | |
8381 | // Given the head of the old chain, ResChain, insert a token factor containing |
8382 | // it and NewResChain, and make users of ResChain now be users of that token |
8383 | // factor. |
8384 | // TODO: Remove and use DAG::makeEquivalentMemoryOrdering() instead. |
8385 | void PPCTargetLowering::spliceIntoChain(SDValue ResChain, |
8386 | SDValue NewResChain, |
8387 | SelectionDAG &DAG) const { |
8388 | if (!ResChain) |
8389 | return; |
8390 | |
8391 | SDLoc dl(NewResChain); |
8392 | |
8393 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
8394 | NewResChain, DAG.getUNDEF(MVT::Other)); |
8395 | assert(TF.getNode() != NewResChain.getNode() &&((TF.getNode() != NewResChain.getNode() && "A new TF really is required here" ) ? static_cast<void> (0) : __assert_fail ("TF.getNode() != NewResChain.getNode() && \"A new TF really is required here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8396, __PRETTY_FUNCTION__)) |
8396 | "A new TF really is required here")((TF.getNode() != NewResChain.getNode() && "A new TF really is required here" ) ? static_cast<void> (0) : __assert_fail ("TF.getNode() != NewResChain.getNode() && \"A new TF really is required here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8396, __PRETTY_FUNCTION__)); |
8397 | |
8398 | DAG.ReplaceAllUsesOfValueWith(ResChain, TF); |
8399 | DAG.UpdateNodeOperands(TF.getNode(), ResChain, NewResChain); |
8400 | } |
8401 | |
8402 | /// Analyze profitability of direct move |
8403 | /// prefer float load to int load plus direct move |
8404 | /// when there is no integer use of int load |
8405 | bool PPCTargetLowering::directMoveIsProfitable(const SDValue &Op) const { |
8406 | SDNode *Origin = Op.getOperand(0).getNode(); |
8407 | if (Origin->getOpcode() != ISD::LOAD) |
8408 | return true; |
8409 | |
8410 | // If there is no LXSIBZX/LXSIHZX, like Power8, |
8411 | // prefer direct move if the memory size is 1 or 2 bytes. |
8412 | MachineMemOperand *MMO = cast<LoadSDNode>(Origin)->getMemOperand(); |
8413 | if (!Subtarget.hasP9Vector() && MMO->getSize() <= 2) |
8414 | return true; |
8415 | |
8416 | for (SDNode::use_iterator UI = Origin->use_begin(), |
8417 | UE = Origin->use_end(); |
8418 | UI != UE; ++UI) { |
8419 | |
8420 | // Only look at the users of the loaded value. |
8421 | if (UI.getUse().get().getResNo() != 0) |
8422 | continue; |
8423 | |
8424 | if (UI->getOpcode() != ISD::SINT_TO_FP && |
8425 | UI->getOpcode() != ISD::UINT_TO_FP && |
8426 | UI->getOpcode() != ISD::STRICT_SINT_TO_FP && |
8427 | UI->getOpcode() != ISD::STRICT_UINT_TO_FP) |
8428 | return true; |
8429 | } |
8430 | |
8431 | return false; |
8432 | } |
8433 | |
8434 | static SDValue convertIntToFP(SDValue Op, SDValue Src, SelectionDAG &DAG, |
8435 | const PPCSubtarget &Subtarget, |
8436 | SDValue Chain = SDValue()) { |
8437 | bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP || |
8438 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8439 | SDLoc dl(Op); |
8440 | |
8441 | // TODO: Any other flags to propagate? |
8442 | SDNodeFlags Flags; |
8443 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8444 | |
8445 | // If we have FCFIDS, then use it when converting to single-precision. |
8446 | // Otherwise, convert to double-precision and then round. |
8447 | bool IsSingle = Op.getValueType() == MVT::f32 && Subtarget.hasFPCVT(); |
8448 | unsigned ConvOpc = IsSingle ? (IsSigned ? PPCISD::FCFIDS : PPCISD::FCFIDUS) |
8449 | : (IsSigned ? PPCISD::FCFID : PPCISD::FCFIDU); |
8450 | EVT ConvTy = IsSingle ? MVT::f32 : MVT::f64; |
8451 | if (Op->isStrictFPOpcode()) { |
8452 | if (!Chain) |
8453 | Chain = Op.getOperand(0); |
8454 | return DAG.getNode(getPPCStrictOpcode(ConvOpc), dl, |
8455 | DAG.getVTList(ConvTy, MVT::Other), {Chain, Src}, Flags); |
8456 | } else |
8457 | return DAG.getNode(ConvOpc, dl, ConvTy, Src); |
8458 | } |
8459 | |
8460 | /// Custom lowers integer to floating point conversions to use |
8461 | /// the direct move instructions available in ISA 2.07 to avoid the |
8462 | /// need for load/store combinations. |
8463 | SDValue PPCTargetLowering::LowerINT_TO_FPDirectMove(SDValue Op, |
8464 | SelectionDAG &DAG, |
8465 | const SDLoc &dl) const { |
8466 | assert((Op.getValueType() == MVT::f32 ||(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8468, __PRETTY_FUNCTION__)) |
8467 | Op.getValueType() == MVT::f64) &&(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8468, __PRETTY_FUNCTION__)) |
8468 | "Invalid floating point type as target of conversion")(((Op.getValueType() == MVT::f32 || Op.getValueType() == MVT:: f64) && "Invalid floating point type as target of conversion" ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::f32 || Op.getValueType() == MVT::f64) && \"Invalid floating point type as target of conversion\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8468, __PRETTY_FUNCTION__)); |
8469 | assert(Subtarget.hasFPCVT() &&((Subtarget.hasFPCVT() && "Int to FP conversions with direct moves require FPCVT" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasFPCVT() && \"Int to FP conversions with direct moves require FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8470, __PRETTY_FUNCTION__)) |
8470 | "Int to FP conversions with direct moves require FPCVT")((Subtarget.hasFPCVT() && "Int to FP conversions with direct moves require FPCVT" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasFPCVT() && \"Int to FP conversions with direct moves require FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8470, __PRETTY_FUNCTION__)); |
8471 | SDValue Src = Op.getOperand(Op->isStrictFPOpcode() ? 1 : 0); |
8472 | bool WordInt = Src.getSimpleValueType().SimpleTy == MVT::i32; |
8473 | bool Signed = Op.getOpcode() == ISD::SINT_TO_FP || |
8474 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8475 | unsigned MovOpc = (WordInt && !Signed) ? PPCISD::MTVSRZ : PPCISD::MTVSRA; |
8476 | SDValue Mov = DAG.getNode(MovOpc, dl, MVT::f64, Src); |
8477 | return convertIntToFP(Op, Mov, DAG, Subtarget); |
8478 | } |
8479 | |
8480 | static SDValue widenVec(SelectionDAG &DAG, SDValue Vec, const SDLoc &dl) { |
8481 | |
8482 | EVT VecVT = Vec.getValueType(); |
8483 | assert(VecVT.isVector() && "Expected a vector type.")((VecVT.isVector() && "Expected a vector type.") ? static_cast <void> (0) : __assert_fail ("VecVT.isVector() && \"Expected a vector type.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8483, __PRETTY_FUNCTION__)); |
8484 | assert(VecVT.getSizeInBits() < 128 && "Vector is already full width.")((VecVT.getSizeInBits() < 128 && "Vector is already full width." ) ? static_cast<void> (0) : __assert_fail ("VecVT.getSizeInBits() < 128 && \"Vector is already full width.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8484, __PRETTY_FUNCTION__)); |
8485 | |
8486 | EVT EltVT = VecVT.getVectorElementType(); |
8487 | unsigned WideNumElts = 128 / EltVT.getSizeInBits(); |
8488 | EVT WideVT = EVT::getVectorVT(*DAG.getContext(), EltVT, WideNumElts); |
8489 | |
8490 | unsigned NumConcat = WideNumElts / VecVT.getVectorNumElements(); |
8491 | SmallVector<SDValue, 16> Ops(NumConcat); |
8492 | Ops[0] = Vec; |
8493 | SDValue UndefVec = DAG.getUNDEF(VecVT); |
8494 | for (unsigned i = 1; i < NumConcat; ++i) |
8495 | Ops[i] = UndefVec; |
8496 | |
8497 | return DAG.getNode(ISD::CONCAT_VECTORS, dl, WideVT, Ops); |
8498 | } |
8499 | |
8500 | SDValue PPCTargetLowering::LowerINT_TO_FPVector(SDValue Op, SelectionDAG &DAG, |
8501 | const SDLoc &dl) const { |
8502 | bool IsStrict = Op->isStrictFPOpcode(); |
8503 | unsigned Opc = Op.getOpcode(); |
8504 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8505 | assert((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP ||(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8507, __PRETTY_FUNCTION__)) |
8506 | Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) &&(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8507, __PRETTY_FUNCTION__)) |
8507 | "Unexpected conversion type")(((Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && "Unexpected conversion type") ? static_cast<void> (0) : __assert_fail ("(Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_UINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP) && \"Unexpected conversion type\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8507, __PRETTY_FUNCTION__)); |
8508 | assert((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) &&(((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT ::v4f32) && "Supports conversions to v2f64/v4f32 only." ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) && \"Supports conversions to v2f64/v4f32 only.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8509, __PRETTY_FUNCTION__)) |
8509 | "Supports conversions to v2f64/v4f32 only.")(((Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT ::v4f32) && "Supports conversions to v2f64/v4f32 only." ) ? static_cast<void> (0) : __assert_fail ("(Op.getValueType() == MVT::v2f64 || Op.getValueType() == MVT::v4f32) && \"Supports conversions to v2f64/v4f32 only.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8509, __PRETTY_FUNCTION__)); |
8510 | |
8511 | // TODO: Any other flags to propagate? |
8512 | SDNodeFlags Flags; |
8513 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8514 | |
8515 | bool SignedConv = Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP; |
8516 | bool FourEltRes = Op.getValueType() == MVT::v4f32; |
8517 | |
8518 | SDValue Wide = widenVec(DAG, Src, dl); |
8519 | EVT WideVT = Wide.getValueType(); |
8520 | unsigned WideNumElts = WideVT.getVectorNumElements(); |
8521 | MVT IntermediateVT = FourEltRes ? MVT::v4i32 : MVT::v2i64; |
8522 | |
8523 | SmallVector<int, 16> ShuffV; |
8524 | for (unsigned i = 0; i < WideNumElts; ++i) |
8525 | ShuffV.push_back(i + WideNumElts); |
8526 | |
8527 | int Stride = FourEltRes ? WideNumElts / 4 : WideNumElts / 2; |
8528 | int SaveElts = FourEltRes ? 4 : 2; |
8529 | if (Subtarget.isLittleEndian()) |
8530 | for (int i = 0; i < SaveElts; i++) |
8531 | ShuffV[i * Stride] = i; |
8532 | else |
8533 | for (int i = 1; i <= SaveElts; i++) |
8534 | ShuffV[i * Stride - 1] = i - 1; |
8535 | |
8536 | SDValue ShuffleSrc2 = |
8537 | SignedConv ? DAG.getUNDEF(WideVT) : DAG.getConstant(0, dl, WideVT); |
8538 | SDValue Arrange = DAG.getVectorShuffle(WideVT, dl, Wide, ShuffleSrc2, ShuffV); |
8539 | |
8540 | SDValue Extend; |
8541 | if (SignedConv) { |
8542 | Arrange = DAG.getBitcast(IntermediateVT, Arrange); |
8543 | EVT ExtVT = Src.getValueType(); |
8544 | if (Subtarget.hasP9Altivec()) |
8545 | ExtVT = EVT::getVectorVT(*DAG.getContext(), WideVT.getVectorElementType(), |
8546 | IntermediateVT.getVectorNumElements()); |
8547 | |
8548 | Extend = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, IntermediateVT, Arrange, |
8549 | DAG.getValueType(ExtVT)); |
8550 | } else |
8551 | Extend = DAG.getNode(ISD::BITCAST, dl, IntermediateVT, Arrange); |
8552 | |
8553 | if (IsStrict) |
8554 | return DAG.getNode(Opc, dl, DAG.getVTList(Op.getValueType(), MVT::Other), |
8555 | {Op.getOperand(0), Extend}, Flags); |
8556 | |
8557 | return DAG.getNode(Opc, dl, Op.getValueType(), Extend); |
8558 | } |
8559 | |
8560 | SDValue PPCTargetLowering::LowerINT_TO_FP(SDValue Op, |
8561 | SelectionDAG &DAG) const { |
8562 | SDLoc dl(Op); |
8563 | bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP || |
8564 | Op.getOpcode() == ISD::STRICT_SINT_TO_FP; |
8565 | bool IsStrict = Op->isStrictFPOpcode(); |
8566 | SDValue Src = Op.getOperand(IsStrict ? 1 : 0); |
8567 | SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode(); |
8568 | |
8569 | // TODO: Any other flags to propagate? |
8570 | SDNodeFlags Flags; |
8571 | Flags.setNoFPExcept(Op->getFlags().hasNoFPExcept()); |
8572 | |
8573 | EVT InVT = Src.getValueType(); |
8574 | EVT OutVT = Op.getValueType(); |
8575 | if (OutVT.isVector() && OutVT.isFloatingPoint() && |
8576 | isOperationCustom(Op.getOpcode(), InVT)) |
8577 | return LowerINT_TO_FPVector(Op, DAG, dl); |
8578 | |
8579 | // Conversions to f128 are legal. |
8580 | if (Op.getValueType() == MVT::f128) |
8581 | return Op; |
8582 | |
8583 | // Don't handle ppc_fp128 here; let it be lowered to a libcall. |
8584 | if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64) |
8585 | return SDValue(); |
8586 | |
8587 | if (Src.getValueType() == MVT::i1) |
8588 | return DAG.getNode(ISD::SELECT, dl, Op.getValueType(), Src, |
8589 | DAG.getConstantFP(1.0, dl, Op.getValueType()), |
8590 | DAG.getConstantFP(0.0, dl, Op.getValueType())); |
8591 | |
8592 | // If we have direct moves, we can do all the conversion, skip the store/load |
8593 | // however, without FPCVT we can't do most conversions. |
8594 | if (Subtarget.hasDirectMove() && directMoveIsProfitable(Op) && |
8595 | Subtarget.isPPC64() && Subtarget.hasFPCVT()) |
8596 | return LowerINT_TO_FPDirectMove(Op, DAG, dl); |
8597 | |
8598 | assert((IsSigned || Subtarget.hasFPCVT()) &&(((IsSigned || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8599, __PRETTY_FUNCTION__)) |
8599 | "UINT_TO_FP is supported only with FPCVT")(((IsSigned || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT" ) ? static_cast<void> (0) : __assert_fail ("(IsSigned || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8599, __PRETTY_FUNCTION__)); |
8600 | |
8601 | if (Src.getValueType() == MVT::i64) { |
8602 | SDValue SINT = Src; |
8603 | // When converting to single-precision, we actually need to convert |
8604 | // to double-precision first and then round to single-precision. |
8605 | // To avoid double-rounding effects during that operation, we have |
8606 | // to prepare the input operand. Bits that might be truncated when |
8607 | // converting to double-precision are replaced by a bit that won't |
8608 | // be lost at this stage, but is below the single-precision rounding |
8609 | // position. |
8610 | // |
8611 | // However, if -enable-unsafe-fp-math is in effect, accept double |
8612 | // rounding to avoid the extra overhead. |
8613 | if (Op.getValueType() == MVT::f32 && |
8614 | !Subtarget.hasFPCVT() && |
8615 | !DAG.getTarget().Options.UnsafeFPMath) { |
8616 | |
8617 | // Twiddle input to make sure the low 11 bits are zero. (If this |
8618 | // is the case, we are guaranteed the value will fit into the 53 bit |
8619 | // mantissa of an IEEE double-precision value without rounding.) |
8620 | // If any of those low 11 bits were not zero originally, make sure |
8621 | // bit 12 (value 2048) is set instead, so that the final rounding |
8622 | // to single-precision gets the correct result. |
8623 | SDValue Round = DAG.getNode(ISD::AND, dl, MVT::i64, |
8624 | SINT, DAG.getConstant(2047, dl, MVT::i64)); |
8625 | Round = DAG.getNode(ISD::ADD, dl, MVT::i64, |
8626 | Round, DAG.getConstant(2047, dl, MVT::i64)); |
8627 | Round = DAG.getNode(ISD::OR, dl, MVT::i64, Round, SINT); |
8628 | Round = DAG.getNode(ISD::AND, dl, MVT::i64, |
8629 | Round, DAG.getConstant(-2048, dl, MVT::i64)); |
8630 | |
8631 | // However, we cannot use that value unconditionally: if the magnitude |
8632 | // of the input value is small, the bit-twiddling we did above might |
8633 | // end up visibly changing the output. Fortunately, in that case, we |
8634 | // don't need to twiddle bits since the original input will convert |
8635 | // exactly to double-precision floating-point already. Therefore, |
8636 | // construct a conditional to use the original value if the top 11 |
8637 | // bits are all sign-bit copies, and use the rounded value computed |
8638 | // above otherwise. |
8639 | SDValue Cond = DAG.getNode(ISD::SRA, dl, MVT::i64, |
8640 | SINT, DAG.getConstant(53, dl, MVT::i32)); |
8641 | Cond = DAG.getNode(ISD::ADD, dl, MVT::i64, |
8642 | Cond, DAG.getConstant(1, dl, MVT::i64)); |
8643 | Cond = DAG.getSetCC( |
8644 | dl, |
8645 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64), |
8646 | Cond, DAG.getConstant(1, dl, MVT::i64), ISD::SETUGT); |
8647 | |
8648 | SINT = DAG.getNode(ISD::SELECT, dl, MVT::i64, Cond, Round, SINT); |
8649 | } |
8650 | |
8651 | ReuseLoadInfo RLI; |
8652 | SDValue Bits; |
8653 | |
8654 | MachineFunction &MF = DAG.getMachineFunction(); |
8655 | if (canReuseLoadAddress(SINT, MVT::i64, RLI, DAG)) { |
8656 | Bits = DAG.getLoad(MVT::f64, dl, RLI.Chain, RLI.Ptr, RLI.MPI, |
8657 | RLI.Alignment, RLI.MMOFlags(), RLI.AAInfo, RLI.Ranges); |
8658 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8659 | } else if (Subtarget.hasLFIWAX() && |
8660 | canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::SEXTLOAD)) { |
8661 | MachineMemOperand *MMO = |
8662 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8663 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8664 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8665 | Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWAX, dl, |
8666 | DAG.getVTList(MVT::f64, MVT::Other), |
8667 | Ops, MVT::i32, MMO); |
8668 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8669 | } else if (Subtarget.hasFPCVT() && |
8670 | canReuseLoadAddress(SINT, MVT::i32, RLI, DAG, ISD::ZEXTLOAD)) { |
8671 | MachineMemOperand *MMO = |
8672 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8673 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8674 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8675 | Bits = DAG.getMemIntrinsicNode(PPCISD::LFIWZX, dl, |
8676 | DAG.getVTList(MVT::f64, MVT::Other), |
8677 | Ops, MVT::i32, MMO); |
8678 | spliceIntoChain(RLI.ResChain, Bits.getValue(1), DAG); |
8679 | } else if (((Subtarget.hasLFIWAX() && |
8680 | SINT.getOpcode() == ISD::SIGN_EXTEND) || |
8681 | (Subtarget.hasFPCVT() && |
8682 | SINT.getOpcode() == ISD::ZERO_EXTEND)) && |
8683 | SINT.getOperand(0).getValueType() == MVT::i32) { |
8684 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
8685 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
8686 | |
8687 | int FrameIdx = MFI.CreateStackObject(4, Align(4), false); |
8688 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8689 | |
8690 | SDValue Store = DAG.getStore(Chain, dl, SINT.getOperand(0), FIdx, |
8691 | MachinePointerInfo::getFixedStack( |
8692 | DAG.getMachineFunction(), FrameIdx)); |
8693 | Chain = Store; |
8694 | |
8695 | assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8696, __PRETTY_FUNCTION__)) |
8696 | "Expected an i32 store")((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8696, __PRETTY_FUNCTION__)); |
8697 | |
8698 | RLI.Ptr = FIdx; |
8699 | RLI.Chain = Chain; |
8700 | RLI.MPI = |
8701 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx); |
8702 | RLI.Alignment = Align(4); |
8703 | |
8704 | MachineMemOperand *MMO = |
8705 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8706 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8707 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8708 | Bits = DAG.getMemIntrinsicNode(SINT.getOpcode() == ISD::ZERO_EXTEND ? |
8709 | PPCISD::LFIWZX : PPCISD::LFIWAX, |
8710 | dl, DAG.getVTList(MVT::f64, MVT::Other), |
8711 | Ops, MVT::i32, MMO); |
8712 | Chain = Bits.getValue(1); |
8713 | } else |
8714 | Bits = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SINT); |
8715 | |
8716 | SDValue FP = convertIntToFP(Op, Bits, DAG, Subtarget, Chain); |
8717 | if (IsStrict) |
8718 | Chain = FP.getValue(1); |
8719 | |
8720 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
8721 | if (IsStrict) |
8722 | FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl, |
8723 | DAG.getVTList(MVT::f32, MVT::Other), |
8724 | {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags); |
8725 | else |
8726 | FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, |
8727 | DAG.getIntPtrConstant(0, dl)); |
8728 | } |
8729 | return FP; |
8730 | } |
8731 | |
8732 | assert(Src.getValueType() == MVT::i32 &&((Src.getValueType() == MVT::i32 && "Unhandled INT_TO_FP type in custom expander!" ) ? static_cast<void> (0) : __assert_fail ("Src.getValueType() == MVT::i32 && \"Unhandled INT_TO_FP type in custom expander!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8733, __PRETTY_FUNCTION__)) |
8733 | "Unhandled INT_TO_FP type in custom expander!")((Src.getValueType() == MVT::i32 && "Unhandled INT_TO_FP type in custom expander!" ) ? static_cast<void> (0) : __assert_fail ("Src.getValueType() == MVT::i32 && \"Unhandled INT_TO_FP type in custom expander!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8733, __PRETTY_FUNCTION__)); |
8734 | // Since we only generate this in 64-bit mode, we can take advantage of |
8735 | // 64-bit registers. In particular, sign extend the input value into the |
8736 | // 64-bit register with extsw, store the WHOLE 64-bit value into the stack |
8737 | // then lfd it and fcfid it. |
8738 | MachineFunction &MF = DAG.getMachineFunction(); |
8739 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
8740 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
8741 | |
8742 | SDValue Ld; |
8743 | if (Subtarget.hasLFIWAX() || Subtarget.hasFPCVT()) { |
8744 | ReuseLoadInfo RLI; |
8745 | bool ReusingLoad; |
8746 | if (!(ReusingLoad = canReuseLoadAddress(Src, MVT::i32, RLI, DAG))) { |
8747 | int FrameIdx = MFI.CreateStackObject(4, Align(4), false); |
8748 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8749 | |
8750 | SDValue Store = DAG.getStore(Chain, dl, Src, FIdx, |
8751 | MachinePointerInfo::getFixedStack( |
8752 | DAG.getMachineFunction(), FrameIdx)); |
8753 | Chain = Store; |
8754 | |
8755 | assert(cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 &&((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8756, __PRETTY_FUNCTION__)) |
8756 | "Expected an i32 store")((cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && "Expected an i32 store") ? static_cast<void> (0) : __assert_fail ("cast<StoreSDNode>(Store)->getMemoryVT() == MVT::i32 && \"Expected an i32 store\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8756, __PRETTY_FUNCTION__)); |
8757 | |
8758 | RLI.Ptr = FIdx; |
8759 | RLI.Chain = Chain; |
8760 | RLI.MPI = |
8761 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx); |
8762 | RLI.Alignment = Align(4); |
8763 | } |
8764 | |
8765 | MachineMemOperand *MMO = |
8766 | MF.getMachineMemOperand(RLI.MPI, MachineMemOperand::MOLoad, 4, |
8767 | RLI.Alignment, RLI.AAInfo, RLI.Ranges); |
8768 | SDValue Ops[] = { RLI.Chain, RLI.Ptr }; |
8769 | Ld = DAG.getMemIntrinsicNode(IsSigned ? PPCISD::LFIWAX : PPCISD::LFIWZX, dl, |
8770 | DAG.getVTList(MVT::f64, MVT::Other), Ops, |
8771 | MVT::i32, MMO); |
8772 | Chain = Ld.getValue(1); |
8773 | if (ReusingLoad) |
8774 | spliceIntoChain(RLI.ResChain, Ld.getValue(1), DAG); |
8775 | } else { |
8776 | assert(Subtarget.isPPC64() &&((Subtarget.isPPC64() && "i32->FP without LFIWAX supported only on PPC64" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"i32->FP without LFIWAX supported only on PPC64\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8777, __PRETTY_FUNCTION__)) |
8777 | "i32->FP without LFIWAX supported only on PPC64")((Subtarget.isPPC64() && "i32->FP without LFIWAX supported only on PPC64" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"i32->FP without LFIWAX supported only on PPC64\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8777, __PRETTY_FUNCTION__)); |
8778 | |
8779 | int FrameIdx = MFI.CreateStackObject(8, Align(8), false); |
8780 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
8781 | |
8782 | SDValue Ext64 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i64, Src); |
8783 | |
8784 | // STD the extended value into the stack slot. |
8785 | SDValue Store = DAG.getStore( |
8786 | Chain, dl, Ext64, FIdx, |
8787 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx)); |
8788 | Chain = Store; |
8789 | |
8790 | // Load the value as a double. |
8791 | Ld = DAG.getLoad( |
8792 | MVT::f64, dl, Chain, FIdx, |
8793 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FrameIdx)); |
8794 | Chain = Ld.getValue(1); |
8795 | } |
8796 | |
8797 | // FCFID it and return it. |
8798 | SDValue FP = convertIntToFP(Op, Ld, DAG, Subtarget, Chain); |
8799 | if (IsStrict) |
8800 | Chain = FP.getValue(1); |
8801 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
8802 | if (IsStrict) |
8803 | FP = DAG.getNode(ISD::STRICT_FP_ROUND, dl, |
8804 | DAG.getVTList(MVT::f32, MVT::Other), |
8805 | {Chain, FP, DAG.getIntPtrConstant(0, dl)}, Flags); |
8806 | else |
8807 | FP = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, FP, |
8808 | DAG.getIntPtrConstant(0, dl)); |
8809 | } |
8810 | return FP; |
8811 | } |
8812 | |
8813 | SDValue PPCTargetLowering::LowerFLT_ROUNDS_(SDValue Op, |
8814 | SelectionDAG &DAG) const { |
8815 | SDLoc dl(Op); |
8816 | /* |
8817 | The rounding mode is in bits 30:31 of FPSR, and has the following |
8818 | settings: |
8819 | 00 Round to nearest |
8820 | 01 Round to 0 |
8821 | 10 Round to +inf |
8822 | 11 Round to -inf |
8823 | |
8824 | FLT_ROUNDS, on the other hand, expects the following: |
8825 | -1 Undefined |
8826 | 0 Round to 0 |
8827 | 1 Round to nearest |
8828 | 2 Round to +inf |
8829 | 3 Round to -inf |
8830 | |
8831 | To perform the conversion, we do: |
8832 | ((FPSCR & 0x3) ^ ((~FPSCR & 0x3) >> 1)) |
8833 | */ |
8834 | |
8835 | MachineFunction &MF = DAG.getMachineFunction(); |
8836 | EVT VT = Op.getValueType(); |
8837 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
8838 | |
8839 | // Save FP Control Word to register |
8840 | SDValue Chain = Op.getOperand(0); |
8841 | SDValue MFFS = DAG.getNode(PPCISD::MFFS, dl, {MVT::f64, MVT::Other}, Chain); |
8842 | Chain = MFFS.getValue(1); |
8843 | |
8844 | // Save FP register to stack slot |
8845 | int SSFI = MF.getFrameInfo().CreateStackObject(8, Align(8), false); |
8846 | SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT); |
8847 | Chain = DAG.getStore(Chain, dl, MFFS, StackSlot, MachinePointerInfo()); |
8848 | |
8849 | // Load FP Control Word from low 32 bits of stack slot. |
8850 | SDValue Four = DAG.getConstant(4, dl, PtrVT); |
8851 | SDValue Addr = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, Four); |
8852 | SDValue CWD = DAG.getLoad(MVT::i32, dl, Chain, Addr, MachinePointerInfo()); |
8853 | Chain = CWD.getValue(1); |
8854 | |
8855 | // Transform as necessary |
8856 | SDValue CWD1 = |
8857 | DAG.getNode(ISD::AND, dl, MVT::i32, |
8858 | CWD, DAG.getConstant(3, dl, MVT::i32)); |
8859 | SDValue CWD2 = |
8860 | DAG.getNode(ISD::SRL, dl, MVT::i32, |
8861 | DAG.getNode(ISD::AND, dl, MVT::i32, |
8862 | DAG.getNode(ISD::XOR, dl, MVT::i32, |
8863 | CWD, DAG.getConstant(3, dl, MVT::i32)), |
8864 | DAG.getConstant(3, dl, MVT::i32)), |
8865 | DAG.getConstant(1, dl, MVT::i32)); |
8866 | |
8867 | SDValue RetVal = |
8868 | DAG.getNode(ISD::XOR, dl, MVT::i32, CWD1, CWD2); |
8869 | |
8870 | RetVal = |
8871 | DAG.getNode((VT.getSizeInBits() < 16 ? ISD::TRUNCATE : ISD::ZERO_EXTEND), |
8872 | dl, VT, RetVal); |
8873 | |
8874 | return DAG.getMergeValues({RetVal, Chain}, dl); |
8875 | } |
8876 | |
8877 | SDValue PPCTargetLowering::LowerSHL_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8878 | EVT VT = Op.getValueType(); |
8879 | unsigned BitWidth = VT.getSizeInBits(); |
8880 | SDLoc dl(Op); |
8881 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8883, __PRETTY_FUNCTION__)) |
8882 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8883, __PRETTY_FUNCTION__)) |
8883 | "Unexpected SHL!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SHL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SHL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8883, __PRETTY_FUNCTION__)); |
8884 | |
8885 | // Expand into a bunch of logical ops. Note that these ops |
8886 | // depend on the PPC behavior for oversized shift amounts. |
8887 | SDValue Lo = Op.getOperand(0); |
8888 | SDValue Hi = Op.getOperand(1); |
8889 | SDValue Amt = Op.getOperand(2); |
8890 | EVT AmtVT = Amt.getValueType(); |
8891 | |
8892 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8893 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8894 | SDValue Tmp2 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Amt); |
8895 | SDValue Tmp3 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Tmp1); |
8896 | SDValue Tmp4 = DAG.getNode(ISD::OR , dl, VT, Tmp2, Tmp3); |
8897 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8898 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8899 | SDValue Tmp6 = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Tmp5); |
8900 | SDValue OutHi = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6); |
8901 | SDValue OutLo = DAG.getNode(PPCISD::SHL, dl, VT, Lo, Amt); |
8902 | SDValue OutOps[] = { OutLo, OutHi }; |
8903 | return DAG.getMergeValues(OutOps, dl); |
8904 | } |
8905 | |
8906 | SDValue PPCTargetLowering::LowerSRL_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8907 | EVT VT = Op.getValueType(); |
8908 | SDLoc dl(Op); |
8909 | unsigned BitWidth = VT.getSizeInBits(); |
8910 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8912, __PRETTY_FUNCTION__)) |
8911 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8912, __PRETTY_FUNCTION__)) |
8912 | "Unexpected SRL!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRL!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRL!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8912, __PRETTY_FUNCTION__)); |
8913 | |
8914 | // Expand into a bunch of logical ops. Note that these ops |
8915 | // depend on the PPC behavior for oversized shift amounts. |
8916 | SDValue Lo = Op.getOperand(0); |
8917 | SDValue Hi = Op.getOperand(1); |
8918 | SDValue Amt = Op.getOperand(2); |
8919 | EVT AmtVT = Amt.getValueType(); |
8920 | |
8921 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8922 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8923 | SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt); |
8924 | SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1); |
8925 | SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); |
8926 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8927 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8928 | SDValue Tmp6 = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Tmp5); |
8929 | SDValue OutLo = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp6); |
8930 | SDValue OutHi = DAG.getNode(PPCISD::SRL, dl, VT, Hi, Amt); |
8931 | SDValue OutOps[] = { OutLo, OutHi }; |
8932 | return DAG.getMergeValues(OutOps, dl); |
8933 | } |
8934 | |
8935 | SDValue PPCTargetLowering::LowerSRA_PARTS(SDValue Op, SelectionDAG &DAG) const { |
8936 | SDLoc dl(Op); |
8937 | EVT VT = Op.getValueType(); |
8938 | unsigned BitWidth = VT.getSizeInBits(); |
8939 | assert(Op.getNumOperands() == 3 &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8941, __PRETTY_FUNCTION__)) |
8940 | VT == Op.getOperand(1).getValueType() &&((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8941, __PRETTY_FUNCTION__)) |
8941 | "Unexpected SRA!")((Op.getNumOperands() == 3 && VT == Op.getOperand(1). getValueType() && "Unexpected SRA!") ? static_cast< void> (0) : __assert_fail ("Op.getNumOperands() == 3 && VT == Op.getOperand(1).getValueType() && \"Unexpected SRA!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 8941, __PRETTY_FUNCTION__)); |
8942 | |
8943 | // Expand into a bunch of logical ops, followed by a select_cc. |
8944 | SDValue Lo = Op.getOperand(0); |
8945 | SDValue Hi = Op.getOperand(1); |
8946 | SDValue Amt = Op.getOperand(2); |
8947 | EVT AmtVT = Amt.getValueType(); |
8948 | |
8949 | SDValue Tmp1 = DAG.getNode(ISD::SUB, dl, AmtVT, |
8950 | DAG.getConstant(BitWidth, dl, AmtVT), Amt); |
8951 | SDValue Tmp2 = DAG.getNode(PPCISD::SRL, dl, VT, Lo, Amt); |
8952 | SDValue Tmp3 = DAG.getNode(PPCISD::SHL, dl, VT, Hi, Tmp1); |
8953 | SDValue Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp3); |
8954 | SDValue Tmp5 = DAG.getNode(ISD::ADD, dl, AmtVT, Amt, |
8955 | DAG.getConstant(-BitWidth, dl, AmtVT)); |
8956 | SDValue Tmp6 = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Tmp5); |
8957 | SDValue OutHi = DAG.getNode(PPCISD::SRA, dl, VT, Hi, Amt); |
8958 | SDValue OutLo = DAG.getSelectCC(dl, Tmp5, DAG.getConstant(0, dl, AmtVT), |
8959 | Tmp4, Tmp6, ISD::SETLE); |
8960 | SDValue OutOps[] = { OutLo, OutHi }; |
8961 | return DAG.getMergeValues(OutOps, dl); |
8962 | } |
8963 | |
8964 | SDValue PPCTargetLowering::LowerFunnelShift(SDValue Op, |
8965 | SelectionDAG &DAG) const { |
8966 | SDLoc dl(Op); |
8967 | EVT VT = Op.getValueType(); |
8968 | unsigned BitWidth = VT.getSizeInBits(); |
8969 | |
8970 | bool IsFSHL = Op.getOpcode() == ISD::FSHL; |
8971 | SDValue X = Op.getOperand(0); |
8972 | SDValue Y = Op.getOperand(1); |
8973 | SDValue Z = Op.getOperand(2); |
8974 | EVT AmtVT = Z.getValueType(); |
8975 | |
8976 | // fshl: (X << (Z % BW)) | (Y >> (BW - (Z % BW))) |
8977 | // fshr: (X << (BW - (Z % BW))) | (Y >> (Z % BW)) |
8978 | // This is simpler than TargetLowering::expandFunnelShift because we can rely |
8979 | // on PowerPC shift by BW being well defined. |
8980 | Z = DAG.getNode(ISD::AND, dl, AmtVT, Z, |
8981 | DAG.getConstant(BitWidth - 1, dl, AmtVT)); |
8982 | SDValue SubZ = |
8983 | DAG.getNode(ISD::SUB, dl, AmtVT, DAG.getConstant(BitWidth, dl, AmtVT), Z); |
8984 | X = DAG.getNode(PPCISD::SHL, dl, VT, X, IsFSHL ? Z : SubZ); |
8985 | Y = DAG.getNode(PPCISD::SRL, dl, VT, Y, IsFSHL ? SubZ : Z); |
8986 | return DAG.getNode(ISD::OR, dl, VT, X, Y); |
8987 | } |
8988 | |
8989 | //===----------------------------------------------------------------------===// |
8990 | // Vector related lowering. |
8991 | // |
8992 | |
8993 | /// getCanonicalConstSplat - Build a canonical splat immediate of Val with an |
8994 | /// element size of SplatSize. Cast the result to VT. |
8995 | static SDValue getCanonicalConstSplat(uint64_t Val, unsigned SplatSize, EVT VT, |
8996 | SelectionDAG &DAG, const SDLoc &dl) { |
8997 | static const MVT VTys[] = { // canonical VT to use for each size. |
8998 | MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32 |
8999 | }; |
9000 | |
9001 | EVT ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1]; |
9002 | |
9003 | // For a splat with all ones, turn it to vspltisb 0xFF to canonicalize. |
9004 | if (Val == ((1LU << (SplatSize * 8)) - 1)) { |
9005 | SplatSize = 1; |
9006 | Val = 0xFF; |
9007 | } |
9008 | |
9009 | EVT CanonicalVT = VTys[SplatSize-1]; |
9010 | |
9011 | // Build a canonical splat for this value. |
9012 | return DAG.getBitcast(ReqVT, DAG.getConstant(Val, dl, CanonicalVT)); |
9013 | } |
9014 | |
9015 | /// BuildIntrinsicOp - Return a unary operator intrinsic node with the |
9016 | /// specified intrinsic ID. |
9017 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op, SelectionDAG &DAG, |
9018 | const SDLoc &dl, EVT DestVT = MVT::Other) { |
9019 | if (DestVT == MVT::Other) DestVT = Op.getValueType(); |
9020 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9021 | DAG.getConstant(IID, dl, MVT::i32), Op); |
9022 | } |
9023 | |
9024 | /// BuildIntrinsicOp - Return a binary operator intrinsic node with the |
9025 | /// specified intrinsic ID. |
9026 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue LHS, SDValue RHS, |
9027 | SelectionDAG &DAG, const SDLoc &dl, |
9028 | EVT DestVT = MVT::Other) { |
9029 | if (DestVT == MVT::Other) DestVT = LHS.getValueType(); |
9030 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9031 | DAG.getConstant(IID, dl, MVT::i32), LHS, RHS); |
9032 | } |
9033 | |
9034 | /// BuildIntrinsicOp - Return a ternary operator intrinsic node with the |
9035 | /// specified intrinsic ID. |
9036 | static SDValue BuildIntrinsicOp(unsigned IID, SDValue Op0, SDValue Op1, |
9037 | SDValue Op2, SelectionDAG &DAG, const SDLoc &dl, |
9038 | EVT DestVT = MVT::Other) { |
9039 | if (DestVT == MVT::Other) DestVT = Op0.getValueType(); |
9040 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT, |
9041 | DAG.getConstant(IID, dl, MVT::i32), Op0, Op1, Op2); |
9042 | } |
9043 | |
9044 | /// BuildVSLDOI - Return a VECTOR_SHUFFLE that is a vsldoi of the specified |
9045 | /// amount. The result has the specified value type. |
9046 | static SDValue BuildVSLDOI(SDValue LHS, SDValue RHS, unsigned Amt, EVT VT, |
9047 | SelectionDAG &DAG, const SDLoc &dl) { |
9048 | // Force LHS/RHS to be the right type. |
9049 | LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, LHS); |
9050 | RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, RHS); |
9051 | |
9052 | int Ops[16]; |
9053 | for (unsigned i = 0; i != 16; ++i) |
9054 | Ops[i] = i + Amt; |
9055 | SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, LHS, RHS, Ops); |
9056 | return DAG.getNode(ISD::BITCAST, dl, VT, T); |
9057 | } |
9058 | |
9059 | /// Do we have an efficient pattern in a .td file for this node? |
9060 | /// |
9061 | /// \param V - pointer to the BuildVectorSDNode being matched |
9062 | /// \param HasDirectMove - does this subtarget have VSR <-> GPR direct moves? |
9063 | /// |
9064 | /// There are some patterns where it is beneficial to keep a BUILD_VECTOR |
9065 | /// node as a BUILD_VECTOR node rather than expanding it. The patterns where |
9066 | /// the opposite is true (expansion is beneficial) are: |
9067 | /// - The node builds a vector out of integers that are not 32 or 64-bits |
9068 | /// - The node builds a vector out of constants |
9069 | /// - The node is a "load-and-splat" |
9070 | /// In all other cases, we will choose to keep the BUILD_VECTOR. |
9071 | static bool haveEfficientBuildVectorPattern(BuildVectorSDNode *V, |
9072 | bool HasDirectMove, |
9073 | bool HasP8Vector) { |
9074 | EVT VecVT = V->getValueType(0); |
9075 | bool RightType = VecVT == MVT::v2f64 || |
9076 | (HasP8Vector && VecVT == MVT::v4f32) || |
9077 | (HasDirectMove && (VecVT == MVT::v2i64 || VecVT == MVT::v4i32)); |
9078 | if (!RightType) |
9079 | return false; |
9080 | |
9081 | bool IsSplat = true; |
9082 | bool IsLoad = false; |
9083 | SDValue Op0 = V->getOperand(0); |
9084 | |
9085 | // This function is called in a block that confirms the node is not a constant |
9086 | // splat. So a constant BUILD_VECTOR here means the vector is built out of |
9087 | // different constants. |
9088 | if (V->isConstant()) |
9089 | return false; |
9090 | for (int i = 0, e = V->getNumOperands(); i < e; ++i) { |
9091 | if (V->getOperand(i).isUndef()) |
9092 | return false; |
9093 | // We want to expand nodes that represent load-and-splat even if the |
9094 | // loaded value is a floating point truncation or conversion to int. |
9095 | if (V->getOperand(i).getOpcode() == ISD::LOAD || |
9096 | (V->getOperand(i).getOpcode() == ISD::FP_ROUND && |
9097 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) || |
9098 | (V->getOperand(i).getOpcode() == ISD::FP_TO_SINT && |
9099 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD) || |
9100 | (V->getOperand(i).getOpcode() == ISD::FP_TO_UINT && |
9101 | V->getOperand(i).getOperand(0).getOpcode() == ISD::LOAD)) |
9102 | IsLoad = true; |
9103 | // If the operands are different or the input is not a load and has more |
9104 | // uses than just this BV node, then it isn't a splat. |
9105 | if (V->getOperand(i) != Op0 || |
9106 | (!IsLoad && !V->isOnlyUserOf(V->getOperand(i).getNode()))) |
9107 | IsSplat = false; |
9108 | } |
9109 | return !(IsSplat && IsLoad); |
9110 | } |
9111 | |
9112 | // Lower BITCAST(f128, (build_pair i64, i64)) to BUILD_FP128. |
9113 | SDValue PPCTargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const { |
9114 | |
9115 | SDLoc dl(Op); |
9116 | SDValue Op0 = Op->getOperand(0); |
9117 | |
9118 | if ((Op.getValueType() != MVT::f128) || |
9119 | (Op0.getOpcode() != ISD::BUILD_PAIR) || |
9120 | (Op0.getOperand(0).getValueType() != MVT::i64) || |
9121 | (Op0.getOperand(1).getValueType() != MVT::i64)) |
9122 | return SDValue(); |
9123 | |
9124 | return DAG.getNode(PPCISD::BUILD_FP128, dl, MVT::f128, Op0.getOperand(0), |
9125 | Op0.getOperand(1)); |
9126 | } |
9127 | |
9128 | static const SDValue *getNormalLoadInput(const SDValue &Op, bool &IsPermuted) { |
9129 | const SDValue *InputLoad = &Op; |
9130 | if (InputLoad->getOpcode() == ISD::BITCAST) |
9131 | InputLoad = &InputLoad->getOperand(0); |
9132 | if (InputLoad->getOpcode() == ISD::SCALAR_TO_VECTOR || |
9133 | InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED) { |
9134 | IsPermuted = InputLoad->getOpcode() == PPCISD::SCALAR_TO_VECTOR_PERMUTED; |
9135 | InputLoad = &InputLoad->getOperand(0); |
9136 | } |
9137 | if (InputLoad->getOpcode() != ISD::LOAD) |
9138 | return nullptr; |
9139 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9140 | return ISD::isNormalLoad(LD) ? InputLoad : nullptr; |
9141 | } |
9142 | |
9143 | // Convert the argument APFloat to a single precision APFloat if there is no |
9144 | // loss in information during the conversion to single precision APFloat and the |
9145 | // resulting number is not a denormal number. Return true if successful. |
9146 | bool llvm::convertToNonDenormSingle(APFloat &ArgAPFloat) { |
9147 | APFloat APFloatToConvert = ArgAPFloat; |
9148 | bool LosesInfo = true; |
9149 | APFloatToConvert.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, |
9150 | &LosesInfo); |
9151 | bool Success = (!LosesInfo && !APFloatToConvert.isDenormal()); |
9152 | if (Success) |
9153 | ArgAPFloat = APFloatToConvert; |
9154 | return Success; |
9155 | } |
9156 | |
9157 | // Bitcast the argument APInt to a double and convert it to a single precision |
9158 | // APFloat, bitcast the APFloat to an APInt and assign it to the original |
9159 | // argument if there is no loss in information during the conversion from |
9160 | // double to single precision APFloat and the resulting number is not a denormal |
9161 | // number. Return true if successful. |
9162 | bool llvm::convertToNonDenormSingle(APInt &ArgAPInt) { |
9163 | double DpValue = ArgAPInt.bitsToDouble(); |
9164 | APFloat APFloatDp(DpValue); |
9165 | bool Success = convertToNonDenormSingle(APFloatDp); |
9166 | if (Success) |
9167 | ArgAPInt = APFloatDp.bitcastToAPInt(); |
9168 | return Success; |
9169 | } |
9170 | |
9171 | // If this is a case we can't handle, return null and let the default |
9172 | // expansion code take care of it. If we CAN select this case, and if it |
9173 | // selects to a single instruction, return Op. Otherwise, if we can codegen |
9174 | // this case more efficiently than a constant pool load, lower it to the |
9175 | // sequence of ops that should be used. |
9176 | SDValue PPCTargetLowering::LowerBUILD_VECTOR(SDValue Op, |
9177 | SelectionDAG &DAG) const { |
9178 | SDLoc dl(Op); |
9179 | BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); |
9180 | assert(BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR")((BVN && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR" ) ? static_cast<void> (0) : __assert_fail ("BVN && \"Expected a BuildVectorSDNode in LowerBUILD_VECTOR\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9180, __PRETTY_FUNCTION__)); |
9181 | |
9182 | // Check if this is a splat of a constant value. |
9183 | APInt APSplatBits, APSplatUndef; |
9184 | unsigned SplatBitSize; |
9185 | bool HasAnyUndefs; |
9186 | bool BVNIsConstantSplat = |
9187 | BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize, |
9188 | HasAnyUndefs, 0, !Subtarget.isLittleEndian()); |
9189 | |
9190 | // If it is a splat of a double, check if we can shrink it to a 32 bit |
9191 | // non-denormal float which when converted back to double gives us the same |
9192 | // double. This is to exploit the XXSPLTIDP instruction. |
9193 | if (BVNIsConstantSplat && Subtarget.hasPrefixInstrs() && |
9194 | (SplatBitSize == 64) && (Op->getValueType(0) == MVT::v2f64) && |
9195 | convertToNonDenormSingle(APSplatBits)) { |
9196 | SDValue SplatNode = DAG.getNode( |
9197 | PPCISD::XXSPLTI_SP_TO_DP, dl, MVT::v2f64, |
9198 | DAG.getTargetConstant(APSplatBits.getZExtValue(), dl, MVT::i32)); |
9199 | return DAG.getBitcast(Op.getValueType(), SplatNode); |
9200 | } |
9201 | |
9202 | if (!BVNIsConstantSplat || SplatBitSize > 32) { |
9203 | |
9204 | bool IsPermutedLoad = false; |
9205 | const SDValue *InputLoad = |
9206 | getNormalLoadInput(Op.getOperand(0), IsPermutedLoad); |
9207 | // Handle load-and-splat patterns as we have instructions that will do this |
9208 | // in one go. |
9209 | if (InputLoad && DAG.isSplatValue(Op, true)) { |
9210 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9211 | |
9212 | // We have handling for 4 and 8 byte elements. |
9213 | unsigned ElementSize = LD->getMemoryVT().getScalarSizeInBits(); |
9214 | |
9215 | // Checking for a single use of this load, we have to check for vector |
9216 | // width (128 bits) / ElementSize uses (since each operand of the |
9217 | // BUILD_VECTOR is a separate use of the value. |
9218 | if (InputLoad->getNode()->hasNUsesOfValue(128 / ElementSize, 0) && |
9219 | ((Subtarget.hasVSX() && ElementSize == 64) || |
9220 | (Subtarget.hasP9Vector() && ElementSize == 32))) { |
9221 | SDValue Ops[] = { |
9222 | LD->getChain(), // Chain |
9223 | LD->getBasePtr(), // Ptr |
9224 | DAG.getValueType(Op.getValueType()) // VT |
9225 | }; |
9226 | return |
9227 | DAG.getMemIntrinsicNode(PPCISD::LD_SPLAT, dl, |
9228 | DAG.getVTList(Op.getValueType(), MVT::Other), |
9229 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
9230 | } |
9231 | } |
9232 | |
9233 | // BUILD_VECTOR nodes that are not constant splats of up to 32-bits can be |
9234 | // lowered to VSX instructions under certain conditions. |
9235 | // Without VSX, there is no pattern more efficient than expanding the node. |
9236 | if (Subtarget.hasVSX() && |
9237 | haveEfficientBuildVectorPattern(BVN, Subtarget.hasDirectMove(), |
9238 | Subtarget.hasP8Vector())) |
9239 | return Op; |
9240 | return SDValue(); |
9241 | } |
9242 | |
9243 | uint64_t SplatBits = APSplatBits.getZExtValue(); |
9244 | uint64_t SplatUndef = APSplatUndef.getZExtValue(); |
9245 | unsigned SplatSize = SplatBitSize / 8; |
9246 | |
9247 | // First, handle single instruction cases. |
9248 | |
9249 | // All zeros? |
9250 | if (SplatBits == 0) { |
9251 | // Canonicalize all zero vectors to be v4i32. |
9252 | if (Op.getValueType() != MVT::v4i32 || HasAnyUndefs) { |
9253 | SDValue Z = DAG.getConstant(0, dl, MVT::v4i32); |
9254 | Op = DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Z); |
9255 | } |
9256 | return Op; |
9257 | } |
9258 | |
9259 | // We have XXSPLTIW for constant splats four bytes wide. |
9260 | // Given vector length is a multiple of 4, 2-byte splats can be replaced |
9261 | // with 4-byte splats. We replicate the SplatBits in case of 2-byte splat to |
9262 | // make a 4-byte splat element. For example: 2-byte splat of 0xABAB can be |
9263 | // turned into a 4-byte splat of 0xABABABAB. |
9264 | if (Subtarget.hasPrefixInstrs() && SplatSize == 2) |
9265 | return getCanonicalConstSplat((SplatBits |= SplatBits << 16), SplatSize * 2, |
Although the value stored to 'SplatBits' is used in the enclosing expression, the value is never actually read from 'SplatBits' | |
9266 | Op.getValueType(), DAG, dl); |
9267 | |
9268 | if (Subtarget.hasPrefixInstrs() && SplatSize == 4) |
9269 | return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG, |
9270 | dl); |
9271 | |
9272 | // We have XXSPLTIB for constant splats one byte wide. |
9273 | if (Subtarget.hasP9Vector() && SplatSize == 1) |
9274 | return getCanonicalConstSplat(SplatBits, SplatSize, Op.getValueType(), DAG, |
9275 | dl); |
9276 | |
9277 | // If the sign extended value is in the range [-16,15], use VSPLTI[bhw]. |
9278 | int32_t SextVal= (int32_t(SplatBits << (32-SplatBitSize)) >> |
9279 | (32-SplatBitSize)); |
9280 | if (SextVal >= -16 && SextVal <= 15) |
9281 | return getCanonicalConstSplat(SextVal, SplatSize, Op.getValueType(), DAG, |
9282 | dl); |
9283 | |
9284 | // Two instruction sequences. |
9285 | |
9286 | // If this value is in the range [-32,30] and is even, use: |
9287 | // VSPLTI[bhw](val/2) + VSPLTI[bhw](val/2) |
9288 | // If this value is in the range [17,31] and is odd, use: |
9289 | // VSPLTI[bhw](val-16) - VSPLTI[bhw](-16) |
9290 | // If this value is in the range [-31,-17] and is odd, use: |
9291 | // VSPLTI[bhw](val+16) + VSPLTI[bhw](-16) |
9292 | // Note the last two are three-instruction sequences. |
9293 | if (SextVal >= -32 && SextVal <= 31) { |
9294 | // To avoid having these optimizations undone by constant folding, |
9295 | // we convert to a pseudo that will be expanded later into one of |
9296 | // the above forms. |
9297 | SDValue Elt = DAG.getConstant(SextVal, dl, MVT::i32); |
9298 | EVT VT = (SplatSize == 1 ? MVT::v16i8 : |
9299 | (SplatSize == 2 ? MVT::v8i16 : MVT::v4i32)); |
9300 | SDValue EltSize = DAG.getConstant(SplatSize, dl, MVT::i32); |
9301 | SDValue RetVal = DAG.getNode(PPCISD::VADD_SPLAT, dl, VT, Elt, EltSize); |
9302 | if (VT == Op.getValueType()) |
9303 | return RetVal; |
9304 | else |
9305 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), RetVal); |
9306 | } |
9307 | |
9308 | // If this is 0x8000_0000 x 4, turn into vspltisw + vslw. If it is |
9309 | // 0x7FFF_FFFF x 4, turn it into not(0x8000_0000). This is important |
9310 | // for fneg/fabs. |
9311 | if (SplatSize == 4 && SplatBits == (0x7FFFFFFF&~SplatUndef)) { |
9312 | // Make -1 and vspltisw -1: |
9313 | SDValue OnesV = getCanonicalConstSplat(-1, 4, MVT::v4i32, DAG, dl); |
9314 | |
9315 | // Make the VSLW intrinsic, computing 0x8000_0000. |
9316 | SDValue Res = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, OnesV, |
9317 | OnesV, DAG, dl); |
9318 | |
9319 | // xor by OnesV to invert it. |
9320 | Res = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Res, OnesV); |
9321 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9322 | } |
9323 | |
9324 | // Check to see if this is a wide variety of vsplti*, binop self cases. |
9325 | static const signed char SplatCsts[] = { |
9326 | -1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7, |
9327 | -8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16 |
9328 | }; |
9329 | |
9330 | for (unsigned idx = 0; idx < array_lengthof(SplatCsts); ++idx) { |
9331 | // Indirect through the SplatCsts array so that we favor 'vsplti -1' for |
9332 | // cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1' |
9333 | int i = SplatCsts[idx]; |
9334 | |
9335 | // Figure out what shift amount will be used by altivec if shifted by i in |
9336 | // this splat size. |
9337 | unsigned TypeShiftAmt = i & (SplatBitSize-1); |
9338 | |
9339 | // vsplti + shl self. |
9340 | if (SextVal == (int)((unsigned)i << TypeShiftAmt)) { |
9341 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9342 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9343 | Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0, |
9344 | Intrinsic::ppc_altivec_vslw |
9345 | }; |
9346 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9347 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9348 | } |
9349 | |
9350 | // vsplti + srl self. |
9351 | if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) { |
9352 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9353 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9354 | Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0, |
9355 | Intrinsic::ppc_altivec_vsrw |
9356 | }; |
9357 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9358 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9359 | } |
9360 | |
9361 | // vsplti + sra self. |
9362 | if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) { |
9363 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9364 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9365 | Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0, |
9366 | Intrinsic::ppc_altivec_vsraw |
9367 | }; |
9368 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9369 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9370 | } |
9371 | |
9372 | // vsplti + rol self. |
9373 | if (SextVal == (int)(((unsigned)i << TypeShiftAmt) | |
9374 | ((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) { |
9375 | SDValue Res = getCanonicalConstSplat(i, SplatSize, MVT::Other, DAG, dl); |
9376 | static const unsigned IIDs[] = { // Intrinsic to use for each size. |
9377 | Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0, |
9378 | Intrinsic::ppc_altivec_vrlw |
9379 | }; |
9380 | Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG, dl); |
9381 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res); |
9382 | } |
9383 | |
9384 | // t = vsplti c, result = vsldoi t, t, 1 |
9385 | if (SextVal == (int)(((unsigned)i << 8) | (i < 0 ? 0xFF : 0))) { |
9386 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9387 | unsigned Amt = Subtarget.isLittleEndian() ? 15 : 1; |
9388 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9389 | } |
9390 | // t = vsplti c, result = vsldoi t, t, 2 |
9391 | if (SextVal == (int)(((unsigned)i << 16) | (i < 0 ? 0xFFFF : 0))) { |
9392 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9393 | unsigned Amt = Subtarget.isLittleEndian() ? 14 : 2; |
9394 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9395 | } |
9396 | // t = vsplti c, result = vsldoi t, t, 3 |
9397 | if (SextVal == (int)(((unsigned)i << 24) | (i < 0 ? 0xFFFFFF : 0))) { |
9398 | SDValue T = getCanonicalConstSplat(i, SplatSize, MVT::v16i8, DAG, dl); |
9399 | unsigned Amt = Subtarget.isLittleEndian() ? 13 : 3; |
9400 | return BuildVSLDOI(T, T, Amt, Op.getValueType(), DAG, dl); |
9401 | } |
9402 | } |
9403 | |
9404 | return SDValue(); |
9405 | } |
9406 | |
9407 | /// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit |
9408 | /// the specified operations to build the shuffle. |
9409 | static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, |
9410 | SDValue RHS, SelectionDAG &DAG, |
9411 | const SDLoc &dl) { |
9412 | unsigned OpNum = (PFEntry >> 26) & 0x0F; |
9413 | unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); |
9414 | unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); |
9415 | |
9416 | enum { |
9417 | OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> |
9418 | OP_VMRGHW, |
9419 | OP_VMRGLW, |
9420 | OP_VSPLTISW0, |
9421 | OP_VSPLTISW1, |
9422 | OP_VSPLTISW2, |
9423 | OP_VSPLTISW3, |
9424 | OP_VSLDOI4, |
9425 | OP_VSLDOI8, |
9426 | OP_VSLDOI12 |
9427 | }; |
9428 | |
9429 | if (OpNum == OP_COPY) { |
9430 | if (LHSID == (1*9+2)*9+3) return LHS; |
9431 | assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!")((LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!") ? static_cast<void> (0) : __assert_fail ("LHSID == ((4*9+5)*9+6)*9+7 && \"Illegal OP_COPY!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9431, __PRETTY_FUNCTION__)); |
9432 | return RHS; |
9433 | } |
9434 | |
9435 | SDValue OpLHS, OpRHS; |
9436 | OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); |
9437 | OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); |
9438 | |
9439 | int ShufIdxs[16]; |
9440 | switch (OpNum) { |
9441 | default: llvm_unreachable("Unknown i32 permute!")::llvm::llvm_unreachable_internal("Unknown i32 permute!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9441); |
9442 | case OP_VMRGHW: |
9443 | ShufIdxs[ 0] = 0; ShufIdxs[ 1] = 1; ShufIdxs[ 2] = 2; ShufIdxs[ 3] = 3; |
9444 | ShufIdxs[ 4] = 16; ShufIdxs[ 5] = 17; ShufIdxs[ 6] = 18; ShufIdxs[ 7] = 19; |
9445 | ShufIdxs[ 8] = 4; ShufIdxs[ 9] = 5; ShufIdxs[10] = 6; ShufIdxs[11] = 7; |
9446 | ShufIdxs[12] = 20; ShufIdxs[13] = 21; ShufIdxs[14] = 22; ShufIdxs[15] = 23; |
9447 | break; |
9448 | case OP_VMRGLW: |
9449 | ShufIdxs[ 0] = 8; ShufIdxs[ 1] = 9; ShufIdxs[ 2] = 10; ShufIdxs[ 3] = 11; |
9450 | ShufIdxs[ 4] = 24; ShufIdxs[ 5] = 25; ShufIdxs[ 6] = 26; ShufIdxs[ 7] = 27; |
9451 | ShufIdxs[ 8] = 12; ShufIdxs[ 9] = 13; ShufIdxs[10] = 14; ShufIdxs[11] = 15; |
9452 | ShufIdxs[12] = 28; ShufIdxs[13] = 29; ShufIdxs[14] = 30; ShufIdxs[15] = 31; |
9453 | break; |
9454 | case OP_VSPLTISW0: |
9455 | for (unsigned i = 0; i != 16; ++i) |
9456 | ShufIdxs[i] = (i&3)+0; |
9457 | break; |
9458 | case OP_VSPLTISW1: |
9459 | for (unsigned i = 0; i != 16; ++i) |
9460 | ShufIdxs[i] = (i&3)+4; |
9461 | break; |
9462 | case OP_VSPLTISW2: |
9463 | for (unsigned i = 0; i != 16; ++i) |
9464 | ShufIdxs[i] = (i&3)+8; |
9465 | break; |
9466 | case OP_VSPLTISW3: |
9467 | for (unsigned i = 0; i != 16; ++i) |
9468 | ShufIdxs[i] = (i&3)+12; |
9469 | break; |
9470 | case OP_VSLDOI4: |
9471 | return BuildVSLDOI(OpLHS, OpRHS, 4, OpLHS.getValueType(), DAG, dl); |
9472 | case OP_VSLDOI8: |
9473 | return BuildVSLDOI(OpLHS, OpRHS, 8, OpLHS.getValueType(), DAG, dl); |
9474 | case OP_VSLDOI12: |
9475 | return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG, dl); |
9476 | } |
9477 | EVT VT = OpLHS.getValueType(); |
9478 | OpLHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLHS); |
9479 | OpRHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpRHS); |
9480 | SDValue T = DAG.getVectorShuffle(MVT::v16i8, dl, OpLHS, OpRHS, ShufIdxs); |
9481 | return DAG.getNode(ISD::BITCAST, dl, VT, T); |
9482 | } |
9483 | |
9484 | /// lowerToVINSERTB - Return the SDValue if this VECTOR_SHUFFLE can be handled |
9485 | /// by the VINSERTB instruction introduced in ISA 3.0, else just return default |
9486 | /// SDValue. |
9487 | SDValue PPCTargetLowering::lowerToVINSERTB(ShuffleVectorSDNode *N, |
9488 | SelectionDAG &DAG) const { |
9489 | const unsigned BytesInVector = 16; |
9490 | bool IsLE = Subtarget.isLittleEndian(); |
9491 | SDLoc dl(N); |
9492 | SDValue V1 = N->getOperand(0); |
9493 | SDValue V2 = N->getOperand(1); |
9494 | unsigned ShiftElts = 0, InsertAtByte = 0; |
9495 | bool Swap = false; |
9496 | |
9497 | // Shifts required to get the byte we want at element 7. |
9498 | unsigned LittleEndianShifts[] = {8, 7, 6, 5, 4, 3, 2, 1, |
9499 | 0, 15, 14, 13, 12, 11, 10, 9}; |
9500 | unsigned BigEndianShifts[] = {9, 10, 11, 12, 13, 14, 15, 0, |
9501 | 1, 2, 3, 4, 5, 6, 7, 8}; |
9502 | |
9503 | ArrayRef<int> Mask = N->getMask(); |
9504 | int OriginalOrder[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}; |
9505 | |
9506 | // For each mask element, find out if we're just inserting something |
9507 | // from V2 into V1 or vice versa. |
9508 | // Possible permutations inserting an element from V2 into V1: |
9509 | // X, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
9510 | // 0, X, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
9511 | // ... |
9512 | // 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, X |
9513 | // Inserting from V1 into V2 will be similar, except mask range will be |
9514 | // [16,31]. |
9515 | |
9516 | bool FoundCandidate = false; |
9517 | // If both vector operands for the shuffle are the same vector, the mask |
9518 | // will contain only elements from the first one and the second one will be |
9519 | // undef. |
9520 | unsigned VINSERTBSrcElem = IsLE ? 8 : 7; |
9521 | // Go through the mask of half-words to find an element that's being moved |
9522 | // from one vector to the other. |
9523 | for (unsigned i = 0; i < BytesInVector; ++i) { |
9524 | unsigned CurrentElement = Mask[i]; |
9525 | // If 2nd operand is undefined, we should only look for element 7 in the |
9526 | // Mask. |
9527 | if (V2.isUndef() && CurrentElement != VINSERTBSrcElem) |
9528 | continue; |
9529 | |
9530 | bool OtherElementsInOrder = true; |
9531 | // Examine the other elements in the Mask to see if they're in original |
9532 | // order. |
9533 | for (unsigned j = 0; j < BytesInVector; ++j) { |
9534 | if (j == i) |
9535 | continue; |
9536 | // If CurrentElement is from V1 [0,15], then we the rest of the Mask to be |
9537 | // from V2 [16,31] and vice versa. Unless the 2nd operand is undefined, |
9538 | // in which we always assume we're always picking from the 1st operand. |
9539 | int MaskOffset = |
9540 | (!V2.isUndef() && CurrentElement < BytesInVector) ? BytesInVector : 0; |
9541 | if (Mask[j] != OriginalOrder[j] + MaskOffset) { |
9542 | OtherElementsInOrder = false; |
9543 | break; |
9544 | } |
9545 | } |
9546 | // If other elements are in original order, we record the number of shifts |
9547 | // we need to get the element we want into element 7. Also record which byte |
9548 | // in the vector we should insert into. |
9549 | if (OtherElementsInOrder) { |
9550 | // If 2nd operand is undefined, we assume no shifts and no swapping. |
9551 | if (V2.isUndef()) { |
9552 | ShiftElts = 0; |
9553 | Swap = false; |
9554 | } else { |
9555 | // Only need the last 4-bits for shifts because operands will be swapped if CurrentElement is >= 2^4. |
9556 | ShiftElts = IsLE ? LittleEndianShifts[CurrentElement & 0xF] |
9557 | : BigEndianShifts[CurrentElement & 0xF]; |
9558 | Swap = CurrentElement < BytesInVector; |
9559 | } |
9560 | InsertAtByte = IsLE ? BytesInVector - (i + 1) : i; |
9561 | FoundCandidate = true; |
9562 | break; |
9563 | } |
9564 | } |
9565 | |
9566 | if (!FoundCandidate) |
9567 | return SDValue(); |
9568 | |
9569 | // Candidate found, construct the proper SDAG sequence with VINSERTB, |
9570 | // optionally with VECSHL if shift is required. |
9571 | if (Swap) |
9572 | std::swap(V1, V2); |
9573 | if (V2.isUndef()) |
9574 | V2 = V1; |
9575 | if (ShiftElts) { |
9576 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2, |
9577 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9578 | return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, Shl, |
9579 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9580 | } |
9581 | return DAG.getNode(PPCISD::VECINSERT, dl, MVT::v16i8, V1, V2, |
9582 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9583 | } |
9584 | |
9585 | /// lowerToVINSERTH - Return the SDValue if this VECTOR_SHUFFLE can be handled |
9586 | /// by the VINSERTH instruction introduced in ISA 3.0, else just return default |
9587 | /// SDValue. |
9588 | SDValue PPCTargetLowering::lowerToVINSERTH(ShuffleVectorSDNode *N, |
9589 | SelectionDAG &DAG) const { |
9590 | const unsigned NumHalfWords = 8; |
9591 | const unsigned BytesInVector = NumHalfWords * 2; |
9592 | // Check that the shuffle is on half-words. |
9593 | if (!isNByteElemShuffleMask(N, 2, 1)) |
9594 | return SDValue(); |
9595 | |
9596 | bool IsLE = Subtarget.isLittleEndian(); |
9597 | SDLoc dl(N); |
9598 | SDValue V1 = N->getOperand(0); |
9599 | SDValue V2 = N->getOperand(1); |
9600 | unsigned ShiftElts = 0, InsertAtByte = 0; |
9601 | bool Swap = false; |
9602 | |
9603 | // Shifts required to get the half-word we want at element 3. |
9604 | unsigned LittleEndianShifts[] = {4, 3, 2, 1, 0, 7, 6, 5}; |
9605 | unsigned BigEndianShifts[] = {5, 6, 7, 0, 1, 2, 3, 4}; |
9606 | |
9607 | uint32_t Mask = 0; |
9608 | uint32_t OriginalOrderLow = 0x1234567; |
9609 | uint32_t OriginalOrderHigh = 0x89ABCDEF; |
9610 | // Now we look at mask elements 0,2,4,6,8,10,12,14. Pack the mask into a |
9611 | // 32-bit space, only need 4-bit nibbles per element. |
9612 | for (unsigned i = 0; i < NumHalfWords; ++i) { |
9613 | unsigned MaskShift = (NumHalfWords - 1 - i) * 4; |
9614 | Mask |= ((uint32_t)(N->getMaskElt(i * 2) / 2) << MaskShift); |
9615 | } |
9616 | |
9617 | // For each mask element, find out if we're just inserting something |
9618 | // from V2 into V1 or vice versa. Possible permutations inserting an element |
9619 | // from V2 into V1: |
9620 | // X, 1, 2, 3, 4, 5, 6, 7 |
9621 | // 0, X, 2, 3, 4, 5, 6, 7 |
9622 | // 0, 1, X, 3, 4, 5, 6, 7 |
9623 | // 0, 1, 2, X, 4, 5, 6, 7 |
9624 | // 0, 1, 2, 3, X, 5, 6, 7 |
9625 | // 0, 1, 2, 3, 4, X, 6, 7 |
9626 | // 0, 1, 2, 3, 4, 5, X, 7 |
9627 | // 0, 1, 2, 3, 4, 5, 6, X |
9628 | // Inserting from V1 into V2 will be similar, except mask range will be [8,15]. |
9629 | |
9630 | bool FoundCandidate = false; |
9631 | // Go through the mask of half-words to find an element that's being moved |
9632 | // from one vector to the other. |
9633 | for (unsigned i = 0; i < NumHalfWords; ++i) { |
9634 | unsigned MaskShift = (NumHalfWords - 1 - i) * 4; |
9635 | uint32_t MaskOneElt = (Mask >> MaskShift) & 0xF; |
9636 | uint32_t MaskOtherElts = ~(0xF << MaskShift); |
9637 | uint32_t TargetOrder = 0x0; |
9638 | |
9639 | // If both vector operands for the shuffle are the same vector, the mask |
9640 | // will contain only elements from the first one and the second one will be |
9641 | // undef. |
9642 | if (V2.isUndef()) { |
9643 | ShiftElts = 0; |
9644 | unsigned VINSERTHSrcElem = IsLE ? 4 : 3; |
9645 | TargetOrder = OriginalOrderLow; |
9646 | Swap = false; |
9647 | // Skip if not the correct element or mask of other elements don't equal |
9648 | // to our expected order. |
9649 | if (MaskOneElt == VINSERTHSrcElem && |
9650 | (Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) { |
9651 | InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2; |
9652 | FoundCandidate = true; |
9653 | break; |
9654 | } |
9655 | } else { // If both operands are defined. |
9656 | // Target order is [8,15] if the current mask is between [0,7]. |
9657 | TargetOrder = |
9658 | (MaskOneElt < NumHalfWords) ? OriginalOrderHigh : OriginalOrderLow; |
9659 | // Skip if mask of other elements don't equal our expected order. |
9660 | if ((Mask & MaskOtherElts) == (TargetOrder & MaskOtherElts)) { |
9661 | // We only need the last 3 bits for the number of shifts. |
9662 | ShiftElts = IsLE ? LittleEndianShifts[MaskOneElt & 0x7] |
9663 | : BigEndianShifts[MaskOneElt & 0x7]; |
9664 | InsertAtByte = IsLE ? BytesInVector - (i + 1) * 2 : i * 2; |
9665 | Swap = MaskOneElt < NumHalfWords; |
9666 | FoundCandidate = true; |
9667 | break; |
9668 | } |
9669 | } |
9670 | } |
9671 | |
9672 | if (!FoundCandidate) |
9673 | return SDValue(); |
9674 | |
9675 | // Candidate found, construct the proper SDAG sequence with VINSERTH, |
9676 | // optionally with VECSHL if shift is required. |
9677 | if (Swap) |
9678 | std::swap(V1, V2); |
9679 | if (V2.isUndef()) |
9680 | V2 = V1; |
9681 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); |
9682 | if (ShiftElts) { |
9683 | // Double ShiftElts because we're left shifting on v16i8 type. |
9684 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v16i8, V2, V2, |
9685 | DAG.getConstant(2 * ShiftElts, dl, MVT::i32)); |
9686 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, Shl); |
9687 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2, |
9688 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9689 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9690 | } |
9691 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2); |
9692 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v8i16, Conv1, Conv2, |
9693 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9694 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9695 | } |
9696 | |
9697 | /// lowerToXXSPLTI32DX - Return the SDValue if this VECTOR_SHUFFLE can be |
9698 | /// handled by the XXSPLTI32DX instruction introduced in ISA 3.1, otherwise |
9699 | /// return the default SDValue. |
9700 | SDValue PPCTargetLowering::lowerToXXSPLTI32DX(ShuffleVectorSDNode *SVN, |
9701 | SelectionDAG &DAG) const { |
9702 | // The LHS and RHS may be bitcasts to v16i8 as we canonicalize shuffles |
9703 | // to v16i8. Peek through the bitcasts to get the actual operands. |
9704 | SDValue LHS = peekThroughBitcasts(SVN->getOperand(0)); |
9705 | SDValue RHS = peekThroughBitcasts(SVN->getOperand(1)); |
9706 | |
9707 | auto ShuffleMask = SVN->getMask(); |
9708 | SDValue VecShuffle(SVN, 0); |
9709 | SDLoc DL(SVN); |
9710 | |
9711 | // Check that we have a four byte shuffle. |
9712 | if (!isNByteElemShuffleMask(SVN, 4, 1)) |
9713 | return SDValue(); |
9714 | |
9715 | // Canonicalize the RHS being a BUILD_VECTOR when lowering to xxsplti32dx. |
9716 | if (RHS->getOpcode() != ISD::BUILD_VECTOR) { |
9717 | std::swap(LHS, RHS); |
9718 | VecShuffle = DAG.getCommutedVectorShuffle(*SVN); |
9719 | ShuffleMask = cast<ShuffleVectorSDNode>(VecShuffle)->getMask(); |
9720 | } |
9721 | |
9722 | // Ensure that the RHS is a vector of constants. |
9723 | BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode()); |
9724 | if (!BVN) |
9725 | return SDValue(); |
9726 | |
9727 | // Check if RHS is a splat of 4-bytes (or smaller). |
9728 | APInt APSplatValue, APSplatUndef; |
9729 | unsigned SplatBitSize; |
9730 | bool HasAnyUndefs; |
9731 | if (!BVN->isConstantSplat(APSplatValue, APSplatUndef, SplatBitSize, |
9732 | HasAnyUndefs, 0, !Subtarget.isLittleEndian()) || |
9733 | SplatBitSize > 32) |
9734 | return SDValue(); |
9735 | |
9736 | // Check that the shuffle mask matches the semantics of XXSPLTI32DX. |
9737 | // The instruction splats a constant C into two words of the source vector |
9738 | // producing { C, Unchanged, C, Unchanged } or { Unchanged, C, Unchanged, C }. |
9739 | // Thus we check that the shuffle mask is the equivalent of |
9740 | // <0, [4-7], 2, [4-7]> or <[4-7], 1, [4-7], 3> respectively. |
9741 | // Note: the check above of isNByteElemShuffleMask() ensures that the bytes |
9742 | // within each word are consecutive, so we only need to check the first byte. |
9743 | SDValue Index; |
9744 | bool IsLE = Subtarget.isLittleEndian(); |
9745 | if ((ShuffleMask[0] == 0 && ShuffleMask[8] == 8) && |
9746 | (ShuffleMask[4] % 4 == 0 && ShuffleMask[12] % 4 == 0 && |
9747 | ShuffleMask[4] > 15 && ShuffleMask[12] > 15)) |
9748 | Index = DAG.getTargetConstant(IsLE ? 0 : 1, DL, MVT::i32); |
9749 | else if ((ShuffleMask[4] == 4 && ShuffleMask[12] == 12) && |
9750 | (ShuffleMask[0] % 4 == 0 && ShuffleMask[8] % 4 == 0 && |
9751 | ShuffleMask[0] > 15 && ShuffleMask[8] > 15)) |
9752 | Index = DAG.getTargetConstant(IsLE ? 1 : 0, DL, MVT::i32); |
9753 | else |
9754 | return SDValue(); |
9755 | |
9756 | // If the splat is narrower than 32-bits, we need to get the 32-bit value |
9757 | // for XXSPLTI32DX. |
9758 | unsigned SplatVal = APSplatValue.getZExtValue(); |
9759 | for (; SplatBitSize < 32; SplatBitSize <<= 1) |
9760 | SplatVal |= (SplatVal << SplatBitSize); |
9761 | |
9762 | SDValue SplatNode = DAG.getNode( |
9763 | PPCISD::XXSPLTI32DX, DL, MVT::v2i64, DAG.getBitcast(MVT::v2i64, LHS), |
9764 | Index, DAG.getTargetConstant(SplatVal, DL, MVT::i32)); |
9765 | return DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, SplatNode); |
9766 | } |
9767 | |
9768 | /// LowerROTL - Custom lowering for ROTL(v1i128) to vector_shuffle(v16i8). |
9769 | /// We lower ROTL(v1i128) to vector_shuffle(v16i8) only if shift amount is |
9770 | /// a multiple of 8. Otherwise convert it to a scalar rotation(i128) |
9771 | /// i.e (or (shl x, C1), (srl x, 128-C1)). |
9772 | SDValue PPCTargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const { |
9773 | assert(Op.getOpcode() == ISD::ROTL && "Should only be called for ISD::ROTL")((Op.getOpcode() == ISD::ROTL && "Should only be called for ISD::ROTL" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ROTL && \"Should only be called for ISD::ROTL\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9773, __PRETTY_FUNCTION__)); |
9774 | assert(Op.getValueType() == MVT::v1i128 &&((Op.getValueType() == MVT::v1i128 && "Only set v1i128 as custom, other type shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::v1i128 && \"Only set v1i128 as custom, other type shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9775, __PRETTY_FUNCTION__)) |
9775 | "Only set v1i128 as custom, other type shouldn't reach here!")((Op.getValueType() == MVT::v1i128 && "Only set v1i128 as custom, other type shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == MVT::v1i128 && \"Only set v1i128 as custom, other type shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9775, __PRETTY_FUNCTION__)); |
9776 | SDLoc dl(Op); |
9777 | SDValue N0 = peekThroughBitcasts(Op.getOperand(0)); |
9778 | SDValue N1 = peekThroughBitcasts(Op.getOperand(1)); |
9779 | unsigned SHLAmt = N1.getConstantOperandVal(0); |
9780 | if (SHLAmt % 8 == 0) { |
9781 | SmallVector<int, 16> Mask(16, 0); |
9782 | std::iota(Mask.begin(), Mask.end(), 0); |
9783 | std::rotate(Mask.begin(), Mask.begin() + SHLAmt / 8, Mask.end()); |
9784 | if (SDValue Shuffle = |
9785 | DAG.getVectorShuffle(MVT::v16i8, dl, DAG.getBitcast(MVT::v16i8, N0), |
9786 | DAG.getUNDEF(MVT::v16i8), Mask)) |
9787 | return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, Shuffle); |
9788 | } |
9789 | SDValue ArgVal = DAG.getBitcast(MVT::i128, N0); |
9790 | SDValue SHLOp = DAG.getNode(ISD::SHL, dl, MVT::i128, ArgVal, |
9791 | DAG.getConstant(SHLAmt, dl, MVT::i32)); |
9792 | SDValue SRLOp = DAG.getNode(ISD::SRL, dl, MVT::i128, ArgVal, |
9793 | DAG.getConstant(128 - SHLAmt, dl, MVT::i32)); |
9794 | SDValue OROp = DAG.getNode(ISD::OR, dl, MVT::i128, SHLOp, SRLOp); |
9795 | return DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, OROp); |
9796 | } |
9797 | |
9798 | /// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this |
9799 | /// is a shuffle we can handle in a single instruction, return it. Otherwise, |
9800 | /// return the code it can be lowered into. Worst case, it can always be |
9801 | /// lowered into a vperm. |
9802 | SDValue PPCTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, |
9803 | SelectionDAG &DAG) const { |
9804 | SDLoc dl(Op); |
9805 | SDValue V1 = Op.getOperand(0); |
9806 | SDValue V2 = Op.getOperand(1); |
9807 | ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op); |
9808 | |
9809 | // Any nodes that were combined in the target-independent combiner prior |
9810 | // to vector legalization will not be sent to the target combine. Try to |
9811 | // combine it here. |
9812 | if (SDValue NewShuffle = combineVectorShuffle(SVOp, DAG)) { |
9813 | if (!isa<ShuffleVectorSDNode>(NewShuffle)) |
9814 | return NewShuffle; |
9815 | Op = NewShuffle; |
9816 | SVOp = cast<ShuffleVectorSDNode>(Op); |
9817 | V1 = Op.getOperand(0); |
9818 | V2 = Op.getOperand(1); |
9819 | } |
9820 | EVT VT = Op.getValueType(); |
9821 | bool isLittleEndian = Subtarget.isLittleEndian(); |
9822 | |
9823 | unsigned ShiftElts, InsertAtByte; |
9824 | bool Swap = false; |
9825 | |
9826 | // If this is a load-and-splat, we can do that with a single instruction |
9827 | // in some cases. However if the load has multiple uses, we don't want to |
9828 | // combine it because that will just produce multiple loads. |
9829 | bool IsPermutedLoad = false; |
9830 | const SDValue *InputLoad = getNormalLoadInput(V1, IsPermutedLoad); |
9831 | if (InputLoad && Subtarget.hasVSX() && V2.isUndef() && |
9832 | (PPC::isSplatShuffleMask(SVOp, 4) || PPC::isSplatShuffleMask(SVOp, 8)) && |
9833 | InputLoad->hasOneUse()) { |
9834 | bool IsFourByte = PPC::isSplatShuffleMask(SVOp, 4); |
9835 | int SplatIdx = |
9836 | PPC::getSplatIdxForPPCMnemonics(SVOp, IsFourByte ? 4 : 8, DAG); |
9837 | |
9838 | // The splat index for permuted loads will be in the left half of the vector |
9839 | // which is strictly wider than the loaded value by 8 bytes. So we need to |
9840 | // adjust the splat index to point to the correct address in memory. |
9841 | if (IsPermutedLoad) { |
9842 | assert(isLittleEndian && "Unexpected permuted load on big endian target")((isLittleEndian && "Unexpected permuted load on big endian target" ) ? static_cast<void> (0) : __assert_fail ("isLittleEndian && \"Unexpected permuted load on big endian target\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9842, __PRETTY_FUNCTION__)); |
9843 | SplatIdx += IsFourByte ? 2 : 1; |
9844 | assert((SplatIdx < (IsFourByte ? 4 : 2)) &&(((SplatIdx < (IsFourByte ? 4 : 2)) && "Splat of a value outside of the loaded memory" ) ? static_cast<void> (0) : __assert_fail ("(SplatIdx < (IsFourByte ? 4 : 2)) && \"Splat of a value outside of the loaded memory\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9845, __PRETTY_FUNCTION__)) |
9845 | "Splat of a value outside of the loaded memory")(((SplatIdx < (IsFourByte ? 4 : 2)) && "Splat of a value outside of the loaded memory" ) ? static_cast<void> (0) : __assert_fail ("(SplatIdx < (IsFourByte ? 4 : 2)) && \"Splat of a value outside of the loaded memory\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 9845, __PRETTY_FUNCTION__)); |
9846 | } |
9847 | |
9848 | LoadSDNode *LD = cast<LoadSDNode>(*InputLoad); |
9849 | // For 4-byte load-and-splat, we need Power9. |
9850 | if ((IsFourByte && Subtarget.hasP9Vector()) || !IsFourByte) { |
9851 | uint64_t Offset = 0; |
9852 | if (IsFourByte) |
9853 | Offset = isLittleEndian ? (3 - SplatIdx) * 4 : SplatIdx * 4; |
9854 | else |
9855 | Offset = isLittleEndian ? (1 - SplatIdx) * 8 : SplatIdx * 8; |
9856 | |
9857 | SDValue BasePtr = LD->getBasePtr(); |
9858 | if (Offset != 0) |
9859 | BasePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()), |
9860 | BasePtr, DAG.getIntPtrConstant(Offset, dl)); |
9861 | SDValue Ops[] = { |
9862 | LD->getChain(), // Chain |
9863 | BasePtr, // BasePtr |
9864 | DAG.getValueType(Op.getValueType()) // VT |
9865 | }; |
9866 | SDVTList VTL = |
9867 | DAG.getVTList(IsFourByte ? MVT::v4i32 : MVT::v2i64, MVT::Other); |
9868 | SDValue LdSplt = |
9869 | DAG.getMemIntrinsicNode(PPCISD::LD_SPLAT, dl, VTL, |
9870 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
9871 | if (LdSplt.getValueType() != SVOp->getValueType(0)) |
9872 | LdSplt = DAG.getBitcast(SVOp->getValueType(0), LdSplt); |
9873 | return LdSplt; |
9874 | } |
9875 | } |
9876 | if (Subtarget.hasP9Vector() && |
9877 | PPC::isXXINSERTWMask(SVOp, ShiftElts, InsertAtByte, Swap, |
9878 | isLittleEndian)) { |
9879 | if (Swap) |
9880 | std::swap(V1, V2); |
9881 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9882 | SDValue Conv2 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2); |
9883 | if (ShiftElts) { |
9884 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv2, Conv2, |
9885 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9886 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Shl, |
9887 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9888 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9889 | } |
9890 | SDValue Ins = DAG.getNode(PPCISD::VECINSERT, dl, MVT::v4i32, Conv1, Conv2, |
9891 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
9892 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Ins); |
9893 | } |
9894 | |
9895 | if (Subtarget.hasPrefixInstrs()) { |
9896 | SDValue SplatInsertNode; |
9897 | if ((SplatInsertNode = lowerToXXSPLTI32DX(SVOp, DAG))) |
9898 | return SplatInsertNode; |
9899 | } |
9900 | |
9901 | if (Subtarget.hasP9Altivec()) { |
9902 | SDValue NewISDNode; |
9903 | if ((NewISDNode = lowerToVINSERTH(SVOp, DAG))) |
9904 | return NewISDNode; |
9905 | |
9906 | if ((NewISDNode = lowerToVINSERTB(SVOp, DAG))) |
9907 | return NewISDNode; |
9908 | } |
9909 | |
9910 | if (Subtarget.hasVSX() && |
9911 | PPC::isXXSLDWIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) { |
9912 | if (Swap) |
9913 | std::swap(V1, V2); |
9914 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9915 | SDValue Conv2 = |
9916 | DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V2.isUndef() ? V1 : V2); |
9917 | |
9918 | SDValue Shl = DAG.getNode(PPCISD::VECSHL, dl, MVT::v4i32, Conv1, Conv2, |
9919 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9920 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Shl); |
9921 | } |
9922 | |
9923 | if (Subtarget.hasVSX() && |
9924 | PPC::isXXPERMDIShuffleMask(SVOp, ShiftElts, Swap, isLittleEndian)) { |
9925 | if (Swap) |
9926 | std::swap(V1, V2); |
9927 | SDValue Conv1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1); |
9928 | SDValue Conv2 = |
9929 | DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2.isUndef() ? V1 : V2); |
9930 | |
9931 | SDValue PermDI = DAG.getNode(PPCISD::XXPERMDI, dl, MVT::v2i64, Conv1, Conv2, |
9932 | DAG.getConstant(ShiftElts, dl, MVT::i32)); |
9933 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, PermDI); |
9934 | } |
9935 | |
9936 | if (Subtarget.hasP9Vector()) { |
9937 | if (PPC::isXXBRHShuffleMask(SVOp)) { |
9938 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1); |
9939 | SDValue ReveHWord = DAG.getNode(ISD::BSWAP, dl, MVT::v8i16, Conv); |
9940 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveHWord); |
9941 | } else if (PPC::isXXBRWShuffleMask(SVOp)) { |
9942 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9943 | SDValue ReveWord = DAG.getNode(ISD::BSWAP, dl, MVT::v4i32, Conv); |
9944 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveWord); |
9945 | } else if (PPC::isXXBRDShuffleMask(SVOp)) { |
9946 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1); |
9947 | SDValue ReveDWord = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Conv); |
9948 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveDWord); |
9949 | } else if (PPC::isXXBRQShuffleMask(SVOp)) { |
9950 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v1i128, V1); |
9951 | SDValue ReveQWord = DAG.getNode(ISD::BSWAP, dl, MVT::v1i128, Conv); |
9952 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, ReveQWord); |
9953 | } |
9954 | } |
9955 | |
9956 | if (Subtarget.hasVSX()) { |
9957 | if (V2.isUndef() && PPC::isSplatShuffleMask(SVOp, 4)) { |
9958 | int SplatIdx = PPC::getSplatIdxForPPCMnemonics(SVOp, 4, DAG); |
9959 | |
9960 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1); |
9961 | SDValue Splat = DAG.getNode(PPCISD::XXSPLT, dl, MVT::v4i32, Conv, |
9962 | DAG.getConstant(SplatIdx, dl, MVT::i32)); |
9963 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Splat); |
9964 | } |
9965 | |
9966 | // Left shifts of 8 bytes are actually swaps. Convert accordingly. |
9967 | if (V2.isUndef() && PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) == 8) { |
9968 | SDValue Conv = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1); |
9969 | SDValue Swap = DAG.getNode(PPCISD::SWAP_NO_CHAIN, dl, MVT::v2f64, Conv); |
9970 | return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, Swap); |
9971 | } |
9972 | } |
9973 | |
9974 | // Cases that are handled by instructions that take permute immediates |
9975 | // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be |
9976 | // selected by the instruction selector. |
9977 | if (V2.isUndef()) { |
9978 | if (PPC::isSplatShuffleMask(SVOp, 1) || |
9979 | PPC::isSplatShuffleMask(SVOp, 2) || |
9980 | PPC::isSplatShuffleMask(SVOp, 4) || |
9981 | PPC::isVPKUWUMShuffleMask(SVOp, 1, DAG) || |
9982 | PPC::isVPKUHUMShuffleMask(SVOp, 1, DAG) || |
9983 | PPC::isVSLDOIShuffleMask(SVOp, 1, DAG) != -1 || |
9984 | PPC::isVMRGLShuffleMask(SVOp, 1, 1, DAG) || |
9985 | PPC::isVMRGLShuffleMask(SVOp, 2, 1, DAG) || |
9986 | PPC::isVMRGLShuffleMask(SVOp, 4, 1, DAG) || |
9987 | PPC::isVMRGHShuffleMask(SVOp, 1, 1, DAG) || |
9988 | PPC::isVMRGHShuffleMask(SVOp, 2, 1, DAG) || |
9989 | PPC::isVMRGHShuffleMask(SVOp, 4, 1, DAG) || |
9990 | (Subtarget.hasP8Altivec() && ( |
9991 | PPC::isVPKUDUMShuffleMask(SVOp, 1, DAG) || |
9992 | PPC::isVMRGEOShuffleMask(SVOp, true, 1, DAG) || |
9993 | PPC::isVMRGEOShuffleMask(SVOp, false, 1, DAG)))) { |
9994 | return Op; |
9995 | } |
9996 | } |
9997 | |
9998 | // Altivec has a variety of "shuffle immediates" that take two vector inputs |
9999 | // and produce a fixed permutation. If any of these match, do not lower to |
10000 | // VPERM. |
10001 | unsigned int ShuffleKind = isLittleEndian ? 2 : 0; |
10002 | if (PPC::isVPKUWUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10003 | PPC::isVPKUHUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10004 | PPC::isVSLDOIShuffleMask(SVOp, ShuffleKind, DAG) != -1 || |
10005 | PPC::isVMRGLShuffleMask(SVOp, 1, ShuffleKind, DAG) || |
10006 | PPC::isVMRGLShuffleMask(SVOp, 2, ShuffleKind, DAG) || |
10007 | PPC::isVMRGLShuffleMask(SVOp, 4, ShuffleKind, DAG) || |
10008 | PPC::isVMRGHShuffleMask(SVOp, 1, ShuffleKind, DAG) || |
10009 | PPC::isVMRGHShuffleMask(SVOp, 2, ShuffleKind, DAG) || |
10010 | PPC::isVMRGHShuffleMask(SVOp, 4, ShuffleKind, DAG) || |
10011 | (Subtarget.hasP8Altivec() && ( |
10012 | PPC::isVPKUDUMShuffleMask(SVOp, ShuffleKind, DAG) || |
10013 | PPC::isVMRGEOShuffleMask(SVOp, true, ShuffleKind, DAG) || |
10014 | PPC::isVMRGEOShuffleMask(SVOp, false, ShuffleKind, DAG)))) |
10015 | return Op; |
10016 | |
10017 | // Check to see if this is a shuffle of 4-byte values. If so, we can use our |
10018 | // perfect shuffle table to emit an optimal matching sequence. |
10019 | ArrayRef<int> PermMask = SVOp->getMask(); |
10020 | |
10021 | unsigned PFIndexes[4]; |
10022 | bool isFourElementShuffle = true; |
10023 | for (unsigned i = 0; i != 4 && isFourElementShuffle; ++i) { // Element number |
10024 | unsigned EltNo = 8; // Start out undef. |
10025 | for (unsigned j = 0; j != 4; ++j) { // Intra-element byte. |
10026 | if (PermMask[i*4+j] < 0) |
10027 | continue; // Undef, ignore it. |
10028 | |
10029 | unsigned ByteSource = PermMask[i*4+j]; |
10030 | if ((ByteSource & 3) != j) { |
10031 | isFourElementShuffle = false; |
10032 | break; |
10033 | } |
10034 | |
10035 | if (EltNo == 8) { |
10036 | EltNo = ByteSource/4; |
10037 | } else if (EltNo != ByteSource/4) { |
10038 | isFourElementShuffle = false; |
10039 | break; |
10040 | } |
10041 | } |
10042 | PFIndexes[i] = EltNo; |
10043 | } |
10044 | |
10045 | // If this shuffle can be expressed as a shuffle of 4-byte elements, use the |
10046 | // perfect shuffle vector to determine if it is cost effective to do this as |
10047 | // discrete instructions, or whether we should use a vperm. |
10048 | // For now, we skip this for little endian until such time as we have a |
10049 | // little-endian perfect shuffle table. |
10050 | if (isFourElementShuffle && !isLittleEndian) { |
10051 | // Compute the index in the perfect shuffle table. |
10052 | unsigned PFTableIndex = |
10053 | PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; |
10054 | |
10055 | unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; |
10056 | unsigned Cost = (PFEntry >> 30); |
10057 | |
10058 | // Determining when to avoid vperm is tricky. Many things affect the cost |
10059 | // of vperm, particularly how many times the perm mask needs to be computed. |
10060 | // For example, if the perm mask can be hoisted out of a loop or is already |
10061 | // used (perhaps because there are multiple permutes with the same shuffle |
10062 | // mask?) the vperm has a cost of 1. OTOH, hoisting the permute mask out of |
10063 | // the loop requires an extra register. |
10064 | // |
10065 | // As a compromise, we only emit discrete instructions if the shuffle can be |
10066 | // generated in 3 or fewer operations. When we have loop information |
10067 | // available, if this block is within a loop, we should avoid using vperm |
10068 | // for 3-operation perms and use a constant pool load instead. |
10069 | if (Cost < 3) |
10070 | return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); |
10071 | } |
10072 | |
10073 | // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant |
10074 | // vector that will get spilled to the constant pool. |
10075 | if (V2.isUndef()) V2 = V1; |
10076 | |
10077 | // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except |
10078 | // that it is in input element units, not in bytes. Convert now. |
10079 | |
10080 | // For little endian, the order of the input vectors is reversed, and |
10081 | // the permutation mask is complemented with respect to 31. This is |
10082 | // necessary to produce proper semantics with the big-endian-biased vperm |
10083 | // instruction. |
10084 | EVT EltVT = V1.getValueType().getVectorElementType(); |
10085 | unsigned BytesPerElement = EltVT.getSizeInBits()/8; |
10086 | |
10087 | SmallVector<SDValue, 16> ResultMask; |
10088 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) { |
10089 | unsigned SrcElt = PermMask[i] < 0 ? 0 : PermMask[i]; |
10090 | |
10091 | for (unsigned j = 0; j != BytesPerElement; ++j) |
10092 | if (isLittleEndian) |
10093 | ResultMask.push_back(DAG.getConstant(31 - (SrcElt*BytesPerElement + j), |
10094 | dl, MVT::i32)); |
10095 | else |
10096 | ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement + j, dl, |
10097 | MVT::i32)); |
10098 | } |
10099 | |
10100 | ShufflesHandledWithVPERM++; |
10101 | SDValue VPermMask = DAG.getBuildVector(MVT::v16i8, dl, ResultMask); |
10102 | LLVM_DEBUG(dbgs() << "Emitting a VPERM for the following shuffle:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "Emitting a VPERM for the following shuffle:\n" ; } } while (false); |
10103 | LLVM_DEBUG(SVOp->dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { SVOp->dump(); } } while (false); |
10104 | LLVM_DEBUG(dbgs() << "With the following permute control vector:\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { dbgs() << "With the following permute control vector:\n" ; } } while (false); |
10105 | LLVM_DEBUG(VPermMask.dump())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("ppc-lowering")) { VPermMask.dump(); } } while (false); |
10106 | |
10107 | if (isLittleEndian) |
10108 | return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), |
10109 | V2, V1, VPermMask); |
10110 | else |
10111 | return DAG.getNode(PPCISD::VPERM, dl, V1.getValueType(), |
10112 | V1, V2, VPermMask); |
10113 | } |
10114 | |
10115 | /// getVectorCompareInfo - Given an intrinsic, return false if it is not a |
10116 | /// vector comparison. If it is, return true and fill in Opc/isDot with |
10117 | /// information about the intrinsic. |
10118 | static bool getVectorCompareInfo(SDValue Intrin, int &CompareOpc, |
10119 | bool &isDot, const PPCSubtarget &Subtarget) { |
10120 | unsigned IntrinsicID = |
10121 | cast<ConstantSDNode>(Intrin.getOperand(0))->getZExtValue(); |
10122 | CompareOpc = -1; |
10123 | isDot = false; |
10124 | switch (IntrinsicID) { |
10125 | default: |
10126 | return false; |
10127 | // Comparison predicates. |
10128 | case Intrinsic::ppc_altivec_vcmpbfp_p: |
10129 | CompareOpc = 966; |
10130 | isDot = true; |
10131 | break; |
10132 | case Intrinsic::ppc_altivec_vcmpeqfp_p: |
10133 | CompareOpc = 198; |
10134 | isDot = true; |
10135 | break; |
10136 | case Intrinsic::ppc_altivec_vcmpequb_p: |
10137 | CompareOpc = 6; |
10138 | isDot = true; |
10139 | break; |
10140 | case Intrinsic::ppc_altivec_vcmpequh_p: |
10141 | CompareOpc = 70; |
10142 | isDot = true; |
10143 | break; |
10144 | case Intrinsic::ppc_altivec_vcmpequw_p: |
10145 | CompareOpc = 134; |
10146 | isDot = true; |
10147 | break; |
10148 | case Intrinsic::ppc_altivec_vcmpequd_p: |
10149 | if (Subtarget.hasP8Altivec()) { |
10150 | CompareOpc = 199; |
10151 | isDot = true; |
10152 | } else |
10153 | return false; |
10154 | break; |
10155 | case Intrinsic::ppc_altivec_vcmpneb_p: |
10156 | case Intrinsic::ppc_altivec_vcmpneh_p: |
10157 | case Intrinsic::ppc_altivec_vcmpnew_p: |
10158 | case Intrinsic::ppc_altivec_vcmpnezb_p: |
10159 | case Intrinsic::ppc_altivec_vcmpnezh_p: |
10160 | case Intrinsic::ppc_altivec_vcmpnezw_p: |
10161 | if (Subtarget.hasP9Altivec()) { |
10162 | switch (IntrinsicID) { |
10163 | default: |
10164 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10164); |
10165 | case Intrinsic::ppc_altivec_vcmpneb_p: |
10166 | CompareOpc = 7; |
10167 | break; |
10168 | case Intrinsic::ppc_altivec_vcmpneh_p: |
10169 | CompareOpc = 71; |
10170 | break; |
10171 | case Intrinsic::ppc_altivec_vcmpnew_p: |
10172 | CompareOpc = 135; |
10173 | break; |
10174 | case Intrinsic::ppc_altivec_vcmpnezb_p: |
10175 | CompareOpc = 263; |
10176 | break; |
10177 | case Intrinsic::ppc_altivec_vcmpnezh_p: |
10178 | CompareOpc = 327; |
10179 | break; |
10180 | case Intrinsic::ppc_altivec_vcmpnezw_p: |
10181 | CompareOpc = 391; |
10182 | break; |
10183 | } |
10184 | isDot = true; |
10185 | } else |
10186 | return false; |
10187 | break; |
10188 | case Intrinsic::ppc_altivec_vcmpgefp_p: |
10189 | CompareOpc = 454; |
10190 | isDot = true; |
10191 | break; |
10192 | case Intrinsic::ppc_altivec_vcmpgtfp_p: |
10193 | CompareOpc = 710; |
10194 | isDot = true; |
10195 | break; |
10196 | case Intrinsic::ppc_altivec_vcmpgtsb_p: |
10197 | CompareOpc = 774; |
10198 | isDot = true; |
10199 | break; |
10200 | case Intrinsic::ppc_altivec_vcmpgtsh_p: |
10201 | CompareOpc = 838; |
10202 | isDot = true; |
10203 | break; |
10204 | case Intrinsic::ppc_altivec_vcmpgtsw_p: |
10205 | CompareOpc = 902; |
10206 | isDot = true; |
10207 | break; |
10208 | case Intrinsic::ppc_altivec_vcmpgtsd_p: |
10209 | if (Subtarget.hasP8Altivec()) { |
10210 | CompareOpc = 967; |
10211 | isDot = true; |
10212 | } else |
10213 | return false; |
10214 | break; |
10215 | case Intrinsic::ppc_altivec_vcmpgtub_p: |
10216 | CompareOpc = 518; |
10217 | isDot = true; |
10218 | break; |
10219 | case Intrinsic::ppc_altivec_vcmpgtuh_p: |
10220 | CompareOpc = 582; |
10221 | isDot = true; |
10222 | break; |
10223 | case Intrinsic::ppc_altivec_vcmpgtuw_p: |
10224 | CompareOpc = 646; |
10225 | isDot = true; |
10226 | break; |
10227 | case Intrinsic::ppc_altivec_vcmpgtud_p: |
10228 | if (Subtarget.hasP8Altivec()) { |
10229 | CompareOpc = 711; |
10230 | isDot = true; |
10231 | } else |
10232 | return false; |
10233 | break; |
10234 | |
10235 | case Intrinsic::ppc_altivec_vcmpequq: |
10236 | case Intrinsic::ppc_altivec_vcmpgtsq: |
10237 | case Intrinsic::ppc_altivec_vcmpgtuq: |
10238 | if (!Subtarget.isISA3_1()) |
10239 | return false; |
10240 | switch (IntrinsicID) { |
10241 | default: |
10242 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10242); |
10243 | case Intrinsic::ppc_altivec_vcmpequq: |
10244 | CompareOpc = 455; |
10245 | break; |
10246 | case Intrinsic::ppc_altivec_vcmpgtsq: |
10247 | CompareOpc = 903; |
10248 | break; |
10249 | case Intrinsic::ppc_altivec_vcmpgtuq: |
10250 | CompareOpc = 647; |
10251 | break; |
10252 | } |
10253 | break; |
10254 | |
10255 | // VSX predicate comparisons use the same infrastructure |
10256 | case Intrinsic::ppc_vsx_xvcmpeqdp_p: |
10257 | case Intrinsic::ppc_vsx_xvcmpgedp_p: |
10258 | case Intrinsic::ppc_vsx_xvcmpgtdp_p: |
10259 | case Intrinsic::ppc_vsx_xvcmpeqsp_p: |
10260 | case Intrinsic::ppc_vsx_xvcmpgesp_p: |
10261 | case Intrinsic::ppc_vsx_xvcmpgtsp_p: |
10262 | if (Subtarget.hasVSX()) { |
10263 | switch (IntrinsicID) { |
10264 | case Intrinsic::ppc_vsx_xvcmpeqdp_p: |
10265 | CompareOpc = 99; |
10266 | break; |
10267 | case Intrinsic::ppc_vsx_xvcmpgedp_p: |
10268 | CompareOpc = 115; |
10269 | break; |
10270 | case Intrinsic::ppc_vsx_xvcmpgtdp_p: |
10271 | CompareOpc = 107; |
10272 | break; |
10273 | case Intrinsic::ppc_vsx_xvcmpeqsp_p: |
10274 | CompareOpc = 67; |
10275 | break; |
10276 | case Intrinsic::ppc_vsx_xvcmpgesp_p: |
10277 | CompareOpc = 83; |
10278 | break; |
10279 | case Intrinsic::ppc_vsx_xvcmpgtsp_p: |
10280 | CompareOpc = 75; |
10281 | break; |
10282 | } |
10283 | isDot = true; |
10284 | } else |
10285 | return false; |
10286 | break; |
10287 | |
10288 | // Normal Comparisons. |
10289 | case Intrinsic::ppc_altivec_vcmpbfp: |
10290 | CompareOpc = 966; |
10291 | break; |
10292 | case Intrinsic::ppc_altivec_vcmpeqfp: |
10293 | CompareOpc = 198; |
10294 | break; |
10295 | case Intrinsic::ppc_altivec_vcmpequb: |
10296 | CompareOpc = 6; |
10297 | break; |
10298 | case Intrinsic::ppc_altivec_vcmpequh: |
10299 | CompareOpc = 70; |
10300 | break; |
10301 | case Intrinsic::ppc_altivec_vcmpequw: |
10302 | CompareOpc = 134; |
10303 | break; |
10304 | case Intrinsic::ppc_altivec_vcmpequd: |
10305 | if (Subtarget.hasP8Altivec()) |
10306 | CompareOpc = 199; |
10307 | else |
10308 | return false; |
10309 | break; |
10310 | case Intrinsic::ppc_altivec_vcmpneb: |
10311 | case Intrinsic::ppc_altivec_vcmpneh: |
10312 | case Intrinsic::ppc_altivec_vcmpnew: |
10313 | case Intrinsic::ppc_altivec_vcmpnezb: |
10314 | case Intrinsic::ppc_altivec_vcmpnezh: |
10315 | case Intrinsic::ppc_altivec_vcmpnezw: |
10316 | if (Subtarget.hasP9Altivec()) |
10317 | switch (IntrinsicID) { |
10318 | default: |
10319 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10319); |
10320 | case Intrinsic::ppc_altivec_vcmpneb: |
10321 | CompareOpc = 7; |
10322 | break; |
10323 | case Intrinsic::ppc_altivec_vcmpneh: |
10324 | CompareOpc = 71; |
10325 | break; |
10326 | case Intrinsic::ppc_altivec_vcmpnew: |
10327 | CompareOpc = 135; |
10328 | break; |
10329 | case Intrinsic::ppc_altivec_vcmpnezb: |
10330 | CompareOpc = 263; |
10331 | break; |
10332 | case Intrinsic::ppc_altivec_vcmpnezh: |
10333 | CompareOpc = 327; |
10334 | break; |
10335 | case Intrinsic::ppc_altivec_vcmpnezw: |
10336 | CompareOpc = 391; |
10337 | break; |
10338 | } |
10339 | else |
10340 | return false; |
10341 | break; |
10342 | case Intrinsic::ppc_altivec_vcmpgefp: |
10343 | CompareOpc = 454; |
10344 | break; |
10345 | case Intrinsic::ppc_altivec_vcmpgtfp: |
10346 | CompareOpc = 710; |
10347 | break; |
10348 | case Intrinsic::ppc_altivec_vcmpgtsb: |
10349 | CompareOpc = 774; |
10350 | break; |
10351 | case Intrinsic::ppc_altivec_vcmpgtsh: |
10352 | CompareOpc = 838; |
10353 | break; |
10354 | case Intrinsic::ppc_altivec_vcmpgtsw: |
10355 | CompareOpc = 902; |
10356 | break; |
10357 | case Intrinsic::ppc_altivec_vcmpgtsd: |
10358 | if (Subtarget.hasP8Altivec()) |
10359 | CompareOpc = 967; |
10360 | else |
10361 | return false; |
10362 | break; |
10363 | case Intrinsic::ppc_altivec_vcmpgtub: |
10364 | CompareOpc = 518; |
10365 | break; |
10366 | case Intrinsic::ppc_altivec_vcmpgtuh: |
10367 | CompareOpc = 582; |
10368 | break; |
10369 | case Intrinsic::ppc_altivec_vcmpgtuw: |
10370 | CompareOpc = 646; |
10371 | break; |
10372 | case Intrinsic::ppc_altivec_vcmpgtud: |
10373 | if (Subtarget.hasP8Altivec()) |
10374 | CompareOpc = 711; |
10375 | else |
10376 | return false; |
10377 | break; |
10378 | case Intrinsic::ppc_altivec_vcmpequq_p: |
10379 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
10380 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
10381 | if (!Subtarget.isISA3_1()) |
10382 | return false; |
10383 | switch (IntrinsicID) { |
10384 | default: |
10385 | llvm_unreachable("Unknown comparison intrinsic.")::llvm::llvm_unreachable_internal("Unknown comparison intrinsic." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10385); |
10386 | case Intrinsic::ppc_altivec_vcmpequq_p: |
10387 | CompareOpc = 455; |
10388 | break; |
10389 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
10390 | CompareOpc = 903; |
10391 | break; |
10392 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
10393 | CompareOpc = 647; |
10394 | break; |
10395 | } |
10396 | isDot = true; |
10397 | break; |
10398 | } |
10399 | return true; |
10400 | } |
10401 | |
10402 | /// LowerINTRINSIC_WO_CHAIN - If this is an intrinsic that we want to custom |
10403 | /// lower, do it, otherwise return null. |
10404 | SDValue PPCTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, |
10405 | SelectionDAG &DAG) const { |
10406 | unsigned IntrinsicID = |
10407 | cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
10408 | |
10409 | SDLoc dl(Op); |
10410 | |
10411 | switch (IntrinsicID) { |
10412 | case Intrinsic::thread_pointer: |
10413 | // Reads the thread pointer register, used for __builtin_thread_pointer. |
10414 | if (Subtarget.isPPC64()) |
10415 | return DAG.getRegister(PPC::X13, MVT::i64); |
10416 | return DAG.getRegister(PPC::R2, MVT::i32); |
10417 | |
10418 | case Intrinsic::ppc_mma_disassemble_acc: |
10419 | case Intrinsic::ppc_mma_disassemble_pair: { |
10420 | int NumVecs = 2; |
10421 | SDValue WideVec = Op.getOperand(1); |
10422 | if (IntrinsicID == Intrinsic::ppc_mma_disassemble_acc) { |
10423 | NumVecs = 4; |
10424 | WideVec = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, WideVec); |
10425 | } |
10426 | SmallVector<SDValue, 4> RetOps; |
10427 | for (int VecNo = 0; VecNo < NumVecs; VecNo++) { |
10428 | SDValue Extract = DAG.getNode( |
10429 | PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, WideVec, |
10430 | DAG.getConstant(Subtarget.isLittleEndian() ? NumVecs - 1 - VecNo |
10431 | : VecNo, |
10432 | dl, MVT::i64)); |
10433 | RetOps.push_back(Extract); |
10434 | } |
10435 | return DAG.getMergeValues(RetOps, dl); |
10436 | } |
10437 | } |
10438 | |
10439 | // If this is a lowered altivec predicate compare, CompareOpc is set to the |
10440 | // opcode number of the comparison. |
10441 | int CompareOpc; |
10442 | bool isDot; |
10443 | if (!getVectorCompareInfo(Op, CompareOpc, isDot, Subtarget)) |
10444 | return SDValue(); // Don't custom lower most intrinsics. |
10445 | |
10446 | // If this is a non-dot comparison, make the VCMP node and we are done. |
10447 | if (!isDot) { |
10448 | SDValue Tmp = DAG.getNode(PPCISD::VCMP, dl, Op.getOperand(2).getValueType(), |
10449 | Op.getOperand(1), Op.getOperand(2), |
10450 | DAG.getConstant(CompareOpc, dl, MVT::i32)); |
10451 | return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Tmp); |
10452 | } |
10453 | |
10454 | // Create the PPCISD altivec 'dot' comparison node. |
10455 | SDValue Ops[] = { |
10456 | Op.getOperand(2), // LHS |
10457 | Op.getOperand(3), // RHS |
10458 | DAG.getConstant(CompareOpc, dl, MVT::i32) |
10459 | }; |
10460 | EVT VTs[] = { Op.getOperand(2).getValueType(), MVT::Glue }; |
10461 | SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops); |
10462 | |
10463 | // Now that we have the comparison, emit a copy from the CR to a GPR. |
10464 | // This is flagged to the above dot comparison. |
10465 | SDValue Flags = DAG.getNode(PPCISD::MFOCRF, dl, MVT::i32, |
10466 | DAG.getRegister(PPC::CR6, MVT::i32), |
10467 | CompNode.getValue(1)); |
10468 | |
10469 | // Unpack the result based on how the target uses it. |
10470 | unsigned BitNo; // Bit # of CR6. |
10471 | bool InvertBit; // Invert result? |
10472 | switch (cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue()) { |
10473 | default: // Can't happen, don't crash on invalid number though. |
10474 | case 0: // Return the value of the EQ bit of CR6. |
10475 | BitNo = 0; InvertBit = false; |
10476 | break; |
10477 | case 1: // Return the inverted value of the EQ bit of CR6. |
10478 | BitNo = 0; InvertBit = true; |
10479 | break; |
10480 | case 2: // Return the value of the LT bit of CR6. |
10481 | BitNo = 2; InvertBit = false; |
10482 | break; |
10483 | case 3: // Return the inverted value of the LT bit of CR6. |
10484 | BitNo = 2; InvertBit = true; |
10485 | break; |
10486 | } |
10487 | |
10488 | // Shift the bit into the low position. |
10489 | Flags = DAG.getNode(ISD::SRL, dl, MVT::i32, Flags, |
10490 | DAG.getConstant(8 - (3 - BitNo), dl, MVT::i32)); |
10491 | // Isolate the bit. |
10492 | Flags = DAG.getNode(ISD::AND, dl, MVT::i32, Flags, |
10493 | DAG.getConstant(1, dl, MVT::i32)); |
10494 | |
10495 | // If we are supposed to, toggle the bit. |
10496 | if (InvertBit) |
10497 | Flags = DAG.getNode(ISD::XOR, dl, MVT::i32, Flags, |
10498 | DAG.getConstant(1, dl, MVT::i32)); |
10499 | return Flags; |
10500 | } |
10501 | |
10502 | SDValue PPCTargetLowering::LowerINTRINSIC_VOID(SDValue Op, |
10503 | SelectionDAG &DAG) const { |
10504 | // SelectionDAGBuilder::visitTargetIntrinsic may insert one extra chain to |
10505 | // the beginning of the argument list. |
10506 | int ArgStart = isa<ConstantSDNode>(Op.getOperand(0)) ? 0 : 1; |
10507 | SDLoc DL(Op); |
10508 | switch (cast<ConstantSDNode>(Op.getOperand(ArgStart))->getZExtValue()) { |
10509 | case Intrinsic::ppc_cfence: { |
10510 | assert(ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument.")((ArgStart == 1 && "llvm.ppc.cfence must carry a chain argument." ) ? static_cast<void> (0) : __assert_fail ("ArgStart == 1 && \"llvm.ppc.cfence must carry a chain argument.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10510, __PRETTY_FUNCTION__)); |
10511 | assert(Subtarget.isPPC64() && "Only 64-bit is supported for now.")((Subtarget.isPPC64() && "Only 64-bit is supported for now." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.isPPC64() && \"Only 64-bit is supported for now.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10511, __PRETTY_FUNCTION__)); |
10512 | return SDValue(DAG.getMachineNode(PPC::CFENCE8, DL, MVT::Other, |
10513 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, |
10514 | Op.getOperand(ArgStart + 1)), |
10515 | Op.getOperand(0)), |
10516 | 0); |
10517 | } |
10518 | default: |
10519 | break; |
10520 | } |
10521 | return SDValue(); |
10522 | } |
10523 | |
10524 | // Lower scalar BSWAP64 to xxbrd. |
10525 | SDValue PPCTargetLowering::LowerBSWAP(SDValue Op, SelectionDAG &DAG) const { |
10526 | SDLoc dl(Op); |
10527 | // MTVSRDD |
10528 | Op = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64, Op.getOperand(0), |
10529 | Op.getOperand(0)); |
10530 | // XXBRD |
10531 | Op = DAG.getNode(ISD::BSWAP, dl, MVT::v2i64, Op); |
10532 | // MFVSRD |
10533 | int VectorIndex = 0; |
10534 | if (Subtarget.isLittleEndian()) |
10535 | VectorIndex = 1; |
10536 | Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Op, |
10537 | DAG.getTargetConstant(VectorIndex, dl, MVT::i32)); |
10538 | return Op; |
10539 | } |
10540 | |
10541 | // ATOMIC_CMP_SWAP for i8/i16 needs to zero-extend its input since it will be |
10542 | // compared to a value that is atomically loaded (atomic loads zero-extend). |
10543 | SDValue PPCTargetLowering::LowerATOMIC_CMP_SWAP(SDValue Op, |
10544 | SelectionDAG &DAG) const { |
10545 | assert(Op.getOpcode() == ISD::ATOMIC_CMP_SWAP &&((Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && "Expecting an atomic compare-and-swap here." ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && \"Expecting an atomic compare-and-swap here.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10546, __PRETTY_FUNCTION__)) |
10546 | "Expecting an atomic compare-and-swap here.")((Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && "Expecting an atomic compare-and-swap here." ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ATOMIC_CMP_SWAP && \"Expecting an atomic compare-and-swap here.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10546, __PRETTY_FUNCTION__)); |
10547 | SDLoc dl(Op); |
10548 | auto *AtomicNode = cast<AtomicSDNode>(Op.getNode()); |
10549 | EVT MemVT = AtomicNode->getMemoryVT(); |
10550 | if (MemVT.getSizeInBits() >= 32) |
10551 | return Op; |
10552 | |
10553 | SDValue CmpOp = Op.getOperand(2); |
10554 | // If this is already correctly zero-extended, leave it alone. |
10555 | auto HighBits = APInt::getHighBitsSet(32, 32 - MemVT.getSizeInBits()); |
10556 | if (DAG.MaskedValueIsZero(CmpOp, HighBits)) |
10557 | return Op; |
10558 | |
10559 | // Clear the high bits of the compare operand. |
10560 | unsigned MaskVal = (1 << MemVT.getSizeInBits()) - 1; |
10561 | SDValue NewCmpOp = |
10562 | DAG.getNode(ISD::AND, dl, MVT::i32, CmpOp, |
10563 | DAG.getConstant(MaskVal, dl, MVT::i32)); |
10564 | |
10565 | // Replace the existing compare operand with the properly zero-extended one. |
10566 | SmallVector<SDValue, 4> Ops; |
10567 | for (int i = 0, e = AtomicNode->getNumOperands(); i < e; i++) |
10568 | Ops.push_back(AtomicNode->getOperand(i)); |
10569 | Ops[2] = NewCmpOp; |
10570 | MachineMemOperand *MMO = AtomicNode->getMemOperand(); |
10571 | SDVTList Tys = DAG.getVTList(MVT::i32, MVT::Other); |
10572 | auto NodeTy = |
10573 | (MemVT == MVT::i8) ? PPCISD::ATOMIC_CMP_SWAP_8 : PPCISD::ATOMIC_CMP_SWAP_16; |
10574 | return DAG.getMemIntrinsicNode(NodeTy, dl, Tys, Ops, MemVT, MMO); |
10575 | } |
10576 | |
10577 | SDValue PPCTargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, |
10578 | SelectionDAG &DAG) const { |
10579 | SDLoc dl(Op); |
10580 | // Create a stack slot that is 16-byte aligned. |
10581 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
10582 | int FrameIdx = MFI.CreateStackObject(16, Align(16), false); |
10583 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
10584 | SDValue FIdx = DAG.getFrameIndex(FrameIdx, PtrVT); |
10585 | |
10586 | // Store the input value into Value#0 of the stack slot. |
10587 | SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0), FIdx, |
10588 | MachinePointerInfo()); |
10589 | // Load it out. |
10590 | return DAG.getLoad(Op.getValueType(), dl, Store, FIdx, MachinePointerInfo()); |
10591 | } |
10592 | |
10593 | SDValue PPCTargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, |
10594 | SelectionDAG &DAG) const { |
10595 | assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT &&((Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Should only be called for ISD::INSERT_VECTOR_ELT" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::INSERT_VECTOR_ELT && \"Should only be called for ISD::INSERT_VECTOR_ELT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10596, __PRETTY_FUNCTION__)) |
10596 | "Should only be called for ISD::INSERT_VECTOR_ELT")((Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Should only be called for ISD::INSERT_VECTOR_ELT" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::INSERT_VECTOR_ELT && \"Should only be called for ISD::INSERT_VECTOR_ELT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10596, __PRETTY_FUNCTION__)); |
10597 | |
10598 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(2)); |
10599 | // We have legal lowering for constant indices but not for variable ones. |
10600 | if (!C) |
10601 | return SDValue(); |
10602 | |
10603 | EVT VT = Op.getValueType(); |
10604 | SDLoc dl(Op); |
10605 | SDValue V1 = Op.getOperand(0); |
10606 | SDValue V2 = Op.getOperand(1); |
10607 | // We can use MTVSRZ + VECINSERT for v8i16 and v16i8 types. |
10608 | if (VT == MVT::v8i16 || VT == MVT::v16i8) { |
10609 | SDValue Mtvsrz = DAG.getNode(PPCISD::MTVSRZ, dl, VT, V2); |
10610 | unsigned BytesInEachElement = VT.getVectorElementType().getSizeInBits() / 8; |
10611 | unsigned InsertAtElement = C->getZExtValue(); |
10612 | unsigned InsertAtByte = InsertAtElement * BytesInEachElement; |
10613 | if (Subtarget.isLittleEndian()) { |
10614 | InsertAtByte = (16 - BytesInEachElement) - InsertAtByte; |
10615 | } |
10616 | return DAG.getNode(PPCISD::VECINSERT, dl, VT, V1, Mtvsrz, |
10617 | DAG.getConstant(InsertAtByte, dl, MVT::i32)); |
10618 | } |
10619 | return Op; |
10620 | } |
10621 | |
10622 | SDValue PPCTargetLowering::LowerVectorLoad(SDValue Op, |
10623 | SelectionDAG &DAG) const { |
10624 | SDLoc dl(Op); |
10625 | LoadSDNode *LN = cast<LoadSDNode>(Op.getNode()); |
10626 | SDValue LoadChain = LN->getChain(); |
10627 | SDValue BasePtr = LN->getBasePtr(); |
10628 | EVT VT = Op.getValueType(); |
10629 | |
10630 | if (VT != MVT::v256i1 && VT != MVT::v512i1) |
10631 | return Op; |
10632 | |
10633 | // Type v256i1 is used for pairs and v512i1 is used for accumulators. |
10634 | // Here we create 2 or 4 v16i8 loads to load the pair or accumulator value in |
10635 | // 2 or 4 vsx registers. |
10636 | assert((VT != MVT::v512i1 || Subtarget.hasMMA()) &&(((VT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10637, __PRETTY_FUNCTION__)) |
10637 | "Type unsupported without MMA")(((VT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10637, __PRETTY_FUNCTION__)); |
10638 | assert((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&(((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10639, __PRETTY_FUNCTION__)) |
10639 | "Type unsupported without paired vector support")(((VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(VT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10639, __PRETTY_FUNCTION__)); |
10640 | Align Alignment = LN->getAlign(); |
10641 | SmallVector<SDValue, 4> Loads; |
10642 | SmallVector<SDValue, 4> LoadChains; |
10643 | unsigned NumVecs = VT.getSizeInBits() / 128; |
10644 | for (unsigned Idx = 0; Idx < NumVecs; ++Idx) { |
10645 | SDValue Load = |
10646 | DAG.getLoad(MVT::v16i8, dl, LoadChain, BasePtr, |
10647 | LN->getPointerInfo().getWithOffset(Idx * 16), |
10648 | commonAlignment(Alignment, Idx * 16), |
10649 | LN->getMemOperand()->getFlags(), LN->getAAInfo()); |
10650 | BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
10651 | DAG.getConstant(16, dl, BasePtr.getValueType())); |
10652 | Loads.push_back(Load); |
10653 | LoadChains.push_back(Load.getValue(1)); |
10654 | } |
10655 | if (Subtarget.isLittleEndian()) { |
10656 | std::reverse(Loads.begin(), Loads.end()); |
10657 | std::reverse(LoadChains.begin(), LoadChains.end()); |
10658 | } |
10659 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, LoadChains); |
10660 | SDValue Value = |
10661 | DAG.getNode(VT == MVT::v512i1 ? PPCISD::ACC_BUILD : PPCISD::PAIR_BUILD, |
10662 | dl, VT, Loads); |
10663 | SDValue RetOps[] = {Value, TF}; |
10664 | return DAG.getMergeValues(RetOps, dl); |
10665 | } |
10666 | |
10667 | SDValue PPCTargetLowering::LowerVectorStore(SDValue Op, |
10668 | SelectionDAG &DAG) const { |
10669 | SDLoc dl(Op); |
10670 | StoreSDNode *SN = cast<StoreSDNode>(Op.getNode()); |
10671 | SDValue StoreChain = SN->getChain(); |
10672 | SDValue BasePtr = SN->getBasePtr(); |
10673 | SDValue Value = SN->getValue(); |
10674 | EVT StoreVT = Value.getValueType(); |
10675 | |
10676 | if (StoreVT != MVT::v256i1 && StoreVT != MVT::v512i1) |
10677 | return Op; |
10678 | |
10679 | // Type v256i1 is used for pairs and v512i1 is used for accumulators. |
10680 | // Here we create 2 or 4 v16i8 stores to store the pair or accumulator |
10681 | // underlying registers individually. |
10682 | assert((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) &&(((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10683, __PRETTY_FUNCTION__)) |
10683 | "Type unsupported without MMA")(((StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && "Type unsupported without MMA" ) ? static_cast<void> (0) : __assert_fail ("(StoreVT != MVT::v512i1 || Subtarget.hasMMA()) && \"Type unsupported without MMA\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10683, __PRETTY_FUNCTION__)); |
10684 | assert((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) &&(((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10685, __PRETTY_FUNCTION__)) |
10685 | "Type unsupported without paired vector support")(((StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && "Type unsupported without paired vector support") ? static_cast <void> (0) : __assert_fail ("(StoreVT != MVT::v256i1 || Subtarget.pairedVectorMemops()) && \"Type unsupported without paired vector support\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10685, __PRETTY_FUNCTION__)); |
10686 | Align Alignment = SN->getAlign(); |
10687 | SmallVector<SDValue, 4> Stores; |
10688 | unsigned NumVecs = 2; |
10689 | if (StoreVT == MVT::v512i1) { |
10690 | Value = DAG.getNode(PPCISD::XXMFACC, dl, MVT::v512i1, Value); |
10691 | NumVecs = 4; |
10692 | } |
10693 | for (unsigned Idx = 0; Idx < NumVecs; ++Idx) { |
10694 | unsigned VecNum = Subtarget.isLittleEndian() ? NumVecs - 1 - Idx : Idx; |
10695 | SDValue Elt = DAG.getNode(PPCISD::EXTRACT_VSX_REG, dl, MVT::v16i8, Value, |
10696 | DAG.getConstant(VecNum, dl, MVT::i64)); |
10697 | SDValue Store = |
10698 | DAG.getStore(StoreChain, dl, Elt, BasePtr, |
10699 | SN->getPointerInfo().getWithOffset(Idx * 16), |
10700 | commonAlignment(Alignment, Idx * 16), |
10701 | SN->getMemOperand()->getFlags(), SN->getAAInfo()); |
10702 | BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
10703 | DAG.getConstant(16, dl, BasePtr.getValueType())); |
10704 | Stores.push_back(Store); |
10705 | } |
10706 | SDValue TF = DAG.getTokenFactor(dl, Stores); |
10707 | return TF; |
10708 | } |
10709 | |
10710 | SDValue PPCTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const { |
10711 | SDLoc dl(Op); |
10712 | if (Op.getValueType() == MVT::v4i32) { |
10713 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
10714 | |
10715 | SDValue Zero = getCanonicalConstSplat(0, 1, MVT::v4i32, DAG, dl); |
10716 | // +16 as shift amt. |
10717 | SDValue Neg16 = getCanonicalConstSplat(-16, 4, MVT::v4i32, DAG, dl); |
10718 | SDValue RHSSwap = // = vrlw RHS, 16 |
10719 | BuildIntrinsicOp(Intrinsic::ppc_altivec_vrlw, RHS, Neg16, DAG, dl); |
10720 | |
10721 | // Shrinkify inputs to v8i16. |
10722 | LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, LHS); |
10723 | RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHS); |
10724 | RHSSwap = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, RHSSwap); |
10725 | |
10726 | // Low parts multiplied together, generating 32-bit results (we ignore the |
10727 | // top parts). |
10728 | SDValue LoProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmulouh, |
10729 | LHS, RHS, DAG, dl, MVT::v4i32); |
10730 | |
10731 | SDValue HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmsumuhm, |
10732 | LHS, RHSSwap, Zero, DAG, dl, MVT::v4i32); |
10733 | // Shift the high parts up 16 bits. |
10734 | HiProd = BuildIntrinsicOp(Intrinsic::ppc_altivec_vslw, HiProd, |
10735 | Neg16, DAG, dl); |
10736 | return DAG.getNode(ISD::ADD, dl, MVT::v4i32, LoProd, HiProd); |
10737 | } else if (Op.getValueType() == MVT::v16i8) { |
10738 | SDValue LHS = Op.getOperand(0), RHS = Op.getOperand(1); |
10739 | bool isLittleEndian = Subtarget.isLittleEndian(); |
10740 | |
10741 | // Multiply the even 8-bit parts, producing 16-bit sums. |
10742 | SDValue EvenParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuleub, |
10743 | LHS, RHS, DAG, dl, MVT::v8i16); |
10744 | EvenParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, EvenParts); |
10745 | |
10746 | // Multiply the odd 8-bit parts, producing 16-bit sums. |
10747 | SDValue OddParts = BuildIntrinsicOp(Intrinsic::ppc_altivec_vmuloub, |
10748 | LHS, RHS, DAG, dl, MVT::v8i16); |
10749 | OddParts = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OddParts); |
10750 | |
10751 | // Merge the results together. Because vmuleub and vmuloub are |
10752 | // instructions with a big-endian bias, we must reverse the |
10753 | // element numbering and reverse the meaning of "odd" and "even" |
10754 | // when generating little endian code. |
10755 | int Ops[16]; |
10756 | for (unsigned i = 0; i != 8; ++i) { |
10757 | if (isLittleEndian) { |
10758 | Ops[i*2 ] = 2*i; |
10759 | Ops[i*2+1] = 2*i+16; |
10760 | } else { |
10761 | Ops[i*2 ] = 2*i+1; |
10762 | Ops[i*2+1] = 2*i+1+16; |
10763 | } |
10764 | } |
10765 | if (isLittleEndian) |
10766 | return DAG.getVectorShuffle(MVT::v16i8, dl, OddParts, EvenParts, Ops); |
10767 | else |
10768 | return DAG.getVectorShuffle(MVT::v16i8, dl, EvenParts, OddParts, Ops); |
10769 | } else { |
10770 | llvm_unreachable("Unknown mul to lower!")::llvm::llvm_unreachable_internal("Unknown mul to lower!", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10770); |
10771 | } |
10772 | } |
10773 | |
10774 | SDValue PPCTargetLowering::LowerABS(SDValue Op, SelectionDAG &DAG) const { |
10775 | |
10776 | assert(Op.getOpcode() == ISD::ABS && "Should only be called for ISD::ABS")((Op.getOpcode() == ISD::ABS && "Should only be called for ISD::ABS" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::ABS && \"Should only be called for ISD::ABS\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10776, __PRETTY_FUNCTION__)); |
10777 | |
10778 | EVT VT = Op.getValueType(); |
10779 | assert(VT.isVector() &&((VT.isVector() && "Only set vector abs as custom, scalar abs shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Only set vector abs as custom, scalar abs shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10780, __PRETTY_FUNCTION__)) |
10780 | "Only set vector abs as custom, scalar abs shouldn't reach here!")((VT.isVector() && "Only set vector abs as custom, scalar abs shouldn't reach here!" ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Only set vector abs as custom, scalar abs shouldn't reach here!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10780, __PRETTY_FUNCTION__)); |
10781 | assert((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10783, __PRETTY_FUNCTION__)) |
10782 | VT == MVT::v16i8) &&(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10783, __PRETTY_FUNCTION__)) |
10783 | "Unexpected vector element type!")(((VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && "Unexpected vector element type!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 || VT == MVT::v16i8) && \"Unexpected vector element type!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10783, __PRETTY_FUNCTION__)); |
10784 | assert((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) &&(((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && "Current subtarget doesn't support smax v2i64!" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && \"Current subtarget doesn't support smax v2i64!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10785, __PRETTY_FUNCTION__)) |
10785 | "Current subtarget doesn't support smax v2i64!")(((VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && "Current subtarget doesn't support smax v2i64!" ) ? static_cast<void> (0) : __assert_fail ("(VT != MVT::v2i64 || Subtarget.hasP8Altivec()) && \"Current subtarget doesn't support smax v2i64!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10785, __PRETTY_FUNCTION__)); |
10786 | |
10787 | // For vector abs, it can be lowered to: |
10788 | // abs x |
10789 | // ==> |
10790 | // y = -x |
10791 | // smax(x, y) |
10792 | |
10793 | SDLoc dl(Op); |
10794 | SDValue X = Op.getOperand(0); |
10795 | SDValue Zero = DAG.getConstant(0, dl, VT); |
10796 | SDValue Y = DAG.getNode(ISD::SUB, dl, VT, Zero, X); |
10797 | |
10798 | // SMAX patch https://reviews.llvm.org/D47332 |
10799 | // hasn't landed yet, so use intrinsic first here. |
10800 | // TODO: Should use SMAX directly once SMAX patch landed |
10801 | Intrinsic::ID BifID = Intrinsic::ppc_altivec_vmaxsw; |
10802 | if (VT == MVT::v2i64) |
10803 | BifID = Intrinsic::ppc_altivec_vmaxsd; |
10804 | else if (VT == MVT::v8i16) |
10805 | BifID = Intrinsic::ppc_altivec_vmaxsh; |
10806 | else if (VT == MVT::v16i8) |
10807 | BifID = Intrinsic::ppc_altivec_vmaxsb; |
10808 | |
10809 | return BuildIntrinsicOp(BifID, X, Y, DAG, dl, VT); |
10810 | } |
10811 | |
10812 | // Custom lowering for fpext vf32 to v2f64 |
10813 | SDValue PPCTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const { |
10814 | |
10815 | assert(Op.getOpcode() == ISD::FP_EXTEND &&((Op.getOpcode() == ISD::FP_EXTEND && "Should only be called for ISD::FP_EXTEND" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::FP_EXTEND && \"Should only be called for ISD::FP_EXTEND\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10816, __PRETTY_FUNCTION__)) |
10816 | "Should only be called for ISD::FP_EXTEND")((Op.getOpcode() == ISD::FP_EXTEND && "Should only be called for ISD::FP_EXTEND" ) ? static_cast<void> (0) : __assert_fail ("Op.getOpcode() == ISD::FP_EXTEND && \"Should only be called for ISD::FP_EXTEND\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10816, __PRETTY_FUNCTION__)); |
10817 | |
10818 | // FIXME: handle extends from half precision float vectors on P9. |
10819 | // We only want to custom lower an extend from v2f32 to v2f64. |
10820 | if (Op.getValueType() != MVT::v2f64 || |
10821 | Op.getOperand(0).getValueType() != MVT::v2f32) |
10822 | return SDValue(); |
10823 | |
10824 | SDLoc dl(Op); |
10825 | SDValue Op0 = Op.getOperand(0); |
10826 | |
10827 | switch (Op0.getOpcode()) { |
10828 | default: |
10829 | return SDValue(); |
10830 | case ISD::EXTRACT_SUBVECTOR: { |
10831 | assert(Op0.getNumOperands() == 2 &&((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10833, __PRETTY_FUNCTION__)) |
10832 | isa<ConstantSDNode>(Op0->getOperand(1)) &&((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10833, __PRETTY_FUNCTION__)) |
10833 | "Node should have 2 operands with second one being a constant!")((Op0.getNumOperands() == 2 && isa<ConstantSDNode> (Op0->getOperand(1)) && "Node should have 2 operands with second one being a constant!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getNumOperands() == 2 && isa<ConstantSDNode>(Op0->getOperand(1)) && \"Node should have 2 operands with second one being a constant!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10833, __PRETTY_FUNCTION__)); |
10834 | |
10835 | if (Op0.getOperand(0).getValueType() != MVT::v4f32) |
10836 | return SDValue(); |
10837 | |
10838 | // Custom lower is only done for high or low doubleword. |
10839 | int Idx = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue(); |
10840 | if (Idx % 2 != 0) |
10841 | return SDValue(); |
10842 | |
10843 | // Since input is v4f32, at this point Idx is either 0 or 2. |
10844 | // Shift to get the doubleword position we want. |
10845 | int DWord = Idx >> 1; |
10846 | |
10847 | // High and low word positions are different on little endian. |
10848 | if (Subtarget.isLittleEndian()) |
10849 | DWord ^= 0x1; |
10850 | |
10851 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, |
10852 | Op0.getOperand(0), DAG.getConstant(DWord, dl, MVT::i32)); |
10853 | } |
10854 | case ISD::FADD: |
10855 | case ISD::FMUL: |
10856 | case ISD::FSUB: { |
10857 | SDValue NewLoad[2]; |
10858 | for (unsigned i = 0, ie = Op0.getNumOperands(); i != ie; ++i) { |
10859 | // Ensure both input are loads. |
10860 | SDValue LdOp = Op0.getOperand(i); |
10861 | if (LdOp.getOpcode() != ISD::LOAD) |
10862 | return SDValue(); |
10863 | // Generate new load node. |
10864 | LoadSDNode *LD = cast<LoadSDNode>(LdOp); |
10865 | SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()}; |
10866 | NewLoad[i] = DAG.getMemIntrinsicNode( |
10867 | PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps, |
10868 | LD->getMemoryVT(), LD->getMemOperand()); |
10869 | } |
10870 | SDValue NewOp = |
10871 | DAG.getNode(Op0.getOpcode(), SDLoc(Op0), MVT::v4f32, NewLoad[0], |
10872 | NewLoad[1], Op0.getNode()->getFlags()); |
10873 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewOp, |
10874 | DAG.getConstant(0, dl, MVT::i32)); |
10875 | } |
10876 | case ISD::LOAD: { |
10877 | LoadSDNode *LD = cast<LoadSDNode>(Op0); |
10878 | SDValue LoadOps[] = {LD->getChain(), LD->getBasePtr()}; |
10879 | SDValue NewLd = DAG.getMemIntrinsicNode( |
10880 | PPCISD::LD_VSX_LH, dl, DAG.getVTList(MVT::v4f32, MVT::Other), LoadOps, |
10881 | LD->getMemoryVT(), LD->getMemOperand()); |
10882 | return DAG.getNode(PPCISD::FP_EXTEND_HALF, dl, MVT::v2f64, NewLd, |
10883 | DAG.getConstant(0, dl, MVT::i32)); |
10884 | } |
10885 | } |
10886 | llvm_unreachable("ERROR:Should return for all cases within swtich.")::llvm::llvm_unreachable_internal("ERROR:Should return for all cases within swtich." , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10886); |
10887 | } |
10888 | |
10889 | /// LowerOperation - Provide custom lowering hooks for some operations. |
10890 | /// |
10891 | SDValue PPCTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
10892 | switch (Op.getOpcode()) { |
10893 | default: llvm_unreachable("Wasn't expecting to be able to lower this!")::llvm::llvm_unreachable_internal("Wasn't expecting to be able to lower this!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10893); |
10894 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
10895 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
10896 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
10897 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
10898 | case ISD::JumpTable: return LowerJumpTable(Op, DAG); |
10899 | case ISD::SETCC: return LowerSETCC(Op, DAG); |
10900 | case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); |
10901 | case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); |
10902 | |
10903 | // Variable argument lowering. |
10904 | case ISD::VASTART: return LowerVASTART(Op, DAG); |
10905 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
10906 | case ISD::VACOPY: return LowerVACOPY(Op, DAG); |
10907 | |
10908 | case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG); |
10909 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); |
10910 | case ISD::GET_DYNAMIC_AREA_OFFSET: |
10911 | return LowerGET_DYNAMIC_AREA_OFFSET(Op, DAG); |
10912 | |
10913 | // Exception handling lowering. |
10914 | case ISD::EH_DWARF_CFA: return LowerEH_DWARF_CFA(Op, DAG); |
10915 | case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG); |
10916 | case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG); |
10917 | |
10918 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
10919 | case ISD::STORE: return LowerSTORE(Op, DAG); |
10920 | case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG); |
10921 | case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG); |
10922 | case ISD::STRICT_FP_TO_UINT: |
10923 | case ISD::STRICT_FP_TO_SINT: |
10924 | case ISD::FP_TO_UINT: |
10925 | case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, SDLoc(Op)); |
10926 | case ISD::STRICT_UINT_TO_FP: |
10927 | case ISD::STRICT_SINT_TO_FP: |
10928 | case ISD::UINT_TO_FP: |
10929 | case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG); |
10930 | case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG); |
10931 | |
10932 | // Lower 64-bit shifts. |
10933 | case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG); |
10934 | case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG); |
10935 | case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG); |
10936 | |
10937 | case ISD::FSHL: return LowerFunnelShift(Op, DAG); |
10938 | case ISD::FSHR: return LowerFunnelShift(Op, DAG); |
10939 | |
10940 | // Vector-related lowering. |
10941 | case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); |
10942 | case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); |
10943 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
10944 | case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG); |
10945 | case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG); |
10946 | case ISD::MUL: return LowerMUL(Op, DAG); |
10947 | case ISD::ABS: return LowerABS(Op, DAG); |
10948 | case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG); |
10949 | case ISD::ROTL: return LowerROTL(Op, DAG); |
10950 | |
10951 | // For counter-based loop handling. |
10952 | case ISD::INTRINSIC_W_CHAIN: return SDValue(); |
10953 | |
10954 | case ISD::BITCAST: return LowerBITCAST(Op, DAG); |
10955 | |
10956 | // Frame & Return address. |
10957 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); |
10958 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); |
10959 | |
10960 | case ISD::INTRINSIC_VOID: |
10961 | return LowerINTRINSIC_VOID(Op, DAG); |
10962 | case ISD::BSWAP: |
10963 | return LowerBSWAP(Op, DAG); |
10964 | case ISD::ATOMIC_CMP_SWAP: |
10965 | return LowerATOMIC_CMP_SWAP(Op, DAG); |
10966 | } |
10967 | } |
10968 | |
10969 | void PPCTargetLowering::LowerOperationWrapper(SDNode *N, |
10970 | SmallVectorImpl<SDValue> &Results, |
10971 | SelectionDAG &DAG) const { |
10972 | SDValue Res = LowerOperation(SDValue(N, 0), DAG); |
10973 | |
10974 | if (!Res.getNode()) |
10975 | return; |
10976 | |
10977 | // Take the return value as-is if original node has only one result. |
10978 | if (N->getNumValues() == 1) { |
10979 | Results.push_back(Res); |
10980 | return; |
10981 | } |
10982 | |
10983 | // New node should have the same number of results. |
10984 | assert((N->getNumValues() == Res->getNumValues()) &&(((N->getNumValues() == Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() == Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10985, __PRETTY_FUNCTION__)) |
10985 | "Lowering returned the wrong number of results!")(((N->getNumValues() == Res->getNumValues()) && "Lowering returned the wrong number of results!") ? static_cast <void> (0) : __assert_fail ("(N->getNumValues() == Res->getNumValues()) && \"Lowering returned the wrong number of results!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10985, __PRETTY_FUNCTION__)); |
10986 | |
10987 | for (unsigned i = 0; i < N->getNumValues(); ++i) |
10988 | Results.push_back(Res.getValue(i)); |
10989 | } |
10990 | |
10991 | void PPCTargetLowering::ReplaceNodeResults(SDNode *N, |
10992 | SmallVectorImpl<SDValue>&Results, |
10993 | SelectionDAG &DAG) const { |
10994 | SDLoc dl(N); |
10995 | switch (N->getOpcode()) { |
10996 | default: |
10997 | llvm_unreachable("Do not know how to custom type legalize this operation!")::llvm::llvm_unreachable_internal("Do not know how to custom type legalize this operation!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 10997); |
10998 | case ISD::READCYCLECOUNTER: { |
10999 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); |
11000 | SDValue RTB = DAG.getNode(PPCISD::READ_TIME_BASE, dl, VTs, N->getOperand(0)); |
11001 | |
11002 | Results.push_back( |
11003 | DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, RTB, RTB.getValue(1))); |
11004 | Results.push_back(RTB.getValue(2)); |
11005 | break; |
11006 | } |
11007 | case ISD::INTRINSIC_W_CHAIN: { |
11008 | if (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != |
11009 | Intrinsic::loop_decrement) |
11010 | break; |
11011 | |
11012 | assert(N->getValueType(0) == MVT::i1 &&((N->getValueType(0) == MVT::i1 && "Unexpected result type for CTR decrement intrinsic" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i1 && \"Unexpected result type for CTR decrement intrinsic\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11013, __PRETTY_FUNCTION__)) |
11013 | "Unexpected result type for CTR decrement intrinsic")((N->getValueType(0) == MVT::i1 && "Unexpected result type for CTR decrement intrinsic" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i1 && \"Unexpected result type for CTR decrement intrinsic\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11013, __PRETTY_FUNCTION__)); |
11014 | EVT SVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), |
11015 | N->getValueType(0)); |
11016 | SDVTList VTs = DAG.getVTList(SVT, MVT::Other); |
11017 | SDValue NewInt = DAG.getNode(N->getOpcode(), dl, VTs, N->getOperand(0), |
11018 | N->getOperand(1)); |
11019 | |
11020 | Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewInt)); |
11021 | Results.push_back(NewInt.getValue(1)); |
11022 | break; |
11023 | } |
11024 | case ISD::VAARG: { |
11025 | if (!Subtarget.isSVR4ABI() || Subtarget.isPPC64()) |
11026 | return; |
11027 | |
11028 | EVT VT = N->getValueType(0); |
11029 | |
11030 | if (VT == MVT::i64) { |
11031 | SDValue NewNode = LowerVAARG(SDValue(N, 1), DAG); |
11032 | |
11033 | Results.push_back(NewNode); |
11034 | Results.push_back(NewNode.getValue(1)); |
11035 | } |
11036 | return; |
11037 | } |
11038 | case ISD::STRICT_FP_TO_SINT: |
11039 | case ISD::STRICT_FP_TO_UINT: |
11040 | case ISD::FP_TO_SINT: |
11041 | case ISD::FP_TO_UINT: |
11042 | // LowerFP_TO_INT() can only handle f32 and f64. |
11043 | if (N->getOperand(N->isStrictFPOpcode() ? 1 : 0).getValueType() == |
11044 | MVT::ppcf128) |
11045 | return; |
11046 | Results.push_back(LowerFP_TO_INT(SDValue(N, 0), DAG, dl)); |
11047 | return; |
11048 | case ISD::TRUNCATE: { |
11049 | if (!N->getValueType(0).isVector()) |
11050 | return; |
11051 | SDValue Lowered = LowerTRUNCATEVector(SDValue(N, 0), DAG); |
11052 | if (Lowered) |
11053 | Results.push_back(Lowered); |
11054 | return; |
11055 | } |
11056 | case ISD::FSHL: |
11057 | case ISD::FSHR: |
11058 | // Don't handle funnel shifts here. |
11059 | return; |
11060 | case ISD::BITCAST: |
11061 | // Don't handle bitcast here. |
11062 | return; |
11063 | case ISD::FP_EXTEND: |
11064 | SDValue Lowered = LowerFP_EXTEND(SDValue(N, 0), DAG); |
11065 | if (Lowered) |
11066 | Results.push_back(Lowered); |
11067 | return; |
11068 | } |
11069 | } |
11070 | |
11071 | //===----------------------------------------------------------------------===// |
11072 | // Other Lowering Code |
11073 | //===----------------------------------------------------------------------===// |
11074 | |
11075 | static Instruction* callIntrinsic(IRBuilder<> &Builder, Intrinsic::ID Id) { |
11076 | Module *M = Builder.GetInsertBlock()->getParent()->getParent(); |
11077 | Function *Func = Intrinsic::getDeclaration(M, Id); |
11078 | return Builder.CreateCall(Func, {}); |
11079 | } |
11080 | |
11081 | // The mappings for emitLeading/TrailingFence is taken from |
11082 | // http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html |
11083 | Instruction *PPCTargetLowering::emitLeadingFence(IRBuilder<> &Builder, |
11084 | Instruction *Inst, |
11085 | AtomicOrdering Ord) const { |
11086 | if (Ord == AtomicOrdering::SequentiallyConsistent) |
11087 | return callIntrinsic(Builder, Intrinsic::ppc_sync); |
11088 | if (isReleaseOrStronger(Ord)) |
11089 | return callIntrinsic(Builder, Intrinsic::ppc_lwsync); |
11090 | return nullptr; |
11091 | } |
11092 | |
11093 | Instruction *PPCTargetLowering::emitTrailingFence(IRBuilder<> &Builder, |
11094 | Instruction *Inst, |
11095 | AtomicOrdering Ord) const { |
11096 | if (Inst->hasAtomicLoad() && isAcquireOrStronger(Ord)) { |
11097 | // See http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and |
11098 | // http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html |
11099 | // and http://www.cl.cam.ac.uk/~pes20/cppppc/ for justification. |
11100 | if (isa<LoadInst>(Inst) && Subtarget.isPPC64()) |
11101 | return Builder.CreateCall( |
11102 | Intrinsic::getDeclaration( |
11103 | Builder.GetInsertBlock()->getParent()->getParent(), |
11104 | Intrinsic::ppc_cfence, {Inst->getType()}), |
11105 | {Inst}); |
11106 | // FIXME: Can use isync for rmw operation. |
11107 | return callIntrinsic(Builder, Intrinsic::ppc_lwsync); |
11108 | } |
11109 | return nullptr; |
11110 | } |
11111 | |
11112 | MachineBasicBlock * |
11113 | PPCTargetLowering::EmitAtomicBinary(MachineInstr &MI, MachineBasicBlock *BB, |
11114 | unsigned AtomicSize, |
11115 | unsigned BinOpcode, |
11116 | unsigned CmpOpcode, |
11117 | unsigned CmpPred) const { |
11118 | // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. |
11119 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11120 | |
11121 | auto LoadMnemonic = PPC::LDARX; |
11122 | auto StoreMnemonic = PPC::STDCX; |
11123 | switch (AtomicSize) { |
11124 | default: |
11125 | llvm_unreachable("Unexpected size of atomic entity")::llvm::llvm_unreachable_internal("Unexpected size of atomic entity" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11125); |
11126 | case 1: |
11127 | LoadMnemonic = PPC::LBARX; |
11128 | StoreMnemonic = PPC::STBCX; |
11129 | assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4")((Subtarget.hasPartwordAtomics() && "Call this only with size >=4" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"Call this only with size >=4\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11129, __PRETTY_FUNCTION__)); |
11130 | break; |
11131 | case 2: |
11132 | LoadMnemonic = PPC::LHARX; |
11133 | StoreMnemonic = PPC::STHCX; |
11134 | assert(Subtarget.hasPartwordAtomics() && "Call this only with size >=4")((Subtarget.hasPartwordAtomics() && "Call this only with size >=4" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"Call this only with size >=4\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11134, __PRETTY_FUNCTION__)); |
11135 | break; |
11136 | case 4: |
11137 | LoadMnemonic = PPC::LWARX; |
11138 | StoreMnemonic = PPC::STWCX; |
11139 | break; |
11140 | case 8: |
11141 | LoadMnemonic = PPC::LDARX; |
11142 | StoreMnemonic = PPC::STDCX; |
11143 | break; |
11144 | } |
11145 | |
11146 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11147 | MachineFunction *F = BB->getParent(); |
11148 | MachineFunction::iterator It = ++BB->getIterator(); |
11149 | |
11150 | Register dest = MI.getOperand(0).getReg(); |
11151 | Register ptrA = MI.getOperand(1).getReg(); |
11152 | Register ptrB = MI.getOperand(2).getReg(); |
11153 | Register incr = MI.getOperand(3).getReg(); |
11154 | DebugLoc dl = MI.getDebugLoc(); |
11155 | |
11156 | MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11157 | MachineBasicBlock *loop2MBB = |
11158 | CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr; |
11159 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11160 | F->insert(It, loopMBB); |
11161 | if (CmpOpcode) |
11162 | F->insert(It, loop2MBB); |
11163 | F->insert(It, exitMBB); |
11164 | exitMBB->splice(exitMBB->begin(), BB, |
11165 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11166 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
11167 | |
11168 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
11169 | Register TmpReg = (!BinOpcode) ? incr : |
11170 | RegInfo.createVirtualRegister( AtomicSize == 8 ? &PPC::G8RCRegClass |
11171 | : &PPC::GPRCRegClass); |
11172 | |
11173 | // thisMBB: |
11174 | // ... |
11175 | // fallthrough --> loopMBB |
11176 | BB->addSuccessor(loopMBB); |
11177 | |
11178 | // loopMBB: |
11179 | // l[wd]arx dest, ptr |
11180 | // add r0, dest, incr |
11181 | // st[wd]cx. r0, ptr |
11182 | // bne- loopMBB |
11183 | // fallthrough --> exitMBB |
11184 | |
11185 | // For max/min... |
11186 | // loopMBB: |
11187 | // l[wd]arx dest, ptr |
11188 | // cmpl?[wd] incr, dest |
11189 | // bgt exitMBB |
11190 | // loop2MBB: |
11191 | // st[wd]cx. dest, ptr |
11192 | // bne- loopMBB |
11193 | // fallthrough --> exitMBB |
11194 | |
11195 | BB = loopMBB; |
11196 | BuildMI(BB, dl, TII->get(LoadMnemonic), dest) |
11197 | .addReg(ptrA).addReg(ptrB); |
11198 | if (BinOpcode) |
11199 | BuildMI(BB, dl, TII->get(BinOpcode), TmpReg).addReg(incr).addReg(dest); |
11200 | if (CmpOpcode) { |
11201 | // Signed comparisons of byte or halfword values must be sign-extended. |
11202 | if (CmpOpcode == PPC::CMPW && AtomicSize < 4) { |
11203 | Register ExtReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass); |
11204 | BuildMI(BB, dl, TII->get(AtomicSize == 1 ? PPC::EXTSB : PPC::EXTSH), |
11205 | ExtReg).addReg(dest); |
11206 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11207 | .addReg(incr).addReg(ExtReg); |
11208 | } else |
11209 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11210 | .addReg(incr).addReg(dest); |
11211 | |
11212 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11213 | .addImm(CmpPred).addReg(PPC::CR0).addMBB(exitMBB); |
11214 | BB->addSuccessor(loop2MBB); |
11215 | BB->addSuccessor(exitMBB); |
11216 | BB = loop2MBB; |
11217 | } |
11218 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
11219 | .addReg(TmpReg).addReg(ptrA).addReg(ptrB); |
11220 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11221 | .addImm(PPC::PRED_NE).addReg(PPC::CR0).addMBB(loopMBB); |
11222 | BB->addSuccessor(loopMBB); |
11223 | BB->addSuccessor(exitMBB); |
11224 | |
11225 | // exitMBB: |
11226 | // ... |
11227 | BB = exitMBB; |
11228 | return BB; |
11229 | } |
11230 | |
11231 | MachineBasicBlock *PPCTargetLowering::EmitPartwordAtomicBinary( |
11232 | MachineInstr &MI, MachineBasicBlock *BB, |
11233 | bool is8bit, // operation |
11234 | unsigned BinOpcode, unsigned CmpOpcode, unsigned CmpPred) const { |
11235 | // If we support part-word atomic mnemonics, just use them |
11236 | if (Subtarget.hasPartwordAtomics()) |
11237 | return EmitAtomicBinary(MI, BB, is8bit ? 1 : 2, BinOpcode, CmpOpcode, |
11238 | CmpPred); |
11239 | |
11240 | // This also handles ATOMIC_SWAP, indicated by BinOpcode==0. |
11241 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11242 | // In 64 bit mode we have to use 64 bits for addresses, even though the |
11243 | // lwarx/stwcx are 32 bits. With the 32-bit atomics we can use address |
11244 | // registers without caring whether they're 32 or 64, but here we're |
11245 | // doing actual arithmetic on the addresses. |
11246 | bool is64bit = Subtarget.isPPC64(); |
11247 | bool isLittleEndian = Subtarget.isLittleEndian(); |
11248 | unsigned ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO; |
11249 | |
11250 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11251 | MachineFunction *F = BB->getParent(); |
11252 | MachineFunction::iterator It = ++BB->getIterator(); |
11253 | |
11254 | Register dest = MI.getOperand(0).getReg(); |
11255 | Register ptrA = MI.getOperand(1).getReg(); |
11256 | Register ptrB = MI.getOperand(2).getReg(); |
11257 | Register incr = MI.getOperand(3).getReg(); |
11258 | DebugLoc dl = MI.getDebugLoc(); |
11259 | |
11260 | MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11261 | MachineBasicBlock *loop2MBB = |
11262 | CmpOpcode ? F->CreateMachineBasicBlock(LLVM_BB) : nullptr; |
11263 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11264 | F->insert(It, loopMBB); |
11265 | if (CmpOpcode) |
11266 | F->insert(It, loop2MBB); |
11267 | F->insert(It, exitMBB); |
11268 | exitMBB->splice(exitMBB->begin(), BB, |
11269 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11270 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
11271 | |
11272 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
11273 | const TargetRegisterClass *RC = |
11274 | is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
11275 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
11276 | |
11277 | Register PtrReg = RegInfo.createVirtualRegister(RC); |
11278 | Register Shift1Reg = RegInfo.createVirtualRegister(GPRC); |
11279 | Register ShiftReg = |
11280 | isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC); |
11281 | Register Incr2Reg = RegInfo.createVirtualRegister(GPRC); |
11282 | Register MaskReg = RegInfo.createVirtualRegister(GPRC); |
11283 | Register Mask2Reg = RegInfo.createVirtualRegister(GPRC); |
11284 | Register Mask3Reg = RegInfo.createVirtualRegister(GPRC); |
11285 | Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC); |
11286 | Register Tmp3Reg = RegInfo.createVirtualRegister(GPRC); |
11287 | Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC); |
11288 | Register TmpDestReg = RegInfo.createVirtualRegister(GPRC); |
11289 | Register Ptr1Reg; |
11290 | Register TmpReg = |
11291 | (!BinOpcode) ? Incr2Reg : RegInfo.createVirtualRegister(GPRC); |
11292 | |
11293 | // thisMBB: |
11294 | // ... |
11295 | // fallthrough --> loopMBB |
11296 | BB->addSuccessor(loopMBB); |
11297 | |
11298 | // The 4-byte load must be aligned, while a char or short may be |
11299 | // anywhere in the word. Hence all this nasty bookkeeping code. |
11300 | // add ptr1, ptrA, ptrB [copy if ptrA==0] |
11301 | // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27] |
11302 | // xori shift, shift1, 24 [16] |
11303 | // rlwinm ptr, ptr1, 0, 0, 29 |
11304 | // slw incr2, incr, shift |
11305 | // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535] |
11306 | // slw mask, mask2, shift |
11307 | // loopMBB: |
11308 | // lwarx tmpDest, ptr |
11309 | // add tmp, tmpDest, incr2 |
11310 | // andc tmp2, tmpDest, mask |
11311 | // and tmp3, tmp, mask |
11312 | // or tmp4, tmp3, tmp2 |
11313 | // stwcx. tmp4, ptr |
11314 | // bne- loopMBB |
11315 | // fallthrough --> exitMBB |
11316 | // srw dest, tmpDest, shift |
11317 | if (ptrA != ZeroReg) { |
11318 | Ptr1Reg = RegInfo.createVirtualRegister(RC); |
11319 | BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg) |
11320 | .addReg(ptrA) |
11321 | .addReg(ptrB); |
11322 | } else { |
11323 | Ptr1Reg = ptrB; |
11324 | } |
11325 | // We need use 32-bit subregister to avoid mismatch register class in 64-bit |
11326 | // mode. |
11327 | BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg) |
11328 | .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0) |
11329 | .addImm(3) |
11330 | .addImm(27) |
11331 | .addImm(is8bit ? 28 : 27); |
11332 | if (!isLittleEndian) |
11333 | BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg) |
11334 | .addReg(Shift1Reg) |
11335 | .addImm(is8bit ? 24 : 16); |
11336 | if (is64bit) |
11337 | BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg) |
11338 | .addReg(Ptr1Reg) |
11339 | .addImm(0) |
11340 | .addImm(61); |
11341 | else |
11342 | BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg) |
11343 | .addReg(Ptr1Reg) |
11344 | .addImm(0) |
11345 | .addImm(0) |
11346 | .addImm(29); |
11347 | BuildMI(BB, dl, TII->get(PPC::SLW), Incr2Reg).addReg(incr).addReg(ShiftReg); |
11348 | if (is8bit) |
11349 | BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255); |
11350 | else { |
11351 | BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0); |
11352 | BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg) |
11353 | .addReg(Mask3Reg) |
11354 | .addImm(65535); |
11355 | } |
11356 | BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg) |
11357 | .addReg(Mask2Reg) |
11358 | .addReg(ShiftReg); |
11359 | |
11360 | BB = loopMBB; |
11361 | BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg) |
11362 | .addReg(ZeroReg) |
11363 | .addReg(PtrReg); |
11364 | if (BinOpcode) |
11365 | BuildMI(BB, dl, TII->get(BinOpcode), TmpReg) |
11366 | .addReg(Incr2Reg) |
11367 | .addReg(TmpDestReg); |
11368 | BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg) |
11369 | .addReg(TmpDestReg) |
11370 | .addReg(MaskReg); |
11371 | BuildMI(BB, dl, TII->get(PPC::AND), Tmp3Reg).addReg(TmpReg).addReg(MaskReg); |
11372 | if (CmpOpcode) { |
11373 | // For unsigned comparisons, we can directly compare the shifted values. |
11374 | // For signed comparisons we shift and sign extend. |
11375 | Register SReg = RegInfo.createVirtualRegister(GPRC); |
11376 | BuildMI(BB, dl, TII->get(PPC::AND), SReg) |
11377 | .addReg(TmpDestReg) |
11378 | .addReg(MaskReg); |
11379 | unsigned ValueReg = SReg; |
11380 | unsigned CmpReg = Incr2Reg; |
11381 | if (CmpOpcode == PPC::CMPW) { |
11382 | ValueReg = RegInfo.createVirtualRegister(GPRC); |
11383 | BuildMI(BB, dl, TII->get(PPC::SRW), ValueReg) |
11384 | .addReg(SReg) |
11385 | .addReg(ShiftReg); |
11386 | Register ValueSReg = RegInfo.createVirtualRegister(GPRC); |
11387 | BuildMI(BB, dl, TII->get(is8bit ? PPC::EXTSB : PPC::EXTSH), ValueSReg) |
11388 | .addReg(ValueReg); |
11389 | ValueReg = ValueSReg; |
11390 | CmpReg = incr; |
11391 | } |
11392 | BuildMI(BB, dl, TII->get(CmpOpcode), PPC::CR0) |
11393 | .addReg(CmpReg) |
11394 | .addReg(ValueReg); |
11395 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11396 | .addImm(CmpPred) |
11397 | .addReg(PPC::CR0) |
11398 | .addMBB(exitMBB); |
11399 | BB->addSuccessor(loop2MBB); |
11400 | BB->addSuccessor(exitMBB); |
11401 | BB = loop2MBB; |
11402 | } |
11403 | BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg).addReg(Tmp3Reg).addReg(Tmp2Reg); |
11404 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
11405 | .addReg(Tmp4Reg) |
11406 | .addReg(ZeroReg) |
11407 | .addReg(PtrReg); |
11408 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11409 | .addImm(PPC::PRED_NE) |
11410 | .addReg(PPC::CR0) |
11411 | .addMBB(loopMBB); |
11412 | BB->addSuccessor(loopMBB); |
11413 | BB->addSuccessor(exitMBB); |
11414 | |
11415 | // exitMBB: |
11416 | // ... |
11417 | BB = exitMBB; |
11418 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest) |
11419 | .addReg(TmpDestReg) |
11420 | .addReg(ShiftReg); |
11421 | return BB; |
11422 | } |
11423 | |
11424 | llvm::MachineBasicBlock * |
11425 | PPCTargetLowering::emitEHSjLjSetJmp(MachineInstr &MI, |
11426 | MachineBasicBlock *MBB) const { |
11427 | DebugLoc DL = MI.getDebugLoc(); |
11428 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11429 | const PPCRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
11430 | |
11431 | MachineFunction *MF = MBB->getParent(); |
11432 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11433 | |
11434 | const BasicBlock *BB = MBB->getBasicBlock(); |
11435 | MachineFunction::iterator I = ++MBB->getIterator(); |
11436 | |
11437 | Register DstReg = MI.getOperand(0).getReg(); |
11438 | const TargetRegisterClass *RC = MRI.getRegClass(DstReg); |
11439 | assert(TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!")((TRI->isTypeLegalForClass(*RC, MVT::i32) && "Invalid destination!" ) ? static_cast<void> (0) : __assert_fail ("TRI->isTypeLegalForClass(*RC, MVT::i32) && \"Invalid destination!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11439, __PRETTY_FUNCTION__)); |
11440 | Register mainDstReg = MRI.createVirtualRegister(RC); |
11441 | Register restoreDstReg = MRI.createVirtualRegister(RC); |
11442 | |
11443 | MVT PVT = getPointerTy(MF->getDataLayout()); |
11444 | assert((PVT == MVT::i64 || PVT == MVT::i32) &&(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11445, __PRETTY_FUNCTION__)) |
11445 | "Invalid Pointer Size!")(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11445, __PRETTY_FUNCTION__)); |
11446 | // For v = setjmp(buf), we generate |
11447 | // |
11448 | // thisMBB: |
11449 | // SjLjSetup mainMBB |
11450 | // bl mainMBB |
11451 | // v_restore = 1 |
11452 | // b sinkMBB |
11453 | // |
11454 | // mainMBB: |
11455 | // buf[LabelOffset] = LR |
11456 | // v_main = 0 |
11457 | // |
11458 | // sinkMBB: |
11459 | // v = phi(main, restore) |
11460 | // |
11461 | |
11462 | MachineBasicBlock *thisMBB = MBB; |
11463 | MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB); |
11464 | MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB); |
11465 | MF->insert(I, mainMBB); |
11466 | MF->insert(I, sinkMBB); |
11467 | |
11468 | MachineInstrBuilder MIB; |
11469 | |
11470 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
11471 | sinkMBB->splice(sinkMBB->begin(), MBB, |
11472 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
11473 | sinkMBB->transferSuccessorsAndUpdatePHIs(MBB); |
11474 | |
11475 | // Note that the structure of the jmp_buf used here is not compatible |
11476 | // with that used by libc, and is not designed to be. Specifically, it |
11477 | // stores only those 'reserved' registers that LLVM does not otherwise |
11478 | // understand how to spill. Also, by convention, by the time this |
11479 | // intrinsic is called, Clang has already stored the frame address in the |
11480 | // first slot of the buffer and stack address in the third. Following the |
11481 | // X86 target code, we'll store the jump address in the second slot. We also |
11482 | // need to save the TOC pointer (R2) to handle jumps between shared |
11483 | // libraries, and that will be stored in the fourth slot. The thread |
11484 | // identifier (R13) is not affected. |
11485 | |
11486 | // thisMBB: |
11487 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
11488 | const int64_t TOCOffset = 3 * PVT.getStoreSize(); |
11489 | const int64_t BPOffset = 4 * PVT.getStoreSize(); |
11490 | |
11491 | // Prepare IP either in reg. |
11492 | const TargetRegisterClass *PtrRC = getRegClassFor(PVT); |
11493 | Register LabelReg = MRI.createVirtualRegister(PtrRC); |
11494 | Register BufReg = MI.getOperand(1).getReg(); |
11495 | |
11496 | if (Subtarget.is64BitELFABI()) { |
11497 | setUsesTOCBasePtr(*MBB->getParent()); |
11498 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::STD)) |
11499 | .addReg(PPC::X2) |
11500 | .addImm(TOCOffset) |
11501 | .addReg(BufReg) |
11502 | .cloneMemRefs(MI); |
11503 | } |
11504 | |
11505 | // Naked functions never have a base pointer, and so we use r1. For all |
11506 | // other functions, this decision must be delayed until during PEI. |
11507 | unsigned BaseReg; |
11508 | if (MF->getFunction().hasFnAttribute(Attribute::Naked)) |
11509 | BaseReg = Subtarget.isPPC64() ? PPC::X1 : PPC::R1; |
11510 | else |
11511 | BaseReg = Subtarget.isPPC64() ? PPC::BP8 : PPC::BP; |
11512 | |
11513 | MIB = BuildMI(*thisMBB, MI, DL, |
11514 | TII->get(Subtarget.isPPC64() ? PPC::STD : PPC::STW)) |
11515 | .addReg(BaseReg) |
11516 | .addImm(BPOffset) |
11517 | .addReg(BufReg) |
11518 | .cloneMemRefs(MI); |
11519 | |
11520 | // Setup |
11521 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::BCLalways)).addMBB(mainMBB); |
11522 | MIB.addRegMask(TRI->getNoPreservedMask()); |
11523 | |
11524 | BuildMI(*thisMBB, MI, DL, TII->get(PPC::LI), restoreDstReg).addImm(1); |
11525 | |
11526 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::EH_SjLj_Setup)) |
11527 | .addMBB(mainMBB); |
11528 | MIB = BuildMI(*thisMBB, MI, DL, TII->get(PPC::B)).addMBB(sinkMBB); |
11529 | |
11530 | thisMBB->addSuccessor(mainMBB, BranchProbability::getZero()); |
11531 | thisMBB->addSuccessor(sinkMBB, BranchProbability::getOne()); |
11532 | |
11533 | // mainMBB: |
11534 | // mainDstReg = 0 |
11535 | MIB = |
11536 | BuildMI(mainMBB, DL, |
11537 | TII->get(Subtarget.isPPC64() ? PPC::MFLR8 : PPC::MFLR), LabelReg); |
11538 | |
11539 | // Store IP |
11540 | if (Subtarget.isPPC64()) { |
11541 | MIB = BuildMI(mainMBB, DL, TII->get(PPC::STD)) |
11542 | .addReg(LabelReg) |
11543 | .addImm(LabelOffset) |
11544 | .addReg(BufReg); |
11545 | } else { |
11546 | MIB = BuildMI(mainMBB, DL, TII->get(PPC::STW)) |
11547 | .addReg(LabelReg) |
11548 | .addImm(LabelOffset) |
11549 | .addReg(BufReg); |
11550 | } |
11551 | MIB.cloneMemRefs(MI); |
11552 | |
11553 | BuildMI(mainMBB, DL, TII->get(PPC::LI), mainDstReg).addImm(0); |
11554 | mainMBB->addSuccessor(sinkMBB); |
11555 | |
11556 | // sinkMBB: |
11557 | BuildMI(*sinkMBB, sinkMBB->begin(), DL, |
11558 | TII->get(PPC::PHI), DstReg) |
11559 | .addReg(mainDstReg).addMBB(mainMBB) |
11560 | .addReg(restoreDstReg).addMBB(thisMBB); |
11561 | |
11562 | MI.eraseFromParent(); |
11563 | return sinkMBB; |
11564 | } |
11565 | |
11566 | MachineBasicBlock * |
11567 | PPCTargetLowering::emitEHSjLjLongJmp(MachineInstr &MI, |
11568 | MachineBasicBlock *MBB) const { |
11569 | DebugLoc DL = MI.getDebugLoc(); |
11570 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11571 | |
11572 | MachineFunction *MF = MBB->getParent(); |
11573 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11574 | |
11575 | MVT PVT = getPointerTy(MF->getDataLayout()); |
11576 | assert((PVT == MVT::i64 || PVT == MVT::i32) &&(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11577, __PRETTY_FUNCTION__)) |
11577 | "Invalid Pointer Size!")(((PVT == MVT::i64 || PVT == MVT::i32) && "Invalid Pointer Size!" ) ? static_cast<void> (0) : __assert_fail ("(PVT == MVT::i64 || PVT == MVT::i32) && \"Invalid Pointer Size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11577, __PRETTY_FUNCTION__)); |
11578 | |
11579 | const TargetRegisterClass *RC = |
11580 | (PVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
11581 | Register Tmp = MRI.createVirtualRegister(RC); |
11582 | // Since FP is only updated here but NOT referenced, it's treated as GPR. |
11583 | unsigned FP = (PVT == MVT::i64) ? PPC::X31 : PPC::R31; |
11584 | unsigned SP = (PVT == MVT::i64) ? PPC::X1 : PPC::R1; |
11585 | unsigned BP = |
11586 | (PVT == MVT::i64) |
11587 | ? PPC::X30 |
11588 | : (Subtarget.isSVR4ABI() && isPositionIndependent() ? PPC::R29 |
11589 | : PPC::R30); |
11590 | |
11591 | MachineInstrBuilder MIB; |
11592 | |
11593 | const int64_t LabelOffset = 1 * PVT.getStoreSize(); |
11594 | const int64_t SPOffset = 2 * PVT.getStoreSize(); |
11595 | const int64_t TOCOffset = 3 * PVT.getStoreSize(); |
11596 | const int64_t BPOffset = 4 * PVT.getStoreSize(); |
11597 | |
11598 | Register BufReg = MI.getOperand(0).getReg(); |
11599 | |
11600 | // Reload FP (the jumped-to function may not have had a |
11601 | // frame pointer, and if so, then its r31 will be restored |
11602 | // as necessary). |
11603 | if (PVT == MVT::i64) { |
11604 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), FP) |
11605 | .addImm(0) |
11606 | .addReg(BufReg); |
11607 | } else { |
11608 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), FP) |
11609 | .addImm(0) |
11610 | .addReg(BufReg); |
11611 | } |
11612 | MIB.cloneMemRefs(MI); |
11613 | |
11614 | // Reload IP |
11615 | if (PVT == MVT::i64) { |
11616 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), Tmp) |
11617 | .addImm(LabelOffset) |
11618 | .addReg(BufReg); |
11619 | } else { |
11620 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), Tmp) |
11621 | .addImm(LabelOffset) |
11622 | .addReg(BufReg); |
11623 | } |
11624 | MIB.cloneMemRefs(MI); |
11625 | |
11626 | // Reload SP |
11627 | if (PVT == MVT::i64) { |
11628 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), SP) |
11629 | .addImm(SPOffset) |
11630 | .addReg(BufReg); |
11631 | } else { |
11632 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), SP) |
11633 | .addImm(SPOffset) |
11634 | .addReg(BufReg); |
11635 | } |
11636 | MIB.cloneMemRefs(MI); |
11637 | |
11638 | // Reload BP |
11639 | if (PVT == MVT::i64) { |
11640 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), BP) |
11641 | .addImm(BPOffset) |
11642 | .addReg(BufReg); |
11643 | } else { |
11644 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LWZ), BP) |
11645 | .addImm(BPOffset) |
11646 | .addReg(BufReg); |
11647 | } |
11648 | MIB.cloneMemRefs(MI); |
11649 | |
11650 | // Reload TOC |
11651 | if (PVT == MVT::i64 && Subtarget.isSVR4ABI()) { |
11652 | setUsesTOCBasePtr(*MBB->getParent()); |
11653 | MIB = BuildMI(*MBB, MI, DL, TII->get(PPC::LD), PPC::X2) |
11654 | .addImm(TOCOffset) |
11655 | .addReg(BufReg) |
11656 | .cloneMemRefs(MI); |
11657 | } |
11658 | |
11659 | // Jump |
11660 | BuildMI(*MBB, MI, DL, |
11661 | TII->get(PVT == MVT::i64 ? PPC::MTCTR8 : PPC::MTCTR)).addReg(Tmp); |
11662 | BuildMI(*MBB, MI, DL, TII->get(PVT == MVT::i64 ? PPC::BCTR8 : PPC::BCTR)); |
11663 | |
11664 | MI.eraseFromParent(); |
11665 | return MBB; |
11666 | } |
11667 | |
11668 | bool PPCTargetLowering::hasInlineStackProbe(MachineFunction &MF) const { |
11669 | // If the function specifically requests inline stack probes, emit them. |
11670 | if (MF.getFunction().hasFnAttribute("probe-stack")) |
11671 | return MF.getFunction().getFnAttribute("probe-stack").getValueAsString() == |
11672 | "inline-asm"; |
11673 | return false; |
11674 | } |
11675 | |
11676 | unsigned PPCTargetLowering::getStackProbeSize(MachineFunction &MF) const { |
11677 | const TargetFrameLowering *TFI = Subtarget.getFrameLowering(); |
11678 | unsigned StackAlign = TFI->getStackAlignment(); |
11679 | assert(StackAlign >= 1 && isPowerOf2_32(StackAlign) &&((StackAlign >= 1 && isPowerOf2_32(StackAlign) && "Unexpected stack alignment") ? static_cast<void> (0) : __assert_fail ("StackAlign >= 1 && isPowerOf2_32(StackAlign) && \"Unexpected stack alignment\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11680, __PRETTY_FUNCTION__)) |
11680 | "Unexpected stack alignment")((StackAlign >= 1 && isPowerOf2_32(StackAlign) && "Unexpected stack alignment") ? static_cast<void> (0) : __assert_fail ("StackAlign >= 1 && isPowerOf2_32(StackAlign) && \"Unexpected stack alignment\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11680, __PRETTY_FUNCTION__)); |
11681 | // The default stack probe size is 4096 if the function has no |
11682 | // stack-probe-size attribute. |
11683 | unsigned StackProbeSize = 4096; |
11684 | const Function &Fn = MF.getFunction(); |
11685 | if (Fn.hasFnAttribute("stack-probe-size")) |
11686 | Fn.getFnAttribute("stack-probe-size") |
11687 | .getValueAsString() |
11688 | .getAsInteger(0, StackProbeSize); |
11689 | // Round down to the stack alignment. |
11690 | StackProbeSize &= ~(StackAlign - 1); |
11691 | return StackProbeSize ? StackProbeSize : StackAlign; |
11692 | } |
11693 | |
11694 | // Lower dynamic stack allocation with probing. `emitProbedAlloca` is splitted |
11695 | // into three phases. In the first phase, it uses pseudo instruction |
11696 | // PREPARE_PROBED_ALLOCA to get the future result of actual FramePointer and |
11697 | // FinalStackPtr. In the second phase, it generates a loop for probing blocks. |
11698 | // At last, it uses pseudo instruction DYNAREAOFFSET to get the future result of |
11699 | // MaxCallFrameSize so that it can calculate correct data area pointer. |
11700 | MachineBasicBlock * |
11701 | PPCTargetLowering::emitProbedAlloca(MachineInstr &MI, |
11702 | MachineBasicBlock *MBB) const { |
11703 | const bool isPPC64 = Subtarget.isPPC64(); |
11704 | MachineFunction *MF = MBB->getParent(); |
11705 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11706 | DebugLoc DL = MI.getDebugLoc(); |
11707 | const unsigned ProbeSize = getStackProbeSize(*MF); |
11708 | const BasicBlock *ProbedBB = MBB->getBasicBlock(); |
11709 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
11710 | // The CFG of probing stack looks as |
11711 | // +-----+ |
11712 | // | MBB | |
11713 | // +--+--+ |
11714 | // | |
11715 | // +----v----+ |
11716 | // +--->+ TestMBB +---+ |
11717 | // | +----+----+ | |
11718 | // | | | |
11719 | // | +-----v----+ | |
11720 | // +---+ BlockMBB | | |
11721 | // +----------+ | |
11722 | // | |
11723 | // +---------+ | |
11724 | // | TailMBB +<--+ |
11725 | // +---------+ |
11726 | // In MBB, calculate previous frame pointer and final stack pointer. |
11727 | // In TestMBB, test if sp is equal to final stack pointer, if so, jump to |
11728 | // TailMBB. In BlockMBB, update the sp atomically and jump back to TestMBB. |
11729 | // TailMBB is spliced via \p MI. |
11730 | MachineBasicBlock *TestMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11731 | MachineBasicBlock *TailMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11732 | MachineBasicBlock *BlockMBB = MF->CreateMachineBasicBlock(ProbedBB); |
11733 | |
11734 | MachineFunction::iterator MBBIter = ++MBB->getIterator(); |
11735 | MF->insert(MBBIter, TestMBB); |
11736 | MF->insert(MBBIter, BlockMBB); |
11737 | MF->insert(MBBIter, TailMBB); |
11738 | |
11739 | const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; |
11740 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
11741 | |
11742 | Register DstReg = MI.getOperand(0).getReg(); |
11743 | Register NegSizeReg = MI.getOperand(1).getReg(); |
11744 | Register SPReg = isPPC64 ? PPC::X1 : PPC::R1; |
11745 | Register FinalStackPtr = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11746 | Register FramePointer = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11747 | Register ActualNegSizeReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11748 | |
11749 | // Since value of NegSizeReg might be realigned in prologepilog, insert a |
11750 | // PREPARE_PROBED_ALLOCA pseudo instruction to get actual FramePointer and |
11751 | // NegSize. |
11752 | unsigned ProbeOpc; |
11753 | if (!MRI.hasOneNonDBGUse(NegSizeReg)) |
11754 | ProbeOpc = |
11755 | isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_64 : PPC::PREPARE_PROBED_ALLOCA_32; |
11756 | else |
11757 | // By introducing PREPARE_PROBED_ALLOCA_NEGSIZE_OPT, ActualNegSizeReg |
11758 | // and NegSizeReg will be allocated in the same phyreg to avoid |
11759 | // redundant copy when NegSizeReg has only one use which is current MI and |
11760 | // will be replaced by PREPARE_PROBED_ALLOCA then. |
11761 | ProbeOpc = isPPC64 ? PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_64 |
11762 | : PPC::PREPARE_PROBED_ALLOCA_NEGSIZE_SAME_REG_32; |
11763 | BuildMI(*MBB, {MI}, DL, TII->get(ProbeOpc), FramePointer) |
11764 | .addDef(ActualNegSizeReg) |
11765 | .addReg(NegSizeReg) |
11766 | .add(MI.getOperand(2)) |
11767 | .add(MI.getOperand(3)); |
11768 | |
11769 | // Calculate final stack pointer, which equals to SP + ActualNegSize. |
11770 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4), |
11771 | FinalStackPtr) |
11772 | .addReg(SPReg) |
11773 | .addReg(ActualNegSizeReg); |
11774 | |
11775 | // Materialize a scratch register for update. |
11776 | int64_t NegProbeSize = -(int64_t)ProbeSize; |
11777 | assert(isInt<32>(NegProbeSize) && "Unhandled probe size!")((isInt<32>(NegProbeSize) && "Unhandled probe size!" ) ? static_cast<void> (0) : __assert_fail ("isInt<32>(NegProbeSize) && \"Unhandled probe size!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 11777, __PRETTY_FUNCTION__)); |
11778 | Register ScratchReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11779 | if (!isInt<16>(NegProbeSize)) { |
11780 | Register TempReg = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11781 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LIS8 : PPC::LIS), TempReg) |
11782 | .addImm(NegProbeSize >> 16); |
11783 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::ORI8 : PPC::ORI), |
11784 | ScratchReg) |
11785 | .addReg(TempReg) |
11786 | .addImm(NegProbeSize & 0xFFFF); |
11787 | } else |
11788 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::LI8 : PPC::LI), ScratchReg) |
11789 | .addImm(NegProbeSize); |
11790 | |
11791 | { |
11792 | // Probing leading residual part. |
11793 | Register Div = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11794 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::DIVD : PPC::DIVW), Div) |
11795 | .addReg(ActualNegSizeReg) |
11796 | .addReg(ScratchReg); |
11797 | Register Mul = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11798 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::MULLD : PPC::MULLW), Mul) |
11799 | .addReg(Div) |
11800 | .addReg(ScratchReg); |
11801 | Register NegMod = MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11802 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::SUBF8 : PPC::SUBF), NegMod) |
11803 | .addReg(Mul) |
11804 | .addReg(ActualNegSizeReg); |
11805 | BuildMI(*MBB, {MI}, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg) |
11806 | .addReg(FramePointer) |
11807 | .addReg(SPReg) |
11808 | .addReg(NegMod); |
11809 | } |
11810 | |
11811 | { |
11812 | // Remaining part should be multiple of ProbeSize. |
11813 | Register CmpResult = MRI.createVirtualRegister(&PPC::CRRCRegClass); |
11814 | BuildMI(TestMBB, DL, TII->get(isPPC64 ? PPC::CMPD : PPC::CMPW), CmpResult) |
11815 | .addReg(SPReg) |
11816 | .addReg(FinalStackPtr); |
11817 | BuildMI(TestMBB, DL, TII->get(PPC::BCC)) |
11818 | .addImm(PPC::PRED_EQ) |
11819 | .addReg(CmpResult) |
11820 | .addMBB(TailMBB); |
11821 | TestMBB->addSuccessor(BlockMBB); |
11822 | TestMBB->addSuccessor(TailMBB); |
11823 | } |
11824 | |
11825 | { |
11826 | // Touch the block. |
11827 | // |P...|P...|P... |
11828 | BuildMI(BlockMBB, DL, TII->get(isPPC64 ? PPC::STDUX : PPC::STWUX), SPReg) |
11829 | .addReg(FramePointer) |
11830 | .addReg(SPReg) |
11831 | .addReg(ScratchReg); |
11832 | BuildMI(BlockMBB, DL, TII->get(PPC::B)).addMBB(TestMBB); |
11833 | BlockMBB->addSuccessor(TestMBB); |
11834 | } |
11835 | |
11836 | // Calculation of MaxCallFrameSize is deferred to prologepilog, use |
11837 | // DYNAREAOFFSET pseudo instruction to get the future result. |
11838 | Register MaxCallFrameSizeReg = |
11839 | MRI.createVirtualRegister(isPPC64 ? G8RC : GPRC); |
11840 | BuildMI(TailMBB, DL, |
11841 | TII->get(isPPC64 ? PPC::DYNAREAOFFSET8 : PPC::DYNAREAOFFSET), |
11842 | MaxCallFrameSizeReg) |
11843 | .add(MI.getOperand(2)) |
11844 | .add(MI.getOperand(3)); |
11845 | BuildMI(TailMBB, DL, TII->get(isPPC64 ? PPC::ADD8 : PPC::ADD4), DstReg) |
11846 | .addReg(SPReg) |
11847 | .addReg(MaxCallFrameSizeReg); |
11848 | |
11849 | // Splice instructions after MI to TailMBB. |
11850 | TailMBB->splice(TailMBB->end(), MBB, |
11851 | std::next(MachineBasicBlock::iterator(MI)), MBB->end()); |
11852 | TailMBB->transferSuccessorsAndUpdatePHIs(MBB); |
11853 | MBB->addSuccessor(TestMBB); |
11854 | |
11855 | // Delete the pseudo instruction. |
11856 | MI.eraseFromParent(); |
11857 | |
11858 | ++NumDynamicAllocaProbed; |
11859 | return TailMBB; |
11860 | } |
11861 | |
11862 | MachineBasicBlock * |
11863 | PPCTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
11864 | MachineBasicBlock *BB) const { |
11865 | if (MI.getOpcode() == TargetOpcode::STACKMAP || |
11866 | MI.getOpcode() == TargetOpcode::PATCHPOINT) { |
11867 | if (Subtarget.is64BitELFABI() && |
11868 | MI.getOpcode() == TargetOpcode::PATCHPOINT && |
11869 | !Subtarget.isUsingPCRelativeCalls()) { |
11870 | // Call lowering should have added an r2 operand to indicate a dependence |
11871 | // on the TOC base pointer value. It can't however, because there is no |
11872 | // way to mark the dependence as implicit there, and so the stackmap code |
11873 | // will confuse it with a regular operand. Instead, add the dependence |
11874 | // here. |
11875 | MI.addOperand(MachineOperand::CreateReg(PPC::X2, false, true)); |
11876 | } |
11877 | |
11878 | return emitPatchPoint(MI, BB); |
11879 | } |
11880 | |
11881 | if (MI.getOpcode() == PPC::EH_SjLj_SetJmp32 || |
11882 | MI.getOpcode() == PPC::EH_SjLj_SetJmp64) { |
11883 | return emitEHSjLjSetJmp(MI, BB); |
11884 | } else if (MI.getOpcode() == PPC::EH_SjLj_LongJmp32 || |
11885 | MI.getOpcode() == PPC::EH_SjLj_LongJmp64) { |
11886 | return emitEHSjLjLongJmp(MI, BB); |
11887 | } |
11888 | |
11889 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
11890 | |
11891 | // To "insert" these instructions we actually have to insert their |
11892 | // control-flow patterns. |
11893 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
11894 | MachineFunction::iterator It = ++BB->getIterator(); |
11895 | |
11896 | MachineFunction *F = BB->getParent(); |
11897 | |
11898 | if (MI.getOpcode() == PPC::SELECT_CC_I4 || |
11899 | MI.getOpcode() == PPC::SELECT_CC_I8 || MI.getOpcode() == PPC::SELECT_I4 || |
11900 | MI.getOpcode() == PPC::SELECT_I8) { |
11901 | SmallVector<MachineOperand, 2> Cond; |
11902 | if (MI.getOpcode() == PPC::SELECT_CC_I4 || |
11903 | MI.getOpcode() == PPC::SELECT_CC_I8) |
11904 | Cond.push_back(MI.getOperand(4)); |
11905 | else |
11906 | Cond.push_back(MachineOperand::CreateImm(PPC::PRED_BIT_SET)); |
11907 | Cond.push_back(MI.getOperand(1)); |
11908 | |
11909 | DebugLoc dl = MI.getDebugLoc(); |
11910 | TII->insertSelect(*BB, MI, dl, MI.getOperand(0).getReg(), Cond, |
11911 | MI.getOperand(2).getReg(), MI.getOperand(3).getReg()); |
11912 | } else if (MI.getOpcode() == PPC::SELECT_CC_F4 || |
11913 | MI.getOpcode() == PPC::SELECT_CC_F8 || |
11914 | MI.getOpcode() == PPC::SELECT_CC_F16 || |
11915 | MI.getOpcode() == PPC::SELECT_CC_VRRC || |
11916 | MI.getOpcode() == PPC::SELECT_CC_VSFRC || |
11917 | MI.getOpcode() == PPC::SELECT_CC_VSSRC || |
11918 | MI.getOpcode() == PPC::SELECT_CC_VSRC || |
11919 | MI.getOpcode() == PPC::SELECT_CC_SPE4 || |
11920 | MI.getOpcode() == PPC::SELECT_CC_SPE || |
11921 | MI.getOpcode() == PPC::SELECT_F4 || |
11922 | MI.getOpcode() == PPC::SELECT_F8 || |
11923 | MI.getOpcode() == PPC::SELECT_F16 || |
11924 | MI.getOpcode() == PPC::SELECT_SPE || |
11925 | MI.getOpcode() == PPC::SELECT_SPE4 || |
11926 | MI.getOpcode() == PPC::SELECT_VRRC || |
11927 | MI.getOpcode() == PPC::SELECT_VSFRC || |
11928 | MI.getOpcode() == PPC::SELECT_VSSRC || |
11929 | MI.getOpcode() == PPC::SELECT_VSRC) { |
11930 | // The incoming instruction knows the destination vreg to set, the |
11931 | // condition code register to branch on, the true/false values to |
11932 | // select between, and a branch opcode to use. |
11933 | |
11934 | // thisMBB: |
11935 | // ... |
11936 | // TrueVal = ... |
11937 | // cmpTY ccX, r1, r2 |
11938 | // bCC copy1MBB |
11939 | // fallthrough --> copy0MBB |
11940 | MachineBasicBlock *thisMBB = BB; |
11941 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); |
11942 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); |
11943 | DebugLoc dl = MI.getDebugLoc(); |
11944 | F->insert(It, copy0MBB); |
11945 | F->insert(It, sinkMBB); |
11946 | |
11947 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
11948 | sinkMBB->splice(sinkMBB->begin(), BB, |
11949 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
11950 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); |
11951 | |
11952 | // Next, add the true and fallthrough blocks as its successors. |
11953 | BB->addSuccessor(copy0MBB); |
11954 | BB->addSuccessor(sinkMBB); |
11955 | |
11956 | if (MI.getOpcode() == PPC::SELECT_I4 || MI.getOpcode() == PPC::SELECT_I8 || |
11957 | MI.getOpcode() == PPC::SELECT_F4 || MI.getOpcode() == PPC::SELECT_F8 || |
11958 | MI.getOpcode() == PPC::SELECT_F16 || |
11959 | MI.getOpcode() == PPC::SELECT_SPE4 || |
11960 | MI.getOpcode() == PPC::SELECT_SPE || |
11961 | MI.getOpcode() == PPC::SELECT_VRRC || |
11962 | MI.getOpcode() == PPC::SELECT_VSFRC || |
11963 | MI.getOpcode() == PPC::SELECT_VSSRC || |
11964 | MI.getOpcode() == PPC::SELECT_VSRC) { |
11965 | BuildMI(BB, dl, TII->get(PPC::BC)) |
11966 | .addReg(MI.getOperand(1).getReg()) |
11967 | .addMBB(sinkMBB); |
11968 | } else { |
11969 | unsigned SelectPred = MI.getOperand(4).getImm(); |
11970 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
11971 | .addImm(SelectPred) |
11972 | .addReg(MI.getOperand(1).getReg()) |
11973 | .addMBB(sinkMBB); |
11974 | } |
11975 | |
11976 | // copy0MBB: |
11977 | // %FalseValue = ... |
11978 | // # fallthrough to sinkMBB |
11979 | BB = copy0MBB; |
11980 | |
11981 | // Update machine-CFG edges |
11982 | BB->addSuccessor(sinkMBB); |
11983 | |
11984 | // sinkMBB: |
11985 | // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] |
11986 | // ... |
11987 | BB = sinkMBB; |
11988 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::PHI), MI.getOperand(0).getReg()) |
11989 | .addReg(MI.getOperand(3).getReg()) |
11990 | .addMBB(copy0MBB) |
11991 | .addReg(MI.getOperand(2).getReg()) |
11992 | .addMBB(thisMBB); |
11993 | } else if (MI.getOpcode() == PPC::ReadTB) { |
11994 | // To read the 64-bit time-base register on a 32-bit target, we read the |
11995 | // two halves. Should the counter have wrapped while it was being read, we |
11996 | // need to try again. |
11997 | // ... |
11998 | // readLoop: |
11999 | // mfspr Rx,TBU # load from TBU |
12000 | // mfspr Ry,TB # load from TB |
12001 | // mfspr Rz,TBU # load from TBU |
12002 | // cmpw crX,Rx,Rz # check if 'old'='new' |
12003 | // bne readLoop # branch if they're not equal |
12004 | // ... |
12005 | |
12006 | MachineBasicBlock *readMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12007 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12008 | DebugLoc dl = MI.getDebugLoc(); |
12009 | F->insert(It, readMBB); |
12010 | F->insert(It, sinkMBB); |
12011 | |
12012 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
12013 | sinkMBB->splice(sinkMBB->begin(), BB, |
12014 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12015 | sinkMBB->transferSuccessorsAndUpdatePHIs(BB); |
12016 | |
12017 | BB->addSuccessor(readMBB); |
12018 | BB = readMBB; |
12019 | |
12020 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12021 | Register ReadAgainReg = RegInfo.createVirtualRegister(&PPC::GPRCRegClass); |
12022 | Register LoReg = MI.getOperand(0).getReg(); |
12023 | Register HiReg = MI.getOperand(1).getReg(); |
12024 | |
12025 | BuildMI(BB, dl, TII->get(PPC::MFSPR), HiReg).addImm(269); |
12026 | BuildMI(BB, dl, TII->get(PPC::MFSPR), LoReg).addImm(268); |
12027 | BuildMI(BB, dl, TII->get(PPC::MFSPR), ReadAgainReg).addImm(269); |
12028 | |
12029 | Register CmpReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass); |
12030 | |
12031 | BuildMI(BB, dl, TII->get(PPC::CMPW), CmpReg) |
12032 | .addReg(HiReg) |
12033 | .addReg(ReadAgainReg); |
12034 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12035 | .addImm(PPC::PRED_NE) |
12036 | .addReg(CmpReg) |
12037 | .addMBB(readMBB); |
12038 | |
12039 | BB->addSuccessor(readMBB); |
12040 | BB->addSuccessor(sinkMBB); |
12041 | } else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I8) |
12042 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::ADD4); |
12043 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I16) |
12044 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::ADD4); |
12045 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I32) |
12046 | BB = EmitAtomicBinary(MI, BB, 4, PPC::ADD4); |
12047 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_ADD_I64) |
12048 | BB = EmitAtomicBinary(MI, BB, 8, PPC::ADD8); |
12049 | |
12050 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I8) |
12051 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::AND); |
12052 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I16) |
12053 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::AND); |
12054 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I32) |
12055 | BB = EmitAtomicBinary(MI, BB, 4, PPC::AND); |
12056 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_AND_I64) |
12057 | BB = EmitAtomicBinary(MI, BB, 8, PPC::AND8); |
12058 | |
12059 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I8) |
12060 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::OR); |
12061 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I16) |
12062 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::OR); |
12063 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I32) |
12064 | BB = EmitAtomicBinary(MI, BB, 4, PPC::OR); |
12065 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_OR_I64) |
12066 | BB = EmitAtomicBinary(MI, BB, 8, PPC::OR8); |
12067 | |
12068 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I8) |
12069 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::XOR); |
12070 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I16) |
12071 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::XOR); |
12072 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I32) |
12073 | BB = EmitAtomicBinary(MI, BB, 4, PPC::XOR); |
12074 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_XOR_I64) |
12075 | BB = EmitAtomicBinary(MI, BB, 8, PPC::XOR8); |
12076 | |
12077 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I8) |
12078 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::NAND); |
12079 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I16) |
12080 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::NAND); |
12081 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I32) |
12082 | BB = EmitAtomicBinary(MI, BB, 4, PPC::NAND); |
12083 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_NAND_I64) |
12084 | BB = EmitAtomicBinary(MI, BB, 8, PPC::NAND8); |
12085 | |
12086 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I8) |
12087 | BB = EmitPartwordAtomicBinary(MI, BB, true, PPC::SUBF); |
12088 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I16) |
12089 | BB = EmitPartwordAtomicBinary(MI, BB, false, PPC::SUBF); |
12090 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I32) |
12091 | BB = EmitAtomicBinary(MI, BB, 4, PPC::SUBF); |
12092 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_SUB_I64) |
12093 | BB = EmitAtomicBinary(MI, BB, 8, PPC::SUBF8); |
12094 | |
12095 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I8) |
12096 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_GE); |
12097 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I16) |
12098 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_GE); |
12099 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I32) |
12100 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_GE); |
12101 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MIN_I64) |
12102 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_GE); |
12103 | |
12104 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I8) |
12105 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPW, PPC::PRED_LE); |
12106 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I16) |
12107 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPW, PPC::PRED_LE); |
12108 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I32) |
12109 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPW, PPC::PRED_LE); |
12110 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_MAX_I64) |
12111 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPD, PPC::PRED_LE); |
12112 | |
12113 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I8) |
12114 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_GE); |
12115 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I16) |
12116 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_GE); |
12117 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I32) |
12118 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_GE); |
12119 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMIN_I64) |
12120 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_GE); |
12121 | |
12122 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I8) |
12123 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0, PPC::CMPLW, PPC::PRED_LE); |
12124 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I16) |
12125 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0, PPC::CMPLW, PPC::PRED_LE); |
12126 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I32) |
12127 | BB = EmitAtomicBinary(MI, BB, 4, 0, PPC::CMPLW, PPC::PRED_LE); |
12128 | else if (MI.getOpcode() == PPC::ATOMIC_LOAD_UMAX_I64) |
12129 | BB = EmitAtomicBinary(MI, BB, 8, 0, PPC::CMPLD, PPC::PRED_LE); |
12130 | |
12131 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I8) |
12132 | BB = EmitPartwordAtomicBinary(MI, BB, true, 0); |
12133 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I16) |
12134 | BB = EmitPartwordAtomicBinary(MI, BB, false, 0); |
12135 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I32) |
12136 | BB = EmitAtomicBinary(MI, BB, 4, 0); |
12137 | else if (MI.getOpcode() == PPC::ATOMIC_SWAP_I64) |
12138 | BB = EmitAtomicBinary(MI, BB, 8, 0); |
12139 | else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I32 || |
12140 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64 || |
12141 | (Subtarget.hasPartwordAtomics() && |
12142 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8) || |
12143 | (Subtarget.hasPartwordAtomics() && |
12144 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16)) { |
12145 | bool is64bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I64; |
12146 | |
12147 | auto LoadMnemonic = PPC::LDARX; |
12148 | auto StoreMnemonic = PPC::STDCX; |
12149 | switch (MI.getOpcode()) { |
12150 | default: |
12151 | llvm_unreachable("Compare and swap of unknown size")::llvm::llvm_unreachable_internal("Compare and swap of unknown size" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12151); |
12152 | case PPC::ATOMIC_CMP_SWAP_I8: |
12153 | LoadMnemonic = PPC::LBARX; |
12154 | StoreMnemonic = PPC::STBCX; |
12155 | assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.")((Subtarget.hasPartwordAtomics() && "No support partword atomics." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"No support partword atomics.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12155, __PRETTY_FUNCTION__)); |
12156 | break; |
12157 | case PPC::ATOMIC_CMP_SWAP_I16: |
12158 | LoadMnemonic = PPC::LHARX; |
12159 | StoreMnemonic = PPC::STHCX; |
12160 | assert(Subtarget.hasPartwordAtomics() && "No support partword atomics.")((Subtarget.hasPartwordAtomics() && "No support partword atomics." ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasPartwordAtomics() && \"No support partword atomics.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12160, __PRETTY_FUNCTION__)); |
12161 | break; |
12162 | case PPC::ATOMIC_CMP_SWAP_I32: |
12163 | LoadMnemonic = PPC::LWARX; |
12164 | StoreMnemonic = PPC::STWCX; |
12165 | break; |
12166 | case PPC::ATOMIC_CMP_SWAP_I64: |
12167 | LoadMnemonic = PPC::LDARX; |
12168 | StoreMnemonic = PPC::STDCX; |
12169 | break; |
12170 | } |
12171 | Register dest = MI.getOperand(0).getReg(); |
12172 | Register ptrA = MI.getOperand(1).getReg(); |
12173 | Register ptrB = MI.getOperand(2).getReg(); |
12174 | Register oldval = MI.getOperand(3).getReg(); |
12175 | Register newval = MI.getOperand(4).getReg(); |
12176 | DebugLoc dl = MI.getDebugLoc(); |
12177 | |
12178 | MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12179 | MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12180 | MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12181 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12182 | F->insert(It, loop1MBB); |
12183 | F->insert(It, loop2MBB); |
12184 | F->insert(It, midMBB); |
12185 | F->insert(It, exitMBB); |
12186 | exitMBB->splice(exitMBB->begin(), BB, |
12187 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12188 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
12189 | |
12190 | // thisMBB: |
12191 | // ... |
12192 | // fallthrough --> loopMBB |
12193 | BB->addSuccessor(loop1MBB); |
12194 | |
12195 | // loop1MBB: |
12196 | // l[bhwd]arx dest, ptr |
12197 | // cmp[wd] dest, oldval |
12198 | // bne- midMBB |
12199 | // loop2MBB: |
12200 | // st[bhwd]cx. newval, ptr |
12201 | // bne- loopMBB |
12202 | // b exitBB |
12203 | // midMBB: |
12204 | // st[bhwd]cx. dest, ptr |
12205 | // exitBB: |
12206 | BB = loop1MBB; |
12207 | BuildMI(BB, dl, TII->get(LoadMnemonic), dest).addReg(ptrA).addReg(ptrB); |
12208 | BuildMI(BB, dl, TII->get(is64bit ? PPC::CMPD : PPC::CMPW), PPC::CR0) |
12209 | .addReg(oldval) |
12210 | .addReg(dest); |
12211 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12212 | .addImm(PPC::PRED_NE) |
12213 | .addReg(PPC::CR0) |
12214 | .addMBB(midMBB); |
12215 | BB->addSuccessor(loop2MBB); |
12216 | BB->addSuccessor(midMBB); |
12217 | |
12218 | BB = loop2MBB; |
12219 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
12220 | .addReg(newval) |
12221 | .addReg(ptrA) |
12222 | .addReg(ptrB); |
12223 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12224 | .addImm(PPC::PRED_NE) |
12225 | .addReg(PPC::CR0) |
12226 | .addMBB(loop1MBB); |
12227 | BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB); |
12228 | BB->addSuccessor(loop1MBB); |
12229 | BB->addSuccessor(exitMBB); |
12230 | |
12231 | BB = midMBB; |
12232 | BuildMI(BB, dl, TII->get(StoreMnemonic)) |
12233 | .addReg(dest) |
12234 | .addReg(ptrA) |
12235 | .addReg(ptrB); |
12236 | BB->addSuccessor(exitMBB); |
12237 | |
12238 | // exitMBB: |
12239 | // ... |
12240 | BB = exitMBB; |
12241 | } else if (MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8 || |
12242 | MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I16) { |
12243 | // We must use 64-bit registers for addresses when targeting 64-bit, |
12244 | // since we're actually doing arithmetic on them. Other registers |
12245 | // can be 32-bit. |
12246 | bool is64bit = Subtarget.isPPC64(); |
12247 | bool isLittleEndian = Subtarget.isLittleEndian(); |
12248 | bool is8bit = MI.getOpcode() == PPC::ATOMIC_CMP_SWAP_I8; |
12249 | |
12250 | Register dest = MI.getOperand(0).getReg(); |
12251 | Register ptrA = MI.getOperand(1).getReg(); |
12252 | Register ptrB = MI.getOperand(2).getReg(); |
12253 | Register oldval = MI.getOperand(3).getReg(); |
12254 | Register newval = MI.getOperand(4).getReg(); |
12255 | DebugLoc dl = MI.getDebugLoc(); |
12256 | |
12257 | MachineBasicBlock *loop1MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12258 | MachineBasicBlock *loop2MBB = F->CreateMachineBasicBlock(LLVM_BB); |
12259 | MachineBasicBlock *midMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12260 | MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB); |
12261 | F->insert(It, loop1MBB); |
12262 | F->insert(It, loop2MBB); |
12263 | F->insert(It, midMBB); |
12264 | F->insert(It, exitMBB); |
12265 | exitMBB->splice(exitMBB->begin(), BB, |
12266 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); |
12267 | exitMBB->transferSuccessorsAndUpdatePHIs(BB); |
12268 | |
12269 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12270 | const TargetRegisterClass *RC = |
12271 | is64bit ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
12272 | const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; |
12273 | |
12274 | Register PtrReg = RegInfo.createVirtualRegister(RC); |
12275 | Register Shift1Reg = RegInfo.createVirtualRegister(GPRC); |
12276 | Register ShiftReg = |
12277 | isLittleEndian ? Shift1Reg : RegInfo.createVirtualRegister(GPRC); |
12278 | Register NewVal2Reg = RegInfo.createVirtualRegister(GPRC); |
12279 | Register NewVal3Reg = RegInfo.createVirtualRegister(GPRC); |
12280 | Register OldVal2Reg = RegInfo.createVirtualRegister(GPRC); |
12281 | Register OldVal3Reg = RegInfo.createVirtualRegister(GPRC); |
12282 | Register MaskReg = RegInfo.createVirtualRegister(GPRC); |
12283 | Register Mask2Reg = RegInfo.createVirtualRegister(GPRC); |
12284 | Register Mask3Reg = RegInfo.createVirtualRegister(GPRC); |
12285 | Register Tmp2Reg = RegInfo.createVirtualRegister(GPRC); |
12286 | Register Tmp4Reg = RegInfo.createVirtualRegister(GPRC); |
12287 | Register TmpDestReg = RegInfo.createVirtualRegister(GPRC); |
12288 | Register Ptr1Reg; |
12289 | Register TmpReg = RegInfo.createVirtualRegister(GPRC); |
12290 | Register ZeroReg = is64bit ? PPC::ZERO8 : PPC::ZERO; |
12291 | // thisMBB: |
12292 | // ... |
12293 | // fallthrough --> loopMBB |
12294 | BB->addSuccessor(loop1MBB); |
12295 | |
12296 | // The 4-byte load must be aligned, while a char or short may be |
12297 | // anywhere in the word. Hence all this nasty bookkeeping code. |
12298 | // add ptr1, ptrA, ptrB [copy if ptrA==0] |
12299 | // rlwinm shift1, ptr1, 3, 27, 28 [3, 27, 27] |
12300 | // xori shift, shift1, 24 [16] |
12301 | // rlwinm ptr, ptr1, 0, 0, 29 |
12302 | // slw newval2, newval, shift |
12303 | // slw oldval2, oldval,shift |
12304 | // li mask2, 255 [li mask3, 0; ori mask2, mask3, 65535] |
12305 | // slw mask, mask2, shift |
12306 | // and newval3, newval2, mask |
12307 | // and oldval3, oldval2, mask |
12308 | // loop1MBB: |
12309 | // lwarx tmpDest, ptr |
12310 | // and tmp, tmpDest, mask |
12311 | // cmpw tmp, oldval3 |
12312 | // bne- midMBB |
12313 | // loop2MBB: |
12314 | // andc tmp2, tmpDest, mask |
12315 | // or tmp4, tmp2, newval3 |
12316 | // stwcx. tmp4, ptr |
12317 | // bne- loop1MBB |
12318 | // b exitBB |
12319 | // midMBB: |
12320 | // stwcx. tmpDest, ptr |
12321 | // exitBB: |
12322 | // srw dest, tmpDest, shift |
12323 | if (ptrA != ZeroReg) { |
12324 | Ptr1Reg = RegInfo.createVirtualRegister(RC); |
12325 | BuildMI(BB, dl, TII->get(is64bit ? PPC::ADD8 : PPC::ADD4), Ptr1Reg) |
12326 | .addReg(ptrA) |
12327 | .addReg(ptrB); |
12328 | } else { |
12329 | Ptr1Reg = ptrB; |
12330 | } |
12331 | |
12332 | // We need use 32-bit subregister to avoid mismatch register class in 64-bit |
12333 | // mode. |
12334 | BuildMI(BB, dl, TII->get(PPC::RLWINM), Shift1Reg) |
12335 | .addReg(Ptr1Reg, 0, is64bit ? PPC::sub_32 : 0) |
12336 | .addImm(3) |
12337 | .addImm(27) |
12338 | .addImm(is8bit ? 28 : 27); |
12339 | if (!isLittleEndian) |
12340 | BuildMI(BB, dl, TII->get(PPC::XORI), ShiftReg) |
12341 | .addReg(Shift1Reg) |
12342 | .addImm(is8bit ? 24 : 16); |
12343 | if (is64bit) |
12344 | BuildMI(BB, dl, TII->get(PPC::RLDICR), PtrReg) |
12345 | .addReg(Ptr1Reg) |
12346 | .addImm(0) |
12347 | .addImm(61); |
12348 | else |
12349 | BuildMI(BB, dl, TII->get(PPC::RLWINM), PtrReg) |
12350 | .addReg(Ptr1Reg) |
12351 | .addImm(0) |
12352 | .addImm(0) |
12353 | .addImm(29); |
12354 | BuildMI(BB, dl, TII->get(PPC::SLW), NewVal2Reg) |
12355 | .addReg(newval) |
12356 | .addReg(ShiftReg); |
12357 | BuildMI(BB, dl, TII->get(PPC::SLW), OldVal2Reg) |
12358 | .addReg(oldval) |
12359 | .addReg(ShiftReg); |
12360 | if (is8bit) |
12361 | BuildMI(BB, dl, TII->get(PPC::LI), Mask2Reg).addImm(255); |
12362 | else { |
12363 | BuildMI(BB, dl, TII->get(PPC::LI), Mask3Reg).addImm(0); |
12364 | BuildMI(BB, dl, TII->get(PPC::ORI), Mask2Reg) |
12365 | .addReg(Mask3Reg) |
12366 | .addImm(65535); |
12367 | } |
12368 | BuildMI(BB, dl, TII->get(PPC::SLW), MaskReg) |
12369 | .addReg(Mask2Reg) |
12370 | .addReg(ShiftReg); |
12371 | BuildMI(BB, dl, TII->get(PPC::AND), NewVal3Reg) |
12372 | .addReg(NewVal2Reg) |
12373 | .addReg(MaskReg); |
12374 | BuildMI(BB, dl, TII->get(PPC::AND), OldVal3Reg) |
12375 | .addReg(OldVal2Reg) |
12376 | .addReg(MaskReg); |
12377 | |
12378 | BB = loop1MBB; |
12379 | BuildMI(BB, dl, TII->get(PPC::LWARX), TmpDestReg) |
12380 | .addReg(ZeroReg) |
12381 | .addReg(PtrReg); |
12382 | BuildMI(BB, dl, TII->get(PPC::AND), TmpReg) |
12383 | .addReg(TmpDestReg) |
12384 | .addReg(MaskReg); |
12385 | BuildMI(BB, dl, TII->get(PPC::CMPW), PPC::CR0) |
12386 | .addReg(TmpReg) |
12387 | .addReg(OldVal3Reg); |
12388 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12389 | .addImm(PPC::PRED_NE) |
12390 | .addReg(PPC::CR0) |
12391 | .addMBB(midMBB); |
12392 | BB->addSuccessor(loop2MBB); |
12393 | BB->addSuccessor(midMBB); |
12394 | |
12395 | BB = loop2MBB; |
12396 | BuildMI(BB, dl, TII->get(PPC::ANDC), Tmp2Reg) |
12397 | .addReg(TmpDestReg) |
12398 | .addReg(MaskReg); |
12399 | BuildMI(BB, dl, TII->get(PPC::OR), Tmp4Reg) |
12400 | .addReg(Tmp2Reg) |
12401 | .addReg(NewVal3Reg); |
12402 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
12403 | .addReg(Tmp4Reg) |
12404 | .addReg(ZeroReg) |
12405 | .addReg(PtrReg); |
12406 | BuildMI(BB, dl, TII->get(PPC::BCC)) |
12407 | .addImm(PPC::PRED_NE) |
12408 | .addReg(PPC::CR0) |
12409 | .addMBB(loop1MBB); |
12410 | BuildMI(BB, dl, TII->get(PPC::B)).addMBB(exitMBB); |
12411 | BB->addSuccessor(loop1MBB); |
12412 | BB->addSuccessor(exitMBB); |
12413 | |
12414 | BB = midMBB; |
12415 | BuildMI(BB, dl, TII->get(PPC::STWCX)) |
12416 | .addReg(TmpDestReg) |
12417 | .addReg(ZeroReg) |
12418 | .addReg(PtrReg); |
12419 | BB->addSuccessor(exitMBB); |
12420 | |
12421 | // exitMBB: |
12422 | // ... |
12423 | BB = exitMBB; |
12424 | BuildMI(*BB, BB->begin(), dl, TII->get(PPC::SRW), dest) |
12425 | .addReg(TmpReg) |
12426 | .addReg(ShiftReg); |
12427 | } else if (MI.getOpcode() == PPC::FADDrtz) { |
12428 | // This pseudo performs an FADD with rounding mode temporarily forced |
12429 | // to round-to-zero. We emit this via custom inserter since the FPSCR |
12430 | // is not modeled at the SelectionDAG level. |
12431 | Register Dest = MI.getOperand(0).getReg(); |
12432 | Register Src1 = MI.getOperand(1).getReg(); |
12433 | Register Src2 = MI.getOperand(2).getReg(); |
12434 | DebugLoc dl = MI.getDebugLoc(); |
12435 | |
12436 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12437 | Register MFFSReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass); |
12438 | |
12439 | // Save FPSCR value. |
12440 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), MFFSReg); |
12441 | |
12442 | // Set rounding mode to round-to-zero. |
12443 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB1)) |
12444 | .addImm(31) |
12445 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12446 | |
12447 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSB0)) |
12448 | .addImm(30) |
12449 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12450 | |
12451 | // Perform addition. |
12452 | auto MIB = BuildMI(*BB, MI, dl, TII->get(PPC::FADD), Dest) |
12453 | .addReg(Src1) |
12454 | .addReg(Src2); |
12455 | if (MI.getFlag(MachineInstr::NoFPExcept)) |
12456 | MIB.setMIFlag(MachineInstr::NoFPExcept); |
12457 | |
12458 | // Restore FPSCR value. |
12459 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSFb)).addImm(1).addReg(MFFSReg); |
12460 | } else if (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT || |
12461 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT || |
12462 | MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 || |
12463 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8) { |
12464 | unsigned Opcode = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8 || |
12465 | MI.getOpcode() == PPC::ANDI_rec_1_GT_BIT8) |
12466 | ? PPC::ANDI8_rec |
12467 | : PPC::ANDI_rec; |
12468 | bool IsEQ = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT || |
12469 | MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT8); |
12470 | |
12471 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12472 | Register Dest = RegInfo.createVirtualRegister( |
12473 | Opcode == PPC::ANDI_rec ? &PPC::GPRCRegClass : &PPC::G8RCRegClass); |
12474 | |
12475 | DebugLoc Dl = MI.getDebugLoc(); |
12476 | BuildMI(*BB, MI, Dl, TII->get(Opcode), Dest) |
12477 | .addReg(MI.getOperand(1).getReg()) |
12478 | .addImm(1); |
12479 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12480 | MI.getOperand(0).getReg()) |
12481 | .addReg(IsEQ ? PPC::CR0EQ : PPC::CR0GT); |
12482 | } else if (MI.getOpcode() == PPC::TCHECK_RET) { |
12483 | DebugLoc Dl = MI.getDebugLoc(); |
12484 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12485 | Register CRReg = RegInfo.createVirtualRegister(&PPC::CRRCRegClass); |
12486 | BuildMI(*BB, MI, Dl, TII->get(PPC::TCHECK), CRReg); |
12487 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12488 | MI.getOperand(0).getReg()) |
12489 | .addReg(CRReg); |
12490 | } else if (MI.getOpcode() == PPC::TBEGIN_RET) { |
12491 | DebugLoc Dl = MI.getDebugLoc(); |
12492 | unsigned Imm = MI.getOperand(1).getImm(); |
12493 | BuildMI(*BB, MI, Dl, TII->get(PPC::TBEGIN)).addImm(Imm); |
12494 | BuildMI(*BB, MI, Dl, TII->get(TargetOpcode::COPY), |
12495 | MI.getOperand(0).getReg()) |
12496 | .addReg(PPC::CR0EQ); |
12497 | } else if (MI.getOpcode() == PPC::SETRNDi) { |
12498 | DebugLoc dl = MI.getDebugLoc(); |
12499 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12500 | |
12501 | // Save FPSCR value. |
12502 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12503 | |
12504 | // The floating point rounding mode is in the bits 62:63 of FPCSR, and has |
12505 | // the following settings: |
12506 | // 00 Round to nearest |
12507 | // 01 Round to 0 |
12508 | // 10 Round to +inf |
12509 | // 11 Round to -inf |
12510 | |
12511 | // When the operand is immediate, using the two least significant bits of |
12512 | // the immediate to set the bits 62:63 of FPSCR. |
12513 | unsigned Mode = MI.getOperand(1).getImm(); |
12514 | BuildMI(*BB, MI, dl, TII->get((Mode & 1) ? PPC::MTFSB1 : PPC::MTFSB0)) |
12515 | .addImm(31) |
12516 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12517 | |
12518 | BuildMI(*BB, MI, dl, TII->get((Mode & 2) ? PPC::MTFSB1 : PPC::MTFSB0)) |
12519 | .addImm(30) |
12520 | .addReg(PPC::RM, RegState::ImplicitDefine); |
12521 | } else if (MI.getOpcode() == PPC::SETRND) { |
12522 | DebugLoc dl = MI.getDebugLoc(); |
12523 | |
12524 | // Copy register from F8RCRegClass::SrcReg to G8RCRegClass::DestReg |
12525 | // or copy register from G8RCRegClass::SrcReg to F8RCRegClass::DestReg. |
12526 | // If the target doesn't have DirectMove, we should use stack to do the |
12527 | // conversion, because the target doesn't have the instructions like mtvsrd |
12528 | // or mfvsrd to do this conversion directly. |
12529 | auto copyRegFromG8RCOrF8RC = [&] (unsigned DestReg, unsigned SrcReg) { |
12530 | if (Subtarget.hasDirectMove()) { |
12531 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), DestReg) |
12532 | .addReg(SrcReg); |
12533 | } else { |
12534 | // Use stack to do the register copy. |
12535 | unsigned StoreOp = PPC::STD, LoadOp = PPC::LFD; |
12536 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12537 | const TargetRegisterClass *RC = RegInfo.getRegClass(SrcReg); |
12538 | if (RC == &PPC::F8RCRegClass) { |
12539 | // Copy register from F8RCRegClass to G8RCRegclass. |
12540 | assert((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) &&(((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12541, __PRETTY_FUNCTION__)) |
12541 | "Unsupported RegClass.")(((RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(DestReg) == &PPC::G8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12541, __PRETTY_FUNCTION__)); |
12542 | |
12543 | StoreOp = PPC::STFD; |
12544 | LoadOp = PPC::LD; |
12545 | } else { |
12546 | // Copy register from G8RCRegClass to F8RCRegclass. |
12547 | assert((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) &&(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12549, __PRETTY_FUNCTION__)) |
12548 | (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) &&(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12549, __PRETTY_FUNCTION__)) |
12549 | "Unsupported RegClass.")(((RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && "Unsupported RegClass.") ? static_cast<void> (0) : __assert_fail ("(RegInfo.getRegClass(SrcReg) == &PPC::G8RCRegClass) && (RegInfo.getRegClass(DestReg) == &PPC::F8RCRegClass) && \"Unsupported RegClass.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12549, __PRETTY_FUNCTION__)); |
12550 | } |
12551 | |
12552 | MachineFrameInfo &MFI = F->getFrameInfo(); |
12553 | int FrameIdx = MFI.CreateStackObject(8, Align(8), false); |
12554 | |
12555 | MachineMemOperand *MMOStore = F->getMachineMemOperand( |
12556 | MachinePointerInfo::getFixedStack(*F, FrameIdx, 0), |
12557 | MachineMemOperand::MOStore, MFI.getObjectSize(FrameIdx), |
12558 | MFI.getObjectAlign(FrameIdx)); |
12559 | |
12560 | // Store the SrcReg into the stack. |
12561 | BuildMI(*BB, MI, dl, TII->get(StoreOp)) |
12562 | .addReg(SrcReg) |
12563 | .addImm(0) |
12564 | .addFrameIndex(FrameIdx) |
12565 | .addMemOperand(MMOStore); |
12566 | |
12567 | MachineMemOperand *MMOLoad = F->getMachineMemOperand( |
12568 | MachinePointerInfo::getFixedStack(*F, FrameIdx, 0), |
12569 | MachineMemOperand::MOLoad, MFI.getObjectSize(FrameIdx), |
12570 | MFI.getObjectAlign(FrameIdx)); |
12571 | |
12572 | // Load from the stack where SrcReg is stored, and save to DestReg, |
12573 | // so we have done the RegClass conversion from RegClass::SrcReg to |
12574 | // RegClass::DestReg. |
12575 | BuildMI(*BB, MI, dl, TII->get(LoadOp), DestReg) |
12576 | .addImm(0) |
12577 | .addFrameIndex(FrameIdx) |
12578 | .addMemOperand(MMOLoad); |
12579 | } |
12580 | }; |
12581 | |
12582 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12583 | |
12584 | // Save FPSCR value. |
12585 | BuildMI(*BB, MI, dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12586 | |
12587 | // When the operand is gprc register, use two least significant bits of the |
12588 | // register and mtfsf instruction to set the bits 62:63 of FPSCR. |
12589 | // |
12590 | // copy OldFPSCRTmpReg, OldFPSCRReg |
12591 | // (INSERT_SUBREG ExtSrcReg, (IMPLICIT_DEF ImDefReg), SrcOp, 1) |
12592 | // rldimi NewFPSCRTmpReg, ExtSrcReg, OldFPSCRReg, 0, 62 |
12593 | // copy NewFPSCRReg, NewFPSCRTmpReg |
12594 | // mtfsf 255, NewFPSCRReg |
12595 | MachineOperand SrcOp = MI.getOperand(1); |
12596 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
12597 | Register OldFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12598 | |
12599 | copyRegFromG8RCOrF8RC(OldFPSCRTmpReg, OldFPSCRReg); |
12600 | |
12601 | Register ImDefReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12602 | Register ExtSrcReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12603 | |
12604 | // The first operand of INSERT_SUBREG should be a register which has |
12605 | // subregisters, we only care about its RegClass, so we should use an |
12606 | // IMPLICIT_DEF register. |
12607 | BuildMI(*BB, MI, dl, TII->get(TargetOpcode::IMPLICIT_DEF), ImDefReg); |
12608 | BuildMI(*BB, MI, dl, TII->get(PPC::INSERT_SUBREG), ExtSrcReg) |
12609 | .addReg(ImDefReg) |
12610 | .add(SrcOp) |
12611 | .addImm(1); |
12612 | |
12613 | Register NewFPSCRTmpReg = RegInfo.createVirtualRegister(&PPC::G8RCRegClass); |
12614 | BuildMI(*BB, MI, dl, TII->get(PPC::RLDIMI), NewFPSCRTmpReg) |
12615 | .addReg(OldFPSCRTmpReg) |
12616 | .addReg(ExtSrcReg) |
12617 | .addImm(0) |
12618 | .addImm(62); |
12619 | |
12620 | Register NewFPSCRReg = RegInfo.createVirtualRegister(&PPC::F8RCRegClass); |
12621 | copyRegFromG8RCOrF8RC(NewFPSCRReg, NewFPSCRTmpReg); |
12622 | |
12623 | // The mask 255 means that put the 32:63 bits of NewFPSCRReg to the 32:63 |
12624 | // bits of FPSCR. |
12625 | BuildMI(*BB, MI, dl, TII->get(PPC::MTFSF)) |
12626 | .addImm(255) |
12627 | .addReg(NewFPSCRReg) |
12628 | .addImm(0) |
12629 | .addImm(0); |
12630 | } else if (MI.getOpcode() == PPC::SETFLM) { |
12631 | DebugLoc Dl = MI.getDebugLoc(); |
12632 | |
12633 | // Result of setflm is previous FPSCR content, so we need to save it first. |
12634 | Register OldFPSCRReg = MI.getOperand(0).getReg(); |
12635 | BuildMI(*BB, MI, Dl, TII->get(PPC::MFFS), OldFPSCRReg); |
12636 | |
12637 | // Put bits in 32:63 to FPSCR. |
12638 | Register NewFPSCRReg = MI.getOperand(1).getReg(); |
12639 | BuildMI(*BB, MI, Dl, TII->get(PPC::MTFSF)) |
12640 | .addImm(255) |
12641 | .addReg(NewFPSCRReg) |
12642 | .addImm(0) |
12643 | .addImm(0); |
12644 | } else if (MI.getOpcode() == PPC::PROBED_ALLOCA_32 || |
12645 | MI.getOpcode() == PPC::PROBED_ALLOCA_64) { |
12646 | return emitProbedAlloca(MI, BB); |
12647 | } else { |
12648 | llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12648); |
12649 | } |
12650 | |
12651 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
12652 | return BB; |
12653 | } |
12654 | |
12655 | //===----------------------------------------------------------------------===// |
12656 | // Target Optimization Hooks |
12657 | //===----------------------------------------------------------------------===// |
12658 | |
12659 | static int getEstimateRefinementSteps(EVT VT, const PPCSubtarget &Subtarget) { |
12660 | // For the estimates, convergence is quadratic, so we essentially double the |
12661 | // number of digits correct after every iteration. For both FRE and FRSQRTE, |
12662 | // the minimum architected relative accuracy is 2^-5. When hasRecipPrec(), |
12663 | // this is 2^-14. IEEE float has 23 digits and double has 52 digits. |
12664 | int RefinementSteps = Subtarget.hasRecipPrec() ? 1 : 3; |
12665 | if (VT.getScalarType() == MVT::f64) |
12666 | RefinementSteps++; |
12667 | return RefinementSteps; |
12668 | } |
12669 | |
12670 | SDValue PPCTargetLowering::getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, |
12671 | int Enabled, int &RefinementSteps, |
12672 | bool &UseOneConstNR, |
12673 | bool Reciprocal) const { |
12674 | EVT VT = Operand.getValueType(); |
12675 | if ((VT == MVT::f32 && Subtarget.hasFRSQRTES()) || |
12676 | (VT == MVT::f64 && Subtarget.hasFRSQRTE()) || |
12677 | (VT == MVT::v4f32 && Subtarget.hasAltivec()) || |
12678 | (VT == MVT::v2f64 && Subtarget.hasVSX())) { |
12679 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
12680 | RefinementSteps = getEstimateRefinementSteps(VT, Subtarget); |
12681 | |
12682 | // The Newton-Raphson computation with a single constant does not provide |
12683 | // enough accuracy on some CPUs. |
12684 | UseOneConstNR = !Subtarget.needsTwoConstNR(); |
12685 | return DAG.getNode(PPCISD::FRSQRTE, SDLoc(Operand), VT, Operand); |
12686 | } |
12687 | return SDValue(); |
12688 | } |
12689 | |
12690 | SDValue PPCTargetLowering::getRecipEstimate(SDValue Operand, SelectionDAG &DAG, |
12691 | int Enabled, |
12692 | int &RefinementSteps) const { |
12693 | EVT VT = Operand.getValueType(); |
12694 | if ((VT == MVT::f32 && Subtarget.hasFRES()) || |
12695 | (VT == MVT::f64 && Subtarget.hasFRE()) || |
12696 | (VT == MVT::v4f32 && Subtarget.hasAltivec()) || |
12697 | (VT == MVT::v2f64 && Subtarget.hasVSX())) { |
12698 | if (RefinementSteps == ReciprocalEstimate::Unspecified) |
12699 | RefinementSteps = getEstimateRefinementSteps(VT, Subtarget); |
12700 | return DAG.getNode(PPCISD::FRE, SDLoc(Operand), VT, Operand); |
12701 | } |
12702 | return SDValue(); |
12703 | } |
12704 | |
12705 | unsigned PPCTargetLowering::combineRepeatedFPDivisors() const { |
12706 | // Note: This functionality is used only when unsafe-fp-math is enabled, and |
12707 | // on cores with reciprocal estimates (which are used when unsafe-fp-math is |
12708 | // enabled for division), this functionality is redundant with the default |
12709 | // combiner logic (once the division -> reciprocal/multiply transformation |
12710 | // has taken place). As a result, this matters more for older cores than for |
12711 | // newer ones. |
12712 | |
12713 | // Combine multiple FDIVs with the same divisor into multiple FMULs by the |
12714 | // reciprocal if there are two or more FDIVs (for embedded cores with only |
12715 | // one FP pipeline) for three or more FDIVs (for generic OOO cores). |
12716 | switch (Subtarget.getCPUDirective()) { |
12717 | default: |
12718 | return 3; |
12719 | case PPC::DIR_440: |
12720 | case PPC::DIR_A2: |
12721 | case PPC::DIR_E500: |
12722 | case PPC::DIR_E500mc: |
12723 | case PPC::DIR_E5500: |
12724 | return 2; |
12725 | } |
12726 | } |
12727 | |
12728 | // isConsecutiveLSLoc needs to work even if all adds have not yet been |
12729 | // collapsed, and so we need to look through chains of them. |
12730 | static void getBaseWithConstantOffset(SDValue Loc, SDValue &Base, |
12731 | int64_t& Offset, SelectionDAG &DAG) { |
12732 | if (DAG.isBaseWithConstantOffset(Loc)) { |
12733 | Base = Loc.getOperand(0); |
12734 | Offset += cast<ConstantSDNode>(Loc.getOperand(1))->getSExtValue(); |
12735 | |
12736 | // The base might itself be a base plus an offset, and if so, accumulate |
12737 | // that as well. |
12738 | getBaseWithConstantOffset(Loc.getOperand(0), Base, Offset, DAG); |
12739 | } |
12740 | } |
12741 | |
12742 | static bool isConsecutiveLSLoc(SDValue Loc, EVT VT, LSBaseSDNode *Base, |
12743 | unsigned Bytes, int Dist, |
12744 | SelectionDAG &DAG) { |
12745 | if (VT.getSizeInBits() / 8 != Bytes) |
12746 | return false; |
12747 | |
12748 | SDValue BaseLoc = Base->getBasePtr(); |
12749 | if (Loc.getOpcode() == ISD::FrameIndex) { |
12750 | if (BaseLoc.getOpcode() != ISD::FrameIndex) |
12751 | return false; |
12752 | const MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
12753 | int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); |
12754 | int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); |
12755 | int FS = MFI.getObjectSize(FI); |
12756 | int BFS = MFI.getObjectSize(BFI); |
12757 | if (FS != BFS || FS != (int)Bytes) return false; |
12758 | return MFI.getObjectOffset(FI) == (MFI.getObjectOffset(BFI) + Dist*Bytes); |
12759 | } |
12760 | |
12761 | SDValue Base1 = Loc, Base2 = BaseLoc; |
12762 | int64_t Offset1 = 0, Offset2 = 0; |
12763 | getBaseWithConstantOffset(Loc, Base1, Offset1, DAG); |
12764 | getBaseWithConstantOffset(BaseLoc, Base2, Offset2, DAG); |
12765 | if (Base1 == Base2 && Offset1 == (Offset2 + Dist * Bytes)) |
12766 | return true; |
12767 | |
12768 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
12769 | const GlobalValue *GV1 = nullptr; |
12770 | const GlobalValue *GV2 = nullptr; |
12771 | Offset1 = 0; |
12772 | Offset2 = 0; |
12773 | bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1); |
12774 | bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); |
12775 | if (isGA1 && isGA2 && GV1 == GV2) |
12776 | return Offset1 == (Offset2 + Dist*Bytes); |
12777 | return false; |
12778 | } |
12779 | |
12780 | // Like SelectionDAG::isConsecutiveLoad, but also works for stores, and does |
12781 | // not enforce equality of the chain operands. |
12782 | static bool isConsecutiveLS(SDNode *N, LSBaseSDNode *Base, |
12783 | unsigned Bytes, int Dist, |
12784 | SelectionDAG &DAG) { |
12785 | if (LSBaseSDNode *LS = dyn_cast<LSBaseSDNode>(N)) { |
12786 | EVT VT = LS->getMemoryVT(); |
12787 | SDValue Loc = LS->getBasePtr(); |
12788 | return isConsecutiveLSLoc(Loc, VT, Base, Bytes, Dist, DAG); |
12789 | } |
12790 | |
12791 | if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) { |
12792 | EVT VT; |
12793 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
12794 | default: return false; |
12795 | case Intrinsic::ppc_altivec_lvx: |
12796 | case Intrinsic::ppc_altivec_lvxl: |
12797 | case Intrinsic::ppc_vsx_lxvw4x: |
12798 | case Intrinsic::ppc_vsx_lxvw4x_be: |
12799 | VT = MVT::v4i32; |
12800 | break; |
12801 | case Intrinsic::ppc_vsx_lxvd2x: |
12802 | case Intrinsic::ppc_vsx_lxvd2x_be: |
12803 | VT = MVT::v2f64; |
12804 | break; |
12805 | case Intrinsic::ppc_altivec_lvebx: |
12806 | VT = MVT::i8; |
12807 | break; |
12808 | case Intrinsic::ppc_altivec_lvehx: |
12809 | VT = MVT::i16; |
12810 | break; |
12811 | case Intrinsic::ppc_altivec_lvewx: |
12812 | VT = MVT::i32; |
12813 | break; |
12814 | } |
12815 | |
12816 | return isConsecutiveLSLoc(N->getOperand(2), VT, Base, Bytes, Dist, DAG); |
12817 | } |
12818 | |
12819 | if (N->getOpcode() == ISD::INTRINSIC_VOID) { |
12820 | EVT VT; |
12821 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
12822 | default: return false; |
12823 | case Intrinsic::ppc_altivec_stvx: |
12824 | case Intrinsic::ppc_altivec_stvxl: |
12825 | case Intrinsic::ppc_vsx_stxvw4x: |
12826 | VT = MVT::v4i32; |
12827 | break; |
12828 | case Intrinsic::ppc_vsx_stxvd2x: |
12829 | VT = MVT::v2f64; |
12830 | break; |
12831 | case Intrinsic::ppc_vsx_stxvw4x_be: |
12832 | VT = MVT::v4i32; |
12833 | break; |
12834 | case Intrinsic::ppc_vsx_stxvd2x_be: |
12835 | VT = MVT::v2f64; |
12836 | break; |
12837 | case Intrinsic::ppc_altivec_stvebx: |
12838 | VT = MVT::i8; |
12839 | break; |
12840 | case Intrinsic::ppc_altivec_stvehx: |
12841 | VT = MVT::i16; |
12842 | break; |
12843 | case Intrinsic::ppc_altivec_stvewx: |
12844 | VT = MVT::i32; |
12845 | break; |
12846 | } |
12847 | |
12848 | return isConsecutiveLSLoc(N->getOperand(3), VT, Base, Bytes, Dist, DAG); |
12849 | } |
12850 | |
12851 | return false; |
12852 | } |
12853 | |
12854 | // Return true is there is a nearyby consecutive load to the one provided |
12855 | // (regardless of alignment). We search up and down the chain, looking though |
12856 | // token factors and other loads (but nothing else). As a result, a true result |
12857 | // indicates that it is safe to create a new consecutive load adjacent to the |
12858 | // load provided. |
12859 | static bool findConsecutiveLoad(LoadSDNode *LD, SelectionDAG &DAG) { |
12860 | SDValue Chain = LD->getChain(); |
12861 | EVT VT = LD->getMemoryVT(); |
12862 | |
12863 | SmallSet<SDNode *, 16> LoadRoots; |
12864 | SmallVector<SDNode *, 8> Queue(1, Chain.getNode()); |
12865 | SmallSet<SDNode *, 16> Visited; |
12866 | |
12867 | // First, search up the chain, branching to follow all token-factor operands. |
12868 | // If we find a consecutive load, then we're done, otherwise, record all |
12869 | // nodes just above the top-level loads and token factors. |
12870 | while (!Queue.empty()) { |
12871 | SDNode *ChainNext = Queue.pop_back_val(); |
12872 | if (!Visited.insert(ChainNext).second) |
12873 | continue; |
12874 | |
12875 | if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(ChainNext)) { |
12876 | if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG)) |
12877 | return true; |
12878 | |
12879 | if (!Visited.count(ChainLD->getChain().getNode())) |
12880 | Queue.push_back(ChainLD->getChain().getNode()); |
12881 | } else if (ChainNext->getOpcode() == ISD::TokenFactor) { |
12882 | for (const SDUse &O : ChainNext->ops()) |
12883 | if (!Visited.count(O.getNode())) |
12884 | Queue.push_back(O.getNode()); |
12885 | } else |
12886 | LoadRoots.insert(ChainNext); |
12887 | } |
12888 | |
12889 | // Second, search down the chain, starting from the top-level nodes recorded |
12890 | // in the first phase. These top-level nodes are the nodes just above all |
12891 | // loads and token factors. Starting with their uses, recursively look though |
12892 | // all loads (just the chain uses) and token factors to find a consecutive |
12893 | // load. |
12894 | Visited.clear(); |
12895 | Queue.clear(); |
12896 | |
12897 | for (SmallSet<SDNode *, 16>::iterator I = LoadRoots.begin(), |
12898 | IE = LoadRoots.end(); I != IE; ++I) { |
12899 | Queue.push_back(*I); |
12900 | |
12901 | while (!Queue.empty()) { |
12902 | SDNode *LoadRoot = Queue.pop_back_val(); |
12903 | if (!Visited.insert(LoadRoot).second) |
12904 | continue; |
12905 | |
12906 | if (MemSDNode *ChainLD = dyn_cast<MemSDNode>(LoadRoot)) |
12907 | if (isConsecutiveLS(ChainLD, LD, VT.getStoreSize(), 1, DAG)) |
12908 | return true; |
12909 | |
12910 | for (SDNode::use_iterator UI = LoadRoot->use_begin(), |
12911 | UE = LoadRoot->use_end(); UI != UE; ++UI) |
12912 | if (((isa<MemSDNode>(*UI) && |
12913 | cast<MemSDNode>(*UI)->getChain().getNode() == LoadRoot) || |
12914 | UI->getOpcode() == ISD::TokenFactor) && !Visited.count(*UI)) |
12915 | Queue.push_back(*UI); |
12916 | } |
12917 | } |
12918 | |
12919 | return false; |
12920 | } |
12921 | |
12922 | /// This function is called when we have proved that a SETCC node can be replaced |
12923 | /// by subtraction (and other supporting instructions) so that the result of |
12924 | /// comparison is kept in a GPR instead of CR. This function is purely for |
12925 | /// codegen purposes and has some flags to guide the codegen process. |
12926 | static SDValue generateEquivalentSub(SDNode *N, int Size, bool Complement, |
12927 | bool Swap, SDLoc &DL, SelectionDAG &DAG) { |
12928 | assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.")((N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected." ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"ISD::SETCC Expected.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12928, __PRETTY_FUNCTION__)); |
12929 | |
12930 | // Zero extend the operands to the largest legal integer. Originally, they |
12931 | // must be of a strictly smaller size. |
12932 | auto Op0 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(0), |
12933 | DAG.getConstant(Size, DL, MVT::i32)); |
12934 | auto Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N->getOperand(1), |
12935 | DAG.getConstant(Size, DL, MVT::i32)); |
12936 | |
12937 | // Swap if needed. Depends on the condition code. |
12938 | if (Swap) |
12939 | std::swap(Op0, Op1); |
12940 | |
12941 | // Subtract extended integers. |
12942 | auto SubNode = DAG.getNode(ISD::SUB, DL, MVT::i64, Op0, Op1); |
12943 | |
12944 | // Move the sign bit to the least significant position and zero out the rest. |
12945 | // Now the least significant bit carries the result of original comparison. |
12946 | auto Shifted = DAG.getNode(ISD::SRL, DL, MVT::i64, SubNode, |
12947 | DAG.getConstant(Size - 1, DL, MVT::i32)); |
12948 | auto Final = Shifted; |
12949 | |
12950 | // Complement the result if needed. Based on the condition code. |
12951 | if (Complement) |
12952 | Final = DAG.getNode(ISD::XOR, DL, MVT::i64, Shifted, |
12953 | DAG.getConstant(1, DL, MVT::i64)); |
12954 | |
12955 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Final); |
12956 | } |
12957 | |
12958 | SDValue PPCTargetLowering::ConvertSETCCToSubtract(SDNode *N, |
12959 | DAGCombinerInfo &DCI) const { |
12960 | assert(N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected.")((N->getOpcode() == ISD::SETCC && "ISD::SETCC Expected." ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"ISD::SETCC Expected.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 12960, __PRETTY_FUNCTION__)); |
12961 | |
12962 | SelectionDAG &DAG = DCI.DAG; |
12963 | SDLoc DL(N); |
12964 | |
12965 | // Size of integers being compared has a critical role in the following |
12966 | // analysis, so we prefer to do this when all types are legal. |
12967 | if (!DCI.isAfterLegalizeDAG()) |
12968 | return SDValue(); |
12969 | |
12970 | // If all users of SETCC extend its value to a legal integer type |
12971 | // then we replace SETCC with a subtraction |
12972 | for (SDNode::use_iterator UI = N->use_begin(), |
12973 | UE = N->use_end(); UI != UE; ++UI) { |
12974 | if (UI->getOpcode() != ISD::ZERO_EXTEND) |
12975 | return SDValue(); |
12976 | } |
12977 | |
12978 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
12979 | auto OpSize = N->getOperand(0).getValueSizeInBits(); |
12980 | |
12981 | unsigned Size = DAG.getDataLayout().getLargestLegalIntTypeSizeInBits(); |
12982 | |
12983 | if (OpSize < Size) { |
12984 | switch (CC) { |
12985 | default: break; |
12986 | case ISD::SETULT: |
12987 | return generateEquivalentSub(N, Size, false, false, DL, DAG); |
12988 | case ISD::SETULE: |
12989 | return generateEquivalentSub(N, Size, true, true, DL, DAG); |
12990 | case ISD::SETUGT: |
12991 | return generateEquivalentSub(N, Size, false, true, DL, DAG); |
12992 | case ISD::SETUGE: |
12993 | return generateEquivalentSub(N, Size, true, false, DL, DAG); |
12994 | } |
12995 | } |
12996 | |
12997 | return SDValue(); |
12998 | } |
12999 | |
13000 | SDValue PPCTargetLowering::DAGCombineTruncBoolExt(SDNode *N, |
13001 | DAGCombinerInfo &DCI) const { |
13002 | SelectionDAG &DAG = DCI.DAG; |
13003 | SDLoc dl(N); |
13004 | |
13005 | assert(Subtarget.useCRBits() && "Expecting to be tracking CR bits")((Subtarget.useCRBits() && "Expecting to be tracking CR bits" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.useCRBits() && \"Expecting to be tracking CR bits\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13005, __PRETTY_FUNCTION__)); |
13006 | // If we're tracking CR bits, we need to be careful that we don't have: |
13007 | // trunc(binary-ops(zext(x), zext(y))) |
13008 | // or |
13009 | // trunc(binary-ops(binary-ops(zext(x), zext(y)), ...) |
13010 | // such that we're unnecessarily moving things into GPRs when it would be |
13011 | // better to keep them in CR bits. |
13012 | |
13013 | // Note that trunc here can be an actual i1 trunc, or can be the effective |
13014 | // truncation that comes from a setcc or select_cc. |
13015 | if (N->getOpcode() == ISD::TRUNCATE && |
13016 | N->getValueType(0) != MVT::i1) |
13017 | return SDValue(); |
13018 | |
13019 | if (N->getOperand(0).getValueType() != MVT::i32 && |
13020 | N->getOperand(0).getValueType() != MVT::i64) |
13021 | return SDValue(); |
13022 | |
13023 | if (N->getOpcode() == ISD::SETCC || |
13024 | N->getOpcode() == ISD::SELECT_CC) { |
13025 | // If we're looking at a comparison, then we need to make sure that the |
13026 | // high bits (all except for the first) don't matter the result. |
13027 | ISD::CondCode CC = |
13028 | cast<CondCodeSDNode>(N->getOperand( |
13029 | N->getOpcode() == ISD::SETCC ? 2 : 4))->get(); |
13030 | unsigned OpBits = N->getOperand(0).getValueSizeInBits(); |
13031 | |
13032 | if (ISD::isSignedIntSetCC(CC)) { |
13033 | if (DAG.ComputeNumSignBits(N->getOperand(0)) != OpBits || |
13034 | DAG.ComputeNumSignBits(N->getOperand(1)) != OpBits) |
13035 | return SDValue(); |
13036 | } else if (ISD::isUnsignedIntSetCC(CC)) { |
13037 | if (!DAG.MaskedValueIsZero(N->getOperand(0), |
13038 | APInt::getHighBitsSet(OpBits, OpBits-1)) || |
13039 | !DAG.MaskedValueIsZero(N->getOperand(1), |
13040 | APInt::getHighBitsSet(OpBits, OpBits-1))) |
13041 | return (N->getOpcode() == ISD::SETCC ? ConvertSETCCToSubtract(N, DCI) |
13042 | : SDValue()); |
13043 | } else { |
13044 | // This is neither a signed nor an unsigned comparison, just make sure |
13045 | // that the high bits are equal. |
13046 | KnownBits Op1Known = DAG.computeKnownBits(N->getOperand(0)); |
13047 | KnownBits Op2Known = DAG.computeKnownBits(N->getOperand(1)); |
13048 | |
13049 | // We don't really care about what is known about the first bit (if |
13050 | // anything), so clear it in all masks prior to comparing them. |
13051 | Op1Known.Zero.clearBit(0); Op1Known.One.clearBit(0); |
13052 | Op2Known.Zero.clearBit(0); Op2Known.One.clearBit(0); |
13053 | |
13054 | if (Op1Known.Zero != Op2Known.Zero || Op1Known.One != Op2Known.One) |
13055 | return SDValue(); |
13056 | } |
13057 | } |
13058 | |
13059 | // We now know that the higher-order bits are irrelevant, we just need to |
13060 | // make sure that all of the intermediate operations are bit operations, and |
13061 | // all inputs are extensions. |
13062 | if (N->getOperand(0).getOpcode() != ISD::AND && |
13063 | N->getOperand(0).getOpcode() != ISD::OR && |
13064 | N->getOperand(0).getOpcode() != ISD::XOR && |
13065 | N->getOperand(0).getOpcode() != ISD::SELECT && |
13066 | N->getOperand(0).getOpcode() != ISD::SELECT_CC && |
13067 | N->getOperand(0).getOpcode() != ISD::TRUNCATE && |
13068 | N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND && |
13069 | N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND && |
13070 | N->getOperand(0).getOpcode() != ISD::ANY_EXTEND) |
13071 | return SDValue(); |
13072 | |
13073 | if ((N->getOpcode() == ISD::SETCC || N->getOpcode() == ISD::SELECT_CC) && |
13074 | N->getOperand(1).getOpcode() != ISD::AND && |
13075 | N->getOperand(1).getOpcode() != ISD::OR && |
13076 | N->getOperand(1).getOpcode() != ISD::XOR && |
13077 | N->getOperand(1).getOpcode() != ISD::SELECT && |
13078 | N->getOperand(1).getOpcode() != ISD::SELECT_CC && |
13079 | N->getOperand(1).getOpcode() != ISD::TRUNCATE && |
13080 | N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND && |
13081 | N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND && |
13082 | N->getOperand(1).getOpcode() != ISD::ANY_EXTEND) |
13083 | return SDValue(); |
13084 | |
13085 | SmallVector<SDValue, 4> Inputs; |
13086 | SmallVector<SDValue, 8> BinOps, PromOps; |
13087 | SmallPtrSet<SDNode *, 16> Visited; |
13088 | |
13089 | for (unsigned i = 0; i < 2; ++i) { |
13090 | if (((N->getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13091 | N->getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13092 | N->getOperand(i).getOpcode() == ISD::ANY_EXTEND) && |
13093 | N->getOperand(i).getOperand(0).getValueType() == MVT::i1) || |
13094 | isa<ConstantSDNode>(N->getOperand(i))) |
13095 | Inputs.push_back(N->getOperand(i)); |
13096 | else |
13097 | BinOps.push_back(N->getOperand(i)); |
13098 | |
13099 | if (N->getOpcode() == ISD::TRUNCATE) |
13100 | break; |
13101 | } |
13102 | |
13103 | // Visit all inputs, collect all binary operations (and, or, xor and |
13104 | // select) that are all fed by extensions. |
13105 | while (!BinOps.empty()) { |
13106 | SDValue BinOp = BinOps.back(); |
13107 | BinOps.pop_back(); |
13108 | |
13109 | if (!Visited.insert(BinOp.getNode()).second) |
13110 | continue; |
13111 | |
13112 | PromOps.push_back(BinOp); |
13113 | |
13114 | for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) { |
13115 | // The condition of the select is not promoted. |
13116 | if (BinOp.getOpcode() == ISD::SELECT && i == 0) |
13117 | continue; |
13118 | if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3) |
13119 | continue; |
13120 | |
13121 | if (((BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13122 | BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13123 | BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) && |
13124 | BinOp.getOperand(i).getOperand(0).getValueType() == MVT::i1) || |
13125 | isa<ConstantSDNode>(BinOp.getOperand(i))) { |
13126 | Inputs.push_back(BinOp.getOperand(i)); |
13127 | } else if (BinOp.getOperand(i).getOpcode() == ISD::AND || |
13128 | BinOp.getOperand(i).getOpcode() == ISD::OR || |
13129 | BinOp.getOperand(i).getOpcode() == ISD::XOR || |
13130 | BinOp.getOperand(i).getOpcode() == ISD::SELECT || |
13131 | BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC || |
13132 | BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE || |
13133 | BinOp.getOperand(i).getOpcode() == ISD::SIGN_EXTEND || |
13134 | BinOp.getOperand(i).getOpcode() == ISD::ZERO_EXTEND || |
13135 | BinOp.getOperand(i).getOpcode() == ISD::ANY_EXTEND) { |
13136 | BinOps.push_back(BinOp.getOperand(i)); |
13137 | } else { |
13138 | // We have an input that is not an extension or another binary |
13139 | // operation; we'll abort this transformation. |
13140 | return SDValue(); |
13141 | } |
13142 | } |
13143 | } |
13144 | |
13145 | // Make sure that this is a self-contained cluster of operations (which |
13146 | // is not quite the same thing as saying that everything has only one |
13147 | // use). |
13148 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13149 | if (isa<ConstantSDNode>(Inputs[i])) |
13150 | continue; |
13151 | |
13152 | for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(), |
13153 | UE = Inputs[i].getNode()->use_end(); |
13154 | UI != UE; ++UI) { |
13155 | SDNode *User = *UI; |
13156 | if (User != N && !Visited.count(User)) |
13157 | return SDValue(); |
13158 | |
13159 | // Make sure that we're not going to promote the non-output-value |
13160 | // operand(s) or SELECT or SELECT_CC. |
13161 | // FIXME: Although we could sometimes handle this, and it does occur in |
13162 | // practice that one of the condition inputs to the select is also one of |
13163 | // the outputs, we currently can't deal with this. |
13164 | if (User->getOpcode() == ISD::SELECT) { |
13165 | if (User->getOperand(0) == Inputs[i]) |
13166 | return SDValue(); |
13167 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13168 | if (User->getOperand(0) == Inputs[i] || |
13169 | User->getOperand(1) == Inputs[i]) |
13170 | return SDValue(); |
13171 | } |
13172 | } |
13173 | } |
13174 | |
13175 | for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) { |
13176 | for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(), |
13177 | UE = PromOps[i].getNode()->use_end(); |
13178 | UI != UE; ++UI) { |
13179 | SDNode *User = *UI; |
13180 | if (User != N && !Visited.count(User)) |
13181 | return SDValue(); |
13182 | |
13183 | // Make sure that we're not going to promote the non-output-value |
13184 | // operand(s) or SELECT or SELECT_CC. |
13185 | // FIXME: Although we could sometimes handle this, and it does occur in |
13186 | // practice that one of the condition inputs to the select is also one of |
13187 | // the outputs, we currently can't deal with this. |
13188 | if (User->getOpcode() == ISD::SELECT) { |
13189 | if (User->getOperand(0) == PromOps[i]) |
13190 | return SDValue(); |
13191 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13192 | if (User->getOperand(0) == PromOps[i] || |
13193 | User->getOperand(1) == PromOps[i]) |
13194 | return SDValue(); |
13195 | } |
13196 | } |
13197 | } |
13198 | |
13199 | // Replace all inputs with the extension operand. |
13200 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13201 | // Constants may have users outside the cluster of to-be-promoted nodes, |
13202 | // and so we need to replace those as we do the promotions. |
13203 | if (isa<ConstantSDNode>(Inputs[i])) |
13204 | continue; |
13205 | else |
13206 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], Inputs[i].getOperand(0)); |
13207 | } |
13208 | |
13209 | std::list<HandleSDNode> PromOpHandles; |
13210 | for (auto &PromOp : PromOps) |
13211 | PromOpHandles.emplace_back(PromOp); |
13212 | |
13213 | // Replace all operations (these are all the same, but have a different |
13214 | // (i1) return type). DAG.getNode will validate that the types of |
13215 | // a binary operator match, so go through the list in reverse so that |
13216 | // we've likely promoted both operands first. Any intermediate truncations or |
13217 | // extensions disappear. |
13218 | while (!PromOpHandles.empty()) { |
13219 | SDValue PromOp = PromOpHandles.back().getValue(); |
13220 | PromOpHandles.pop_back(); |
13221 | |
13222 | if (PromOp.getOpcode() == ISD::TRUNCATE || |
13223 | PromOp.getOpcode() == ISD::SIGN_EXTEND || |
13224 | PromOp.getOpcode() == ISD::ZERO_EXTEND || |
13225 | PromOp.getOpcode() == ISD::ANY_EXTEND) { |
13226 | if (!isa<ConstantSDNode>(PromOp.getOperand(0)) && |
13227 | PromOp.getOperand(0).getValueType() != MVT::i1) { |
13228 | // The operand is not yet ready (see comment below). |
13229 | PromOpHandles.emplace_front(PromOp); |
13230 | continue; |
13231 | } |
13232 | |
13233 | SDValue RepValue = PromOp.getOperand(0); |
13234 | if (isa<ConstantSDNode>(RepValue)) |
13235 | RepValue = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, RepValue); |
13236 | |
13237 | DAG.ReplaceAllUsesOfValueWith(PromOp, RepValue); |
13238 | continue; |
13239 | } |
13240 | |
13241 | unsigned C; |
13242 | switch (PromOp.getOpcode()) { |
13243 | default: C = 0; break; |
13244 | case ISD::SELECT: C = 1; break; |
13245 | case ISD::SELECT_CC: C = 2; break; |
13246 | } |
13247 | |
13248 | if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) && |
13249 | PromOp.getOperand(C).getValueType() != MVT::i1) || |
13250 | (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) && |
13251 | PromOp.getOperand(C+1).getValueType() != MVT::i1)) { |
13252 | // The to-be-promoted operands of this node have not yet been |
13253 | // promoted (this should be rare because we're going through the |
13254 | // list backward, but if one of the operands has several users in |
13255 | // this cluster of to-be-promoted nodes, it is possible). |
13256 | PromOpHandles.emplace_front(PromOp); |
13257 | continue; |
13258 | } |
13259 | |
13260 | SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(), |
13261 | PromOp.getNode()->op_end()); |
13262 | |
13263 | // If there are any constant inputs, make sure they're replaced now. |
13264 | for (unsigned i = 0; i < 2; ++i) |
13265 | if (isa<ConstantSDNode>(Ops[C+i])) |
13266 | Ops[C+i] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ops[C+i]); |
13267 | |
13268 | DAG.ReplaceAllUsesOfValueWith(PromOp, |
13269 | DAG.getNode(PromOp.getOpcode(), dl, MVT::i1, Ops)); |
13270 | } |
13271 | |
13272 | // Now we're left with the initial truncation itself. |
13273 | if (N->getOpcode() == ISD::TRUNCATE) |
13274 | return N->getOperand(0); |
13275 | |
13276 | // Otherwise, this is a comparison. The operands to be compared have just |
13277 | // changed type (to i1), but everything else is the same. |
13278 | return SDValue(N, 0); |
13279 | } |
13280 | |
13281 | SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N, |
13282 | DAGCombinerInfo &DCI) const { |
13283 | SelectionDAG &DAG = DCI.DAG; |
13284 | SDLoc dl(N); |
13285 | |
13286 | // If we're tracking CR bits, we need to be careful that we don't have: |
13287 | // zext(binary-ops(trunc(x), trunc(y))) |
13288 | // or |
13289 | // zext(binary-ops(binary-ops(trunc(x), trunc(y)), ...) |
13290 | // such that we're unnecessarily moving things into CR bits that can more |
13291 | // efficiently stay in GPRs. Note that if we're not certain that the high |
13292 | // bits are set as required by the final extension, we still may need to do |
13293 | // some masking to get the proper behavior. |
13294 | |
13295 | // This same functionality is important on PPC64 when dealing with |
13296 | // 32-to-64-bit extensions; these occur often when 32-bit values are used as |
13297 | // the return values of functions. Because it is so similar, it is handled |
13298 | // here as well. |
13299 | |
13300 | if (N->getValueType(0) != MVT::i32 && |
13301 | N->getValueType(0) != MVT::i64) |
13302 | return SDValue(); |
13303 | |
13304 | if (!((N->getOperand(0).getValueType() == MVT::i1 && Subtarget.useCRBits()) || |
13305 | (N->getOperand(0).getValueType() == MVT::i32 && Subtarget.isPPC64()))) |
13306 | return SDValue(); |
13307 | |
13308 | if (N->getOperand(0).getOpcode() != ISD::AND && |
13309 | N->getOperand(0).getOpcode() != ISD::OR && |
13310 | N->getOperand(0).getOpcode() != ISD::XOR && |
13311 | N->getOperand(0).getOpcode() != ISD::SELECT && |
13312 | N->getOperand(0).getOpcode() != ISD::SELECT_CC) |
13313 | return SDValue(); |
13314 | |
13315 | SmallVector<SDValue, 4> Inputs; |
13316 | SmallVector<SDValue, 8> BinOps(1, N->getOperand(0)), PromOps; |
13317 | SmallPtrSet<SDNode *, 16> Visited; |
13318 | |
13319 | // Visit all inputs, collect all binary operations (and, or, xor and |
13320 | // select) that are all fed by truncations. |
13321 | while (!BinOps.empty()) { |
13322 | SDValue BinOp = BinOps.back(); |
13323 | BinOps.pop_back(); |
13324 | |
13325 | if (!Visited.insert(BinOp.getNode()).second) |
13326 | continue; |
13327 | |
13328 | PromOps.push_back(BinOp); |
13329 | |
13330 | for (unsigned i = 0, ie = BinOp.getNumOperands(); i != ie; ++i) { |
13331 | // The condition of the select is not promoted. |
13332 | if (BinOp.getOpcode() == ISD::SELECT && i == 0) |
13333 | continue; |
13334 | if (BinOp.getOpcode() == ISD::SELECT_CC && i != 2 && i != 3) |
13335 | continue; |
13336 | |
13337 | if (BinOp.getOperand(i).getOpcode() == ISD::TRUNCATE || |
13338 | isa<ConstantSDNode>(BinOp.getOperand(i))) { |
13339 | Inputs.push_back(BinOp.getOperand(i)); |
13340 | } else if (BinOp.getOperand(i).getOpcode() == ISD::AND || |
13341 | BinOp.getOperand(i).getOpcode() == ISD::OR || |
13342 | BinOp.getOperand(i).getOpcode() == ISD::XOR || |
13343 | BinOp.getOperand(i).getOpcode() == ISD::SELECT || |
13344 | BinOp.getOperand(i).getOpcode() == ISD::SELECT_CC) { |
13345 | BinOps.push_back(BinOp.getOperand(i)); |
13346 | } else { |
13347 | // We have an input that is not a truncation or another binary |
13348 | // operation; we'll abort this transformation. |
13349 | return SDValue(); |
13350 | } |
13351 | } |
13352 | } |
13353 | |
13354 | // The operands of a select that must be truncated when the select is |
13355 | // promoted because the operand is actually part of the to-be-promoted set. |
13356 | DenseMap<SDNode *, EVT> SelectTruncOp[2]; |
13357 | |
13358 | // Make sure that this is a self-contained cluster of operations (which |
13359 | // is not quite the same thing as saying that everything has only one |
13360 | // use). |
13361 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13362 | if (isa<ConstantSDNode>(Inputs[i])) |
13363 | continue; |
13364 | |
13365 | for (SDNode::use_iterator UI = Inputs[i].getNode()->use_begin(), |
13366 | UE = Inputs[i].getNode()->use_end(); |
13367 | UI != UE; ++UI) { |
13368 | SDNode *User = *UI; |
13369 | if (User != N && !Visited.count(User)) |
13370 | return SDValue(); |
13371 | |
13372 | // If we're going to promote the non-output-value operand(s) or SELECT or |
13373 | // SELECT_CC, record them for truncation. |
13374 | if (User->getOpcode() == ISD::SELECT) { |
13375 | if (User->getOperand(0) == Inputs[i]) |
13376 | SelectTruncOp[0].insert(std::make_pair(User, |
13377 | User->getOperand(0).getValueType())); |
13378 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13379 | if (User->getOperand(0) == Inputs[i]) |
13380 | SelectTruncOp[0].insert(std::make_pair(User, |
13381 | User->getOperand(0).getValueType())); |
13382 | if (User->getOperand(1) == Inputs[i]) |
13383 | SelectTruncOp[1].insert(std::make_pair(User, |
13384 | User->getOperand(1).getValueType())); |
13385 | } |
13386 | } |
13387 | } |
13388 | |
13389 | for (unsigned i = 0, ie = PromOps.size(); i != ie; ++i) { |
13390 | for (SDNode::use_iterator UI = PromOps[i].getNode()->use_begin(), |
13391 | UE = PromOps[i].getNode()->use_end(); |
13392 | UI != UE; ++UI) { |
13393 | SDNode *User = *UI; |
13394 | if (User != N && !Visited.count(User)) |
13395 | return SDValue(); |
13396 | |
13397 | // If we're going to promote the non-output-value operand(s) or SELECT or |
13398 | // SELECT_CC, record them for truncation. |
13399 | if (User->getOpcode() == ISD::SELECT) { |
13400 | if (User->getOperand(0) == PromOps[i]) |
13401 | SelectTruncOp[0].insert(std::make_pair(User, |
13402 | User->getOperand(0).getValueType())); |
13403 | } else if (User->getOpcode() == ISD::SELECT_CC) { |
13404 | if (User->getOperand(0) == PromOps[i]) |
13405 | SelectTruncOp[0].insert(std::make_pair(User, |
13406 | User->getOperand(0).getValueType())); |
13407 | if (User->getOperand(1) == PromOps[i]) |
13408 | SelectTruncOp[1].insert(std::make_pair(User, |
13409 | User->getOperand(1).getValueType())); |
13410 | } |
13411 | } |
13412 | } |
13413 | |
13414 | unsigned PromBits = N->getOperand(0).getValueSizeInBits(); |
13415 | bool ReallyNeedsExt = false; |
13416 | if (N->getOpcode() != ISD::ANY_EXTEND) { |
13417 | // If all of the inputs are not already sign/zero extended, then |
13418 | // we'll still need to do that at the end. |
13419 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13420 | if (isa<ConstantSDNode>(Inputs[i])) |
13421 | continue; |
13422 | |
13423 | unsigned OpBits = |
13424 | Inputs[i].getOperand(0).getValueSizeInBits(); |
13425 | assert(PromBits < OpBits && "Truncation not to a smaller bit count?")((PromBits < OpBits && "Truncation not to a smaller bit count?" ) ? static_cast<void> (0) : __assert_fail ("PromBits < OpBits && \"Truncation not to a smaller bit count?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13425, __PRETTY_FUNCTION__)); |
13426 | |
13427 | if ((N->getOpcode() == ISD::ZERO_EXTEND && |
13428 | !DAG.MaskedValueIsZero(Inputs[i].getOperand(0), |
13429 | APInt::getHighBitsSet(OpBits, |
13430 | OpBits-PromBits))) || |
13431 | (N->getOpcode() == ISD::SIGN_EXTEND && |
13432 | DAG.ComputeNumSignBits(Inputs[i].getOperand(0)) < |
13433 | (OpBits-(PromBits-1)))) { |
13434 | ReallyNeedsExt = true; |
13435 | break; |
13436 | } |
13437 | } |
13438 | } |
13439 | |
13440 | // Replace all inputs, either with the truncation operand, or a |
13441 | // truncation or extension to the final output type. |
13442 | for (unsigned i = 0, ie = Inputs.size(); i != ie; ++i) { |
13443 | // Constant inputs need to be replaced with the to-be-promoted nodes that |
13444 | // use them because they might have users outside of the cluster of |
13445 | // promoted nodes. |
13446 | if (isa<ConstantSDNode>(Inputs[i])) |
13447 | continue; |
13448 | |
13449 | SDValue InSrc = Inputs[i].getOperand(0); |
13450 | if (Inputs[i].getValueType() == N->getValueType(0)) |
13451 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], InSrc); |
13452 | else if (N->getOpcode() == ISD::SIGN_EXTEND) |
13453 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13454 | DAG.getSExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13455 | else if (N->getOpcode() == ISD::ZERO_EXTEND) |
13456 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13457 | DAG.getZExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13458 | else |
13459 | DAG.ReplaceAllUsesOfValueWith(Inputs[i], |
13460 | DAG.getAnyExtOrTrunc(InSrc, dl, N->getValueType(0))); |
13461 | } |
13462 | |
13463 | std::list<HandleSDNode> PromOpHandles; |
13464 | for (auto &PromOp : PromOps) |
13465 | PromOpHandles.emplace_back(PromOp); |
13466 | |
13467 | // Replace all operations (these are all the same, but have a different |
13468 | // (promoted) return type). DAG.getNode will validate that the types of |
13469 | // a binary operator match, so go through the list in reverse so that |
13470 | // we've likely promoted both operands first. |
13471 | while (!PromOpHandles.empty()) { |
13472 | SDValue PromOp = PromOpHandles.back().getValue(); |
13473 | PromOpHandles.pop_back(); |
13474 | |
13475 | unsigned C; |
13476 | switch (PromOp.getOpcode()) { |
13477 | default: C = 0; break; |
13478 | case ISD::SELECT: C = 1; break; |
13479 | case ISD::SELECT_CC: C = 2; break; |
13480 | } |
13481 | |
13482 | if ((!isa<ConstantSDNode>(PromOp.getOperand(C)) && |
13483 | PromOp.getOperand(C).getValueType() != N->getValueType(0)) || |
13484 | (!isa<ConstantSDNode>(PromOp.getOperand(C+1)) && |
13485 | PromOp.getOperand(C+1).getValueType() != N->getValueType(0))) { |
13486 | // The to-be-promoted operands of this node have not yet been |
13487 | // promoted (this should be rare because we're going through the |
13488 | // list backward, but if one of the operands has several users in |
13489 | // this cluster of to-be-promoted nodes, it is possible). |
13490 | PromOpHandles.emplace_front(PromOp); |
13491 | continue; |
13492 | } |
13493 | |
13494 | // For SELECT and SELECT_CC nodes, we do a similar check for any |
13495 | // to-be-promoted comparison inputs. |
13496 | if (PromOp.getOpcode() == ISD::SELECT || |
13497 | PromOp.getOpcode() == ISD::SELECT_CC) { |
13498 | if ((SelectTruncOp[0].count(PromOp.getNode()) && |
13499 | PromOp.getOperand(0).getValueType() != N->getValueType(0)) || |
13500 | (SelectTruncOp[1].count(PromOp.getNode()) && |
13501 | PromOp.getOperand(1).getValueType() != N->getValueType(0))) { |
13502 | PromOpHandles.emplace_front(PromOp); |
13503 | continue; |
13504 | } |
13505 | } |
13506 | |
13507 | SmallVector<SDValue, 3> Ops(PromOp.getNode()->op_begin(), |
13508 | PromOp.getNode()->op_end()); |
13509 | |
13510 | // If this node has constant inputs, then they'll need to be promoted here. |
13511 | for (unsigned i = 0; i < 2; ++i) { |
13512 | if (!isa<ConstantSDNode>(Ops[C+i])) |
13513 | continue; |
13514 | if (Ops[C+i].getValueType() == N->getValueType(0)) |
13515 | continue; |
13516 | |
13517 | if (N->getOpcode() == ISD::SIGN_EXTEND) |
13518 | Ops[C+i] = DAG.getSExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13519 | else if (N->getOpcode() == ISD::ZERO_EXTEND) |
13520 | Ops[C+i] = DAG.getZExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13521 | else |
13522 | Ops[C+i] = DAG.getAnyExtOrTrunc(Ops[C+i], dl, N->getValueType(0)); |
13523 | } |
13524 | |
13525 | // If we've promoted the comparison inputs of a SELECT or SELECT_CC, |
13526 | // truncate them again to the original value type. |
13527 | if (PromOp.getOpcode() == ISD::SELECT || |
13528 | PromOp.getOpcode() == ISD::SELECT_CC) { |
13529 | auto SI0 = SelectTruncOp[0].find(PromOp.getNode()); |
13530 | if (SI0 != SelectTruncOp[0].end()) |
13531 | Ops[0] = DAG.getNode(ISD::TRUNCATE, dl, SI0->second, Ops[0]); |
13532 | auto SI1 = SelectTruncOp[1].find(PromOp.getNode()); |
13533 | if (SI1 != SelectTruncOp[1].end()) |
13534 | Ops[1] = DAG.getNode(ISD::TRUNCATE, dl, SI1->second, Ops[1]); |
13535 | } |
13536 | |
13537 | DAG.ReplaceAllUsesOfValueWith(PromOp, |
13538 | DAG.getNode(PromOp.getOpcode(), dl, N->getValueType(0), Ops)); |
13539 | } |
13540 | |
13541 | // Now we're left with the initial extension itself. |
13542 | if (!ReallyNeedsExt) |
13543 | return N->getOperand(0); |
13544 | |
13545 | // To zero extend, just mask off everything except for the first bit (in the |
13546 | // i1 case). |
13547 | if (N->getOpcode() == ISD::ZERO_EXTEND) |
13548 | return DAG.getNode(ISD::AND, dl, N->getValueType(0), N->getOperand(0), |
13549 | DAG.getConstant(APInt::getLowBitsSet( |
13550 | N->getValueSizeInBits(0), PromBits), |
13551 | dl, N->getValueType(0))); |
13552 | |
13553 | assert(N->getOpcode() == ISD::SIGN_EXTEND &&((N->getOpcode() == ISD::SIGN_EXTEND && "Invalid extension type" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SIGN_EXTEND && \"Invalid extension type\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13554, __PRETTY_FUNCTION__)) |
13554 | "Invalid extension type")((N->getOpcode() == ISD::SIGN_EXTEND && "Invalid extension type" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SIGN_EXTEND && \"Invalid extension type\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13554, __PRETTY_FUNCTION__)); |
13555 | EVT ShiftAmountTy = getShiftAmountTy(N->getValueType(0), DAG.getDataLayout()); |
13556 | SDValue ShiftCst = |
13557 | DAG.getConstant(N->getValueSizeInBits(0) - PromBits, dl, ShiftAmountTy); |
13558 | return DAG.getNode( |
13559 | ISD::SRA, dl, N->getValueType(0), |
13560 | DAG.getNode(ISD::SHL, dl, N->getValueType(0), N->getOperand(0), ShiftCst), |
13561 | ShiftCst); |
13562 | } |
13563 | |
13564 | SDValue PPCTargetLowering::combineSetCC(SDNode *N, |
13565 | DAGCombinerInfo &DCI) const { |
13566 | assert(N->getOpcode() == ISD::SETCC &&((N->getOpcode() == ISD::SETCC && "Should be called with a SETCC node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"Should be called with a SETCC node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13567, __PRETTY_FUNCTION__)) |
13567 | "Should be called with a SETCC node")((N->getOpcode() == ISD::SETCC && "Should be called with a SETCC node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::SETCC && \"Should be called with a SETCC node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13567, __PRETTY_FUNCTION__)); |
13568 | |
13569 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); |
13570 | if (CC == ISD::SETNE || CC == ISD::SETEQ) { |
13571 | SDValue LHS = N->getOperand(0); |
13572 | SDValue RHS = N->getOperand(1); |
13573 | |
13574 | // If there is a '0 - y' pattern, canonicalize the pattern to the RHS. |
13575 | if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) && |
13576 | LHS.hasOneUse()) |
13577 | std::swap(LHS, RHS); |
13578 | |
13579 | // x == 0-y --> x+y == 0 |
13580 | // x != 0-y --> x+y != 0 |
13581 | if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) && |
13582 | RHS.hasOneUse()) { |
13583 | SDLoc DL(N); |
13584 | SelectionDAG &DAG = DCI.DAG; |
13585 | EVT VT = N->getValueType(0); |
13586 | EVT OpVT = LHS.getValueType(); |
13587 | SDValue Add = DAG.getNode(ISD::ADD, DL, OpVT, LHS, RHS.getOperand(1)); |
13588 | return DAG.getSetCC(DL, VT, Add, DAG.getConstant(0, DL, OpVT), CC); |
13589 | } |
13590 | } |
13591 | |
13592 | return DAGCombineTruncBoolExt(N, DCI); |
13593 | } |
13594 | |
13595 | // Is this an extending load from an f32 to an f64? |
13596 | static bool isFPExtLoad(SDValue Op) { |
13597 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op.getNode())) |
13598 | return LD->getExtensionType() == ISD::EXTLOAD && |
13599 | Op.getValueType() == MVT::f64; |
13600 | return false; |
13601 | } |
13602 | |
13603 | /// Reduces the number of fp-to-int conversion when building a vector. |
13604 | /// |
13605 | /// If this vector is built out of floating to integer conversions, |
13606 | /// transform it to a vector built out of floating point values followed by a |
13607 | /// single floating to integer conversion of the vector. |
13608 | /// Namely (build_vector (fptosi $A), (fptosi $B), ...) |
13609 | /// becomes (fptosi (build_vector ($A, $B, ...))) |
13610 | SDValue PPCTargetLowering:: |
13611 | combineElementTruncationToVectorTruncation(SDNode *N, |
13612 | DAGCombinerInfo &DCI) const { |
13613 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13614, __PRETTY_FUNCTION__)) |
13614 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13614, __PRETTY_FUNCTION__)); |
13615 | |
13616 | SelectionDAG &DAG = DCI.DAG; |
13617 | SDLoc dl(N); |
13618 | |
13619 | SDValue FirstInput = N->getOperand(0); |
13620 | assert(FirstInput.getOpcode() == PPCISD::MFVSR &&((FirstInput.getOpcode() == PPCISD::MFVSR && "The input operand must be an fp-to-int conversion." ) ? static_cast<void> (0) : __assert_fail ("FirstInput.getOpcode() == PPCISD::MFVSR && \"The input operand must be an fp-to-int conversion.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13621, __PRETTY_FUNCTION__)) |
13621 | "The input operand must be an fp-to-int conversion.")((FirstInput.getOpcode() == PPCISD::MFVSR && "The input operand must be an fp-to-int conversion." ) ? static_cast<void> (0) : __assert_fail ("FirstInput.getOpcode() == PPCISD::MFVSR && \"The input operand must be an fp-to-int conversion.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13621, __PRETTY_FUNCTION__)); |
13622 | |
13623 | // This combine happens after legalization so the fp_to_[su]i nodes are |
13624 | // already converted to PPCSISD nodes. |
13625 | unsigned FirstConversion = FirstInput.getOperand(0).getOpcode(); |
13626 | if (FirstConversion == PPCISD::FCTIDZ || |
13627 | FirstConversion == PPCISD::FCTIDUZ || |
13628 | FirstConversion == PPCISD::FCTIWZ || |
13629 | FirstConversion == PPCISD::FCTIWUZ) { |
13630 | bool IsSplat = true; |
13631 | bool Is32Bit = FirstConversion == PPCISD::FCTIWZ || |
13632 | FirstConversion == PPCISD::FCTIWUZ; |
13633 | EVT SrcVT = FirstInput.getOperand(0).getValueType(); |
13634 | SmallVector<SDValue, 4> Ops; |
13635 | EVT TargetVT = N->getValueType(0); |
13636 | for (int i = 0, e = N->getNumOperands(); i < e; ++i) { |
13637 | SDValue NextOp = N->getOperand(i); |
13638 | if (NextOp.getOpcode() != PPCISD::MFVSR) |
13639 | return SDValue(); |
13640 | unsigned NextConversion = NextOp.getOperand(0).getOpcode(); |
13641 | if (NextConversion != FirstConversion) |
13642 | return SDValue(); |
13643 | // If we are converting to 32-bit integers, we need to add an FP_ROUND. |
13644 | // This is not valid if the input was originally double precision. It is |
13645 | // also not profitable to do unless this is an extending load in which |
13646 | // case doing this combine will allow us to combine consecutive loads. |
13647 | if (Is32Bit && !isFPExtLoad(NextOp.getOperand(0).getOperand(0))) |
13648 | return SDValue(); |
13649 | if (N->getOperand(i) != FirstInput) |
13650 | IsSplat = false; |
13651 | } |
13652 | |
13653 | // If this is a splat, we leave it as-is since there will be only a single |
13654 | // fp-to-int conversion followed by a splat of the integer. This is better |
13655 | // for 32-bit and smaller ints and neutral for 64-bit ints. |
13656 | if (IsSplat) |
13657 | return SDValue(); |
13658 | |
13659 | // Now that we know we have the right type of node, get its operands |
13660 | for (int i = 0, e = N->getNumOperands(); i < e; ++i) { |
13661 | SDValue In = N->getOperand(i).getOperand(0); |
13662 | if (Is32Bit) { |
13663 | // For 32-bit values, we need to add an FP_ROUND node (if we made it |
13664 | // here, we know that all inputs are extending loads so this is safe). |
13665 | if (In.isUndef()) |
13666 | Ops.push_back(DAG.getUNDEF(SrcVT)); |
13667 | else { |
13668 | SDValue Trunc = DAG.getNode(ISD::FP_ROUND, dl, |
13669 | MVT::f32, In.getOperand(0), |
13670 | DAG.getIntPtrConstant(1, dl)); |
13671 | Ops.push_back(Trunc); |
13672 | } |
13673 | } else |
13674 | Ops.push_back(In.isUndef() ? DAG.getUNDEF(SrcVT) : In.getOperand(0)); |
13675 | } |
13676 | |
13677 | unsigned Opcode; |
13678 | if (FirstConversion == PPCISD::FCTIDZ || |
13679 | FirstConversion == PPCISD::FCTIWZ) |
13680 | Opcode = ISD::FP_TO_SINT; |
13681 | else |
13682 | Opcode = ISD::FP_TO_UINT; |
13683 | |
13684 | EVT NewVT = TargetVT == MVT::v2i64 ? MVT::v2f64 : MVT::v4f32; |
13685 | SDValue BV = DAG.getBuildVector(NewVT, dl, Ops); |
13686 | return DAG.getNode(Opcode, dl, TargetVT, BV); |
13687 | } |
13688 | return SDValue(); |
13689 | } |
13690 | |
13691 | /// Reduce the number of loads when building a vector. |
13692 | /// |
13693 | /// Building a vector out of multiple loads can be converted to a load |
13694 | /// of the vector type if the loads are consecutive. If the loads are |
13695 | /// consecutive but in descending order, a shuffle is added at the end |
13696 | /// to reorder the vector. |
13697 | static SDValue combineBVOfConsecutiveLoads(SDNode *N, SelectionDAG &DAG) { |
13698 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13699, __PRETTY_FUNCTION__)) |
13699 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13699, __PRETTY_FUNCTION__)); |
13700 | |
13701 | SDLoc dl(N); |
13702 | |
13703 | // Return early for non byte-sized type, as they can't be consecutive. |
13704 | if (!N->getValueType(0).getVectorElementType().isByteSized()) |
13705 | return SDValue(); |
13706 | |
13707 | bool InputsAreConsecutiveLoads = true; |
13708 | bool InputsAreReverseConsecutive = true; |
13709 | unsigned ElemSize = N->getValueType(0).getScalarType().getStoreSize(); |
13710 | SDValue FirstInput = N->getOperand(0); |
13711 | bool IsRoundOfExtLoad = false; |
13712 | |
13713 | if (FirstInput.getOpcode() == ISD::FP_ROUND && |
13714 | FirstInput.getOperand(0).getOpcode() == ISD::LOAD) { |
13715 | LoadSDNode *LD = dyn_cast<LoadSDNode>(FirstInput.getOperand(0)); |
13716 | IsRoundOfExtLoad = LD->getExtensionType() == ISD::EXTLOAD; |
13717 | } |
13718 | // Not a build vector of (possibly fp_rounded) loads. |
13719 | if ((!IsRoundOfExtLoad && FirstInput.getOpcode() != ISD::LOAD) || |
13720 | N->getNumOperands() == 1) |
13721 | return SDValue(); |
13722 | |
13723 | for (int i = 1, e = N->getNumOperands(); i < e; ++i) { |
13724 | // If any inputs are fp_round(extload), they all must be. |
13725 | if (IsRoundOfExtLoad && N->getOperand(i).getOpcode() != ISD::FP_ROUND) |
13726 | return SDValue(); |
13727 | |
13728 | SDValue NextInput = IsRoundOfExtLoad ? N->getOperand(i).getOperand(0) : |
13729 | N->getOperand(i); |
13730 | if (NextInput.getOpcode() != ISD::LOAD) |
13731 | return SDValue(); |
13732 | |
13733 | SDValue PreviousInput = |
13734 | IsRoundOfExtLoad ? N->getOperand(i-1).getOperand(0) : N->getOperand(i-1); |
13735 | LoadSDNode *LD1 = dyn_cast<LoadSDNode>(PreviousInput); |
13736 | LoadSDNode *LD2 = dyn_cast<LoadSDNode>(NextInput); |
13737 | |
13738 | // If any inputs are fp_round(extload), they all must be. |
13739 | if (IsRoundOfExtLoad && LD2->getExtensionType() != ISD::EXTLOAD) |
13740 | return SDValue(); |
13741 | |
13742 | if (!isConsecutiveLS(LD2, LD1, ElemSize, 1, DAG)) |
13743 | InputsAreConsecutiveLoads = false; |
13744 | if (!isConsecutiveLS(LD1, LD2, ElemSize, 1, DAG)) |
13745 | InputsAreReverseConsecutive = false; |
13746 | |
13747 | // Exit early if the loads are neither consecutive nor reverse consecutive. |
13748 | if (!InputsAreConsecutiveLoads && !InputsAreReverseConsecutive) |
13749 | return SDValue(); |
13750 | } |
13751 | |
13752 | assert(!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) &&((!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive ) && "The loads cannot be both consecutive and reverse consecutive." ) ? static_cast<void> (0) : __assert_fail ("!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) && \"The loads cannot be both consecutive and reverse consecutive.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13753, __PRETTY_FUNCTION__)) |
13753 | "The loads cannot be both consecutive and reverse consecutive.")((!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive ) && "The loads cannot be both consecutive and reverse consecutive." ) ? static_cast<void> (0) : __assert_fail ("!(InputsAreConsecutiveLoads && InputsAreReverseConsecutive) && \"The loads cannot be both consecutive and reverse consecutive.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13753, __PRETTY_FUNCTION__)); |
13754 | |
13755 | SDValue FirstLoadOp = |
13756 | IsRoundOfExtLoad ? FirstInput.getOperand(0) : FirstInput; |
13757 | SDValue LastLoadOp = |
13758 | IsRoundOfExtLoad ? N->getOperand(N->getNumOperands()-1).getOperand(0) : |
13759 | N->getOperand(N->getNumOperands()-1); |
13760 | |
13761 | LoadSDNode *LD1 = dyn_cast<LoadSDNode>(FirstLoadOp); |
13762 | LoadSDNode *LDL = dyn_cast<LoadSDNode>(LastLoadOp); |
13763 | if (InputsAreConsecutiveLoads) { |
13764 | assert(LD1 && "Input needs to be a LoadSDNode.")((LD1 && "Input needs to be a LoadSDNode.") ? static_cast <void> (0) : __assert_fail ("LD1 && \"Input needs to be a LoadSDNode.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13764, __PRETTY_FUNCTION__)); |
13765 | return DAG.getLoad(N->getValueType(0), dl, LD1->getChain(), |
13766 | LD1->getBasePtr(), LD1->getPointerInfo(), |
13767 | LD1->getAlignment()); |
13768 | } |
13769 | if (InputsAreReverseConsecutive) { |
13770 | assert(LDL && "Input needs to be a LoadSDNode.")((LDL && "Input needs to be a LoadSDNode.") ? static_cast <void> (0) : __assert_fail ("LDL && \"Input needs to be a LoadSDNode.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13770, __PRETTY_FUNCTION__)); |
13771 | SDValue Load = DAG.getLoad(N->getValueType(0), dl, LDL->getChain(), |
13772 | LDL->getBasePtr(), LDL->getPointerInfo(), |
13773 | LDL->getAlignment()); |
13774 | SmallVector<int, 16> Ops; |
13775 | for (int i = N->getNumOperands() - 1; i >= 0; i--) |
13776 | Ops.push_back(i); |
13777 | |
13778 | return DAG.getVectorShuffle(N->getValueType(0), dl, Load, |
13779 | DAG.getUNDEF(N->getValueType(0)), Ops); |
13780 | } |
13781 | return SDValue(); |
13782 | } |
13783 | |
13784 | // This function adds the required vector_shuffle needed to get |
13785 | // the elements of the vector extract in the correct position |
13786 | // as specified by the CorrectElems encoding. |
13787 | static SDValue addShuffleForVecExtend(SDNode *N, SelectionDAG &DAG, |
13788 | SDValue Input, uint64_t Elems, |
13789 | uint64_t CorrectElems) { |
13790 | SDLoc dl(N); |
13791 | |
13792 | unsigned NumElems = Input.getValueType().getVectorNumElements(); |
13793 | SmallVector<int, 16> ShuffleMask(NumElems, -1); |
13794 | |
13795 | // Knowing the element indices being extracted from the original |
13796 | // vector and the order in which they're being inserted, just put |
13797 | // them at element indices required for the instruction. |
13798 | for (unsigned i = 0; i < N->getNumOperands(); i++) { |
13799 | if (DAG.getDataLayout().isLittleEndian()) |
13800 | ShuffleMask[CorrectElems & 0xF] = Elems & 0xF; |
13801 | else |
13802 | ShuffleMask[(CorrectElems & 0xF0) >> 4] = (Elems & 0xF0) >> 4; |
13803 | CorrectElems = CorrectElems >> 8; |
13804 | Elems = Elems >> 8; |
13805 | } |
13806 | |
13807 | SDValue Shuffle = |
13808 | DAG.getVectorShuffle(Input.getValueType(), dl, Input, |
13809 | DAG.getUNDEF(Input.getValueType()), ShuffleMask); |
13810 | |
13811 | EVT VT = N->getValueType(0); |
13812 | SDValue Conv = DAG.getBitcast(VT, Shuffle); |
13813 | |
13814 | EVT ExtVT = EVT::getVectorVT(*DAG.getContext(), |
13815 | Input.getValueType().getVectorElementType(), |
13816 | VT.getVectorNumElements()); |
13817 | return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, VT, Conv, |
13818 | DAG.getValueType(ExtVT)); |
13819 | } |
13820 | |
13821 | // Look for build vector patterns where input operands come from sign |
13822 | // extended vector_extract elements of specific indices. If the correct indices |
13823 | // aren't used, add a vector shuffle to fix up the indices and create |
13824 | // SIGN_EXTEND_INREG node which selects the vector sign extend instructions |
13825 | // during instruction selection. |
13826 | static SDValue combineBVOfVecSExt(SDNode *N, SelectionDAG &DAG) { |
13827 | // This array encodes the indices that the vector sign extend instructions |
13828 | // extract from when extending from one type to another for both BE and LE. |
13829 | // The right nibble of each byte corresponds to the LE incides. |
13830 | // and the left nibble of each byte corresponds to the BE incides. |
13831 | // For example: 0x3074B8FC byte->word |
13832 | // For LE: the allowed indices are: 0x0,0x4,0x8,0xC |
13833 | // For BE: the allowed indices are: 0x3,0x7,0xB,0xF |
13834 | // For example: 0x000070F8 byte->double word |
13835 | // For LE: the allowed indices are: 0x0,0x8 |
13836 | // For BE: the allowed indices are: 0x7,0xF |
13837 | uint64_t TargetElems[] = { |
13838 | 0x3074B8FC, // b->w |
13839 | 0x000070F8, // b->d |
13840 | 0x10325476, // h->w |
13841 | 0x00003074, // h->d |
13842 | 0x00001032, // w->d |
13843 | }; |
13844 | |
13845 | uint64_t Elems = 0; |
13846 | int Index; |
13847 | SDValue Input; |
13848 | |
13849 | auto isSExtOfVecExtract = [&](SDValue Op) -> bool { |
13850 | if (!Op) |
13851 | return false; |
13852 | if (Op.getOpcode() != ISD::SIGN_EXTEND && |
13853 | Op.getOpcode() != ISD::SIGN_EXTEND_INREG) |
13854 | return false; |
13855 | |
13856 | // A SIGN_EXTEND_INREG might be fed by an ANY_EXTEND to produce a value |
13857 | // of the right width. |
13858 | SDValue Extract = Op.getOperand(0); |
13859 | if (Extract.getOpcode() == ISD::ANY_EXTEND) |
13860 | Extract = Extract.getOperand(0); |
13861 | if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
13862 | return false; |
13863 | |
13864 | ConstantSDNode *ExtOp = dyn_cast<ConstantSDNode>(Extract.getOperand(1)); |
13865 | if (!ExtOp) |
13866 | return false; |
13867 | |
13868 | Index = ExtOp->getZExtValue(); |
13869 | if (Input && Input != Extract.getOperand(0)) |
13870 | return false; |
13871 | |
13872 | if (!Input) |
13873 | Input = Extract.getOperand(0); |
13874 | |
13875 | Elems = Elems << 8; |
13876 | Index = DAG.getDataLayout().isLittleEndian() ? Index : Index << 4; |
13877 | Elems |= Index; |
13878 | |
13879 | return true; |
13880 | }; |
13881 | |
13882 | // If the build vector operands aren't sign extended vector extracts, |
13883 | // of the same input vector, then return. |
13884 | for (unsigned i = 0; i < N->getNumOperands(); i++) { |
13885 | if (!isSExtOfVecExtract(N->getOperand(i))) { |
13886 | return SDValue(); |
13887 | } |
13888 | } |
13889 | |
13890 | // If the vector extract indicies are not correct, add the appropriate |
13891 | // vector_shuffle. |
13892 | int TgtElemArrayIdx; |
13893 | int InputSize = Input.getValueType().getScalarSizeInBits(); |
13894 | int OutputSize = N->getValueType(0).getScalarSizeInBits(); |
13895 | if (InputSize + OutputSize == 40) |
13896 | TgtElemArrayIdx = 0; |
13897 | else if (InputSize + OutputSize == 72) |
13898 | TgtElemArrayIdx = 1; |
13899 | else if (InputSize + OutputSize == 48) |
13900 | TgtElemArrayIdx = 2; |
13901 | else if (InputSize + OutputSize == 80) |
13902 | TgtElemArrayIdx = 3; |
13903 | else if (InputSize + OutputSize == 96) |
13904 | TgtElemArrayIdx = 4; |
13905 | else |
13906 | return SDValue(); |
13907 | |
13908 | uint64_t CorrectElems = TargetElems[TgtElemArrayIdx]; |
13909 | CorrectElems = DAG.getDataLayout().isLittleEndian() |
13910 | ? CorrectElems & 0x0F0F0F0F0F0F0F0F |
13911 | : CorrectElems & 0xF0F0F0F0F0F0F0F0; |
13912 | if (Elems != CorrectElems) { |
13913 | return addShuffleForVecExtend(N, DAG, Input, Elems, CorrectElems); |
13914 | } |
13915 | |
13916 | // Regular lowering will catch cases where a shuffle is not needed. |
13917 | return SDValue(); |
13918 | } |
13919 | |
13920 | // Look for the pattern of a load from a narrow width to i128, feeding |
13921 | // into a BUILD_VECTOR of v1i128. Replace this sequence with a PPCISD node |
13922 | // (LXVRZX). This node represents a zero extending load that will be matched |
13923 | // to the Load VSX Vector Rightmost instructions. |
13924 | static SDValue combineBVZEXTLOAD(SDNode *N, SelectionDAG &DAG) { |
13925 | SDLoc DL(N); |
13926 | |
13927 | // This combine is only eligible for a BUILD_VECTOR of v1i128. |
13928 | if (N->getValueType(0) != MVT::v1i128) |
13929 | return SDValue(); |
13930 | |
13931 | SDValue Operand = N->getOperand(0); |
13932 | // Proceed with the transformation if the operand to the BUILD_VECTOR |
13933 | // is a load instruction. |
13934 | if (Operand.getOpcode() != ISD::LOAD) |
13935 | return SDValue(); |
13936 | |
13937 | LoadSDNode *LD = dyn_cast<LoadSDNode>(Operand); |
13938 | EVT MemoryType = LD->getMemoryVT(); |
13939 | |
13940 | // This transformation is only valid if the we are loading either a byte, |
13941 | // halfword, word, or doubleword. |
13942 | bool ValidLDType = MemoryType == MVT::i8 || MemoryType == MVT::i16 || |
13943 | MemoryType == MVT::i32 || MemoryType == MVT::i64; |
13944 | |
13945 | // Ensure that the load from the narrow width is being zero extended to i128. |
13946 | if (!ValidLDType || |
13947 | (LD->getExtensionType() != ISD::ZEXTLOAD && |
13948 | LD->getExtensionType() != ISD::EXTLOAD)) |
13949 | return SDValue(); |
13950 | |
13951 | SDValue LoadOps[] = { |
13952 | LD->getChain(), LD->getBasePtr(), |
13953 | DAG.getIntPtrConstant(MemoryType.getScalarSizeInBits(), DL)}; |
13954 | |
13955 | return DAG.getMemIntrinsicNode(PPCISD::LXVRZX, DL, |
13956 | DAG.getVTList(MVT::v1i128, MVT::Other), |
13957 | LoadOps, MemoryType, LD->getMemOperand()); |
13958 | } |
13959 | |
13960 | SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N, |
13961 | DAGCombinerInfo &DCI) const { |
13962 | assert(N->getOpcode() == ISD::BUILD_VECTOR &&((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13963, __PRETTY_FUNCTION__)) |
13963 | "Should be called with a BUILD_VECTOR node")((N->getOpcode() == ISD::BUILD_VECTOR && "Should be called with a BUILD_VECTOR node" ) ? static_cast<void> (0) : __assert_fail ("N->getOpcode() == ISD::BUILD_VECTOR && \"Should be called with a BUILD_VECTOR node\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 13963, __PRETTY_FUNCTION__)); |
13964 | |
13965 | SelectionDAG &DAG = DCI.DAG; |
13966 | SDLoc dl(N); |
13967 | |
13968 | if (!Subtarget.hasVSX()) |
13969 | return SDValue(); |
13970 | |
13971 | // The target independent DAG combiner will leave a build_vector of |
13972 | // float-to-int conversions intact. We can generate MUCH better code for |
13973 | // a float-to-int conversion of a vector of floats. |
13974 | SDValue FirstInput = N->getOperand(0); |
13975 | if (FirstInput.getOpcode() == PPCISD::MFVSR) { |
13976 | SDValue Reduced = combineElementTruncationToVectorTruncation(N, DCI); |
13977 | if (Reduced) |
13978 | return Reduced; |
13979 | } |
13980 | |
13981 | // If we're building a vector out of consecutive loads, just load that |
13982 | // vector type. |
13983 | SDValue Reduced = combineBVOfConsecutiveLoads(N, DAG); |
13984 | if (Reduced) |
13985 | return Reduced; |
13986 | |
13987 | // If we're building a vector out of extended elements from another vector |
13988 | // we have P9 vector integer extend instructions. The code assumes legal |
13989 | // input types (i.e. it can't handle things like v4i16) so do not run before |
13990 | // legalization. |
13991 | if (Subtarget.hasP9Altivec() && !DCI.isBeforeLegalize()) { |
13992 | Reduced = combineBVOfVecSExt(N, DAG); |
13993 | if (Reduced) |
13994 | return Reduced; |
13995 | } |
13996 | |
13997 | // On Power10, the Load VSX Vector Rightmost instructions can be utilized |
13998 | // if this is a BUILD_VECTOR of v1i128, and if the operand to the BUILD_VECTOR |
13999 | // is a load from <valid narrow width> to i128. |
14000 | if (Subtarget.isISA3_1()) { |
14001 | SDValue BVOfZLoad = combineBVZEXTLOAD(N, DAG); |
14002 | if (BVOfZLoad) |
14003 | return BVOfZLoad; |
14004 | } |
14005 | |
14006 | if (N->getValueType(0) != MVT::v2f64) |
14007 | return SDValue(); |
14008 | |
14009 | // Looking for: |
14010 | // (build_vector ([su]int_to_fp (extractelt 0)), [su]int_to_fp (extractelt 1)) |
14011 | if (FirstInput.getOpcode() != ISD::SINT_TO_FP && |
14012 | FirstInput.getOpcode() != ISD::UINT_TO_FP) |
14013 | return SDValue(); |
14014 | if (N->getOperand(1).getOpcode() != ISD::SINT_TO_FP && |
14015 | N->getOperand(1).getOpcode() != ISD::UINT_TO_FP) |
14016 | return SDValue(); |
14017 | if (FirstInput.getOpcode() != N->getOperand(1).getOpcode()) |
14018 | return SDValue(); |
14019 | |
14020 | SDValue Ext1 = FirstInput.getOperand(0); |
14021 | SDValue Ext2 = N->getOperand(1).getOperand(0); |
14022 | if(Ext1.getOpcode() != ISD::EXTRACT_VECTOR_ELT || |
14023 | Ext2.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
14024 | return SDValue(); |
14025 | |
14026 | ConstantSDNode *Ext1Op = dyn_cast<ConstantSDNode>(Ext1.getOperand(1)); |
14027 | ConstantSDNode *Ext2Op = dyn_cast<ConstantSDNode>(Ext2.getOperand(1)); |
14028 | if (!Ext1Op || !Ext2Op) |
14029 | return SDValue(); |
14030 | if (Ext1.getOperand(0).getValueType() != MVT::v4i32 || |
14031 | Ext1.getOperand(0) != Ext2.getOperand(0)) |
14032 | return SDValue(); |
14033 | |
14034 | int FirstElem = Ext1Op->getZExtValue(); |
14035 | int SecondElem = Ext2Op->getZExtValue(); |
14036 | int SubvecIdx; |
14037 | if (FirstElem == 0 && SecondElem == 1) |
14038 | SubvecIdx = Subtarget.isLittleEndian() ? 1 : 0; |
14039 | else if (FirstElem == 2 && SecondElem == 3) |
14040 | SubvecIdx = Subtarget.isLittleEndian() ? 0 : 1; |
14041 | else |
14042 | return SDValue(); |
14043 | |
14044 | SDValue SrcVec = Ext1.getOperand(0); |
14045 | auto NodeType = (N->getOperand(1).getOpcode() == ISD::SINT_TO_FP) ? |
14046 | PPCISD::SINT_VEC_TO_FP : PPCISD::UINT_VEC_TO_FP; |
14047 | return DAG.getNode(NodeType, dl, MVT::v2f64, |
14048 | SrcVec, DAG.getIntPtrConstant(SubvecIdx, dl)); |
14049 | } |
14050 | |
14051 | SDValue PPCTargetLowering::combineFPToIntToFP(SDNode *N, |
14052 | DAGCombinerInfo &DCI) const { |
14053 | assert((N->getOpcode() == ISD::SINT_TO_FP ||(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14055, __PRETTY_FUNCTION__)) |
14054 | N->getOpcode() == ISD::UINT_TO_FP) &&(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14055, __PRETTY_FUNCTION__)) |
14055 | "Need an int -> FP conversion node here")(((N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && "Need an int -> FP conversion node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::SINT_TO_FP || N->getOpcode() == ISD::UINT_TO_FP) && \"Need an int -> FP conversion node here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14055, __PRETTY_FUNCTION__)); |
14056 | |
14057 | if (useSoftFloat() || !Subtarget.has64BitSupport()) |
14058 | return SDValue(); |
14059 | |
14060 | SelectionDAG &DAG = DCI.DAG; |
14061 | SDLoc dl(N); |
14062 | SDValue Op(N, 0); |
14063 | |
14064 | // Don't handle ppc_fp128 here or conversions that are out-of-range capable |
14065 | // from the hardware. |
14066 | if (Op.getValueType() != MVT::f32 && Op.getValueType() != MVT::f64) |
14067 | return SDValue(); |
14068 | if (!Op.getOperand(0).getValueType().isSimple()) |
14069 | return SDValue(); |
14070 | if (Op.getOperand(0).getValueType().getSimpleVT() <= MVT(MVT::i1) || |
14071 | Op.getOperand(0).getValueType().getSimpleVT() > MVT(MVT::i64)) |
14072 | return SDValue(); |
14073 | |
14074 | SDValue FirstOperand(Op.getOperand(0)); |
14075 | bool SubWordLoad = FirstOperand.getOpcode() == ISD::LOAD && |
14076 | (FirstOperand.getValueType() == MVT::i8 || |
14077 | FirstOperand.getValueType() == MVT::i16); |
14078 | if (Subtarget.hasP9Vector() && Subtarget.hasP9Altivec() && SubWordLoad) { |
14079 | bool Signed = N->getOpcode() == ISD::SINT_TO_FP; |
14080 | bool DstDouble = Op.getValueType() == MVT::f64; |
14081 | unsigned ConvOp = Signed ? |
14082 | (DstDouble ? PPCISD::FCFID : PPCISD::FCFIDS) : |
14083 | (DstDouble ? PPCISD::FCFIDU : PPCISD::FCFIDUS); |
14084 | SDValue WidthConst = |
14085 | DAG.getIntPtrConstant(FirstOperand.getValueType() == MVT::i8 ? 1 : 2, |
14086 | dl, false); |
14087 | LoadSDNode *LDN = cast<LoadSDNode>(FirstOperand.getNode()); |
14088 | SDValue Ops[] = { LDN->getChain(), LDN->getBasePtr(), WidthConst }; |
14089 | SDValue Ld = DAG.getMemIntrinsicNode(PPCISD::LXSIZX, dl, |
14090 | DAG.getVTList(MVT::f64, MVT::Other), |
14091 | Ops, MVT::i8, LDN->getMemOperand()); |
14092 | |
14093 | // For signed conversion, we need to sign-extend the value in the VSR |
14094 | if (Signed) { |
14095 | SDValue ExtOps[] = { Ld, WidthConst }; |
14096 | SDValue Ext = DAG.getNode(PPCISD::VEXTS, dl, MVT::f64, ExtOps); |
14097 | return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ext); |
14098 | } else |
14099 | return DAG.getNode(ConvOp, dl, DstDouble ? MVT::f64 : MVT::f32, Ld); |
14100 | } |
14101 | |
14102 | |
14103 | // For i32 intermediate values, unfortunately, the conversion functions |
14104 | // leave the upper 32 bits of the value are undefined. Within the set of |
14105 | // scalar instructions, we have no method for zero- or sign-extending the |
14106 | // value. Thus, we cannot handle i32 intermediate values here. |
14107 | if (Op.getOperand(0).getValueType() == MVT::i32) |
14108 | return SDValue(); |
14109 | |
14110 | assert((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) &&(((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14111, __PRETTY_FUNCTION__)) |
14111 | "UINT_TO_FP is supported only with FPCVT")(((Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && "UINT_TO_FP is supported only with FPCVT") ? static_cast <void> (0) : __assert_fail ("(Op.getOpcode() == ISD::SINT_TO_FP || Subtarget.hasFPCVT()) && \"UINT_TO_FP is supported only with FPCVT\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14111, __PRETTY_FUNCTION__)); |
14112 | |
14113 | // If we have FCFIDS, then use it when converting to single-precision. |
14114 | // Otherwise, convert to double-precision and then round. |
14115 | unsigned FCFOp = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) |
14116 | ? (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDUS |
14117 | : PPCISD::FCFIDS) |
14118 | : (Op.getOpcode() == ISD::UINT_TO_FP ? PPCISD::FCFIDU |
14119 | : PPCISD::FCFID); |
14120 | MVT FCFTy = (Subtarget.hasFPCVT() && Op.getValueType() == MVT::f32) |
14121 | ? MVT::f32 |
14122 | : MVT::f64; |
14123 | |
14124 | // If we're converting from a float, to an int, and back to a float again, |
14125 | // then we don't need the store/load pair at all. |
14126 | if ((Op.getOperand(0).getOpcode() == ISD::FP_TO_UINT && |
14127 | Subtarget.hasFPCVT()) || |
14128 | (Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT)) { |
14129 | SDValue Src = Op.getOperand(0).getOperand(0); |
14130 | if (Src.getValueType() == MVT::f32) { |
14131 | Src = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Src); |
14132 | DCI.AddToWorklist(Src.getNode()); |
14133 | } else if (Src.getValueType() != MVT::f64) { |
14134 | // Make sure that we don't pick up a ppc_fp128 source value. |
14135 | return SDValue(); |
14136 | } |
14137 | |
14138 | unsigned FCTOp = |
14139 | Op.getOperand(0).getOpcode() == ISD::FP_TO_SINT ? PPCISD::FCTIDZ : |
14140 | PPCISD::FCTIDUZ; |
14141 | |
14142 | SDValue Tmp = DAG.getNode(FCTOp, dl, MVT::f64, Src); |
14143 | SDValue FP = DAG.getNode(FCFOp, dl, FCFTy, Tmp); |
14144 | |
14145 | if (Op.getValueType() == MVT::f32 && !Subtarget.hasFPCVT()) { |
14146 | FP = DAG.getNode(ISD::FP_ROUND, dl, |
14147 | MVT::f32, FP, DAG.getIntPtrConstant(0, dl)); |
14148 | DCI.AddToWorklist(FP.getNode()); |
14149 | } |
14150 | |
14151 | return FP; |
14152 | } |
14153 | |
14154 | return SDValue(); |
14155 | } |
14156 | |
14157 | // expandVSXLoadForLE - Convert VSX loads (which may be intrinsics for |
14158 | // builtins) into loads with swaps. |
14159 | SDValue PPCTargetLowering::expandVSXLoadForLE(SDNode *N, |
14160 | DAGCombinerInfo &DCI) const { |
14161 | SelectionDAG &DAG = DCI.DAG; |
14162 | SDLoc dl(N); |
14163 | SDValue Chain; |
14164 | SDValue Base; |
14165 | MachineMemOperand *MMO; |
14166 | |
14167 | switch (N->getOpcode()) { |
14168 | default: |
14169 | llvm_unreachable("Unexpected opcode for little endian VSX load")::llvm::llvm_unreachable_internal("Unexpected opcode for little endian VSX load" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14169); |
14170 | case ISD::LOAD: { |
14171 | LoadSDNode *LD = cast<LoadSDNode>(N); |
14172 | Chain = LD->getChain(); |
14173 | Base = LD->getBasePtr(); |
14174 | MMO = LD->getMemOperand(); |
14175 | // If the MMO suggests this isn't a load of a full vector, leave |
14176 | // things alone. For a built-in, we have to make the change for |
14177 | // correctness, so if there is a size problem that will be a bug. |
14178 | if (MMO->getSize() < 16) |
14179 | return SDValue(); |
14180 | break; |
14181 | } |
14182 | case ISD::INTRINSIC_W_CHAIN: { |
14183 | MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N); |
14184 | Chain = Intrin->getChain(); |
14185 | // Similarly to the store case below, Intrin->getBasePtr() doesn't get |
14186 | // us what we want. Get operand 2 instead. |
14187 | Base = Intrin->getOperand(2); |
14188 | MMO = Intrin->getMemOperand(); |
14189 | break; |
14190 | } |
14191 | } |
14192 | |
14193 | MVT VecTy = N->getValueType(0).getSimpleVT(); |
14194 | |
14195 | // Do not expand to PPCISD::LXVD2X + PPCISD::XXSWAPD when the load is |
14196 | // aligned and the type is a vector with elements up to 4 bytes |
14197 | if (Subtarget.needsSwapsForVSXMemOps() && MMO->getAlign() >= Align(16) && |
14198 | VecTy.getScalarSizeInBits() <= 32) { |
14199 | return SDValue(); |
14200 | } |
14201 | |
14202 | SDValue LoadOps[] = { Chain, Base }; |
14203 | SDValue Load = DAG.getMemIntrinsicNode(PPCISD::LXVD2X, dl, |
14204 | DAG.getVTList(MVT::v2f64, MVT::Other), |
14205 | LoadOps, MVT::v2f64, MMO); |
14206 | |
14207 | DCI.AddToWorklist(Load.getNode()); |
14208 | Chain = Load.getValue(1); |
14209 | SDValue Swap = DAG.getNode( |
14210 | PPCISD::XXSWAPD, dl, DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Load); |
14211 | DCI.AddToWorklist(Swap.getNode()); |
14212 | |
14213 | // Add a bitcast if the resulting load type doesn't match v2f64. |
14214 | if (VecTy != MVT::v2f64) { |
14215 | SDValue N = DAG.getNode(ISD::BITCAST, dl, VecTy, Swap); |
14216 | DCI.AddToWorklist(N.getNode()); |
14217 | // Package {bitcast value, swap's chain} to match Load's shape. |
14218 | return DAG.getNode(ISD::MERGE_VALUES, dl, DAG.getVTList(VecTy, MVT::Other), |
14219 | N, Swap.getValue(1)); |
14220 | } |
14221 | |
14222 | return Swap; |
14223 | } |
14224 | |
14225 | // expandVSXStoreForLE - Convert VSX stores (which may be intrinsics for |
14226 | // builtins) into stores with swaps. |
14227 | SDValue PPCTargetLowering::expandVSXStoreForLE(SDNode *N, |
14228 | DAGCombinerInfo &DCI) const { |
14229 | SelectionDAG &DAG = DCI.DAG; |
14230 | SDLoc dl(N); |
14231 | SDValue Chain; |
14232 | SDValue Base; |
14233 | unsigned SrcOpnd; |
14234 | MachineMemOperand *MMO; |
14235 | |
14236 | switch (N->getOpcode()) { |
14237 | default: |
14238 | llvm_unreachable("Unexpected opcode for little endian VSX store")::llvm::llvm_unreachable_internal("Unexpected opcode for little endian VSX store" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14238); |
14239 | case ISD::STORE: { |
14240 | StoreSDNode *ST = cast<StoreSDNode>(N); |
14241 | Chain = ST->getChain(); |
14242 | Base = ST->getBasePtr(); |
14243 | MMO = ST->getMemOperand(); |
14244 | SrcOpnd = 1; |
14245 | // If the MMO suggests this isn't a store of a full vector, leave |
14246 | // things alone. For a built-in, we have to make the change for |
14247 | // correctness, so if there is a size problem that will be a bug. |
14248 | if (MMO->getSize() < 16) |
14249 | return SDValue(); |
14250 | break; |
14251 | } |
14252 | case ISD::INTRINSIC_VOID: { |
14253 | MemIntrinsicSDNode *Intrin = cast<MemIntrinsicSDNode>(N); |
14254 | Chain = Intrin->getChain(); |
14255 | // Intrin->getBasePtr() oddly does not get what we want. |
14256 | Base = Intrin->getOperand(3); |
14257 | MMO = Intrin->getMemOperand(); |
14258 | SrcOpnd = 2; |
14259 | break; |
14260 | } |
14261 | } |
14262 | |
14263 | SDValue Src = N->getOperand(SrcOpnd); |
14264 | MVT VecTy = Src.getValueType().getSimpleVT(); |
14265 | |
14266 | // Do not expand to PPCISD::XXSWAPD and PPCISD::STXVD2X when the load is |
14267 | // aligned and the type is a vector with elements up to 4 bytes |
14268 | if (Subtarget.needsSwapsForVSXMemOps() && MMO->getAlign() >= Align(16) && |
14269 | VecTy.getScalarSizeInBits() <= 32) { |
14270 | return SDValue(); |
14271 | } |
14272 | |
14273 | // All stores are done as v2f64 and possible bit cast. |
14274 | if (VecTy != MVT::v2f64) { |
14275 | Src = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Src); |
14276 | DCI.AddToWorklist(Src.getNode()); |
14277 | } |
14278 | |
14279 | SDValue Swap = DAG.getNode(PPCISD::XXSWAPD, dl, |
14280 | DAG.getVTList(MVT::v2f64, MVT::Other), Chain, Src); |
14281 | DCI.AddToWorklist(Swap.getNode()); |
14282 | Chain = Swap.getValue(1); |
14283 | SDValue StoreOps[] = { Chain, Swap, Base }; |
14284 | SDValue Store = DAG.getMemIntrinsicNode(PPCISD::STXVD2X, dl, |
14285 | DAG.getVTList(MVT::Other), |
14286 | StoreOps, VecTy, MMO); |
14287 | DCI.AddToWorklist(Store.getNode()); |
14288 | return Store; |
14289 | } |
14290 | |
14291 | // Handle DAG combine for STORE (FP_TO_INT F). |
14292 | SDValue PPCTargetLowering::combineStoreFPToInt(SDNode *N, |
14293 | DAGCombinerInfo &DCI) const { |
14294 | |
14295 | SelectionDAG &DAG = DCI.DAG; |
14296 | SDLoc dl(N); |
14297 | unsigned Opcode = N->getOperand(1).getOpcode(); |
14298 | |
14299 | assert((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT)(((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && "Not a FP_TO_INT Instruction!") ? static_cast<void> (0 ) : __assert_fail ("(Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && \"Not a FP_TO_INT Instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14300, __PRETTY_FUNCTION__)) |
14300 | && "Not a FP_TO_INT Instruction!")(((Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && "Not a FP_TO_INT Instruction!") ? static_cast<void> (0 ) : __assert_fail ("(Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) && \"Not a FP_TO_INT Instruction!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14300, __PRETTY_FUNCTION__)); |
14301 | |
14302 | SDValue Val = N->getOperand(1).getOperand(0); |
14303 | EVT Op1VT = N->getOperand(1).getValueType(); |
14304 | EVT ResVT = Val.getValueType(); |
14305 | |
14306 | if (!isTypeLegal(ResVT)) |
14307 | return SDValue(); |
14308 | |
14309 | // Only perform combine for conversion to i64/i32 or power9 i16/i8. |
14310 | bool ValidTypeForStoreFltAsInt = |
14311 | (Op1VT == MVT::i32 || Op1VT == MVT::i64 || |
14312 | (Subtarget.hasP9Vector() && (Op1VT == MVT::i16 || Op1VT == MVT::i8))); |
14313 | |
14314 | if (ResVT == MVT::ppcf128 || !Subtarget.hasP8Vector() || |
14315 | cast<StoreSDNode>(N)->isTruncatingStore() || !ValidTypeForStoreFltAsInt) |
14316 | return SDValue(); |
14317 | |
14318 | // Extend f32 values to f64 |
14319 | if (ResVT.getScalarSizeInBits() == 32) { |
14320 | Val = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f64, Val); |
14321 | DCI.AddToWorklist(Val.getNode()); |
14322 | } |
14323 | |
14324 | // Set signed or unsigned conversion opcode. |
14325 | unsigned ConvOpcode = (Opcode == ISD::FP_TO_SINT) ? |
14326 | PPCISD::FP_TO_SINT_IN_VSR : |
14327 | PPCISD::FP_TO_UINT_IN_VSR; |
14328 | |
14329 | Val = DAG.getNode(ConvOpcode, |
14330 | dl, ResVT == MVT::f128 ? MVT::f128 : MVT::f64, Val); |
14331 | DCI.AddToWorklist(Val.getNode()); |
14332 | |
14333 | // Set number of bytes being converted. |
14334 | unsigned ByteSize = Op1VT.getScalarSizeInBits() / 8; |
14335 | SDValue Ops[] = { N->getOperand(0), Val, N->getOperand(2), |
14336 | DAG.getIntPtrConstant(ByteSize, dl, false), |
14337 | DAG.getValueType(Op1VT) }; |
14338 | |
14339 | Val = DAG.getMemIntrinsicNode(PPCISD::ST_VSR_SCAL_INT, dl, |
14340 | DAG.getVTList(MVT::Other), Ops, |
14341 | cast<StoreSDNode>(N)->getMemoryVT(), |
14342 | cast<StoreSDNode>(N)->getMemOperand()); |
14343 | |
14344 | DCI.AddToWorklist(Val.getNode()); |
14345 | return Val; |
14346 | } |
14347 | |
14348 | static bool isAlternatingShuffMask(const ArrayRef<int> &Mask, int NumElts) { |
14349 | // Check that the source of the element keeps flipping |
14350 | // (i.e. Mask[i] < NumElts -> Mask[i+i] >= NumElts). |
14351 | bool PrevElemFromFirstVec = Mask[0] < NumElts; |
14352 | for (int i = 1, e = Mask.size(); i < e; i++) { |
14353 | if (PrevElemFromFirstVec && Mask[i] < NumElts) |
14354 | return false; |
14355 | if (!PrevElemFromFirstVec && Mask[i] >= NumElts) |
14356 | return false; |
14357 | PrevElemFromFirstVec = !PrevElemFromFirstVec; |
14358 | } |
14359 | return true; |
14360 | } |
14361 | |
14362 | static bool isSplatBV(SDValue Op) { |
14363 | if (Op.getOpcode() != ISD::BUILD_VECTOR) |
14364 | return false; |
14365 | SDValue FirstOp; |
14366 | |
14367 | // Find first non-undef input. |
14368 | for (int i = 0, e = Op.getNumOperands(); i < e; i++) { |
14369 | FirstOp = Op.getOperand(i); |
14370 | if (!FirstOp.isUndef()) |
14371 | break; |
14372 | } |
14373 | |
14374 | // All inputs are undef or the same as the first non-undef input. |
14375 | for (int i = 1, e = Op.getNumOperands(); i < e; i++) |
14376 | if (Op.getOperand(i) != FirstOp && !Op.getOperand(i).isUndef()) |
14377 | return false; |
14378 | return true; |
14379 | } |
14380 | |
14381 | static SDValue isScalarToVec(SDValue Op) { |
14382 | if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR) |
14383 | return Op; |
14384 | if (Op.getOpcode() != ISD::BITCAST) |
14385 | return SDValue(); |
14386 | Op = Op.getOperand(0); |
14387 | if (Op.getOpcode() == ISD::SCALAR_TO_VECTOR) |
14388 | return Op; |
14389 | return SDValue(); |
14390 | } |
14391 | |
14392 | static void fixupShuffleMaskForPermutedSToV(SmallVectorImpl<int> &ShuffV, |
14393 | int LHSMaxIdx, int RHSMinIdx, |
14394 | int RHSMaxIdx, int HalfVec) { |
14395 | for (int i = 0, e = ShuffV.size(); i < e; i++) { |
14396 | int Idx = ShuffV[i]; |
14397 | if ((Idx >= 0 && Idx < LHSMaxIdx) || (Idx >= RHSMinIdx && Idx < RHSMaxIdx)) |
14398 | ShuffV[i] += HalfVec; |
14399 | } |
14400 | return; |
14401 | } |
14402 | |
14403 | // Replace a SCALAR_TO_VECTOR with a SCALAR_TO_VECTOR_PERMUTED except if |
14404 | // the original is: |
14405 | // (<n x Ty> (scalar_to_vector (Ty (extract_elt <n x Ty> %a, C)))) |
14406 | // In such a case, just change the shuffle mask to extract the element |
14407 | // from the permuted index. |
14408 | static SDValue getSToVPermuted(SDValue OrigSToV, SelectionDAG &DAG) { |
14409 | SDLoc dl(OrigSToV); |
14410 | EVT VT = OrigSToV.getValueType(); |
14411 | assert(OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR &&((OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && "Expecting a SCALAR_TO_VECTOR here" ) ? static_cast<void> (0) : __assert_fail ("OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && \"Expecting a SCALAR_TO_VECTOR here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14412, __PRETTY_FUNCTION__)) |
14412 | "Expecting a SCALAR_TO_VECTOR here")((OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && "Expecting a SCALAR_TO_VECTOR here" ) ? static_cast<void> (0) : __assert_fail ("OrigSToV.getOpcode() == ISD::SCALAR_TO_VECTOR && \"Expecting a SCALAR_TO_VECTOR here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14412, __PRETTY_FUNCTION__)); |
14413 | SDValue Input = OrigSToV.getOperand(0); |
14414 | |
14415 | if (Input.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { |
14416 | ConstantSDNode *Idx = dyn_cast<ConstantSDNode>(Input.getOperand(1)); |
14417 | SDValue OrigVector = Input.getOperand(0); |
14418 | |
14419 | // Can't handle non-const element indices or different vector types |
14420 | // for the input to the extract and the output of the scalar_to_vector. |
14421 | if (Idx && VT == OrigVector.getValueType()) { |
14422 | SmallVector<int, 16> NewMask(VT.getVectorNumElements(), -1); |
14423 | NewMask[VT.getVectorNumElements() / 2] = Idx->getZExtValue(); |
14424 | return DAG.getVectorShuffle(VT, dl, OrigVector, OrigVector, NewMask); |
14425 | } |
14426 | } |
14427 | return DAG.getNode(PPCISD::SCALAR_TO_VECTOR_PERMUTED, dl, VT, |
14428 | OrigSToV.getOperand(0)); |
14429 | } |
14430 | |
14431 | // On little endian subtargets, combine shuffles such as: |
14432 | // vector_shuffle<16,1,17,3,18,5,19,7,20,9,21,11,22,13,23,15>, <zero>, %b |
14433 | // into: |
14434 | // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7>, <zero>, %b |
14435 | // because the latter can be matched to a single instruction merge. |
14436 | // Furthermore, SCALAR_TO_VECTOR on little endian always involves a permute |
14437 | // to put the value into element zero. Adjust the shuffle mask so that the |
14438 | // vector can remain in permuted form (to prevent a swap prior to a shuffle). |
14439 | SDValue PPCTargetLowering::combineVectorShuffle(ShuffleVectorSDNode *SVN, |
14440 | SelectionDAG &DAG) const { |
14441 | SDValue LHS = SVN->getOperand(0); |
14442 | SDValue RHS = SVN->getOperand(1); |
14443 | auto Mask = SVN->getMask(); |
14444 | int NumElts = LHS.getValueType().getVectorNumElements(); |
14445 | SDValue Res(SVN, 0); |
14446 | SDLoc dl(SVN); |
14447 | |
14448 | // None of these combines are useful on big endian systems since the ISA |
14449 | // already has a big endian bias. |
14450 | if (!Subtarget.isLittleEndian() || !Subtarget.hasVSX()) |
14451 | return Res; |
14452 | |
14453 | // If this is not a shuffle of a shuffle and the first element comes from |
14454 | // the second vector, canonicalize to the commuted form. This will make it |
14455 | // more likely to match one of the single instruction patterns. |
14456 | if (Mask[0] >= NumElts && LHS.getOpcode() != ISD::VECTOR_SHUFFLE && |
14457 | RHS.getOpcode() != ISD::VECTOR_SHUFFLE) { |
14458 | std::swap(LHS, RHS); |
14459 | Res = DAG.getCommutedVectorShuffle(*SVN); |
14460 | Mask = cast<ShuffleVectorSDNode>(Res)->getMask(); |
14461 | } |
14462 | |
14463 | // Adjust the shuffle mask if either input vector comes from a |
14464 | // SCALAR_TO_VECTOR and keep the respective input vector in permuted |
14465 | // form (to prevent the need for a swap). |
14466 | SmallVector<int, 16> ShuffV(Mask.begin(), Mask.end()); |
14467 | SDValue SToVLHS = isScalarToVec(LHS); |
14468 | SDValue SToVRHS = isScalarToVec(RHS); |
14469 | if (SToVLHS || SToVRHS) { |
14470 | int NumEltsIn = SToVLHS ? SToVLHS.getValueType().getVectorNumElements() |
14471 | : SToVRHS.getValueType().getVectorNumElements(); |
14472 | int NumEltsOut = ShuffV.size(); |
14473 | |
14474 | // Initially assume that neither input is permuted. These will be adjusted |
14475 | // accordingly if either input is. |
14476 | int LHSMaxIdx = -1; |
14477 | int RHSMinIdx = -1; |
14478 | int RHSMaxIdx = -1; |
14479 | int HalfVec = LHS.getValueType().getVectorNumElements() / 2; |
14480 | |
14481 | // Get the permuted scalar to vector nodes for the source(s) that come from |
14482 | // ISD::SCALAR_TO_VECTOR. |
14483 | if (SToVLHS) { |
14484 | // Set up the values for the shuffle vector fixup. |
14485 | LHSMaxIdx = NumEltsOut / NumEltsIn; |
14486 | SToVLHS = getSToVPermuted(SToVLHS, DAG); |
14487 | if (SToVLHS.getValueType() != LHS.getValueType()) |
14488 | SToVLHS = DAG.getBitcast(LHS.getValueType(), SToVLHS); |
14489 | LHS = SToVLHS; |
14490 | } |
14491 | if (SToVRHS) { |
14492 | RHSMinIdx = NumEltsOut; |
14493 | RHSMaxIdx = NumEltsOut / NumEltsIn + RHSMinIdx; |
14494 | SToVRHS = getSToVPermuted(SToVRHS, DAG); |
14495 | if (SToVRHS.getValueType() != RHS.getValueType()) |
14496 | SToVRHS = DAG.getBitcast(RHS.getValueType(), SToVRHS); |
14497 | RHS = SToVRHS; |
14498 | } |
14499 | |
14500 | // Fix up the shuffle mask to reflect where the desired element actually is. |
14501 | // The minimum and maximum indices that correspond to element zero for both |
14502 | // the LHS and RHS are computed and will control which shuffle mask entries |
14503 | // are to be changed. For example, if the RHS is permuted, any shuffle mask |
14504 | // entries in the range [RHSMinIdx,RHSMaxIdx) will be incremented by |
14505 | // HalfVec to refer to the corresponding element in the permuted vector. |
14506 | fixupShuffleMaskForPermutedSToV(ShuffV, LHSMaxIdx, RHSMinIdx, RHSMaxIdx, |
14507 | HalfVec); |
14508 | Res = DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV); |
14509 | |
14510 | // We may have simplified away the shuffle. We won't be able to do anything |
14511 | // further with it here. |
14512 | if (!isa<ShuffleVectorSDNode>(Res)) |
14513 | return Res; |
14514 | Mask = cast<ShuffleVectorSDNode>(Res)->getMask(); |
14515 | } |
14516 | |
14517 | // The common case after we commuted the shuffle is that the RHS is a splat |
14518 | // and we have elements coming in from the splat at indices that are not |
14519 | // conducive to using a merge. |
14520 | // Example: |
14521 | // vector_shuffle<0,17,1,19,2,21,3,23,4,25,5,27,6,29,7,31> t1, <zero> |
14522 | if (!isSplatBV(RHS)) |
14523 | return Res; |
14524 | |
14525 | // We are looking for a mask such that all even elements are from |
14526 | // one vector and all odd elements from the other. |
14527 | if (!isAlternatingShuffMask(Mask, NumElts)) |
14528 | return Res; |
14529 | |
14530 | // Adjust the mask so we are pulling in the same index from the splat |
14531 | // as the index from the interesting vector in consecutive elements. |
14532 | // Example (even elements from first vector): |
14533 | // vector_shuffle<0,16,1,17,2,18,3,19,4,20,5,21,6,22,7,23> t1, <zero> |
14534 | if (Mask[0] < NumElts) |
14535 | for (int i = 1, e = Mask.size(); i < e; i += 2) |
14536 | ShuffV[i] = (ShuffV[i - 1] + NumElts); |
14537 | // Example (odd elements from first vector): |
14538 | // vector_shuffle<16,0,17,1,18,2,19,3,20,4,21,5,22,6,23,7> t1, <zero> |
14539 | else |
14540 | for (int i = 0, e = Mask.size(); i < e; i += 2) |
14541 | ShuffV[i] = (ShuffV[i + 1] + NumElts); |
14542 | |
14543 | // If the RHS has undefs, we need to remove them since we may have created |
14544 | // a shuffle that adds those instead of the splat value. |
14545 | SDValue SplatVal = cast<BuildVectorSDNode>(RHS.getNode())->getSplatValue(); |
14546 | RHS = DAG.getSplatBuildVector(RHS.getValueType(), dl, SplatVal); |
14547 | |
14548 | Res = DAG.getVectorShuffle(SVN->getValueType(0), dl, LHS, RHS, ShuffV); |
14549 | return Res; |
14550 | } |
14551 | |
14552 | SDValue PPCTargetLowering::combineVReverseMemOP(ShuffleVectorSDNode *SVN, |
14553 | LSBaseSDNode *LSBase, |
14554 | DAGCombinerInfo &DCI) const { |
14555 | assert((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) &&(((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && "Not a reverse memop pattern!") ? static_cast<void> (0 ) : __assert_fail ("(ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && \"Not a reverse memop pattern!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14556, __PRETTY_FUNCTION__)) |
14556 | "Not a reverse memop pattern!")(((ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && "Not a reverse memop pattern!") ? static_cast<void> (0 ) : __assert_fail ("(ISD::isNormalLoad(LSBase) || ISD::isNormalStore(LSBase)) && \"Not a reverse memop pattern!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14556, __PRETTY_FUNCTION__)); |
14557 | |
14558 | auto IsElementReverse = [](const ShuffleVectorSDNode *SVN) -> bool { |
14559 | auto Mask = SVN->getMask(); |
14560 | int i = 0; |
14561 | auto I = Mask.rbegin(); |
14562 | auto E = Mask.rend(); |
14563 | |
14564 | for (; I != E; ++I) { |
14565 | if (*I != i) |
14566 | return false; |
14567 | i++; |
14568 | } |
14569 | return true; |
14570 | }; |
14571 | |
14572 | SelectionDAG &DAG = DCI.DAG; |
14573 | EVT VT = SVN->getValueType(0); |
14574 | |
14575 | if (!isTypeLegal(VT) || !Subtarget.isLittleEndian() || !Subtarget.hasVSX()) |
14576 | return SDValue(); |
14577 | |
14578 | // Before P9, we have PPCVSXSwapRemoval pass to hack the element order. |
14579 | // See comment in PPCVSXSwapRemoval.cpp. |
14580 | // It is conflict with PPCVSXSwapRemoval opt. So we don't do it. |
14581 | if (!Subtarget.hasP9Vector()) |
14582 | return SDValue(); |
14583 | |
14584 | if(!IsElementReverse(SVN)) |
14585 | return SDValue(); |
14586 | |
14587 | if (LSBase->getOpcode() == ISD::LOAD) { |
14588 | SDLoc dl(SVN); |
14589 | SDValue LoadOps[] = {LSBase->getChain(), LSBase->getBasePtr()}; |
14590 | return DAG.getMemIntrinsicNode( |
14591 | PPCISD::LOAD_VEC_BE, dl, DAG.getVTList(VT, MVT::Other), LoadOps, |
14592 | LSBase->getMemoryVT(), LSBase->getMemOperand()); |
14593 | } |
14594 | |
14595 | if (LSBase->getOpcode() == ISD::STORE) { |
14596 | SDLoc dl(LSBase); |
14597 | SDValue StoreOps[] = {LSBase->getChain(), SVN->getOperand(0), |
14598 | LSBase->getBasePtr()}; |
14599 | return DAG.getMemIntrinsicNode( |
14600 | PPCISD::STORE_VEC_BE, dl, DAG.getVTList(MVT::Other), StoreOps, |
14601 | LSBase->getMemoryVT(), LSBase->getMemOperand()); |
14602 | } |
14603 | |
14604 | llvm_unreachable("Expected a load or store node here")::llvm::llvm_unreachable_internal("Expected a load or store node here" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14604); |
14605 | } |
14606 | |
14607 | SDValue PPCTargetLowering::PerformDAGCombine(SDNode *N, |
14608 | DAGCombinerInfo &DCI) const { |
14609 | SelectionDAG &DAG = DCI.DAG; |
14610 | SDLoc dl(N); |
14611 | switch (N->getOpcode()) { |
14612 | default: break; |
14613 | case ISD::ADD: |
14614 | return combineADD(N, DCI); |
14615 | case ISD::SHL: |
14616 | return combineSHL(N, DCI); |
14617 | case ISD::SRA: |
14618 | return combineSRA(N, DCI); |
14619 | case ISD::SRL: |
14620 | return combineSRL(N, DCI); |
14621 | case ISD::MUL: |
14622 | return combineMUL(N, DCI); |
14623 | case ISD::FMA: |
14624 | case PPCISD::FNMSUB: |
14625 | return combineFMALike(N, DCI); |
14626 | case PPCISD::SHL: |
14627 | if (isNullConstant(N->getOperand(0))) // 0 << V -> 0. |
14628 | return N->getOperand(0); |
14629 | break; |
14630 | case PPCISD::SRL: |
14631 | if (isNullConstant(N->getOperand(0))) // 0 >>u V -> 0. |
14632 | return N->getOperand(0); |
14633 | break; |
14634 | case PPCISD::SRA: |
14635 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) { |
14636 | if (C->isNullValue() || // 0 >>s V -> 0. |
14637 | C->isAllOnesValue()) // -1 >>s V -> -1. |
14638 | return N->getOperand(0); |
14639 | } |
14640 | break; |
14641 | case ISD::SIGN_EXTEND: |
14642 | case ISD::ZERO_EXTEND: |
14643 | case ISD::ANY_EXTEND: |
14644 | return DAGCombineExtBoolTrunc(N, DCI); |
14645 | case ISD::TRUNCATE: |
14646 | return combineTRUNCATE(N, DCI); |
14647 | case ISD::SETCC: |
14648 | if (SDValue CSCC = combineSetCC(N, DCI)) |
14649 | return CSCC; |
14650 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
14651 | case ISD::SELECT_CC: |
14652 | return DAGCombineTruncBoolExt(N, DCI); |
14653 | case ISD::SINT_TO_FP: |
14654 | case ISD::UINT_TO_FP: |
14655 | return combineFPToIntToFP(N, DCI); |
14656 | case ISD::VECTOR_SHUFFLE: |
14657 | if (ISD::isNormalLoad(N->getOperand(0).getNode())) { |
14658 | LSBaseSDNode* LSBase = cast<LSBaseSDNode>(N->getOperand(0)); |
14659 | return combineVReverseMemOP(cast<ShuffleVectorSDNode>(N), LSBase, DCI); |
14660 | } |
14661 | return combineVectorShuffle(cast<ShuffleVectorSDNode>(N), DCI.DAG); |
14662 | case ISD::STORE: { |
14663 | |
14664 | EVT Op1VT = N->getOperand(1).getValueType(); |
14665 | unsigned Opcode = N->getOperand(1).getOpcode(); |
14666 | |
14667 | if (Opcode == ISD::FP_TO_SINT || Opcode == ISD::FP_TO_UINT) { |
14668 | SDValue Val= combineStoreFPToInt(N, DCI); |
14669 | if (Val) |
14670 | return Val; |
14671 | } |
14672 | |
14673 | if (Opcode == ISD::VECTOR_SHUFFLE && ISD::isNormalStore(N)) { |
14674 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N->getOperand(1)); |
14675 | SDValue Val= combineVReverseMemOP(SVN, cast<LSBaseSDNode>(N), DCI); |
14676 | if (Val) |
14677 | return Val; |
14678 | } |
14679 | |
14680 | // Turn STORE (BSWAP) -> sthbrx/stwbrx. |
14681 | if (cast<StoreSDNode>(N)->isUnindexed() && Opcode == ISD::BSWAP && |
14682 | N->getOperand(1).getNode()->hasOneUse() && |
14683 | (Op1VT == MVT::i32 || Op1VT == MVT::i16 || |
14684 | (Subtarget.hasLDBRX() && Subtarget.isPPC64() && Op1VT == MVT::i64))) { |
14685 | |
14686 | // STBRX can only handle simple types and it makes no sense to store less |
14687 | // two bytes in byte-reversed order. |
14688 | EVT mVT = cast<StoreSDNode>(N)->getMemoryVT(); |
14689 | if (mVT.isExtended() || mVT.getSizeInBits() < 16) |
14690 | break; |
14691 | |
14692 | SDValue BSwapOp = N->getOperand(1).getOperand(0); |
14693 | // Do an any-extend to 32-bits if this is a half-word input. |
14694 | if (BSwapOp.getValueType() == MVT::i16) |
14695 | BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp); |
14696 | |
14697 | // If the type of BSWAP operand is wider than stored memory width |
14698 | // it need to be shifted to the right side before STBRX. |
14699 | if (Op1VT.bitsGT(mVT)) { |
14700 | int Shift = Op1VT.getSizeInBits() - mVT.getSizeInBits(); |
14701 | BSwapOp = DAG.getNode(ISD::SRL, dl, Op1VT, BSwapOp, |
14702 | DAG.getConstant(Shift, dl, MVT::i32)); |
14703 | // Need to truncate if this is a bswap of i64 stored as i32/i16. |
14704 | if (Op1VT == MVT::i64) |
14705 | BSwapOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BSwapOp); |
14706 | } |
14707 | |
14708 | SDValue Ops[] = { |
14709 | N->getOperand(0), BSwapOp, N->getOperand(2), DAG.getValueType(mVT) |
14710 | }; |
14711 | return |
14712 | DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other), |
14713 | Ops, cast<StoreSDNode>(N)->getMemoryVT(), |
14714 | cast<StoreSDNode>(N)->getMemOperand()); |
14715 | } |
14716 | |
14717 | // STORE Constant:i32<0> -> STORE<trunc to i32> Constant:i64<0> |
14718 | // So it can increase the chance of CSE constant construction. |
14719 | if (Subtarget.isPPC64() && !DCI.isBeforeLegalize() && |
14720 | isa<ConstantSDNode>(N->getOperand(1)) && Op1VT == MVT::i32) { |
14721 | // Need to sign-extended to 64-bits to handle negative values. |
14722 | EVT MemVT = cast<StoreSDNode>(N)->getMemoryVT(); |
14723 | uint64_t Val64 = SignExtend64(N->getConstantOperandVal(1), |
14724 | MemVT.getSizeInBits()); |
14725 | SDValue Const64 = DAG.getConstant(Val64, dl, MVT::i64); |
14726 | |
14727 | // DAG.getTruncStore() can't be used here because it doesn't accept |
14728 | // the general (base + offset) addressing mode. |
14729 | // So we use UpdateNodeOperands and setTruncatingStore instead. |
14730 | DAG.UpdateNodeOperands(N, N->getOperand(0), Const64, N->getOperand(2), |
14731 | N->getOperand(3)); |
14732 | cast<StoreSDNode>(N)->setTruncatingStore(true); |
14733 | return SDValue(N, 0); |
14734 | } |
14735 | |
14736 | // For little endian, VSX stores require generating xxswapd/lxvd2x. |
14737 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting store. |
14738 | if (Op1VT.isSimple()) { |
14739 | MVT StoreVT = Op1VT.getSimpleVT(); |
14740 | if (Subtarget.needsSwapsForVSXMemOps() && |
14741 | (StoreVT == MVT::v2f64 || StoreVT == MVT::v2i64 || |
14742 | StoreVT == MVT::v4f32 || StoreVT == MVT::v4i32)) |
14743 | return expandVSXStoreForLE(N, DCI); |
14744 | } |
14745 | break; |
14746 | } |
14747 | case ISD::LOAD: { |
14748 | LoadSDNode *LD = cast<LoadSDNode>(N); |
14749 | EVT VT = LD->getValueType(0); |
14750 | |
14751 | // For little endian, VSX loads require generating lxvd2x/xxswapd. |
14752 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting load. |
14753 | if (VT.isSimple()) { |
14754 | MVT LoadVT = VT.getSimpleVT(); |
14755 | if (Subtarget.needsSwapsForVSXMemOps() && |
14756 | (LoadVT == MVT::v2f64 || LoadVT == MVT::v2i64 || |
14757 | LoadVT == MVT::v4f32 || LoadVT == MVT::v4i32)) |
14758 | return expandVSXLoadForLE(N, DCI); |
14759 | } |
14760 | |
14761 | // We sometimes end up with a 64-bit integer load, from which we extract |
14762 | // two single-precision floating-point numbers. This happens with |
14763 | // std::complex<float>, and other similar structures, because of the way we |
14764 | // canonicalize structure copies. However, if we lack direct moves, |
14765 | // then the final bitcasts from the extracted integer values to the |
14766 | // floating-point numbers turn into store/load pairs. Even with direct moves, |
14767 | // just loading the two floating-point numbers is likely better. |
14768 | auto ReplaceTwoFloatLoad = [&]() { |
14769 | if (VT != MVT::i64) |
14770 | return false; |
14771 | |
14772 | if (LD->getExtensionType() != ISD::NON_EXTLOAD || |
14773 | LD->isVolatile()) |
14774 | return false; |
14775 | |
14776 | // We're looking for a sequence like this: |
14777 | // t13: i64,ch = load<LD8[%ref.tmp]> t0, t6, undef:i64 |
14778 | // t16: i64 = srl t13, Constant:i32<32> |
14779 | // t17: i32 = truncate t16 |
14780 | // t18: f32 = bitcast t17 |
14781 | // t19: i32 = truncate t13 |
14782 | // t20: f32 = bitcast t19 |
14783 | |
14784 | if (!LD->hasNUsesOfValue(2, 0)) |
14785 | return false; |
14786 | |
14787 | auto UI = LD->use_begin(); |
14788 | while (UI.getUse().getResNo() != 0) ++UI; |
14789 | SDNode *Trunc = *UI++; |
14790 | while (UI.getUse().getResNo() != 0) ++UI; |
14791 | SDNode *RightShift = *UI; |
14792 | if (Trunc->getOpcode() != ISD::TRUNCATE) |
14793 | std::swap(Trunc, RightShift); |
14794 | |
14795 | if (Trunc->getOpcode() != ISD::TRUNCATE || |
14796 | Trunc->getValueType(0) != MVT::i32 || |
14797 | !Trunc->hasOneUse()) |
14798 | return false; |
14799 | if (RightShift->getOpcode() != ISD::SRL || |
14800 | !isa<ConstantSDNode>(RightShift->getOperand(1)) || |
14801 | RightShift->getConstantOperandVal(1) != 32 || |
14802 | !RightShift->hasOneUse()) |
14803 | return false; |
14804 | |
14805 | SDNode *Trunc2 = *RightShift->use_begin(); |
14806 | if (Trunc2->getOpcode() != ISD::TRUNCATE || |
14807 | Trunc2->getValueType(0) != MVT::i32 || |
14808 | !Trunc2->hasOneUse()) |
14809 | return false; |
14810 | |
14811 | SDNode *Bitcast = *Trunc->use_begin(); |
14812 | SDNode *Bitcast2 = *Trunc2->use_begin(); |
14813 | |
14814 | if (Bitcast->getOpcode() != ISD::BITCAST || |
14815 | Bitcast->getValueType(0) != MVT::f32) |
14816 | return false; |
14817 | if (Bitcast2->getOpcode() != ISD::BITCAST || |
14818 | Bitcast2->getValueType(0) != MVT::f32) |
14819 | return false; |
14820 | |
14821 | if (Subtarget.isLittleEndian()) |
14822 | std::swap(Bitcast, Bitcast2); |
14823 | |
14824 | // Bitcast has the second float (in memory-layout order) and Bitcast2 |
14825 | // has the first one. |
14826 | |
14827 | SDValue BasePtr = LD->getBasePtr(); |
14828 | if (LD->isIndexed()) { |
14829 | assert(LD->getAddressingMode() == ISD::PRE_INC &&((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14830, __PRETTY_FUNCTION__)) |
14830 | "Non-pre-inc AM on PPC?")((LD->getAddressingMode() == ISD::PRE_INC && "Non-pre-inc AM on PPC?" ) ? static_cast<void> (0) : __assert_fail ("LD->getAddressingMode() == ISD::PRE_INC && \"Non-pre-inc AM on PPC?\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 14830, __PRETTY_FUNCTION__)); |
14831 | BasePtr = |
14832 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, |
14833 | LD->getOffset()); |
14834 | } |
14835 | |
14836 | auto MMOFlags = |
14837 | LD->getMemOperand()->getFlags() & ~MachineMemOperand::MOVolatile; |
14838 | SDValue FloatLoad = DAG.getLoad(MVT::f32, dl, LD->getChain(), BasePtr, |
14839 | LD->getPointerInfo(), LD->getAlignment(), |
14840 | MMOFlags, LD->getAAInfo()); |
14841 | SDValue AddPtr = |
14842 | DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), |
14843 | BasePtr, DAG.getIntPtrConstant(4, dl)); |
14844 | SDValue FloatLoad2 = DAG.getLoad( |
14845 | MVT::f32, dl, SDValue(FloatLoad.getNode(), 1), AddPtr, |
14846 | LD->getPointerInfo().getWithOffset(4), |
14847 | MinAlign(LD->getAlignment(), 4), MMOFlags, LD->getAAInfo()); |
14848 | |
14849 | if (LD->isIndexed()) { |
14850 | // Note that DAGCombine should re-form any pre-increment load(s) from |
14851 | // what is produced here if that makes sense. |
14852 | DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), BasePtr); |
14853 | } |
14854 | |
14855 | DCI.CombineTo(Bitcast2, FloatLoad); |
14856 | DCI.CombineTo(Bitcast, FloatLoad2); |
14857 | |
14858 | DAG.ReplaceAllUsesOfValueWith(SDValue(LD, LD->isIndexed() ? 2 : 1), |
14859 | SDValue(FloatLoad2.getNode(), 1)); |
14860 | return true; |
14861 | }; |
14862 | |
14863 | if (ReplaceTwoFloatLoad()) |
14864 | return SDValue(N, 0); |
14865 | |
14866 | EVT MemVT = LD->getMemoryVT(); |
14867 | Type *Ty = MemVT.getTypeForEVT(*DAG.getContext()); |
14868 | Align ABIAlignment = DAG.getDataLayout().getABITypeAlign(Ty); |
14869 | if (LD->isUnindexed() && VT.isVector() && |
14870 | ((Subtarget.hasAltivec() && ISD::isNON_EXTLoad(N) && |
14871 | // P8 and later hardware should just use LOAD. |
14872 | !Subtarget.hasP8Vector() && |
14873 | (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 || |
14874 | VT == MVT::v4f32))) && |
14875 | LD->getAlign() < ABIAlignment) { |
14876 | // This is a type-legal unaligned Altivec load. |
14877 | SDValue Chain = LD->getChain(); |
14878 | SDValue Ptr = LD->getBasePtr(); |
14879 | bool isLittleEndian = Subtarget.isLittleEndian(); |
14880 | |
14881 | // This implements the loading of unaligned vectors as described in |
14882 | // the venerable Apple Velocity Engine overview. Specifically: |
14883 | // https://developer.apple.com/hardwaredrivers/ve/alignment.html |
14884 | // https://developer.apple.com/hardwaredrivers/ve/code_optimization.html |
14885 | // |
14886 | // The general idea is to expand a sequence of one or more unaligned |
14887 | // loads into an alignment-based permutation-control instruction (lvsl |
14888 | // or lvsr), a series of regular vector loads (which always truncate |
14889 | // their input address to an aligned address), and a series of |
14890 | // permutations. The results of these permutations are the requested |
14891 | // loaded values. The trick is that the last "extra" load is not taken |
14892 | // from the address you might suspect (sizeof(vector) bytes after the |
14893 | // last requested load), but rather sizeof(vector) - 1 bytes after the |
14894 | // last requested vector. The point of this is to avoid a page fault if |
14895 | // the base address happened to be aligned. This works because if the |
14896 | // base address is aligned, then adding less than a full vector length |
14897 | // will cause the last vector in the sequence to be (re)loaded. |
14898 | // Otherwise, the next vector will be fetched as you might suspect was |
14899 | // necessary. |
14900 | |
14901 | // We might be able to reuse the permutation generation from |
14902 | // a different base address offset from this one by an aligned amount. |
14903 | // The INTRINSIC_WO_CHAIN DAG combine will attempt to perform this |
14904 | // optimization later. |
14905 | Intrinsic::ID Intr, IntrLD, IntrPerm; |
14906 | MVT PermCntlTy, PermTy, LDTy; |
14907 | Intr = isLittleEndian ? Intrinsic::ppc_altivec_lvsr |
14908 | : Intrinsic::ppc_altivec_lvsl; |
14909 | IntrLD = Intrinsic::ppc_altivec_lvx; |
14910 | IntrPerm = Intrinsic::ppc_altivec_vperm; |
14911 | PermCntlTy = MVT::v16i8; |
14912 | PermTy = MVT::v4i32; |
14913 | LDTy = MVT::v4i32; |
14914 | |
14915 | SDValue PermCntl = BuildIntrinsicOp(Intr, Ptr, DAG, dl, PermCntlTy); |
14916 | |
14917 | // Create the new MMO for the new base load. It is like the original MMO, |
14918 | // but represents an area in memory almost twice the vector size centered |
14919 | // on the original address. If the address is unaligned, we might start |
14920 | // reading up to (sizeof(vector)-1) bytes below the address of the |
14921 | // original unaligned load. |
14922 | MachineFunction &MF = DAG.getMachineFunction(); |
14923 | MachineMemOperand *BaseMMO = |
14924 | MF.getMachineMemOperand(LD->getMemOperand(), |
14925 | -(long)MemVT.getStoreSize()+1, |
14926 | 2*MemVT.getStoreSize()-1); |
14927 | |
14928 | // Create the new base load. |
14929 | SDValue LDXIntID = |
14930 | DAG.getTargetConstant(IntrLD, dl, getPointerTy(MF.getDataLayout())); |
14931 | SDValue BaseLoadOps[] = { Chain, LDXIntID, Ptr }; |
14932 | SDValue BaseLoad = |
14933 | DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl, |
14934 | DAG.getVTList(PermTy, MVT::Other), |
14935 | BaseLoadOps, LDTy, BaseMMO); |
14936 | |
14937 | // Note that the value of IncOffset (which is provided to the next |
14938 | // load's pointer info offset value, and thus used to calculate the |
14939 | // alignment), and the value of IncValue (which is actually used to |
14940 | // increment the pointer value) are different! This is because we |
14941 | // require the next load to appear to be aligned, even though it |
14942 | // is actually offset from the base pointer by a lesser amount. |
14943 | int IncOffset = VT.getSizeInBits() / 8; |
14944 | int IncValue = IncOffset; |
14945 | |
14946 | // Walk (both up and down) the chain looking for another load at the real |
14947 | // (aligned) offset (the alignment of the other load does not matter in |
14948 | // this case). If found, then do not use the offset reduction trick, as |
14949 | // that will prevent the loads from being later combined (as they would |
14950 | // otherwise be duplicates). |
14951 | if (!findConsecutiveLoad(LD, DAG)) |
14952 | --IncValue; |
14953 | |
14954 | SDValue Increment = |
14955 | DAG.getConstant(IncValue, dl, getPointerTy(MF.getDataLayout())); |
14956 | Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); |
14957 | |
14958 | MachineMemOperand *ExtraMMO = |
14959 | MF.getMachineMemOperand(LD->getMemOperand(), |
14960 | 1, 2*MemVT.getStoreSize()-1); |
14961 | SDValue ExtraLoadOps[] = { Chain, LDXIntID, Ptr }; |
14962 | SDValue ExtraLoad = |
14963 | DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, dl, |
14964 | DAG.getVTList(PermTy, MVT::Other), |
14965 | ExtraLoadOps, LDTy, ExtraMMO); |
14966 | |
14967 | SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, |
14968 | BaseLoad.getValue(1), ExtraLoad.getValue(1)); |
14969 | |
14970 | // Because vperm has a big-endian bias, we must reverse the order |
14971 | // of the input vectors and complement the permute control vector |
14972 | // when generating little endian code. We have already handled the |
14973 | // latter by using lvsr instead of lvsl, so just reverse BaseLoad |
14974 | // and ExtraLoad here. |
14975 | SDValue Perm; |
14976 | if (isLittleEndian) |
14977 | Perm = BuildIntrinsicOp(IntrPerm, |
14978 | ExtraLoad, BaseLoad, PermCntl, DAG, dl); |
14979 | else |
14980 | Perm = BuildIntrinsicOp(IntrPerm, |
14981 | BaseLoad, ExtraLoad, PermCntl, DAG, dl); |
14982 | |
14983 | if (VT != PermTy) |
14984 | Perm = Subtarget.hasAltivec() |
14985 | ? DAG.getNode(ISD::BITCAST, dl, VT, Perm) |
14986 | : DAG.getNode(ISD::FP_ROUND, dl, VT, Perm, |
14987 | DAG.getTargetConstant(1, dl, MVT::i64)); |
14988 | // second argument is 1 because this rounding |
14989 | // is always exact. |
14990 | |
14991 | // The output of the permutation is our loaded result, the TokenFactor is |
14992 | // our new chain. |
14993 | DCI.CombineTo(N, Perm, TF); |
14994 | return SDValue(N, 0); |
14995 | } |
14996 | } |
14997 | break; |
14998 | case ISD::INTRINSIC_WO_CHAIN: { |
14999 | bool isLittleEndian = Subtarget.isLittleEndian(); |
15000 | unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); |
15001 | Intrinsic::ID Intr = (isLittleEndian ? Intrinsic::ppc_altivec_lvsr |
15002 | : Intrinsic::ppc_altivec_lvsl); |
15003 | if (IID == Intr && N->getOperand(1)->getOpcode() == ISD::ADD) { |
15004 | SDValue Add = N->getOperand(1); |
15005 | |
15006 | int Bits = 4 /* 16 byte alignment */; |
15007 | |
15008 | if (DAG.MaskedValueIsZero(Add->getOperand(1), |
15009 | APInt::getAllOnesValue(Bits /* alignment */) |
15010 | .zext(Add.getScalarValueSizeInBits()))) { |
15011 | SDNode *BasePtr = Add->getOperand(0).getNode(); |
15012 | for (SDNode::use_iterator UI = BasePtr->use_begin(), |
15013 | UE = BasePtr->use_end(); |
15014 | UI != UE; ++UI) { |
15015 | if (UI->getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15016 | cast<ConstantSDNode>(UI->getOperand(0))->getZExtValue() == |
15017 | IID) { |
15018 | // We've found another LVSL/LVSR, and this address is an aligned |
15019 | // multiple of that one. The results will be the same, so use the |
15020 | // one we've just found instead. |
15021 | |
15022 | return SDValue(*UI, 0); |
15023 | } |
15024 | } |
15025 | } |
15026 | |
15027 | if (isa<ConstantSDNode>(Add->getOperand(1))) { |
15028 | SDNode *BasePtr = Add->getOperand(0).getNode(); |
15029 | for (SDNode::use_iterator UI = BasePtr->use_begin(), |
15030 | UE = BasePtr->use_end(); UI != UE; ++UI) { |
15031 | if (UI->getOpcode() == ISD::ADD && |
15032 | isa<ConstantSDNode>(UI->getOperand(1)) && |
15033 | (cast<ConstantSDNode>(Add->getOperand(1))->getZExtValue() - |
15034 | cast<ConstantSDNode>(UI->getOperand(1))->getZExtValue()) % |
15035 | (1ULL << Bits) == 0) { |
15036 | SDNode *OtherAdd = *UI; |
15037 | for (SDNode::use_iterator VI = OtherAdd->use_begin(), |
15038 | VE = OtherAdd->use_end(); VI != VE; ++VI) { |
15039 | if (VI->getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15040 | cast<ConstantSDNode>(VI->getOperand(0))->getZExtValue() == IID) { |
15041 | return SDValue(*VI, 0); |
15042 | } |
15043 | } |
15044 | } |
15045 | } |
15046 | } |
15047 | } |
15048 | |
15049 | // Combine vmaxsw/h/b(a, a's negation) to abs(a) |
15050 | // Expose the vabsduw/h/b opportunity for down stream |
15051 | if (!DCI.isAfterLegalizeDAG() && Subtarget.hasP9Altivec() && |
15052 | (IID == Intrinsic::ppc_altivec_vmaxsw || |
15053 | IID == Intrinsic::ppc_altivec_vmaxsh || |
15054 | IID == Intrinsic::ppc_altivec_vmaxsb)) { |
15055 | SDValue V1 = N->getOperand(1); |
15056 | SDValue V2 = N->getOperand(2); |
15057 | if ((V1.getSimpleValueType() == MVT::v4i32 || |
15058 | V1.getSimpleValueType() == MVT::v8i16 || |
15059 | V1.getSimpleValueType() == MVT::v16i8) && |
15060 | V1.getSimpleValueType() == V2.getSimpleValueType()) { |
15061 | // (0-a, a) |
15062 | if (V1.getOpcode() == ISD::SUB && |
15063 | ISD::isBuildVectorAllZeros(V1.getOperand(0).getNode()) && |
15064 | V1.getOperand(1) == V2) { |
15065 | return DAG.getNode(ISD::ABS, dl, V2.getValueType(), V2); |
15066 | } |
15067 | // (a, 0-a) |
15068 | if (V2.getOpcode() == ISD::SUB && |
15069 | ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()) && |
15070 | V2.getOperand(1) == V1) { |
15071 | return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1); |
15072 | } |
15073 | // (x-y, y-x) |
15074 | if (V1.getOpcode() == ISD::SUB && V2.getOpcode() == ISD::SUB && |
15075 | V1.getOperand(0) == V2.getOperand(1) && |
15076 | V1.getOperand(1) == V2.getOperand(0)) { |
15077 | return DAG.getNode(ISD::ABS, dl, V1.getValueType(), V1); |
15078 | } |
15079 | } |
15080 | } |
15081 | } |
15082 | |
15083 | break; |
15084 | case ISD::INTRINSIC_W_CHAIN: |
15085 | // For little endian, VSX loads require generating lxvd2x/xxswapd. |
15086 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting load. |
15087 | if (Subtarget.needsSwapsForVSXMemOps()) { |
15088 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
15089 | default: |
15090 | break; |
15091 | case Intrinsic::ppc_vsx_lxvw4x: |
15092 | case Intrinsic::ppc_vsx_lxvd2x: |
15093 | return expandVSXLoadForLE(N, DCI); |
15094 | } |
15095 | } |
15096 | break; |
15097 | case ISD::INTRINSIC_VOID: |
15098 | // For little endian, VSX stores require generating xxswapd/stxvd2x. |
15099 | // Not needed on ISA 3.0 based CPUs since we have a non-permuting store. |
15100 | if (Subtarget.needsSwapsForVSXMemOps()) { |
15101 | switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) { |
15102 | default: |
15103 | break; |
15104 | case Intrinsic::ppc_vsx_stxvw4x: |
15105 | case Intrinsic::ppc_vsx_stxvd2x: |
15106 | return expandVSXStoreForLE(N, DCI); |
15107 | } |
15108 | } |
15109 | break; |
15110 | case ISD::BSWAP: |
15111 | // Turn BSWAP (LOAD) -> lhbrx/lwbrx. |
15112 | if (ISD::isNON_EXTLoad(N->getOperand(0).getNode()) && |
15113 | N->getOperand(0).hasOneUse() && |
15114 | (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16 || |
15115 | (Subtarget.hasLDBRX() && Subtarget.isPPC64() && |
15116 | N->getValueType(0) == MVT::i64))) { |
15117 | SDValue Load = N->getOperand(0); |
15118 | LoadSDNode *LD = cast<LoadSDNode>(Load); |
15119 | // Create the byte-swapping load. |
15120 | SDValue Ops[] = { |
15121 | LD->getChain(), // Chain |
15122 | LD->getBasePtr(), // Ptr |
15123 | DAG.getValueType(N->getValueType(0)) // VT |
15124 | }; |
15125 | SDValue BSLoad = |
15126 | DAG.getMemIntrinsicNode(PPCISD::LBRX, dl, |
15127 | DAG.getVTList(N->getValueType(0) == MVT::i64 ? |
15128 | MVT::i64 : MVT::i32, MVT::Other), |
15129 | Ops, LD->getMemoryVT(), LD->getMemOperand()); |
15130 | |
15131 | // If this is an i16 load, insert the truncate. |
15132 | SDValue ResVal = BSLoad; |
15133 | if (N->getValueType(0) == MVT::i16) |
15134 | ResVal = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, BSLoad); |
15135 | |
15136 | // First, combine the bswap away. This makes the value produced by the |
15137 | // load dead. |
15138 | DCI.CombineTo(N, ResVal); |
15139 | |
15140 | // Next, combine the load away, we give it a bogus result value but a real |
15141 | // chain result. The result value is dead because the bswap is dead. |
15142 | DCI.CombineTo(Load.getNode(), ResVal, BSLoad.getValue(1)); |
15143 | |
15144 | // Return N so it doesn't get rechecked! |
15145 | return SDValue(N, 0); |
15146 | } |
15147 | break; |
15148 | case PPCISD::VCMP: |
15149 | // If a VCMPo node already exists with exactly the same operands as this |
15150 | // node, use its result instead of this node (VCMPo computes both a CR6 and |
15151 | // a normal output). |
15152 | // |
15153 | if (!N->getOperand(0).hasOneUse() && |
15154 | !N->getOperand(1).hasOneUse() && |
15155 | !N->getOperand(2).hasOneUse()) { |
15156 | |
15157 | // Scan all of the users of the LHS, looking for VCMPo's that match. |
15158 | SDNode *VCMPoNode = nullptr; |
15159 | |
15160 | SDNode *LHSN = N->getOperand(0).getNode(); |
15161 | for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end(); |
15162 | UI != E; ++UI) |
15163 | if (UI->getOpcode() == PPCISD::VCMPo && |
15164 | UI->getOperand(1) == N->getOperand(1) && |
15165 | UI->getOperand(2) == N->getOperand(2) && |
15166 | UI->getOperand(0) == N->getOperand(0)) { |
15167 | VCMPoNode = *UI; |
15168 | break; |
15169 | } |
15170 | |
15171 | // If there is no VCMPo node, or if the flag value has a single use, don't |
15172 | // transform this. |
15173 | if (!VCMPoNode || VCMPoNode->hasNUsesOfValue(0, 1)) |
15174 | break; |
15175 | |
15176 | // Look at the (necessarily single) use of the flag value. If it has a |
15177 | // chain, this transformation is more complex. Note that multiple things |
15178 | // could use the value result, which we should ignore. |
15179 | SDNode *FlagUser = nullptr; |
15180 | for (SDNode::use_iterator UI = VCMPoNode->use_begin(); |
15181 | FlagUser == nullptr; ++UI) { |
15182 | assert(UI != VCMPoNode->use_end() && "Didn't find user!")((UI != VCMPoNode->use_end() && "Didn't find user!" ) ? static_cast<void> (0) : __assert_fail ("UI != VCMPoNode->use_end() && \"Didn't find user!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15182, __PRETTY_FUNCTION__)); |
15183 | SDNode *User = *UI; |
15184 | for (unsigned i = 0, e = User->getNumOperands(); i != e; ++i) { |
15185 | if (User->getOperand(i) == SDValue(VCMPoNode, 1)) { |
15186 | FlagUser = User; |
15187 | break; |
15188 | } |
15189 | } |
15190 | } |
15191 | |
15192 | // If the user is a MFOCRF instruction, we know this is safe. |
15193 | // Otherwise we give up for right now. |
15194 | if (FlagUser->getOpcode() == PPCISD::MFOCRF) |
15195 | return SDValue(VCMPoNode, 0); |
15196 | } |
15197 | break; |
15198 | case ISD::BRCOND: { |
15199 | SDValue Cond = N->getOperand(1); |
15200 | SDValue Target = N->getOperand(2); |
15201 | |
15202 | if (Cond.getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15203 | cast<ConstantSDNode>(Cond.getOperand(1))->getZExtValue() == |
15204 | Intrinsic::loop_decrement) { |
15205 | |
15206 | // We now need to make the intrinsic dead (it cannot be instruction |
15207 | // selected). |
15208 | DAG.ReplaceAllUsesOfValueWith(Cond.getValue(1), Cond.getOperand(0)); |
15209 | assert(Cond.getNode()->hasOneUse() &&((Cond.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("Cond.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15210, __PRETTY_FUNCTION__)) |
15210 | "Counter decrement has more than one use")((Cond.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("Cond.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15210, __PRETTY_FUNCTION__)); |
15211 | |
15212 | return DAG.getNode(PPCISD::BDNZ, dl, MVT::Other, |
15213 | N->getOperand(0), Target); |
15214 | } |
15215 | } |
15216 | break; |
15217 | case ISD::BR_CC: { |
15218 | // If this is a branch on an altivec predicate comparison, lower this so |
15219 | // that we don't have to do a MFOCRF: instead, branch directly on CR6. This |
15220 | // lowering is done pre-legalize, because the legalizer lowers the predicate |
15221 | // compare down to code that is difficult to reassemble. |
15222 | ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get(); |
15223 | SDValue LHS = N->getOperand(2), RHS = N->getOperand(3); |
15224 | |
15225 | // Sometimes the promoted value of the intrinsic is ANDed by some non-zero |
15226 | // value. If so, pass-through the AND to get to the intrinsic. |
15227 | if (LHS.getOpcode() == ISD::AND && |
15228 | LHS.getOperand(0).getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15229 | cast<ConstantSDNode>(LHS.getOperand(0).getOperand(1))->getZExtValue() == |
15230 | Intrinsic::loop_decrement && |
15231 | isa<ConstantSDNode>(LHS.getOperand(1)) && |
15232 | !isNullConstant(LHS.getOperand(1))) |
15233 | LHS = LHS.getOperand(0); |
15234 | |
15235 | if (LHS.getOpcode() == ISD::INTRINSIC_W_CHAIN && |
15236 | cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue() == |
15237 | Intrinsic::loop_decrement && |
15238 | isa<ConstantSDNode>(RHS)) { |
15239 | assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&(((CC == ISD::SETEQ || CC == ISD::SETNE) && "Counter decrement comparison is not EQ or NE" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETEQ || CC == ISD::SETNE) && \"Counter decrement comparison is not EQ or NE\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15240, __PRETTY_FUNCTION__)) |
15240 | "Counter decrement comparison is not EQ or NE")(((CC == ISD::SETEQ || CC == ISD::SETNE) && "Counter decrement comparison is not EQ or NE" ) ? static_cast<void> (0) : __assert_fail ("(CC == ISD::SETEQ || CC == ISD::SETNE) && \"Counter decrement comparison is not EQ or NE\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15240, __PRETTY_FUNCTION__)); |
15241 | |
15242 | unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue(); |
15243 | bool isBDNZ = (CC == ISD::SETEQ && Val) || |
15244 | (CC == ISD::SETNE && !Val); |
15245 | |
15246 | // We now need to make the intrinsic dead (it cannot be instruction |
15247 | // selected). |
15248 | DAG.ReplaceAllUsesOfValueWith(LHS.getValue(1), LHS.getOperand(0)); |
15249 | assert(LHS.getNode()->hasOneUse() &&((LHS.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("LHS.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15250, __PRETTY_FUNCTION__)) |
15250 | "Counter decrement has more than one use")((LHS.getNode()->hasOneUse() && "Counter decrement has more than one use" ) ? static_cast<void> (0) : __assert_fail ("LHS.getNode()->hasOneUse() && \"Counter decrement has more than one use\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15250, __PRETTY_FUNCTION__)); |
15251 | |
15252 | return DAG.getNode(isBDNZ ? PPCISD::BDNZ : PPCISD::BDZ, dl, MVT::Other, |
15253 | N->getOperand(0), N->getOperand(4)); |
15254 | } |
15255 | |
15256 | int CompareOpc; |
15257 | bool isDot; |
15258 | |
15259 | if (LHS.getOpcode() == ISD::INTRINSIC_WO_CHAIN && |
15260 | isa<ConstantSDNode>(RHS) && (CC == ISD::SETEQ || CC == ISD::SETNE) && |
15261 | getVectorCompareInfo(LHS, CompareOpc, isDot, Subtarget)) { |
15262 | assert(isDot && "Can't compare against a vector result!")((isDot && "Can't compare against a vector result!") ? static_cast<void> (0) : __assert_fail ("isDot && \"Can't compare against a vector result!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15262, __PRETTY_FUNCTION__)); |
15263 | |
15264 | // If this is a comparison against something other than 0/1, then we know |
15265 | // that the condition is never/always true. |
15266 | unsigned Val = cast<ConstantSDNode>(RHS)->getZExtValue(); |
15267 | if (Val != 0 && Val != 1) { |
15268 | if (CC == ISD::SETEQ) // Cond never true, remove branch. |
15269 | return N->getOperand(0); |
15270 | // Always !=, turn it into an unconditional branch. |
15271 | return DAG.getNode(ISD::BR, dl, MVT::Other, |
15272 | N->getOperand(0), N->getOperand(4)); |
15273 | } |
15274 | |
15275 | bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0); |
15276 | |
15277 | // Create the PPCISD altivec 'dot' comparison node. |
15278 | SDValue Ops[] = { |
15279 | LHS.getOperand(2), // LHS of compare |
15280 | LHS.getOperand(3), // RHS of compare |
15281 | DAG.getConstant(CompareOpc, dl, MVT::i32) |
15282 | }; |
15283 | EVT VTs[] = { LHS.getOperand(2).getValueType(), MVT::Glue }; |
15284 | SDValue CompNode = DAG.getNode(PPCISD::VCMPo, dl, VTs, Ops); |
15285 | |
15286 | // Unpack the result based on how the target uses it. |
15287 | PPC::Predicate CompOpc; |
15288 | switch (cast<ConstantSDNode>(LHS.getOperand(1))->getZExtValue()) { |
15289 | default: // Can't happen, don't crash on invalid number though. |
15290 | case 0: // Branch on the value of the EQ bit of CR6. |
15291 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE; |
15292 | break; |
15293 | case 1: // Branch on the inverted value of the EQ bit of CR6. |
15294 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ; |
15295 | break; |
15296 | case 2: // Branch on the value of the LT bit of CR6. |
15297 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE; |
15298 | break; |
15299 | case 3: // Branch on the inverted value of the LT bit of CR6. |
15300 | CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT; |
15301 | break; |
15302 | } |
15303 | |
15304 | return DAG.getNode(PPCISD::COND_BRANCH, dl, MVT::Other, N->getOperand(0), |
15305 | DAG.getConstant(CompOpc, dl, MVT::i32), |
15306 | DAG.getRegister(PPC::CR6, MVT::i32), |
15307 | N->getOperand(4), CompNode.getValue(1)); |
15308 | } |
15309 | break; |
15310 | } |
15311 | case ISD::BUILD_VECTOR: |
15312 | return DAGCombineBuildVector(N, DCI); |
15313 | case ISD::ABS: |
15314 | return combineABS(N, DCI); |
15315 | case ISD::VSELECT: |
15316 | return combineVSelect(N, DCI); |
15317 | } |
15318 | |
15319 | return SDValue(); |
15320 | } |
15321 | |
15322 | SDValue |
15323 | PPCTargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, |
15324 | SelectionDAG &DAG, |
15325 | SmallVectorImpl<SDNode *> &Created) const { |
15326 | // fold (sdiv X, pow2) |
15327 | EVT VT = N->getValueType(0); |
15328 | if (VT == MVT::i64 && !Subtarget.isPPC64()) |
15329 | return SDValue(); |
15330 | if ((VT != MVT::i32 && VT != MVT::i64) || |
15331 | !(Divisor.isPowerOf2() || (-Divisor).isPowerOf2())) |
15332 | return SDValue(); |
15333 | |
15334 | SDLoc DL(N); |
15335 | SDValue N0 = N->getOperand(0); |
15336 | |
15337 | bool IsNegPow2 = (-Divisor).isPowerOf2(); |
15338 | unsigned Lg2 = (IsNegPow2 ? -Divisor : Divisor).countTrailingZeros(); |
15339 | SDValue ShiftAmt = DAG.getConstant(Lg2, DL, VT); |
15340 | |
15341 | SDValue Op = DAG.getNode(PPCISD::SRA_ADDZE, DL, VT, N0, ShiftAmt); |
15342 | Created.push_back(Op.getNode()); |
15343 | |
15344 | if (IsNegPow2) { |
15345 | Op = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Op); |
15346 | Created.push_back(Op.getNode()); |
15347 | } |
15348 | |
15349 | return Op; |
15350 | } |
15351 | |
15352 | //===----------------------------------------------------------------------===// |
15353 | // Inline Assembly Support |
15354 | //===----------------------------------------------------------------------===// |
15355 | |
15356 | void PPCTargetLowering::computeKnownBitsForTargetNode(const SDValue Op, |
15357 | KnownBits &Known, |
15358 | const APInt &DemandedElts, |
15359 | const SelectionDAG &DAG, |
15360 | unsigned Depth) const { |
15361 | Known.resetAll(); |
15362 | switch (Op.getOpcode()) { |
15363 | default: break; |
15364 | case PPCISD::LBRX: { |
15365 | // lhbrx is known to have the top bits cleared out. |
15366 | if (cast<VTSDNode>(Op.getOperand(2))->getVT() == MVT::i16) |
15367 | Known.Zero = 0xFFFF0000; |
15368 | break; |
15369 | } |
15370 | case ISD::INTRINSIC_WO_CHAIN: { |
15371 | switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) { |
15372 | default: break; |
15373 | case Intrinsic::ppc_altivec_vcmpbfp_p: |
15374 | case Intrinsic::ppc_altivec_vcmpeqfp_p: |
15375 | case Intrinsic::ppc_altivec_vcmpequb_p: |
15376 | case Intrinsic::ppc_altivec_vcmpequh_p: |
15377 | case Intrinsic::ppc_altivec_vcmpequw_p: |
15378 | case Intrinsic::ppc_altivec_vcmpequd_p: |
15379 | case Intrinsic::ppc_altivec_vcmpequq_p: |
15380 | case Intrinsic::ppc_altivec_vcmpgefp_p: |
15381 | case Intrinsic::ppc_altivec_vcmpgtfp_p: |
15382 | case Intrinsic::ppc_altivec_vcmpgtsb_p: |
15383 | case Intrinsic::ppc_altivec_vcmpgtsh_p: |
15384 | case Intrinsic::ppc_altivec_vcmpgtsw_p: |
15385 | case Intrinsic::ppc_altivec_vcmpgtsd_p: |
15386 | case Intrinsic::ppc_altivec_vcmpgtsq_p: |
15387 | case Intrinsic::ppc_altivec_vcmpgtub_p: |
15388 | case Intrinsic::ppc_altivec_vcmpgtuh_p: |
15389 | case Intrinsic::ppc_altivec_vcmpgtuw_p: |
15390 | case Intrinsic::ppc_altivec_vcmpgtud_p: |
15391 | case Intrinsic::ppc_altivec_vcmpgtuq_p: |
15392 | Known.Zero = ~1U; // All bits but the low one are known to be zero. |
15393 | break; |
15394 | } |
15395 | } |
15396 | } |
15397 | } |
15398 | |
15399 | Align PPCTargetLowering::getPrefLoopAlignment(MachineLoop *ML) const { |
15400 | switch (Subtarget.getCPUDirective()) { |
15401 | default: break; |
15402 | case PPC::DIR_970: |
15403 | case PPC::DIR_PWR4: |
15404 | case PPC::DIR_PWR5: |
15405 | case PPC::DIR_PWR5X: |
15406 | case PPC::DIR_PWR6: |
15407 | case PPC::DIR_PWR6X: |
15408 | case PPC::DIR_PWR7: |
15409 | case PPC::DIR_PWR8: |
15410 | case PPC::DIR_PWR9: |
15411 | case PPC::DIR_PWR10: |
15412 | case PPC::DIR_PWR_FUTURE: { |
15413 | if (!ML) |
15414 | break; |
15415 | |
15416 | if (!DisableInnermostLoopAlign32) { |
15417 | // If the nested loop is an innermost loop, prefer to a 32-byte alignment, |
15418 | // so that we can decrease cache misses and branch-prediction misses. |
15419 | // Actual alignment of the loop will depend on the hotness check and other |
15420 | // logic in alignBlocks. |
15421 | if (ML->getLoopDepth() > 1 && ML->getSubLoops().empty()) |
15422 | return Align(32); |
15423 | } |
15424 | |
15425 | const PPCInstrInfo *TII = Subtarget.getInstrInfo(); |
15426 | |
15427 | // For small loops (between 5 and 8 instructions), align to a 32-byte |
15428 | // boundary so that the entire loop fits in one instruction-cache line. |
15429 | uint64_t LoopSize = 0; |
15430 | for (auto I = ML->block_begin(), IE = ML->block_end(); I != IE; ++I) |
15431 | for (auto J = (*I)->begin(), JE = (*I)->end(); J != JE; ++J) { |
15432 | LoopSize += TII->getInstSizeInBytes(*J); |
15433 | if (LoopSize > 32) |
15434 | break; |
15435 | } |
15436 | |
15437 | if (LoopSize > 16 && LoopSize <= 32) |
15438 | return Align(32); |
15439 | |
15440 | break; |
15441 | } |
15442 | } |
15443 | |
15444 | return TargetLowering::getPrefLoopAlignment(ML); |
15445 | } |
15446 | |
15447 | /// getConstraintType - Given a constraint, return the type of |
15448 | /// constraint it is for this target. |
15449 | PPCTargetLowering::ConstraintType |
15450 | PPCTargetLowering::getConstraintType(StringRef Constraint) const { |
15451 | if (Constraint.size() == 1) { |
15452 | switch (Constraint[0]) { |
15453 | default: break; |
15454 | case 'b': |
15455 | case 'r': |
15456 | case 'f': |
15457 | case 'd': |
15458 | case 'v': |
15459 | case 'y': |
15460 | return C_RegisterClass; |
15461 | case 'Z': |
15462 | // FIXME: While Z does indicate a memory constraint, it specifically |
15463 | // indicates an r+r address (used in conjunction with the 'y' modifier |
15464 | // in the replacement string). Currently, we're forcing the base |
15465 | // register to be r0 in the asm printer (which is interpreted as zero) |
15466 | // and forming the complete address in the second register. This is |
15467 | // suboptimal. |
15468 | return C_Memory; |
15469 | } |
15470 | } else if (Constraint == "wc") { // individual CR bits. |
15471 | return C_RegisterClass; |
15472 | } else if (Constraint == "wa" || Constraint == "wd" || |
15473 | Constraint == "wf" || Constraint == "ws" || |
15474 | Constraint == "wi" || Constraint == "ww") { |
15475 | return C_RegisterClass; // VSX registers. |
15476 | } |
15477 | return TargetLowering::getConstraintType(Constraint); |
15478 | } |
15479 | |
15480 | /// Examine constraint type and operand type and determine a weight value. |
15481 | /// This object must already have been set up with the operand type |
15482 | /// and the current alternative constraint selected. |
15483 | TargetLowering::ConstraintWeight |
15484 | PPCTargetLowering::getSingleConstraintMatchWeight( |
15485 | AsmOperandInfo &info, const char *constraint) const { |
15486 | ConstraintWeight weight = CW_Invalid; |
15487 | Value *CallOperandVal = info.CallOperandVal; |
15488 | // If we don't have a value, we can't do a match, |
15489 | // but allow it at the lowest weight. |
15490 | if (!CallOperandVal) |
15491 | return CW_Default; |
15492 | Type *type = CallOperandVal->getType(); |
15493 | |
15494 | // Look at the constraint type. |
15495 | if (StringRef(constraint) == "wc" && type->isIntegerTy(1)) |
15496 | return CW_Register; // an individual CR bit. |
15497 | else if ((StringRef(constraint) == "wa" || |
15498 | StringRef(constraint) == "wd" || |
15499 | StringRef(constraint) == "wf") && |
15500 | type->isVectorTy()) |
15501 | return CW_Register; |
15502 | else if (StringRef(constraint) == "wi" && type->isIntegerTy(64)) |
15503 | return CW_Register; // just hold 64-bit integers data. |
15504 | else if (StringRef(constraint) == "ws" && type->isDoubleTy()) |
15505 | return CW_Register; |
15506 | else if (StringRef(constraint) == "ww" && type->isFloatTy()) |
15507 | return CW_Register; |
15508 | |
15509 | switch (*constraint) { |
15510 | default: |
15511 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
15512 | break; |
15513 | case 'b': |
15514 | if (type->isIntegerTy()) |
15515 | weight = CW_Register; |
15516 | break; |
15517 | case 'f': |
15518 | if (type->isFloatTy()) |
15519 | weight = CW_Register; |
15520 | break; |
15521 | case 'd': |
15522 | if (type->isDoubleTy()) |
15523 | weight = CW_Register; |
15524 | break; |
15525 | case 'v': |
15526 | if (type->isVectorTy()) |
15527 | weight = CW_Register; |
15528 | break; |
15529 | case 'y': |
15530 | weight = CW_Register; |
15531 | break; |
15532 | case 'Z': |
15533 | weight = CW_Memory; |
15534 | break; |
15535 | } |
15536 | return weight; |
15537 | } |
15538 | |
15539 | std::pair<unsigned, const TargetRegisterClass *> |
15540 | PPCTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
15541 | StringRef Constraint, |
15542 | MVT VT) const { |
15543 | if (Constraint.size() == 1) { |
15544 | // GCC RS6000 Constraint Letters |
15545 | switch (Constraint[0]) { |
15546 | case 'b': // R1-R31 |
15547 | if (VT == MVT::i64 && Subtarget.isPPC64()) |
15548 | return std::make_pair(0U, &PPC::G8RC_NOX0RegClass); |
15549 | return std::make_pair(0U, &PPC::GPRC_NOR0RegClass); |
15550 | case 'r': // R0-R31 |
15551 | if (VT == MVT::i64 && Subtarget.isPPC64()) |
15552 | return std::make_pair(0U, &PPC::G8RCRegClass); |
15553 | return std::make_pair(0U, &PPC::GPRCRegClass); |
15554 | // 'd' and 'f' constraints are both defined to be "the floating point |
15555 | // registers", where one is for 32-bit and the other for 64-bit. We don't |
15556 | // really care overly much here so just give them all the same reg classes. |
15557 | case 'd': |
15558 | case 'f': |
15559 | if (Subtarget.hasSPE()) { |
15560 | if (VT == MVT::f32 || VT == MVT::i32) |
15561 | return std::make_pair(0U, &PPC::GPRCRegClass); |
15562 | if (VT == MVT::f64 || VT == MVT::i64) |
15563 | return std::make_pair(0U, &PPC::SPERCRegClass); |
15564 | } else { |
15565 | if (VT == MVT::f32 || VT == MVT::i32) |
15566 | return std::make_pair(0U, &PPC::F4RCRegClass); |
15567 | if (VT == MVT::f64 || VT == MVT::i64) |
15568 | return std::make_pair(0U, &PPC::F8RCRegClass); |
15569 | } |
15570 | break; |
15571 | case 'v': |
15572 | if (Subtarget.hasAltivec()) |
15573 | return std::make_pair(0U, &PPC::VRRCRegClass); |
15574 | break; |
15575 | case 'y': // crrc |
15576 | return std::make_pair(0U, &PPC::CRRCRegClass); |
15577 | } |
15578 | } else if (Constraint == "wc" && Subtarget.useCRBits()) { |
15579 | // An individual CR bit. |
15580 | return std::make_pair(0U, &PPC::CRBITRCRegClass); |
15581 | } else if ((Constraint == "wa" || Constraint == "wd" || |
15582 | Constraint == "wf" || Constraint == "wi") && |
15583 | Subtarget.hasVSX()) { |
15584 | return std::make_pair(0U, &PPC::VSRCRegClass); |
15585 | } else if ((Constraint == "ws" || Constraint == "ww") && Subtarget.hasVSX()) { |
15586 | if (VT == MVT::f32 && Subtarget.hasP8Vector()) |
15587 | return std::make_pair(0U, &PPC::VSSRCRegClass); |
15588 | else |
15589 | return std::make_pair(0U, &PPC::VSFRCRegClass); |
15590 | } |
15591 | |
15592 | // If we name a VSX register, we can't defer to the base class because it |
15593 | // will not recognize the correct register (their names will be VSL{0-31} |
15594 | // and V{0-31} so they won't match). So we match them here. |
15595 | if (Constraint.size() > 3 && Constraint[1] == 'v' && Constraint[2] == 's') { |
15596 | int VSNum = atoi(Constraint.data() + 3); |
15597 | assert(VSNum >= 0 && VSNum <= 63 &&((VSNum >= 0 && VSNum <= 63 && "Attempted to access a vsr out of range" ) ? static_cast<void> (0) : __assert_fail ("VSNum >= 0 && VSNum <= 63 && \"Attempted to access a vsr out of range\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15598, __PRETTY_FUNCTION__)) |
15598 | "Attempted to access a vsr out of range")((VSNum >= 0 && VSNum <= 63 && "Attempted to access a vsr out of range" ) ? static_cast<void> (0) : __assert_fail ("VSNum >= 0 && VSNum <= 63 && \"Attempted to access a vsr out of range\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15598, __PRETTY_FUNCTION__)); |
15599 | if (VSNum < 32) |
15600 | return std::make_pair(PPC::VSL0 + VSNum, &PPC::VSRCRegClass); |
15601 | return std::make_pair(PPC::V0 + VSNum - 32, &PPC::VSRCRegClass); |
15602 | } |
15603 | std::pair<unsigned, const TargetRegisterClass *> R = |
15604 | TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
15605 | |
15606 | // r[0-9]+ are used, on PPC64, to refer to the corresponding 64-bit registers |
15607 | // (which we call X[0-9]+). If a 64-bit value has been requested, and a |
15608 | // 32-bit GPR has been selected, then 'upgrade' it to the 64-bit parent |
15609 | // register. |
15610 | // FIXME: If TargetLowering::getRegForInlineAsmConstraint could somehow use |
15611 | // the AsmName field from *RegisterInfo.td, then this would not be necessary. |
15612 | if (R.first && VT == MVT::i64 && Subtarget.isPPC64() && |
15613 | PPC::GPRCRegClass.contains(R.first)) |
15614 | return std::make_pair(TRI->getMatchingSuperReg(R.first, |
15615 | PPC::sub_32, &PPC::G8RCRegClass), |
15616 | &PPC::G8RCRegClass); |
15617 | |
15618 | // GCC accepts 'cc' as an alias for 'cr0', and we need to do the same. |
15619 | if (!R.second && StringRef("{cc}").equals_lower(Constraint)) { |
15620 | R.first = PPC::CR0; |
15621 | R.second = &PPC::CRRCRegClass; |
15622 | } |
15623 | |
15624 | return R; |
15625 | } |
15626 | |
15627 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
15628 | /// vector. If it is invalid, don't add anything to Ops. |
15629 | void PPCTargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
15630 | std::string &Constraint, |
15631 | std::vector<SDValue>&Ops, |
15632 | SelectionDAG &DAG) const { |
15633 | SDValue Result; |
15634 | |
15635 | // Only support length 1 constraints. |
15636 | if (Constraint.length() > 1) return; |
15637 | |
15638 | char Letter = Constraint[0]; |
15639 | switch (Letter) { |
15640 | default: break; |
15641 | case 'I': |
15642 | case 'J': |
15643 | case 'K': |
15644 | case 'L': |
15645 | case 'M': |
15646 | case 'N': |
15647 | case 'O': |
15648 | case 'P': { |
15649 | ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op); |
15650 | if (!CST) return; // Must be an immediate to match. |
15651 | SDLoc dl(Op); |
15652 | int64_t Value = CST->getSExtValue(); |
15653 | EVT TCVT = MVT::i64; // All constants taken to be 64 bits so that negative |
15654 | // numbers are printed as such. |
15655 | switch (Letter) { |
15656 | default: llvm_unreachable("Unknown constraint letter!")::llvm::llvm_unreachable_internal("Unknown constraint letter!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15656); |
15657 | case 'I': // "I" is a signed 16-bit constant. |
15658 | if (isInt<16>(Value)) |
15659 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15660 | break; |
15661 | case 'J': // "J" is a constant with only the high-order 16 bits nonzero. |
15662 | if (isShiftedUInt<16, 16>(Value)) |
15663 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15664 | break; |
15665 | case 'L': // "L" is a signed 16-bit constant shifted left 16 bits. |
15666 | if (isShiftedInt<16, 16>(Value)) |
15667 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15668 | break; |
15669 | case 'K': // "K" is a constant with only the low-order 16 bits nonzero. |
15670 | if (isUInt<16>(Value)) |
15671 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15672 | break; |
15673 | case 'M': // "M" is a constant that is greater than 31. |
15674 | if (Value > 31) |
15675 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15676 | break; |
15677 | case 'N': // "N" is a positive constant that is an exact power of two. |
15678 | if (Value > 0 && isPowerOf2_64(Value)) |
15679 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15680 | break; |
15681 | case 'O': // "O" is the constant zero. |
15682 | if (Value == 0) |
15683 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15684 | break; |
15685 | case 'P': // "P" is a constant whose negation is a signed 16-bit constant. |
15686 | if (isInt<16>(-Value)) |
15687 | Result = DAG.getTargetConstant(Value, dl, TCVT); |
15688 | break; |
15689 | } |
15690 | break; |
15691 | } |
15692 | } |
15693 | |
15694 | if (Result.getNode()) { |
15695 | Ops.push_back(Result); |
15696 | return; |
15697 | } |
15698 | |
15699 | // Handle standard constraint letters. |
15700 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
15701 | } |
15702 | |
15703 | // isLegalAddressingMode - Return true if the addressing mode represented |
15704 | // by AM is legal for this target, for a load/store of the specified type. |
15705 | bool PPCTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
15706 | const AddrMode &AM, Type *Ty, |
15707 | unsigned AS, |
15708 | Instruction *I) const { |
15709 | // Vector type r+i form is supported since power9 as DQ form. We don't check |
15710 | // the offset matching DQ form requirement(off % 16 == 0), because on PowerPC, |
15711 | // imm form is preferred and the offset can be adjusted to use imm form later |
15712 | // in pass PPCLoopInstrFormPrep. Also in LSR, for one LSRUse, it uses min and |
15713 | // max offset to check legal addressing mode, we should be a little aggressive |
15714 | // to contain other offsets for that LSRUse. |
15715 | if (Ty->isVectorTy() && AM.BaseOffs != 0 && !Subtarget.hasP9Vector()) |
15716 | return false; |
15717 | |
15718 | // PPC allows a sign-extended 16-bit immediate field. |
15719 | if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) |
15720 | return false; |
15721 | |
15722 | // No global is ever allowed as a base. |
15723 | if (AM.BaseGV) |
15724 | return false; |
15725 | |
15726 | // PPC only support r+r, |
15727 | switch (AM.Scale) { |
15728 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
15729 | break; |
15730 | case 1: |
15731 | if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. |
15732 | return false; |
15733 | // Otherwise we have r+r or r+i. |
15734 | break; |
15735 | case 2: |
15736 | if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. |
15737 | return false; |
15738 | // Allow 2*r as r+r. |
15739 | break; |
15740 | default: |
15741 | // No other scales are supported. |
15742 | return false; |
15743 | } |
15744 | |
15745 | return true; |
15746 | } |
15747 | |
15748 | SDValue PPCTargetLowering::LowerRETURNADDR(SDValue Op, |
15749 | SelectionDAG &DAG) const { |
15750 | MachineFunction &MF = DAG.getMachineFunction(); |
15751 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
15752 | MFI.setReturnAddressIsTaken(true); |
15753 | |
15754 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
15755 | return SDValue(); |
15756 | |
15757 | SDLoc dl(Op); |
15758 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
15759 | |
15760 | // Make sure the function does not optimize away the store of the RA to |
15761 | // the stack. |
15762 | PPCFunctionInfo *FuncInfo = MF.getInfo<PPCFunctionInfo>(); |
15763 | FuncInfo->setLRStoreRequired(); |
15764 | bool isPPC64 = Subtarget.isPPC64(); |
15765 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
15766 | |
15767 | if (Depth > 0) { |
15768 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); |
15769 | SDValue Offset = |
15770 | DAG.getConstant(Subtarget.getFrameLowering()->getReturnSaveOffset(), dl, |
15771 | isPPC64 ? MVT::i64 : MVT::i32); |
15772 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), |
15773 | DAG.getNode(ISD::ADD, dl, PtrVT, FrameAddr, Offset), |
15774 | MachinePointerInfo()); |
15775 | } |
15776 | |
15777 | // Just load the return address off the stack. |
15778 | SDValue RetAddrFI = getReturnAddrFrameIndex(DAG); |
15779 | return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), RetAddrFI, |
15780 | MachinePointerInfo()); |
15781 | } |
15782 | |
15783 | SDValue PPCTargetLowering::LowerFRAMEADDR(SDValue Op, |
15784 | SelectionDAG &DAG) const { |
15785 | SDLoc dl(Op); |
15786 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
15787 | |
15788 | MachineFunction &MF = DAG.getMachineFunction(); |
15789 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
15790 | MFI.setFrameAddressIsTaken(true); |
15791 | |
15792 | EVT PtrVT = getPointerTy(MF.getDataLayout()); |
15793 | bool isPPC64 = PtrVT == MVT::i64; |
15794 | |
15795 | // Naked functions never have a frame pointer, and so we use r1. For all |
15796 | // other functions, this decision must be delayed until during PEI. |
15797 | unsigned FrameReg; |
15798 | if (MF.getFunction().hasFnAttribute(Attribute::Naked)) |
15799 | FrameReg = isPPC64 ? PPC::X1 : PPC::R1; |
15800 | else |
15801 | FrameReg = isPPC64 ? PPC::FP8 : PPC::FP; |
15802 | |
15803 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, |
15804 | PtrVT); |
15805 | while (Depth--) |
15806 | FrameAddr = DAG.getLoad(Op.getValueType(), dl, DAG.getEntryNode(), |
15807 | FrameAddr, MachinePointerInfo()); |
15808 | return FrameAddr; |
15809 | } |
15810 | |
15811 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
15812 | // this table could be generated automatically from RegInfo. |
15813 | Register PPCTargetLowering::getRegisterByName(const char* RegName, LLT VT, |
15814 | const MachineFunction &MF) const { |
15815 | bool isPPC64 = Subtarget.isPPC64(); |
15816 | |
15817 | bool is64Bit = isPPC64 && VT == LLT::scalar(64); |
15818 | if (!is64Bit && VT != LLT::scalar(32)) |
15819 | report_fatal_error("Invalid register global variable type"); |
15820 | |
15821 | Register Reg = StringSwitch<Register>(RegName) |
15822 | .Case("r1", is64Bit ? PPC::X1 : PPC::R1) |
15823 | .Case("r2", isPPC64 ? Register() : PPC::R2) |
15824 | .Case("r13", (is64Bit ? PPC::X13 : PPC::R13)) |
15825 | .Default(Register()); |
15826 | |
15827 | if (Reg) |
15828 | return Reg; |
15829 | report_fatal_error("Invalid register name global variable"); |
15830 | } |
15831 | |
15832 | bool PPCTargetLowering::isAccessedAsGotIndirect(SDValue GA) const { |
15833 | // 32-bit SVR4 ABI access everything as got-indirect. |
15834 | if (Subtarget.is32BitELFABI()) |
15835 | return true; |
15836 | |
15837 | // AIX accesses everything indirectly through the TOC, which is similar to |
15838 | // the GOT. |
15839 | if (Subtarget.isAIXABI()) |
15840 | return true; |
15841 | |
15842 | CodeModel::Model CModel = getTargetMachine().getCodeModel(); |
15843 | // If it is small or large code model, module locals are accessed |
15844 | // indirectly by loading their address from .toc/.got. |
15845 | if (CModel == CodeModel::Small || CModel == CodeModel::Large) |
15846 | return true; |
15847 | |
15848 | // JumpTable and BlockAddress are accessed as got-indirect. |
15849 | if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA)) |
15850 | return true; |
15851 | |
15852 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) |
15853 | return Subtarget.isGVIndirectSymbol(G->getGlobal()); |
15854 | |
15855 | return false; |
15856 | } |
15857 | |
15858 | bool |
15859 | PPCTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
15860 | // The PowerPC target isn't yet aware of offsets. |
15861 | return false; |
15862 | } |
15863 | |
15864 | bool PPCTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
15865 | const CallInst &I, |
15866 | MachineFunction &MF, |
15867 | unsigned Intrinsic) const { |
15868 | switch (Intrinsic) { |
15869 | case Intrinsic::ppc_altivec_lvx: |
15870 | case Intrinsic::ppc_altivec_lvxl: |
15871 | case Intrinsic::ppc_altivec_lvebx: |
15872 | case Intrinsic::ppc_altivec_lvehx: |
15873 | case Intrinsic::ppc_altivec_lvewx: |
15874 | case Intrinsic::ppc_vsx_lxvd2x: |
15875 | case Intrinsic::ppc_vsx_lxvw4x: |
15876 | case Intrinsic::ppc_vsx_lxvd2x_be: |
15877 | case Intrinsic::ppc_vsx_lxvw4x_be: |
15878 | case Intrinsic::ppc_vsx_lxvl: |
15879 | case Intrinsic::ppc_vsx_lxvll: { |
15880 | EVT VT; |
15881 | switch (Intrinsic) { |
15882 | case Intrinsic::ppc_altivec_lvebx: |
15883 | VT = MVT::i8; |
15884 | break; |
15885 | case Intrinsic::ppc_altivec_lvehx: |
15886 | VT = MVT::i16; |
15887 | break; |
15888 | case Intrinsic::ppc_altivec_lvewx: |
15889 | VT = MVT::i32; |
15890 | break; |
15891 | case Intrinsic::ppc_vsx_lxvd2x: |
15892 | case Intrinsic::ppc_vsx_lxvd2x_be: |
15893 | VT = MVT::v2f64; |
15894 | break; |
15895 | default: |
15896 | VT = MVT::v4i32; |
15897 | break; |
15898 | } |
15899 | |
15900 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
15901 | Info.memVT = VT; |
15902 | Info.ptrVal = I.getArgOperand(0); |
15903 | Info.offset = -VT.getStoreSize()+1; |
15904 | Info.size = 2*VT.getStoreSize()-1; |
15905 | Info.align = Align(1); |
15906 | Info.flags = MachineMemOperand::MOLoad; |
15907 | return true; |
15908 | } |
15909 | case Intrinsic::ppc_altivec_stvx: |
15910 | case Intrinsic::ppc_altivec_stvxl: |
15911 | case Intrinsic::ppc_altivec_stvebx: |
15912 | case Intrinsic::ppc_altivec_stvehx: |
15913 | case Intrinsic::ppc_altivec_stvewx: |
15914 | case Intrinsic::ppc_vsx_stxvd2x: |
15915 | case Intrinsic::ppc_vsx_stxvw4x: |
15916 | case Intrinsic::ppc_vsx_stxvd2x_be: |
15917 | case Intrinsic::ppc_vsx_stxvw4x_be: |
15918 | case Intrinsic::ppc_vsx_stxvl: |
15919 | case Intrinsic::ppc_vsx_stxvll: { |
15920 | EVT VT; |
15921 | switch (Intrinsic) { |
15922 | case Intrinsic::ppc_altivec_stvebx: |
15923 | VT = MVT::i8; |
15924 | break; |
15925 | case Intrinsic::ppc_altivec_stvehx: |
15926 | VT = MVT::i16; |
15927 | break; |
15928 | case Intrinsic::ppc_altivec_stvewx: |
15929 | VT = MVT::i32; |
15930 | break; |
15931 | case Intrinsic::ppc_vsx_stxvd2x: |
15932 | case Intrinsic::ppc_vsx_stxvd2x_be: |
15933 | VT = MVT::v2f64; |
15934 | break; |
15935 | default: |
15936 | VT = MVT::v4i32; |
15937 | break; |
15938 | } |
15939 | |
15940 | Info.opc = ISD::INTRINSIC_VOID; |
15941 | Info.memVT = VT; |
15942 | Info.ptrVal = I.getArgOperand(1); |
15943 | Info.offset = -VT.getStoreSize()+1; |
15944 | Info.size = 2*VT.getStoreSize()-1; |
15945 | Info.align = Align(1); |
15946 | Info.flags = MachineMemOperand::MOStore; |
15947 | return true; |
15948 | } |
15949 | default: |
15950 | break; |
15951 | } |
15952 | |
15953 | return false; |
15954 | } |
15955 | |
15956 | /// It returns EVT::Other if the type should be determined using generic |
15957 | /// target-independent logic. |
15958 | EVT PPCTargetLowering::getOptimalMemOpType( |
15959 | const MemOp &Op, const AttributeList &FuncAttributes) const { |
15960 | if (getTargetMachine().getOptLevel() != CodeGenOpt::None) { |
15961 | // We should use Altivec/VSX loads and stores when available. For unaligned |
15962 | // addresses, unaligned VSX loads are only fast starting with the P8. |
15963 | if (Subtarget.hasAltivec() && Op.size() >= 16 && |
15964 | (Op.isAligned(Align(16)) || |
15965 | ((Op.isMemset() && Subtarget.hasVSX()) || Subtarget.hasP8Vector()))) |
15966 | return MVT::v4i32; |
15967 | } |
15968 | |
15969 | if (Subtarget.isPPC64()) { |
15970 | return MVT::i64; |
15971 | } |
15972 | |
15973 | return MVT::i32; |
15974 | } |
15975 | |
15976 | /// Returns true if it is beneficial to convert a load of a constant |
15977 | /// to just the constant itself. |
15978 | bool PPCTargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm, |
15979 | Type *Ty) const { |
15980 | assert(Ty->isIntegerTy())((Ty->isIntegerTy()) ? static_cast<void> (0) : __assert_fail ("Ty->isIntegerTy()", "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 15980, __PRETTY_FUNCTION__)); |
15981 | |
15982 | unsigned BitSize = Ty->getPrimitiveSizeInBits(); |
15983 | return !(BitSize == 0 || BitSize > 64); |
15984 | } |
15985 | |
15986 | bool PPCTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { |
15987 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
15988 | return false; |
15989 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); |
15990 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); |
15991 | return NumBits1 == 64 && NumBits2 == 32; |
15992 | } |
15993 | |
15994 | bool PPCTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { |
15995 | if (!VT1.isInteger() || !VT2.isInteger()) |
15996 | return false; |
15997 | unsigned NumBits1 = VT1.getSizeInBits(); |
15998 | unsigned NumBits2 = VT2.getSizeInBits(); |
15999 | return NumBits1 == 64 && NumBits2 == 32; |
16000 | } |
16001 | |
16002 | bool PPCTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { |
16003 | // Generally speaking, zexts are not free, but they are free when they can be |
16004 | // folded with other operations. |
16005 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val)) { |
16006 | EVT MemVT = LD->getMemoryVT(); |
16007 | if ((MemVT == MVT::i1 || MemVT == MVT::i8 || MemVT == MVT::i16 || |
16008 | (Subtarget.isPPC64() && MemVT == MVT::i32)) && |
16009 | (LD->getExtensionType() == ISD::NON_EXTLOAD || |
16010 | LD->getExtensionType() == ISD::ZEXTLOAD)) |
16011 | return true; |
16012 | } |
16013 | |
16014 | // FIXME: Add other cases... |
16015 | // - 32-bit shifts with a zext to i64 |
16016 | // - zext after ctlz, bswap, etc. |
16017 | // - zext after and by a constant mask |
16018 | |
16019 | return TargetLowering::isZExtFree(Val, VT2); |
16020 | } |
16021 | |
16022 | bool PPCTargetLowering::isFPExtFree(EVT DestVT, EVT SrcVT) const { |
16023 | assert(DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() &&((DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && "invalid fpext types") ? static_cast<void> ( 0) : __assert_fail ("DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && \"invalid fpext types\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16024, __PRETTY_FUNCTION__)) |
16024 | "invalid fpext types")((DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && "invalid fpext types") ? static_cast<void> ( 0) : __assert_fail ("DestVT.isFloatingPoint() && SrcVT.isFloatingPoint() && \"invalid fpext types\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16024, __PRETTY_FUNCTION__)); |
16025 | // Extending to float128 is not free. |
16026 | if (DestVT == MVT::f128) |
16027 | return false; |
16028 | return true; |
16029 | } |
16030 | |
16031 | bool PPCTargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
16032 | return isInt<16>(Imm) || isUInt<16>(Imm); |
16033 | } |
16034 | |
16035 | bool PPCTargetLowering::isLegalAddImmediate(int64_t Imm) const { |
16036 | return isInt<16>(Imm) || isUInt<16>(Imm); |
16037 | } |
16038 | |
16039 | bool PPCTargetLowering::allowsMisalignedMemoryAccesses(EVT VT, |
16040 | unsigned, |
16041 | unsigned, |
16042 | MachineMemOperand::Flags, |
16043 | bool *Fast) const { |
16044 | if (DisablePPCUnaligned) |
16045 | return false; |
16046 | |
16047 | // PowerPC supports unaligned memory access for simple non-vector types. |
16048 | // Although accessing unaligned addresses is not as efficient as accessing |
16049 | // aligned addresses, it is generally more efficient than manual expansion, |
16050 | // and generally only traps for software emulation when crossing page |
16051 | // boundaries. |
16052 | |
16053 | if (!VT.isSimple()) |
16054 | return false; |
16055 | |
16056 | if (VT.isFloatingPoint() && !VT.isVector() && |
16057 | !Subtarget.allowsUnalignedFPAccess()) |
16058 | return false; |
16059 | |
16060 | if (VT.getSimpleVT().isVector()) { |
16061 | if (Subtarget.hasVSX()) { |
16062 | if (VT != MVT::v2f64 && VT != MVT::v2i64 && |
16063 | VT != MVT::v4f32 && VT != MVT::v4i32) |
16064 | return false; |
16065 | } else { |
16066 | return false; |
16067 | } |
16068 | } |
16069 | |
16070 | if (VT == MVT::ppcf128) |
16071 | return false; |
16072 | |
16073 | if (Fast) |
16074 | *Fast = true; |
16075 | |
16076 | return true; |
16077 | } |
16078 | |
16079 | bool PPCTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT, |
16080 | SDValue C) const { |
16081 | // Check integral scalar types. |
16082 | if (!VT.isScalarInteger()) |
16083 | return false; |
16084 | if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) { |
16085 | if (!ConstNode->getAPIntValue().isSignedIntN(64)) |
16086 | return false; |
16087 | // This transformation will generate >= 2 operations. But the following |
16088 | // cases will generate <= 2 instructions during ISEL. So exclude them. |
16089 | // 1. If the constant multiplier fits 16 bits, it can be handled by one |
16090 | // HW instruction, ie. MULLI |
16091 | // 2. If the multiplier after shifted fits 16 bits, an extra shift |
16092 | // instruction is needed than case 1, ie. MULLI and RLDICR |
16093 | int64_t Imm = ConstNode->getSExtValue(); |
16094 | unsigned Shift = countTrailingZeros<uint64_t>(Imm); |
16095 | Imm >>= Shift; |
16096 | if (isInt<16>(Imm)) |
16097 | return false; |
16098 | uint64_t UImm = static_cast<uint64_t>(Imm); |
16099 | if (isPowerOf2_64(UImm + 1) || isPowerOf2_64(UImm - 1) || |
16100 | isPowerOf2_64(1 - UImm) || isPowerOf2_64(-1 - UImm)) |
16101 | return true; |
16102 | } |
16103 | return false; |
16104 | } |
16105 | |
16106 | bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, |
16107 | EVT VT) const { |
16108 | return isFMAFasterThanFMulAndFAdd( |
16109 | MF.getFunction(), VT.getTypeForEVT(MF.getFunction().getContext())); |
16110 | } |
16111 | |
16112 | bool PPCTargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F, |
16113 | Type *Ty) const { |
16114 | switch (Ty->getScalarType()->getTypeID()) { |
16115 | case Type::FloatTyID: |
16116 | case Type::DoubleTyID: |
16117 | return true; |
16118 | case Type::FP128TyID: |
16119 | return Subtarget.hasP9Vector(); |
16120 | default: |
16121 | return false; |
16122 | } |
16123 | } |
16124 | |
16125 | // FIXME: add more patterns which are not profitable to hoist. |
16126 | bool PPCTargetLowering::isProfitableToHoist(Instruction *I) const { |
16127 | if (!I->hasOneUse()) |
16128 | return true; |
16129 | |
16130 | Instruction *User = I->user_back(); |
16131 | assert(User && "A single use instruction with no uses.")((User && "A single use instruction with no uses.") ? static_cast<void> (0) : __assert_fail ("User && \"A single use instruction with no uses.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16131, __PRETTY_FUNCTION__)); |
16132 | |
16133 | switch (I->getOpcode()) { |
16134 | case Instruction::FMul: { |
16135 | // Don't break FMA, PowerPC prefers FMA. |
16136 | if (User->getOpcode() != Instruction::FSub && |
16137 | User->getOpcode() != Instruction::FAdd) |
16138 | return true; |
16139 | |
16140 | const TargetOptions &Options = getTargetMachine().Options; |
16141 | const Function *F = I->getFunction(); |
16142 | const DataLayout &DL = F->getParent()->getDataLayout(); |
16143 | Type *Ty = User->getOperand(0)->getType(); |
16144 | |
16145 | return !( |
16146 | isFMAFasterThanFMulAndFAdd(*F, Ty) && |
16147 | isOperationLegalOrCustom(ISD::FMA, getValueType(DL, Ty)) && |
16148 | (Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath)); |
16149 | } |
16150 | case Instruction::Load: { |
16151 | // Don't break "store (load float*)" pattern, this pattern will be combined |
16152 | // to "store (load int32)" in later InstCombine pass. See function |
16153 | // combineLoadToOperationType. On PowerPC, loading a float point takes more |
16154 | // cycles than loading a 32 bit integer. |
16155 | LoadInst *LI = cast<LoadInst>(I); |
16156 | // For the loads that combineLoadToOperationType does nothing, like |
16157 | // ordered load, it should be profitable to hoist them. |
16158 | // For swifterror load, it can only be used for pointer to pointer type, so |
16159 | // later type check should get rid of this case. |
16160 | if (!LI->isUnordered()) |
16161 | return true; |
16162 | |
16163 | if (User->getOpcode() != Instruction::Store) |
16164 | return true; |
16165 | |
16166 | if (I->getType()->getTypeID() != Type::FloatTyID) |
16167 | return true; |
16168 | |
16169 | return false; |
16170 | } |
16171 | default: |
16172 | return true; |
16173 | } |
16174 | return true; |
16175 | } |
16176 | |
16177 | const MCPhysReg * |
16178 | PPCTargetLowering::getScratchRegisters(CallingConv::ID) const { |
16179 | // LR is a callee-save register, but we must treat it as clobbered by any call |
16180 | // site. Hence we include LR in the scratch registers, which are in turn added |
16181 | // as implicit-defs for stackmaps and patchpoints. The same reasoning applies |
16182 | // to CTR, which is used by any indirect call. |
16183 | static const MCPhysReg ScratchRegs[] = { |
16184 | PPC::X12, PPC::LR8, PPC::CTR8, 0 |
16185 | }; |
16186 | |
16187 | return ScratchRegs; |
16188 | } |
16189 | |
16190 | Register PPCTargetLowering::getExceptionPointerRegister( |
16191 | const Constant *PersonalityFn) const { |
16192 | return Subtarget.isPPC64() ? PPC::X3 : PPC::R3; |
16193 | } |
16194 | |
16195 | Register PPCTargetLowering::getExceptionSelectorRegister( |
16196 | const Constant *PersonalityFn) const { |
16197 | return Subtarget.isPPC64() ? PPC::X4 : PPC::R4; |
16198 | } |
16199 | |
16200 | bool |
16201 | PPCTargetLowering::shouldExpandBuildVectorWithShuffles( |
16202 | EVT VT , unsigned DefinedValues) const { |
16203 | if (VT == MVT::v2i64) |
16204 | return Subtarget.hasDirectMove(); // Don't need stack ops with direct moves |
16205 | |
16206 | if (Subtarget.hasVSX()) |
16207 | return true; |
16208 | |
16209 | return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues); |
16210 | } |
16211 | |
16212 | Sched::Preference PPCTargetLowering::getSchedulingPreference(SDNode *N) const { |
16213 | if (DisableILPPref || Subtarget.enableMachineScheduler()) |
16214 | return TargetLowering::getSchedulingPreference(N); |
16215 | |
16216 | return Sched::ILP; |
16217 | } |
16218 | |
16219 | // Create a fast isel object. |
16220 | FastISel * |
16221 | PPCTargetLowering::createFastISel(FunctionLoweringInfo &FuncInfo, |
16222 | const TargetLibraryInfo *LibInfo) const { |
16223 | return PPC::createFastISel(FuncInfo, LibInfo); |
16224 | } |
16225 | |
16226 | // 'Inverted' means the FMA opcode after negating one multiplicand. |
16227 | // For example, (fma -a b c) = (fnmsub a b c) |
16228 | static unsigned invertFMAOpcode(unsigned Opc) { |
16229 | switch (Opc) { |
16230 | default: |
16231 | llvm_unreachable("Invalid FMA opcode for PowerPC!")::llvm::llvm_unreachable_internal("Invalid FMA opcode for PowerPC!" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16231); |
16232 | case ISD::FMA: |
16233 | return PPCISD::FNMSUB; |
16234 | case PPCISD::FNMSUB: |
16235 | return ISD::FMA; |
16236 | } |
16237 | } |
16238 | |
16239 | SDValue PPCTargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG, |
16240 | bool LegalOps, bool OptForSize, |
16241 | NegatibleCost &Cost, |
16242 | unsigned Depth) const { |
16243 | if (Depth > SelectionDAG::MaxRecursionDepth) |
16244 | return SDValue(); |
16245 | |
16246 | unsigned Opc = Op.getOpcode(); |
16247 | EVT VT = Op.getValueType(); |
16248 | SDNodeFlags Flags = Op.getNode()->getFlags(); |
16249 | |
16250 | switch (Opc) { |
16251 | case PPCISD::FNMSUB: |
16252 | if (!Op.hasOneUse() || !isTypeLegal(VT)) |
16253 | break; |
16254 | |
16255 | const TargetOptions &Options = getTargetMachine().Options; |
16256 | SDValue N0 = Op.getOperand(0); |
16257 | SDValue N1 = Op.getOperand(1); |
16258 | SDValue N2 = Op.getOperand(2); |
16259 | SDLoc Loc(Op); |
16260 | |
16261 | NegatibleCost N2Cost = NegatibleCost::Expensive; |
16262 | SDValue NegN2 = |
16263 | getNegatedExpression(N2, DAG, LegalOps, OptForSize, N2Cost, Depth + 1); |
16264 | |
16265 | if (!NegN2) |
16266 | return SDValue(); |
16267 | |
16268 | // (fneg (fnmsub a b c)) => (fnmsub (fneg a) b (fneg c)) |
16269 | // (fneg (fnmsub a b c)) => (fnmsub a (fneg b) (fneg c)) |
16270 | // These transformations may change sign of zeroes. For example, |
16271 | // -(-ab-(-c))=-0 while -(-(ab-c))=+0 when a=b=c=1. |
16272 | if (Flags.hasNoSignedZeros() || Options.NoSignedZerosFPMath) { |
16273 | // Try and choose the cheaper one to negate. |
16274 | NegatibleCost N0Cost = NegatibleCost::Expensive; |
16275 | SDValue NegN0 = getNegatedExpression(N0, DAG, LegalOps, OptForSize, |
16276 | N0Cost, Depth + 1); |
16277 | |
16278 | NegatibleCost N1Cost = NegatibleCost::Expensive; |
16279 | SDValue NegN1 = getNegatedExpression(N1, DAG, LegalOps, OptForSize, |
16280 | N1Cost, Depth + 1); |
16281 | |
16282 | if (NegN0 && N0Cost <= N1Cost) { |
16283 | Cost = std::min(N0Cost, N2Cost); |
16284 | return DAG.getNode(Opc, Loc, VT, NegN0, N1, NegN2, Flags); |
16285 | } else if (NegN1) { |
16286 | Cost = std::min(N1Cost, N2Cost); |
16287 | return DAG.getNode(Opc, Loc, VT, N0, NegN1, NegN2, Flags); |
16288 | } |
16289 | } |
16290 | |
16291 | // (fneg (fnmsub a b c)) => (fma a b (fneg c)) |
16292 | if (isOperationLegal(ISD::FMA, VT)) { |
16293 | Cost = N2Cost; |
16294 | return DAG.getNode(ISD::FMA, Loc, VT, N0, N1, NegN2, Flags); |
16295 | } |
16296 | |
16297 | break; |
16298 | } |
16299 | |
16300 | return TargetLowering::getNegatedExpression(Op, DAG, LegalOps, OptForSize, |
16301 | Cost, Depth); |
16302 | } |
16303 | |
16304 | // Override to enable LOAD_STACK_GUARD lowering on Linux. |
16305 | bool PPCTargetLowering::useLoadStackGuardNode() const { |
16306 | if (!Subtarget.isTargetLinux()) |
16307 | return TargetLowering::useLoadStackGuardNode(); |
16308 | return true; |
16309 | } |
16310 | |
16311 | // Override to disable global variable loading on Linux. |
16312 | void PPCTargetLowering::insertSSPDeclarations(Module &M) const { |
16313 | if (!Subtarget.isTargetLinux()) |
16314 | return TargetLowering::insertSSPDeclarations(M); |
16315 | } |
16316 | |
16317 | bool PPCTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
16318 | bool ForCodeSize) const { |
16319 | if (!VT.isSimple() || !Subtarget.hasVSX()) |
16320 | return false; |
16321 | |
16322 | switch(VT.getSimpleVT().SimpleTy) { |
16323 | default: |
16324 | // For FP types that are currently not supported by PPC backend, return |
16325 | // false. Examples: f16, f80. |
16326 | return false; |
16327 | case MVT::f32: |
16328 | case MVT::f64: |
16329 | if (Subtarget.hasPrefixInstrs()) { |
16330 | // With prefixed instructions, we can materialize anything that can be |
16331 | // represented with a 32-bit immediate, not just positive zero. |
16332 | APFloat APFloatOfImm = Imm; |
16333 | return convertToNonDenormSingle(APFloatOfImm); |
16334 | } |
16335 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
16336 | case MVT::ppcf128: |
16337 | return Imm.isPosZero(); |
16338 | } |
16339 | } |
16340 | |
16341 | // For vector shift operation op, fold |
16342 | // (op x, (and y, ((1 << numbits(x)) - 1))) -> (target op x, y) |
16343 | static SDValue stripModuloOnShift(const TargetLowering &TLI, SDNode *N, |
16344 | SelectionDAG &DAG) { |
16345 | SDValue N0 = N->getOperand(0); |
16346 | SDValue N1 = N->getOperand(1); |
16347 | EVT VT = N0.getValueType(); |
16348 | unsigned OpSizeInBits = VT.getScalarSizeInBits(); |
16349 | unsigned Opcode = N->getOpcode(); |
16350 | unsigned TargetOpcode; |
16351 | |
16352 | switch (Opcode) { |
16353 | default: |
16354 | llvm_unreachable("Unexpected shift operation")::llvm::llvm_unreachable_internal("Unexpected shift operation" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16354); |
16355 | case ISD::SHL: |
16356 | TargetOpcode = PPCISD::SHL; |
16357 | break; |
16358 | case ISD::SRL: |
16359 | TargetOpcode = PPCISD::SRL; |
16360 | break; |
16361 | case ISD::SRA: |
16362 | TargetOpcode = PPCISD::SRA; |
16363 | break; |
16364 | } |
16365 | |
16366 | if (VT.isVector() && TLI.isOperationLegal(Opcode, VT) && |
16367 | N1->getOpcode() == ISD::AND) |
16368 | if (ConstantSDNode *Mask = isConstOrConstSplat(N1->getOperand(1))) |
16369 | if (Mask->getZExtValue() == OpSizeInBits - 1) |
16370 | return DAG.getNode(TargetOpcode, SDLoc(N), VT, N0, N1->getOperand(0)); |
16371 | |
16372 | return SDValue(); |
16373 | } |
16374 | |
16375 | SDValue PPCTargetLowering::combineSHL(SDNode *N, DAGCombinerInfo &DCI) const { |
16376 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16377 | return Value; |
16378 | |
16379 | SDValue N0 = N->getOperand(0); |
16380 | ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1)); |
16381 | if (!Subtarget.isISA3_0() || !Subtarget.isPPC64() || |
16382 | N0.getOpcode() != ISD::SIGN_EXTEND || |
16383 | N0.getOperand(0).getValueType() != MVT::i32 || CN1 == nullptr || |
16384 | N->getValueType(0) != MVT::i64) |
16385 | return SDValue(); |
16386 | |
16387 | // We can't save an operation here if the value is already extended, and |
16388 | // the existing shift is easier to combine. |
16389 | SDValue ExtsSrc = N0.getOperand(0); |
16390 | if (ExtsSrc.getOpcode() == ISD::TRUNCATE && |
16391 | ExtsSrc.getOperand(0).getOpcode() == ISD::AssertSext) |
16392 | return SDValue(); |
16393 | |
16394 | SDLoc DL(N0); |
16395 | SDValue ShiftBy = SDValue(CN1, 0); |
16396 | // We want the shift amount to be i32 on the extswli, but the shift could |
16397 | // have an i64. |
16398 | if (ShiftBy.getValueType() == MVT::i64) |
16399 | ShiftBy = DCI.DAG.getConstant(CN1->getZExtValue(), DL, MVT::i32); |
16400 | |
16401 | return DCI.DAG.getNode(PPCISD::EXTSWSLI, DL, MVT::i64, N0->getOperand(0), |
16402 | ShiftBy); |
16403 | } |
16404 | |
16405 | SDValue PPCTargetLowering::combineSRA(SDNode *N, DAGCombinerInfo &DCI) const { |
16406 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16407 | return Value; |
16408 | |
16409 | return SDValue(); |
16410 | } |
16411 | |
16412 | SDValue PPCTargetLowering::combineSRL(SDNode *N, DAGCombinerInfo &DCI) const { |
16413 | if (auto Value = stripModuloOnShift(*this, N, DCI.DAG)) |
16414 | return Value; |
16415 | |
16416 | return SDValue(); |
16417 | } |
16418 | |
16419 | // Transform (add X, (zext(setne Z, C))) -> (addze X, (addic (addi Z, -C), -1)) |
16420 | // Transform (add X, (zext(sete Z, C))) -> (addze X, (subfic (addi Z, -C), 0)) |
16421 | // When C is zero, the equation (addi Z, -C) can be simplified to Z |
16422 | // Requirement: -C in [-32768, 32767], X and Z are MVT::i64 types |
16423 | static SDValue combineADDToADDZE(SDNode *N, SelectionDAG &DAG, |
16424 | const PPCSubtarget &Subtarget) { |
16425 | if (!Subtarget.isPPC64()) |
16426 | return SDValue(); |
16427 | |
16428 | SDValue LHS = N->getOperand(0); |
16429 | SDValue RHS = N->getOperand(1); |
16430 | |
16431 | auto isZextOfCompareWithConstant = [](SDValue Op) { |
16432 | if (Op.getOpcode() != ISD::ZERO_EXTEND || !Op.hasOneUse() || |
16433 | Op.getValueType() != MVT::i64) |
16434 | return false; |
16435 | |
16436 | SDValue Cmp = Op.getOperand(0); |
16437 | if (Cmp.getOpcode() != ISD::SETCC || !Cmp.hasOneUse() || |
16438 | Cmp.getOperand(0).getValueType() != MVT::i64) |
16439 | return false; |
16440 | |
16441 | if (auto *Constant = dyn_cast<ConstantSDNode>(Cmp.getOperand(1))) { |
16442 | int64_t NegConstant = 0 - Constant->getSExtValue(); |
16443 | // Due to the limitations of the addi instruction, |
16444 | // -C is required to be [-32768, 32767]. |
16445 | return isInt<16>(NegConstant); |
16446 | } |
16447 | |
16448 | return false; |
16449 | }; |
16450 | |
16451 | bool LHSHasPattern = isZextOfCompareWithConstant(LHS); |
16452 | bool RHSHasPattern = isZextOfCompareWithConstant(RHS); |
16453 | |
16454 | // If there is a pattern, canonicalize a zext operand to the RHS. |
16455 | if (LHSHasPattern && !RHSHasPattern) |
16456 | std::swap(LHS, RHS); |
16457 | else if (!LHSHasPattern && !RHSHasPattern) |
16458 | return SDValue(); |
16459 | |
16460 | SDLoc DL(N); |
16461 | SDVTList VTs = DAG.getVTList(MVT::i64, MVT::Glue); |
16462 | SDValue Cmp = RHS.getOperand(0); |
16463 | SDValue Z = Cmp.getOperand(0); |
16464 | auto *Constant = dyn_cast<ConstantSDNode>(Cmp.getOperand(1)); |
16465 | |
16466 | assert(Constant && "Constant Should not be a null pointer.")((Constant && "Constant Should not be a null pointer." ) ? static_cast<void> (0) : __assert_fail ("Constant && \"Constant Should not be a null pointer.\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16466, __PRETTY_FUNCTION__)); |
16467 | int64_t NegConstant = 0 - Constant->getSExtValue(); |
16468 | |
16469 | switch(cast<CondCodeSDNode>(Cmp.getOperand(2))->get()) { |
16470 | default: break; |
16471 | case ISD::SETNE: { |
16472 | // when C == 0 |
16473 | // --> addze X, (addic Z, -1).carry |
16474 | // / |
16475 | // add X, (zext(setne Z, C))-- |
16476 | // \ when -32768 <= -C <= 32767 && C != 0 |
16477 | // --> addze X, (addic (addi Z, -C), -1).carry |
16478 | SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z, |
16479 | DAG.getConstant(NegConstant, DL, MVT::i64)); |
16480 | SDValue AddOrZ = NegConstant != 0 ? Add : Z; |
16481 | SDValue Addc = DAG.getNode(ISD::ADDC, DL, DAG.getVTList(MVT::i64, MVT::Glue), |
16482 | AddOrZ, DAG.getConstant(-1ULL, DL, MVT::i64)); |
16483 | return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64), |
16484 | SDValue(Addc.getNode(), 1)); |
16485 | } |
16486 | case ISD::SETEQ: { |
16487 | // when C == 0 |
16488 | // --> addze X, (subfic Z, 0).carry |
16489 | // / |
16490 | // add X, (zext(sete Z, C))-- |
16491 | // \ when -32768 <= -C <= 32767 && C != 0 |
16492 | // --> addze X, (subfic (addi Z, -C), 0).carry |
16493 | SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Z, |
16494 | DAG.getConstant(NegConstant, DL, MVT::i64)); |
16495 | SDValue AddOrZ = NegConstant != 0 ? Add : Z; |
16496 | SDValue Subc = DAG.getNode(ISD::SUBC, DL, DAG.getVTList(MVT::i64, MVT::Glue), |
16497 | DAG.getConstant(0, DL, MVT::i64), AddOrZ); |
16498 | return DAG.getNode(ISD::ADDE, DL, VTs, LHS, DAG.getConstant(0, DL, MVT::i64), |
16499 | SDValue(Subc.getNode(), 1)); |
16500 | } |
16501 | } |
16502 | |
16503 | return SDValue(); |
16504 | } |
16505 | |
16506 | // Transform |
16507 | // (add C1, (MAT_PCREL_ADDR GlobalAddr+C2)) to |
16508 | // (MAT_PCREL_ADDR GlobalAddr+(C1+C2)) |
16509 | // In this case both C1 and C2 must be known constants. |
16510 | // C1+C2 must fit into a 34 bit signed integer. |
16511 | static SDValue combineADDToMAT_PCREL_ADDR(SDNode *N, SelectionDAG &DAG, |
16512 | const PPCSubtarget &Subtarget) { |
16513 | if (!Subtarget.isUsingPCRelativeCalls()) |
16514 | return SDValue(); |
16515 | |
16516 | // Check both Operand 0 and Operand 1 of the ADD node for the PCRel node. |
16517 | // If we find that node try to cast the Global Address and the Constant. |
16518 | SDValue LHS = N->getOperand(0); |
16519 | SDValue RHS = N->getOperand(1); |
16520 | |
16521 | if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR) |
16522 | std::swap(LHS, RHS); |
16523 | |
16524 | if (LHS.getOpcode() != PPCISD::MAT_PCREL_ADDR) |
16525 | return SDValue(); |
16526 | |
16527 | // Operand zero of PPCISD::MAT_PCREL_ADDR is the GA node. |
16528 | GlobalAddressSDNode *GSDN = dyn_cast<GlobalAddressSDNode>(LHS.getOperand(0)); |
16529 | ConstantSDNode* ConstNode = dyn_cast<ConstantSDNode>(RHS); |
16530 | |
16531 | // Check that both casts succeeded. |
16532 | if (!GSDN || !ConstNode) |
16533 | return SDValue(); |
16534 | |
16535 | int64_t NewOffset = GSDN->getOffset() + ConstNode->getSExtValue(); |
16536 | SDLoc DL(GSDN); |
16537 | |
16538 | // The signed int offset needs to fit in 34 bits. |
16539 | if (!isInt<34>(NewOffset)) |
16540 | return SDValue(); |
16541 | |
16542 | // The new global address is a copy of the old global address except |
16543 | // that it has the updated Offset. |
16544 | SDValue GA = |
16545 | DAG.getTargetGlobalAddress(GSDN->getGlobal(), DL, GSDN->getValueType(0), |
16546 | NewOffset, GSDN->getTargetFlags()); |
16547 | SDValue MatPCRel = |
16548 | DAG.getNode(PPCISD::MAT_PCREL_ADDR, DL, GSDN->getValueType(0), GA); |
16549 | return MatPCRel; |
16550 | } |
16551 | |
16552 | SDValue PPCTargetLowering::combineADD(SDNode *N, DAGCombinerInfo &DCI) const { |
16553 | if (auto Value = combineADDToADDZE(N, DCI.DAG, Subtarget)) |
16554 | return Value; |
16555 | |
16556 | if (auto Value = combineADDToMAT_PCREL_ADDR(N, DCI.DAG, Subtarget)) |
16557 | return Value; |
16558 | |
16559 | return SDValue(); |
16560 | } |
16561 | |
16562 | // Detect TRUNCATE operations on bitcasts of float128 values. |
16563 | // What we are looking for here is the situtation where we extract a subset |
16564 | // of bits from a 128 bit float. |
16565 | // This can be of two forms: |
16566 | // 1) BITCAST of f128 feeding TRUNCATE |
16567 | // 2) BITCAST of f128 feeding SRL (a shift) feeding TRUNCATE |
16568 | // The reason this is required is because we do not have a legal i128 type |
16569 | // and so we want to prevent having to store the f128 and then reload part |
16570 | // of it. |
16571 | SDValue PPCTargetLowering::combineTRUNCATE(SDNode *N, |
16572 | DAGCombinerInfo &DCI) const { |
16573 | // If we are using CRBits then try that first. |
16574 | if (Subtarget.useCRBits()) { |
16575 | // Check if CRBits did anything and return that if it did. |
16576 | if (SDValue CRTruncValue = DAGCombineTruncBoolExt(N, DCI)) |
16577 | return CRTruncValue; |
16578 | } |
16579 | |
16580 | SDLoc dl(N); |
16581 | SDValue Op0 = N->getOperand(0); |
16582 | |
16583 | // fold (truncate (abs (sub (zext a), (zext b)))) -> (vabsd a, b) |
16584 | if (Subtarget.hasP9Altivec() && Op0.getOpcode() == ISD::ABS) { |
16585 | EVT VT = N->getValueType(0); |
16586 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16587 | return SDValue(); |
16588 | SDValue Sub = Op0.getOperand(0); |
16589 | if (Sub.getOpcode() == ISD::SUB) { |
16590 | SDValue SubOp0 = Sub.getOperand(0); |
16591 | SDValue SubOp1 = Sub.getOperand(1); |
16592 | if ((SubOp0.getOpcode() == ISD::ZERO_EXTEND) && |
16593 | (SubOp1.getOpcode() == ISD::ZERO_EXTEND)) { |
16594 | return DCI.DAG.getNode(PPCISD::VABSD, dl, VT, SubOp0.getOperand(0), |
16595 | SubOp1.getOperand(0), |
16596 | DCI.DAG.getTargetConstant(0, dl, MVT::i32)); |
16597 | } |
16598 | } |
16599 | } |
16600 | |
16601 | // Looking for a truncate of i128 to i64. |
16602 | if (Op0.getValueType() != MVT::i128 || N->getValueType(0) != MVT::i64) |
16603 | return SDValue(); |
16604 | |
16605 | int EltToExtract = DCI.DAG.getDataLayout().isBigEndian() ? 1 : 0; |
16606 | |
16607 | // SRL feeding TRUNCATE. |
16608 | if (Op0.getOpcode() == ISD::SRL) { |
16609 | ConstantSDNode *ConstNode = dyn_cast<ConstantSDNode>(Op0.getOperand(1)); |
16610 | // The right shift has to be by 64 bits. |
16611 | if (!ConstNode || ConstNode->getZExtValue() != 64) |
16612 | return SDValue(); |
16613 | |
16614 | // Switch the element number to extract. |
16615 | EltToExtract = EltToExtract ? 0 : 1; |
16616 | // Update Op0 past the SRL. |
16617 | Op0 = Op0.getOperand(0); |
16618 | } |
16619 | |
16620 | // BITCAST feeding a TRUNCATE possibly via SRL. |
16621 | if (Op0.getOpcode() == ISD::BITCAST && |
16622 | Op0.getValueType() == MVT::i128 && |
16623 | Op0.getOperand(0).getValueType() == MVT::f128) { |
16624 | SDValue Bitcast = DCI.DAG.getBitcast(MVT::v2i64, Op0.getOperand(0)); |
16625 | return DCI.DAG.getNode( |
16626 | ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Bitcast, |
16627 | DCI.DAG.getTargetConstant(EltToExtract, dl, MVT::i32)); |
16628 | } |
16629 | return SDValue(); |
16630 | } |
16631 | |
16632 | SDValue PPCTargetLowering::combineMUL(SDNode *N, DAGCombinerInfo &DCI) const { |
16633 | SelectionDAG &DAG = DCI.DAG; |
16634 | |
16635 | ConstantSDNode *ConstOpOrElement = isConstOrConstSplat(N->getOperand(1)); |
16636 | if (!ConstOpOrElement) |
16637 | return SDValue(); |
16638 | |
16639 | // An imul is usually smaller than the alternative sequence for legal type. |
16640 | if (DAG.getMachineFunction().getFunction().hasMinSize() && |
16641 | isOperationLegal(ISD::MUL, N->getValueType(0))) |
16642 | return SDValue(); |
16643 | |
16644 | auto IsProfitable = [this](bool IsNeg, bool IsAddOne, EVT VT) -> bool { |
16645 | switch (this->Subtarget.getCPUDirective()) { |
16646 | default: |
16647 | // TODO: enhance the condition for subtarget before pwr8 |
16648 | return false; |
16649 | case PPC::DIR_PWR8: |
16650 | // type mul add shl |
16651 | // scalar 4 1 1 |
16652 | // vector 7 2 2 |
16653 | return true; |
16654 | case PPC::DIR_PWR9: |
16655 | case PPC::DIR_PWR10: |
16656 | case PPC::DIR_PWR_FUTURE: |
16657 | // type mul add shl |
16658 | // scalar 5 2 2 |
16659 | // vector 7 2 2 |
16660 | |
16661 | // The cycle RATIO of related operations are showed as a table above. |
16662 | // Because mul is 5(scalar)/7(vector), add/sub/shl are all 2 for both |
16663 | // scalar and vector type. For 2 instrs patterns, add/sub + shl |
16664 | // are 4, it is always profitable; but for 3 instrs patterns |
16665 | // (mul x, -(2^N + 1)) => -(add (shl x, N), x), sub + add + shl are 6. |
16666 | // So we should only do it for vector type. |
16667 | return IsAddOne && IsNeg ? VT.isVector() : true; |
16668 | } |
16669 | }; |
16670 | |
16671 | EVT VT = N->getValueType(0); |
16672 | SDLoc DL(N); |
16673 | |
16674 | const APInt &MulAmt = ConstOpOrElement->getAPIntValue(); |
16675 | bool IsNeg = MulAmt.isNegative(); |
16676 | APInt MulAmtAbs = MulAmt.abs(); |
16677 | |
16678 | if ((MulAmtAbs - 1).isPowerOf2()) { |
16679 | // (mul x, 2^N + 1) => (add (shl x, N), x) |
16680 | // (mul x, -(2^N + 1)) => -(add (shl x, N), x) |
16681 | |
16682 | if (!IsProfitable(IsNeg, true, VT)) |
16683 | return SDValue(); |
16684 | |
16685 | SDValue Op0 = N->getOperand(0); |
16686 | SDValue Op1 = |
16687 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
16688 | DAG.getConstant((MulAmtAbs - 1).logBase2(), DL, VT)); |
16689 | SDValue Res = DAG.getNode(ISD::ADD, DL, VT, Op0, Op1); |
16690 | |
16691 | if (!IsNeg) |
16692 | return Res; |
16693 | |
16694 | return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Res); |
16695 | } else if ((MulAmtAbs + 1).isPowerOf2()) { |
16696 | // (mul x, 2^N - 1) => (sub (shl x, N), x) |
16697 | // (mul x, -(2^N - 1)) => (sub x, (shl x, N)) |
16698 | |
16699 | if (!IsProfitable(IsNeg, false, VT)) |
16700 | return SDValue(); |
16701 | |
16702 | SDValue Op0 = N->getOperand(0); |
16703 | SDValue Op1 = |
16704 | DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0), |
16705 | DAG.getConstant((MulAmtAbs + 1).logBase2(), DL, VT)); |
16706 | |
16707 | if (!IsNeg) |
16708 | return DAG.getNode(ISD::SUB, DL, VT, Op1, Op0); |
16709 | else |
16710 | return DAG.getNode(ISD::SUB, DL, VT, Op0, Op1); |
16711 | |
16712 | } else { |
16713 | return SDValue(); |
16714 | } |
16715 | } |
16716 | |
16717 | // Combine fma-like op (like fnmsub) with fnegs to appropriate op. Do this |
16718 | // in combiner since we need to check SD flags and other subtarget features. |
16719 | SDValue PPCTargetLowering::combineFMALike(SDNode *N, |
16720 | DAGCombinerInfo &DCI) const { |
16721 | SDValue N0 = N->getOperand(0); |
16722 | SDValue N1 = N->getOperand(1); |
16723 | SDValue N2 = N->getOperand(2); |
16724 | SDNodeFlags Flags = N->getFlags(); |
16725 | EVT VT = N->getValueType(0); |
16726 | SelectionDAG &DAG = DCI.DAG; |
16727 | const TargetOptions &Options = getTargetMachine().Options; |
16728 | unsigned Opc = N->getOpcode(); |
16729 | bool CodeSize = DAG.getMachineFunction().getFunction().hasOptSize(); |
16730 | bool LegalOps = !DCI.isBeforeLegalizeOps(); |
16731 | SDLoc Loc(N); |
16732 | |
16733 | if (!isOperationLegal(ISD::FMA, VT)) |
16734 | return SDValue(); |
16735 | |
16736 | // Allowing transformation to FNMSUB may change sign of zeroes when ab-c=0 |
16737 | // since (fnmsub a b c)=-0 while c-ab=+0. |
16738 | if (!Flags.hasNoSignedZeros() && !Options.NoSignedZerosFPMath) |
16739 | return SDValue(); |
16740 | |
16741 | // (fma (fneg a) b c) => (fnmsub a b c) |
16742 | // (fnmsub (fneg a) b c) => (fma a b c) |
16743 | if (SDValue NegN0 = getCheaperNegatedExpression(N0, DAG, LegalOps, CodeSize)) |
16744 | return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, NegN0, N1, N2, Flags); |
16745 | |
16746 | // (fma a (fneg b) c) => (fnmsub a b c) |
16747 | // (fnmsub a (fneg b) c) => (fma a b c) |
16748 | if (SDValue NegN1 = getCheaperNegatedExpression(N1, DAG, LegalOps, CodeSize)) |
16749 | return DAG.getNode(invertFMAOpcode(Opc), Loc, VT, N0, NegN1, N2, Flags); |
16750 | |
16751 | return SDValue(); |
16752 | } |
16753 | |
16754 | bool PPCTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { |
16755 | // Only duplicate to increase tail-calls for the 64bit SysV ABIs. |
16756 | if (!Subtarget.is64BitELFABI()) |
16757 | return false; |
16758 | |
16759 | // If not a tail call then no need to proceed. |
16760 | if (!CI->isTailCall()) |
16761 | return false; |
16762 | |
16763 | // If sibling calls have been disabled and tail-calls aren't guaranteed |
16764 | // there is no reason to duplicate. |
16765 | auto &TM = getTargetMachine(); |
16766 | if (!TM.Options.GuaranteedTailCallOpt && DisableSCO) |
16767 | return false; |
16768 | |
16769 | // Can't tail call a function called indirectly, or if it has variadic args. |
16770 | const Function *Callee = CI->getCalledFunction(); |
16771 | if (!Callee || Callee->isVarArg()) |
16772 | return false; |
16773 | |
16774 | // Make sure the callee and caller calling conventions are eligible for tco. |
16775 | const Function *Caller = CI->getParent()->getParent(); |
16776 | if (!areCallingConvEligibleForTCO_64SVR4(Caller->getCallingConv(), |
16777 | CI->getCallingConv())) |
16778 | return false; |
16779 | |
16780 | // If the function is local then we have a good chance at tail-calling it |
16781 | return getTargetMachine().shouldAssumeDSOLocal(*Caller->getParent(), Callee); |
16782 | } |
16783 | |
16784 | bool PPCTargetLowering::hasBitPreservingFPLogic(EVT VT) const { |
16785 | if (!Subtarget.hasVSX()) |
16786 | return false; |
16787 | if (Subtarget.hasP9Vector() && VT == MVT::f128) |
16788 | return true; |
16789 | return VT == MVT::f32 || VT == MVT::f64 || |
16790 | VT == MVT::v4f32 || VT == MVT::v2f64; |
16791 | } |
16792 | |
16793 | bool PPCTargetLowering:: |
16794 | isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const { |
16795 | const Value *Mask = AndI.getOperand(1); |
16796 | // If the mask is suitable for andi. or andis. we should sink the and. |
16797 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Mask)) { |
16798 | // Can't handle constants wider than 64-bits. |
16799 | if (CI->getBitWidth() > 64) |
16800 | return false; |
16801 | int64_t ConstVal = CI->getZExtValue(); |
16802 | return isUInt<16>(ConstVal) || |
16803 | (isUInt<16>(ConstVal >> 16) && !(ConstVal & 0xFFFF)); |
16804 | } |
16805 | |
16806 | // For non-constant masks, we can always use the record-form and. |
16807 | return true; |
16808 | } |
16809 | |
16810 | // Transform (abs (sub (zext a), (zext b))) to (vabsd a b 0) |
16811 | // Transform (abs (sub (zext a), (zext_invec b))) to (vabsd a b 0) |
16812 | // Transform (abs (sub (zext_invec a), (zext_invec b))) to (vabsd a b 0) |
16813 | // Transform (abs (sub (zext_invec a), (zext b))) to (vabsd a b 0) |
16814 | // Transform (abs (sub a, b) to (vabsd a b 1)) if a & b of type v4i32 |
16815 | SDValue PPCTargetLowering::combineABS(SDNode *N, DAGCombinerInfo &DCI) const { |
16816 | assert((N->getOpcode() == ISD::ABS) && "Need ABS node here")(((N->getOpcode() == ISD::ABS) && "Need ABS node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::ABS) && \"Need ABS node here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16816, __PRETTY_FUNCTION__)); |
16817 | assert(Subtarget.hasP9Altivec() &&((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16818, __PRETTY_FUNCTION__)) |
16818 | "Only combine this when P9 altivec supported!")((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16818, __PRETTY_FUNCTION__)); |
16819 | EVT VT = N->getValueType(0); |
16820 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16821 | return SDValue(); |
16822 | |
16823 | SelectionDAG &DAG = DCI.DAG; |
16824 | SDLoc dl(N); |
16825 | if (N->getOperand(0).getOpcode() == ISD::SUB) { |
16826 | // Even for signed integers, if it's known to be positive (as signed |
16827 | // integer) due to zero-extended inputs. |
16828 | unsigned SubOpcd0 = N->getOperand(0)->getOperand(0).getOpcode(); |
16829 | unsigned SubOpcd1 = N->getOperand(0)->getOperand(1).getOpcode(); |
16830 | if ((SubOpcd0 == ISD::ZERO_EXTEND || |
16831 | SubOpcd0 == ISD::ZERO_EXTEND_VECTOR_INREG) && |
16832 | (SubOpcd1 == ISD::ZERO_EXTEND || |
16833 | SubOpcd1 == ISD::ZERO_EXTEND_VECTOR_INREG)) { |
16834 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(0).getValueType(), |
16835 | N->getOperand(0)->getOperand(0), |
16836 | N->getOperand(0)->getOperand(1), |
16837 | DAG.getTargetConstant(0, dl, MVT::i32)); |
16838 | } |
16839 | |
16840 | // For type v4i32, it can be optimized with xvnegsp + vabsduw |
16841 | if (N->getOperand(0).getValueType() == MVT::v4i32 && |
16842 | N->getOperand(0).hasOneUse()) { |
16843 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(0).getValueType(), |
16844 | N->getOperand(0)->getOperand(0), |
16845 | N->getOperand(0)->getOperand(1), |
16846 | DAG.getTargetConstant(1, dl, MVT::i32)); |
16847 | } |
16848 | } |
16849 | |
16850 | return SDValue(); |
16851 | } |
16852 | |
16853 | // For type v4i32/v8ii16/v16i8, transform |
16854 | // from (vselect (setcc a, b, setugt), (sub a, b), (sub b, a)) to (vabsd a, b) |
16855 | // from (vselect (setcc a, b, setuge), (sub a, b), (sub b, a)) to (vabsd a, b) |
16856 | // from (vselect (setcc a, b, setult), (sub b, a), (sub a, b)) to (vabsd a, b) |
16857 | // from (vselect (setcc a, b, setule), (sub b, a), (sub a, b)) to (vabsd a, b) |
16858 | SDValue PPCTargetLowering::combineVSelect(SDNode *N, |
16859 | DAGCombinerInfo &DCI) const { |
16860 | assert((N->getOpcode() == ISD::VSELECT) && "Need VSELECT node here")(((N->getOpcode() == ISD::VSELECT) && "Need VSELECT node here" ) ? static_cast<void> (0) : __assert_fail ("(N->getOpcode() == ISD::VSELECT) && \"Need VSELECT node here\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16860, __PRETTY_FUNCTION__)); |
16861 | assert(Subtarget.hasP9Altivec() &&((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16862, __PRETTY_FUNCTION__)) |
16862 | "Only combine this when P9 altivec supported!")((Subtarget.hasP9Altivec() && "Only combine this when P9 altivec supported!" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasP9Altivec() && \"Only combine this when P9 altivec supported!\"" , "/build/llvm-toolchain-snapshot-12~++20201026111116+d3205bbca3e/llvm/lib/Target/PowerPC/PPCISelLowering.cpp" , 16862, __PRETTY_FUNCTION__)); |
16863 | |
16864 | SelectionDAG &DAG = DCI.DAG; |
16865 | SDLoc dl(N); |
16866 | SDValue Cond = N->getOperand(0); |
16867 | SDValue TrueOpnd = N->getOperand(1); |
16868 | SDValue FalseOpnd = N->getOperand(2); |
16869 | EVT VT = N->getOperand(1).getValueType(); |
16870 | |
16871 | if (Cond.getOpcode() != ISD::SETCC || TrueOpnd.getOpcode() != ISD::SUB || |
16872 | FalseOpnd.getOpcode() != ISD::SUB) |
16873 | return SDValue(); |
16874 | |
16875 | // ABSD only available for type v4i32/v8i16/v16i8 |
16876 | if (VT != MVT::v4i32 && VT != MVT::v8i16 && VT != MVT::v16i8) |
16877 | return SDValue(); |
16878 | |
16879 | // At least to save one more dependent computation |
16880 | if (!(Cond.hasOneUse() || TrueOpnd.hasOneUse() || FalseOpnd.hasOneUse())) |
16881 | return SDValue(); |
16882 | |
16883 | ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get(); |
16884 | |
16885 | // Can only handle unsigned comparison here |
16886 | switch (CC) { |
16887 | default: |
16888 | return SDValue(); |
16889 | case ISD::SETUGT: |
16890 | case ISD::SETUGE: |
16891 | break; |
16892 | case ISD::SETULT: |
16893 | case ISD::SETULE: |
16894 | std::swap(TrueOpnd, FalseOpnd); |
16895 | break; |
16896 | } |
16897 | |
16898 | SDValue CmpOpnd1 = Cond.getOperand(0); |
16899 | SDValue CmpOpnd2 = Cond.getOperand(1); |
16900 | |
16901 | // SETCC CmpOpnd1 CmpOpnd2 cond |
16902 | // TrueOpnd = CmpOpnd1 - CmpOpnd2 |
16903 | // FalseOpnd = CmpOpnd2 - CmpOpnd1 |
16904 | if (TrueOpnd.getOperand(0) == CmpOpnd1 && |
16905 | TrueOpnd.getOperand(1) == CmpOpnd2 && |
16906 | FalseOpnd.getOperand(0) == CmpOpnd2 && |
16907 | FalseOpnd.getOperand(1) == CmpOpnd1) { |
16908 | return DAG.getNode(PPCISD::VABSD, dl, N->getOperand(1).getValueType(), |
16909 | CmpOpnd1, CmpOpnd2, |
16910 | DAG.getTargetConstant(0, dl, MVT::i32)); |
16911 | } |
16912 | |
16913 | return SDValue(); |
16914 | } |