File: | lib/Target/Hexagon/HexagonISelLowering.cpp |
Warning: | line 2824, column 16 Value stored to 'Subreg' during its initialization is never read |
1 | //===-- HexagonISelLowering.cpp - Hexagon DAG Lowering Implementation -----===// |
2 | // |
3 | // The LLVM Compiler Infrastructure |
4 | // |
5 | // This file is distributed under the University of Illinois Open Source |
6 | // License. See LICENSE.TXT for details. |
7 | // |
8 | //===----------------------------------------------------------------------===// |
9 | // |
10 | // This file implements the interfaces that Hexagon uses to lower LLVM code |
11 | // into a selection DAG. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "HexagonISelLowering.h" |
16 | #include "Hexagon.h" |
17 | #include "HexagonMachineFunctionInfo.h" |
18 | #include "HexagonRegisterInfo.h" |
19 | #include "HexagonSubtarget.h" |
20 | #include "HexagonTargetMachine.h" |
21 | #include "HexagonTargetObjectFile.h" |
22 | #include "llvm/ADT/APInt.h" |
23 | #include "llvm/ADT/ArrayRef.h" |
24 | #include "llvm/ADT/SmallVector.h" |
25 | #include "llvm/CodeGen/CallingConvLower.h" |
26 | #include "llvm/CodeGen/MachineFrameInfo.h" |
27 | #include "llvm/CodeGen/MachineFunction.h" |
28 | #include "llvm/CodeGen/MachineMemOperand.h" |
29 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
30 | #include "llvm/CodeGen/RuntimeLibcalls.h" |
31 | #include "llvm/CodeGen/SelectionDAG.h" |
32 | #include "llvm/CodeGen/ValueTypes.h" |
33 | #include "llvm/IR/BasicBlock.h" |
34 | #include "llvm/IR/CallingConv.h" |
35 | #include "llvm/IR/DataLayout.h" |
36 | #include "llvm/IR/DerivedTypes.h" |
37 | #include "llvm/IR/Function.h" |
38 | #include "llvm/IR/GlobalValue.h" |
39 | #include "llvm/IR/InlineAsm.h" |
40 | #include "llvm/IR/Instructions.h" |
41 | #include "llvm/IR/Intrinsics.h" |
42 | #include "llvm/IR/Module.h" |
43 | #include "llvm/IR/Type.h" |
44 | #include "llvm/IR/Value.h" |
45 | #include "llvm/MC/MCRegisterInfo.h" |
46 | #include "llvm/Support/Casting.h" |
47 | #include "llvm/Support/CodeGen.h" |
48 | #include "llvm/Support/CommandLine.h" |
49 | #include "llvm/Support/Debug.h" |
50 | #include "llvm/Support/ErrorHandling.h" |
51 | #include "llvm/Support/MathExtras.h" |
52 | #include "llvm/Support/raw_ostream.h" |
53 | #include "llvm/Target/TargetCallingConv.h" |
54 | #include "llvm/Target/TargetMachine.h" |
55 | #include <algorithm> |
56 | #include <cassert> |
57 | #include <cstddef> |
58 | #include <cstdint> |
59 | #include <limits> |
60 | #include <utility> |
61 | |
62 | using namespace llvm; |
63 | |
64 | #define DEBUG_TYPE"hexagon-lowering" "hexagon-lowering" |
65 | |
66 | static cl::opt<bool> EmitJumpTables("hexagon-emit-jump-tables", |
67 | cl::init(true), cl::Hidden, |
68 | cl::desc("Control jump table emission on Hexagon target")); |
69 | |
70 | static cl::opt<bool> EnableHexSDNodeSched("enable-hexagon-sdnode-sched", |
71 | cl::Hidden, cl::ZeroOrMore, cl::init(false), |
72 | cl::desc("Enable Hexagon SDNode scheduling")); |
73 | |
74 | static cl::opt<bool> EnableFastMath("ffast-math", |
75 | cl::Hidden, cl::ZeroOrMore, cl::init(false), |
76 | cl::desc("Enable Fast Math processing")); |
77 | |
78 | static cl::opt<int> MinimumJumpTables("minimum-jump-tables", |
79 | cl::Hidden, cl::ZeroOrMore, cl::init(5), |
80 | cl::desc("Set minimum jump tables")); |
81 | |
82 | static cl::opt<int> MaxStoresPerMemcpyCL("max-store-memcpy", |
83 | cl::Hidden, cl::ZeroOrMore, cl::init(6), |
84 | cl::desc("Max #stores to inline memcpy")); |
85 | |
86 | static cl::opt<int> MaxStoresPerMemcpyOptSizeCL("max-store-memcpy-Os", |
87 | cl::Hidden, cl::ZeroOrMore, cl::init(4), |
88 | cl::desc("Max #stores to inline memcpy")); |
89 | |
90 | static cl::opt<int> MaxStoresPerMemmoveCL("max-store-memmove", |
91 | cl::Hidden, cl::ZeroOrMore, cl::init(6), |
92 | cl::desc("Max #stores to inline memmove")); |
93 | |
94 | static cl::opt<int> MaxStoresPerMemmoveOptSizeCL("max-store-memmove-Os", |
95 | cl::Hidden, cl::ZeroOrMore, cl::init(4), |
96 | cl::desc("Max #stores to inline memmove")); |
97 | |
98 | static cl::opt<int> MaxStoresPerMemsetCL("max-store-memset", |
99 | cl::Hidden, cl::ZeroOrMore, cl::init(8), |
100 | cl::desc("Max #stores to inline memset")); |
101 | |
102 | static cl::opt<int> MaxStoresPerMemsetOptSizeCL("max-store-memset-Os", |
103 | cl::Hidden, cl::ZeroOrMore, cl::init(4), |
104 | cl::desc("Max #stores to inline memset")); |
105 | |
106 | |
107 | namespace { |
108 | |
109 | class HexagonCCState : public CCState { |
110 | unsigned NumNamedVarArgParams; |
111 | |
112 | public: |
113 | HexagonCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF, |
114 | SmallVectorImpl<CCValAssign> &locs, LLVMContext &C, |
115 | int NumNamedVarArgParams) |
116 | : CCState(CC, isVarArg, MF, locs, C), |
117 | NumNamedVarArgParams(NumNamedVarArgParams) {} |
118 | |
119 | unsigned getNumNamedVarArgParams() const { return NumNamedVarArgParams; } |
120 | }; |
121 | |
122 | enum StridedLoadKind { |
123 | Even = 0, |
124 | Odd, |
125 | NoPattern |
126 | }; |
127 | |
128 | } // end anonymous namespace |
129 | |
130 | // Implement calling convention for Hexagon. |
131 | |
132 | static bool isHvxVectorType(MVT ty); |
133 | |
134 | static bool |
135 | CC_Hexagon(unsigned ValNo, MVT ValVT, |
136 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
137 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
138 | |
139 | static bool |
140 | CC_Hexagon32(unsigned ValNo, MVT ValVT, |
141 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
142 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
143 | |
144 | static bool |
145 | CC_Hexagon64(unsigned ValNo, MVT ValVT, |
146 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
147 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
148 | |
149 | static bool |
150 | CC_HexagonVector(unsigned ValNo, MVT ValVT, |
151 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
152 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
153 | |
154 | static bool |
155 | RetCC_Hexagon(unsigned ValNo, MVT ValVT, |
156 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
157 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
158 | |
159 | static bool |
160 | RetCC_Hexagon32(unsigned ValNo, MVT ValVT, |
161 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
162 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
163 | |
164 | static bool |
165 | RetCC_Hexagon64(unsigned ValNo, MVT ValVT, |
166 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
167 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
168 | |
169 | static bool |
170 | RetCC_HexagonVector(unsigned ValNo, MVT ValVT, |
171 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
172 | ISD::ArgFlagsTy ArgFlags, CCState &State); |
173 | |
174 | static bool |
175 | CC_Hexagon_VarArg (unsigned ValNo, MVT ValVT, |
176 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
177 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
178 | HexagonCCState &HState = static_cast<HexagonCCState &>(State); |
179 | |
180 | if (ValNo < HState.getNumNamedVarArgParams()) { |
181 | // Deal with named arguments. |
182 | return CC_Hexagon(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State); |
183 | } |
184 | |
185 | // Deal with un-named arguments. |
186 | unsigned Offset; |
187 | if (ArgFlags.isByVal()) { |
188 | // If pass-by-value, the size allocated on stack is decided |
189 | // by ArgFlags.getByValSize(), not by the size of LocVT. |
190 | Offset = State.AllocateStack(ArgFlags.getByValSize(), |
191 | ArgFlags.getByValAlign()); |
192 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
193 | return false; |
194 | } |
195 | if (LocVT == MVT::i1 || LocVT == MVT::i8 || LocVT == MVT::i16) { |
196 | LocVT = MVT::i32; |
197 | ValVT = MVT::i32; |
198 | if (ArgFlags.isSExt()) |
199 | LocInfo = CCValAssign::SExt; |
200 | else if (ArgFlags.isZExt()) |
201 | LocInfo = CCValAssign::ZExt; |
202 | else |
203 | LocInfo = CCValAssign::AExt; |
204 | } |
205 | if (LocVT == MVT::i32 || LocVT == MVT::f32) { |
206 | Offset = State.AllocateStack(4, 4); |
207 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
208 | return false; |
209 | } |
210 | if (LocVT == MVT::i64 || LocVT == MVT::f64) { |
211 | Offset = State.AllocateStack(8, 8); |
212 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
213 | return false; |
214 | } |
215 | if (LocVT == MVT::v2i64 || LocVT == MVT::v4i32 || LocVT == MVT::v8i16 || |
216 | LocVT == MVT::v16i8) { |
217 | Offset = State.AllocateStack(16, 16); |
218 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
219 | return false; |
220 | } |
221 | if (LocVT == MVT::v4i64 || LocVT == MVT::v8i32 || LocVT == MVT::v16i16 || |
222 | LocVT == MVT::v32i8) { |
223 | Offset = State.AllocateStack(32, 32); |
224 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
225 | return false; |
226 | } |
227 | if (LocVT == MVT::v8i64 || LocVT == MVT::v16i32 || LocVT == MVT::v32i16 || |
228 | LocVT == MVT::v64i8 || LocVT == MVT::v512i1) { |
229 | Offset = State.AllocateStack(64, 64); |
230 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
231 | return false; |
232 | } |
233 | if (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 || |
234 | LocVT == MVT::v128i8 || LocVT == MVT::v1024i1) { |
235 | Offset = State.AllocateStack(128, 128); |
236 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
237 | return false; |
238 | } |
239 | if (LocVT == MVT::v32i64 || LocVT == MVT::v64i32 || LocVT == MVT::v128i16 || |
240 | LocVT == MVT::v256i8) { |
241 | Offset = State.AllocateStack(256, 256); |
242 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
243 | return false; |
244 | } |
245 | |
246 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 246); |
247 | } |
248 | |
249 | static bool CC_Hexagon (unsigned ValNo, MVT ValVT, MVT LocVT, |
250 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, CCState &State) { |
251 | if (ArgFlags.isByVal()) { |
252 | // Passed on stack. |
253 | unsigned Offset = State.AllocateStack(ArgFlags.getByValSize(), |
254 | ArgFlags.getByValAlign()); |
255 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
256 | return false; |
257 | } |
258 | |
259 | if (LocVT == MVT::i1) { |
260 | LocVT = MVT::i32; |
261 | } else if (LocVT == MVT::i8 || LocVT == MVT::i16) { |
262 | LocVT = MVT::i32; |
263 | ValVT = MVT::i32; |
264 | if (ArgFlags.isSExt()) |
265 | LocInfo = CCValAssign::SExt; |
266 | else if (ArgFlags.isZExt()) |
267 | LocInfo = CCValAssign::ZExt; |
268 | else |
269 | LocInfo = CCValAssign::AExt; |
270 | } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) { |
271 | LocVT = MVT::i32; |
272 | LocInfo = CCValAssign::BCvt; |
273 | } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) { |
274 | LocVT = MVT::i64; |
275 | LocInfo = CCValAssign::BCvt; |
276 | } |
277 | |
278 | if (LocVT == MVT::i32 || LocVT == MVT::f32) { |
279 | if (!CC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
280 | return false; |
281 | } |
282 | |
283 | if (LocVT == MVT::i64 || LocVT == MVT::f64) { |
284 | if (!CC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
285 | return false; |
286 | } |
287 | |
288 | if (LocVT == MVT::v8i32 || LocVT == MVT::v16i16 || LocVT == MVT::v32i8) { |
289 | unsigned Offset = State.AllocateStack(ArgFlags.getByValSize(), 32); |
290 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
291 | return false; |
292 | } |
293 | |
294 | if (isHvxVectorType(LocVT)) { |
295 | if (!CC_HexagonVector(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
296 | return false; |
297 | } |
298 | |
299 | return true; // CC didn't match. |
300 | } |
301 | |
302 | |
303 | static bool CC_Hexagon32(unsigned ValNo, MVT ValVT, |
304 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
305 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
306 | static const MCPhysReg RegList[] = { |
307 | Hexagon::R0, Hexagon::R1, Hexagon::R2, Hexagon::R3, Hexagon::R4, |
308 | Hexagon::R5 |
309 | }; |
310 | if (unsigned Reg = State.AllocateReg(RegList)) { |
311 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
312 | return false; |
313 | } |
314 | |
315 | unsigned Offset = State.AllocateStack(4, 4); |
316 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
317 | return false; |
318 | } |
319 | |
320 | static bool CC_Hexagon64(unsigned ValNo, MVT ValVT, |
321 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
322 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
323 | if (unsigned Reg = State.AllocateReg(Hexagon::D0)) { |
324 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
325 | return false; |
326 | } |
327 | |
328 | static const MCPhysReg RegList1[] = { |
329 | Hexagon::D1, Hexagon::D2 |
330 | }; |
331 | static const MCPhysReg RegList2[] = { |
332 | Hexagon::R1, Hexagon::R3 |
333 | }; |
334 | if (unsigned Reg = State.AllocateReg(RegList1, RegList2)) { |
335 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
336 | return false; |
337 | } |
338 | |
339 | unsigned Offset = State.AllocateStack(8, 8, Hexagon::D2); |
340 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
341 | return false; |
342 | } |
343 | |
344 | static bool CC_HexagonVector(unsigned ValNo, MVT ValVT, |
345 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
346 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
347 | static const MCPhysReg VecLstS[] = { |
348 | Hexagon::V0, Hexagon::V1, Hexagon::V2, Hexagon::V3, Hexagon::V4, |
349 | Hexagon::V5, Hexagon::V6, Hexagon::V7, Hexagon::V8, Hexagon::V9, |
350 | Hexagon::V10, Hexagon::V11, Hexagon::V12, Hexagon::V13, Hexagon::V14, |
351 | Hexagon::V15 |
352 | }; |
353 | static const MCPhysReg VecLstD[] = { |
354 | Hexagon::W0, Hexagon::W1, Hexagon::W2, Hexagon::W3, Hexagon::W4, |
355 | Hexagon::W5, Hexagon::W6, Hexagon::W7 |
356 | }; |
357 | auto &MF = State.getMachineFunction(); |
358 | auto &HST = MF.getSubtarget<HexagonSubtarget>(); |
359 | bool UseHVX = HST.useHVXOps(); |
360 | bool UseHVXDbl = HST.useHVXDblOps(); |
361 | |
362 | if ((UseHVX && !UseHVXDbl) && |
363 | (LocVT == MVT::v8i64 || LocVT == MVT::v16i32 || LocVT == MVT::v32i16 || |
364 | LocVT == MVT::v64i8 || LocVT == MVT::v512i1)) { |
365 | if (unsigned Reg = State.AllocateReg(VecLstS)) { |
366 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
367 | return false; |
368 | } |
369 | unsigned Offset = State.AllocateStack(64, 64); |
370 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
371 | return false; |
372 | } |
373 | if ((UseHVX && !UseHVXDbl) && |
374 | (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 || |
375 | LocVT == MVT::v128i8)) { |
376 | if (unsigned Reg = State.AllocateReg(VecLstD)) { |
377 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
378 | return false; |
379 | } |
380 | unsigned Offset = State.AllocateStack(128, 128); |
381 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
382 | return false; |
383 | } |
384 | // 128B Mode |
385 | if ((UseHVX && UseHVXDbl) && |
386 | (LocVT == MVT::v32i64 || LocVT == MVT::v64i32 || LocVT == MVT::v128i16 || |
387 | LocVT == MVT::v256i8)) { |
388 | if (unsigned Reg = State.AllocateReg(VecLstD)) { |
389 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
390 | return false; |
391 | } |
392 | unsigned Offset = State.AllocateStack(256, 256); |
393 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
394 | return false; |
395 | } |
396 | if ((UseHVX && UseHVXDbl) && |
397 | (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 || |
398 | LocVT == MVT::v128i8 || LocVT == MVT::v1024i1)) { |
399 | if (unsigned Reg = State.AllocateReg(VecLstS)) { |
400 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
401 | return false; |
402 | } |
403 | unsigned Offset = State.AllocateStack(128, 128); |
404 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo)); |
405 | return false; |
406 | } |
407 | return true; |
408 | } |
409 | |
410 | static bool RetCC_Hexagon(unsigned ValNo, MVT ValVT, |
411 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
412 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
413 | auto &MF = State.getMachineFunction(); |
414 | auto &HST = MF.getSubtarget<HexagonSubtarget>(); |
415 | bool UseHVX = HST.useHVXOps(); |
416 | bool UseHVXDbl = HST.useHVXDblOps(); |
417 | |
418 | if (LocVT == MVT::i1) { |
419 | // Return values of type MVT::i1 still need to be assigned to R0, but |
420 | // the value type needs to remain i1. LowerCallResult will deal with it, |
421 | // but it needs to recognize i1 as the value type. |
422 | LocVT = MVT::i32; |
423 | } else if (LocVT == MVT::i8 || LocVT == MVT::i16) { |
424 | LocVT = MVT::i32; |
425 | ValVT = MVT::i32; |
426 | if (ArgFlags.isSExt()) |
427 | LocInfo = CCValAssign::SExt; |
428 | else if (ArgFlags.isZExt()) |
429 | LocInfo = CCValAssign::ZExt; |
430 | else |
431 | LocInfo = CCValAssign::AExt; |
432 | } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) { |
433 | LocVT = MVT::i32; |
434 | LocInfo = CCValAssign::BCvt; |
435 | } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) { |
436 | LocVT = MVT::i64; |
437 | LocInfo = CCValAssign::BCvt; |
438 | } else if (LocVT == MVT::v64i8 || LocVT == MVT::v32i16 || |
439 | LocVT == MVT::v16i32 || LocVT == MVT::v8i64 || |
440 | LocVT == MVT::v512i1) { |
441 | LocVT = MVT::v16i32; |
442 | ValVT = MVT::v16i32; |
443 | LocInfo = CCValAssign::Full; |
444 | } else if (LocVT == MVT::v128i8 || LocVT == MVT::v64i16 || |
445 | LocVT == MVT::v32i32 || LocVT == MVT::v16i64 || |
446 | (LocVT == MVT::v1024i1 && UseHVX && UseHVXDbl)) { |
447 | LocVT = MVT::v32i32; |
448 | ValVT = MVT::v32i32; |
449 | LocInfo = CCValAssign::Full; |
450 | } else if (LocVT == MVT::v256i8 || LocVT == MVT::v128i16 || |
451 | LocVT == MVT::v64i32 || LocVT == MVT::v32i64) { |
452 | LocVT = MVT::v64i32; |
453 | ValVT = MVT::v64i32; |
454 | LocInfo = CCValAssign::Full; |
455 | } |
456 | if (LocVT == MVT::i32 || LocVT == MVT::f32) { |
457 | if (!RetCC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
458 | return false; |
459 | } |
460 | |
461 | if (LocVT == MVT::i64 || LocVT == MVT::f64) { |
462 | if (!RetCC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
463 | return false; |
464 | } |
465 | if (LocVT == MVT::v16i32 || LocVT == MVT::v32i32 || LocVT == MVT::v64i32) { |
466 | if (!RetCC_HexagonVector(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State)) |
467 | return false; |
468 | } |
469 | return true; // CC didn't match. |
470 | } |
471 | |
472 | static bool RetCC_Hexagon32(unsigned ValNo, MVT ValVT, |
473 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
474 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
475 | if (LocVT == MVT::i32 || LocVT == MVT::f32) { |
476 | // Note that use of registers beyond R1 is not ABI compliant. However there |
477 | // are (experimental) IR passes which generate internal functions that |
478 | // return structs using these additional registers. |
479 | static const uint16_t RegList[] = { Hexagon::R0, Hexagon::R1, |
480 | Hexagon::R2, Hexagon::R3, |
481 | Hexagon::R4, Hexagon::R5 }; |
482 | if (unsigned Reg = State.AllocateReg(RegList)) { |
483 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
484 | return false; |
485 | } |
486 | } |
487 | |
488 | return true; |
489 | } |
490 | |
491 | static bool RetCC_Hexagon64(unsigned ValNo, MVT ValVT, |
492 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
493 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
494 | if (LocVT == MVT::i64 || LocVT == MVT::f64) { |
495 | if (unsigned Reg = State.AllocateReg(Hexagon::D0)) { |
496 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
497 | return false; |
498 | } |
499 | } |
500 | |
501 | return true; |
502 | } |
503 | |
504 | static bool RetCC_HexagonVector(unsigned ValNo, MVT ValVT, |
505 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
506 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
507 | auto &MF = State.getMachineFunction(); |
508 | auto &HST = MF.getSubtarget<HexagonSubtarget>(); |
509 | bool UseHVX = HST.useHVXOps(); |
510 | bool UseHVXDbl = HST.useHVXDblOps(); |
511 | |
512 | if (LocVT == MVT::v16i32) { |
513 | if (unsigned Reg = State.AllocateReg(Hexagon::V0)) { |
514 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
515 | return false; |
516 | } |
517 | } else if (LocVT == MVT::v32i32) { |
518 | unsigned Req = (UseHVX && UseHVXDbl) ? Hexagon::V0 : Hexagon::W0; |
519 | if (unsigned Reg = State.AllocateReg(Req)) { |
520 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
521 | return false; |
522 | } |
523 | } else if (LocVT == MVT::v64i32) { |
524 | if (unsigned Reg = State.AllocateReg(Hexagon::W0)) { |
525 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); |
526 | return false; |
527 | } |
528 | } |
529 | |
530 | return true; |
531 | } |
532 | |
533 | void HexagonTargetLowering::promoteLdStType(MVT VT, MVT PromotedLdStVT) { |
534 | if (VT != PromotedLdStVT) { |
535 | setOperationAction(ISD::LOAD, VT, Promote); |
536 | AddPromotedToType(ISD::LOAD, VT, PromotedLdStVT); |
537 | |
538 | setOperationAction(ISD::STORE, VT, Promote); |
539 | AddPromotedToType(ISD::STORE, VT, PromotedLdStVT); |
540 | } |
541 | } |
542 | |
543 | SDValue |
544 | HexagonTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) |
545 | const { |
546 | return SDValue(); |
547 | } |
548 | |
549 | /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified |
550 | /// by "Src" to address "Dst" of size "Size". Alignment information is |
551 | /// specified by the specific parameter attribute. The copy will be passed as |
552 | /// a byval function parameter. Sometimes what we are copying is the end of a |
553 | /// larger object, the part that does not fit in registers. |
554 | static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, |
555 | SDValue Chain, ISD::ArgFlagsTy Flags, |
556 | SelectionDAG &DAG, const SDLoc &dl) { |
557 | SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32); |
558 | return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), |
559 | /*isVolatile=*/false, /*AlwaysInline=*/false, |
560 | /*isTailCall=*/false, |
561 | MachinePointerInfo(), MachinePointerInfo()); |
562 | } |
563 | |
564 | static bool isHvxVectorType(MVT Ty) { |
565 | switch (Ty.SimpleTy) { |
566 | case MVT::v8i64: |
567 | case MVT::v16i32: |
568 | case MVT::v32i16: |
569 | case MVT::v64i8: |
570 | case MVT::v16i64: |
571 | case MVT::v32i32: |
572 | case MVT::v64i16: |
573 | case MVT::v128i8: |
574 | case MVT::v32i64: |
575 | case MVT::v64i32: |
576 | case MVT::v128i16: |
577 | case MVT::v256i8: |
578 | case MVT::v512i1: |
579 | case MVT::v1024i1: |
580 | return true; |
581 | default: |
582 | return false; |
583 | } |
584 | } |
585 | |
586 | bool |
587 | HexagonTargetLowering::CanLowerReturn( |
588 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, |
589 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
590 | LLVMContext &Context) const { |
591 | SmallVector<CCValAssign, 16> RVLocs; |
592 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
593 | return CCInfo.CheckReturn(Outs, RetCC_Hexagon); |
594 | } |
595 | |
596 | // LowerReturn - Lower ISD::RET. If a struct is larger than 8 bytes and is |
597 | // passed by value, the function prototype is modified to return void and |
598 | // the value is stored in memory pointed by a pointer passed by caller. |
599 | SDValue |
600 | HexagonTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
601 | bool isVarArg, |
602 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
603 | const SmallVectorImpl<SDValue> &OutVals, |
604 | const SDLoc &dl, SelectionDAG &DAG) const { |
605 | // CCValAssign - represent the assignment of the return value to locations. |
606 | SmallVector<CCValAssign, 16> RVLocs; |
607 | |
608 | // CCState - Info about the registers and stack slot. |
609 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
610 | *DAG.getContext()); |
611 | |
612 | // Analyze return values of ISD::RET |
613 | CCInfo.AnalyzeReturn(Outs, RetCC_Hexagon); |
614 | |
615 | SDValue Flag; |
616 | SmallVector<SDValue, 4> RetOps(1, Chain); |
617 | |
618 | // Copy the result values into the output registers. |
619 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
620 | CCValAssign &VA = RVLocs[i]; |
621 | |
622 | Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag); |
623 | |
624 | // Guarantee that all emitted copies are stuck together with flags. |
625 | Flag = Chain.getValue(1); |
626 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); |
627 | } |
628 | |
629 | RetOps[0] = Chain; // Update chain. |
630 | |
631 | // Add the flag if we have it. |
632 | if (Flag.getNode()) |
633 | RetOps.push_back(Flag); |
634 | |
635 | return DAG.getNode(HexagonISD::RET_FLAG, dl, MVT::Other, RetOps); |
636 | } |
637 | |
638 | bool HexagonTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { |
639 | // If either no tail call or told not to tail call at all, don't. |
640 | auto Attr = |
641 | CI->getParent()->getParent()->getFnAttribute("disable-tail-calls"); |
642 | if (!CI->isTailCall() || Attr.getValueAsString() == "true") |
643 | return false; |
644 | |
645 | return true; |
646 | } |
647 | |
648 | /// LowerCallResult - Lower the result values of an ISD::CALL into the |
649 | /// appropriate copies out of appropriate physical registers. This assumes that |
650 | /// Chain/Glue are the input chain/glue to use, and that TheCall is the call |
651 | /// being lowered. Returns a SDNode with the same number of values as the |
652 | /// ISD::CALL. |
653 | SDValue HexagonTargetLowering::LowerCallResult( |
654 | SDValue Chain, SDValue Glue, CallingConv::ID CallConv, bool isVarArg, |
655 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
656 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
657 | const SmallVectorImpl<SDValue> &OutVals, SDValue Callee) const { |
658 | // Assign locations to each value returned by this call. |
659 | SmallVector<CCValAssign, 16> RVLocs; |
660 | |
661 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
662 | *DAG.getContext()); |
663 | |
664 | CCInfo.AnalyzeCallResult(Ins, RetCC_Hexagon); |
665 | |
666 | // Copy all of the result registers out of their specified physreg. |
667 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
668 | SDValue RetVal; |
669 | if (RVLocs[i].getValVT() == MVT::i1) { |
670 | // Return values of type MVT::i1 require special handling. The reason |
671 | // is that MVT::i1 is associated with the PredRegs register class, but |
672 | // values of that type are still returned in R0. Generate an explicit |
673 | // copy into a predicate register from R0, and treat the value of the |
674 | // predicate register as the call result. |
675 | auto &MRI = DAG.getMachineFunction().getRegInfo(); |
676 | SDValue FR0 = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(), |
677 | MVT::i32, Glue); |
678 | // FR0 = (Value, Chain, Glue) |
679 | unsigned PredR = MRI.createVirtualRegister(&Hexagon::PredRegsRegClass); |
680 | SDValue TPR = DAG.getCopyToReg(FR0.getValue(1), dl, PredR, |
681 | FR0.getValue(0), FR0.getValue(2)); |
682 | // TPR = (Chain, Glue) |
683 | // Don't glue this CopyFromReg, because it copies from a virtual |
684 | // register. If it is glued to the call, InstrEmitter will add it |
685 | // as an implicit def to the call (EmitMachineNode). |
686 | RetVal = DAG.getCopyFromReg(TPR.getValue(0), dl, PredR, MVT::i1); |
687 | Glue = TPR.getValue(1); |
688 | } else { |
689 | RetVal = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(), |
690 | RVLocs[i].getValVT(), Glue); |
691 | Glue = RetVal.getValue(2); |
692 | } |
693 | InVals.push_back(RetVal.getValue(0)); |
694 | Chain = RetVal.getValue(1); |
695 | } |
696 | |
697 | return Chain; |
698 | } |
699 | |
700 | /// LowerCall - Functions arguments are copied from virtual regs to |
701 | /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. |
702 | SDValue |
703 | HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
704 | SmallVectorImpl<SDValue> &InVals) const { |
705 | SelectionDAG &DAG = CLI.DAG; |
706 | SDLoc &dl = CLI.DL; |
707 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
708 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
709 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
710 | SDValue Chain = CLI.Chain; |
711 | SDValue Callee = CLI.Callee; |
712 | bool &IsTailCall = CLI.IsTailCall; |
713 | CallingConv::ID CallConv = CLI.CallConv; |
714 | bool IsVarArg = CLI.IsVarArg; |
715 | bool DoesNotReturn = CLI.DoesNotReturn; |
716 | |
717 | bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); |
718 | MachineFunction &MF = DAG.getMachineFunction(); |
719 | auto PtrVT = getPointerTy(MF.getDataLayout()); |
720 | |
721 | // Check for varargs. |
722 | unsigned NumNamedVarArgParams = -1U; |
723 | if (GlobalAddressSDNode *GAN = dyn_cast<GlobalAddressSDNode>(Callee)) { |
724 | const GlobalValue *GV = GAN->getGlobal(); |
725 | Callee = DAG.getTargetGlobalAddress(GV, dl, MVT::i32); |
726 | if (const Function* F = dyn_cast<Function>(GV)) { |
727 | // If a function has zero args and is a vararg function, that's |
728 | // disallowed so it must be an undeclared function. Do not assume |
729 | // varargs if the callee is undefined. |
730 | if (F->isVarArg() && F->getFunctionType()->getNumParams() != 0) |
731 | NumNamedVarArgParams = F->getFunctionType()->getNumParams(); |
732 | } |
733 | } |
734 | |
735 | // Analyze operands of the call, assigning locations to each operand. |
736 | SmallVector<CCValAssign, 16> ArgLocs; |
737 | HexagonCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, |
738 | *DAG.getContext(), NumNamedVarArgParams); |
739 | |
740 | if (IsVarArg) |
741 | CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon_VarArg); |
742 | else |
743 | CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon); |
744 | |
745 | auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls"); |
746 | if (Attr.getValueAsString() == "true") |
747 | IsTailCall = false; |
748 | |
749 | if (IsTailCall) { |
750 | bool StructAttrFlag = MF.getFunction()->hasStructRetAttr(); |
751 | IsTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, |
752 | IsVarArg, IsStructRet, |
753 | StructAttrFlag, |
754 | Outs, OutVals, Ins, DAG); |
755 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
756 | CCValAssign &VA = ArgLocs[i]; |
757 | if (VA.isMemLoc()) { |
758 | IsTailCall = false; |
759 | break; |
760 | } |
761 | } |
762 | DEBUG(dbgs() << (IsTailCall ? "Eligible for Tail Call\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { dbgs() << (IsTailCall ? "Eligible for Tail Call\n" : "Argument must be passed on stack. " "Not eligible for Tail Call\n" ); } } while (false) |
763 | : "Argument must be passed on stack. "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { dbgs() << (IsTailCall ? "Eligible for Tail Call\n" : "Argument must be passed on stack. " "Not eligible for Tail Call\n" ); } } while (false) |
764 | "Not eligible for Tail Call\n"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { dbgs() << (IsTailCall ? "Eligible for Tail Call\n" : "Argument must be passed on stack. " "Not eligible for Tail Call\n" ); } } while (false); |
765 | } |
766 | // Get a count of how many bytes are to be pushed on the stack. |
767 | unsigned NumBytes = CCInfo.getNextStackOffset(); |
768 | SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass; |
769 | SmallVector<SDValue, 8> MemOpChains; |
770 | |
771 | auto &HRI = *Subtarget.getRegisterInfo(); |
772 | SDValue StackPtr = |
773 | DAG.getCopyFromReg(Chain, dl, HRI.getStackRegister(), PtrVT); |
774 | |
775 | bool NeedsArgAlign = false; |
776 | unsigned LargestAlignSeen = 0; |
777 | // Walk the register/memloc assignments, inserting copies/loads. |
778 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
779 | CCValAssign &VA = ArgLocs[i]; |
780 | SDValue Arg = OutVals[i]; |
781 | ISD::ArgFlagsTy Flags = Outs[i].Flags; |
782 | // Record if we need > 8 byte alignment on an argument. |
783 | bool ArgAlign = isHvxVectorType(VA.getValVT()); |
784 | NeedsArgAlign |= ArgAlign; |
785 | |
786 | // Promote the value if needed. |
787 | switch (VA.getLocInfo()) { |
788 | default: |
789 | // Loc info must be one of Full, SExt, ZExt, or AExt. |
790 | llvm_unreachable("Unknown loc info!")::llvm::llvm_unreachable_internal("Unknown loc info!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 790); |
791 | case CCValAssign::BCvt: |
792 | case CCValAssign::Full: |
793 | break; |
794 | case CCValAssign::SExt: |
795 | Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); |
796 | break; |
797 | case CCValAssign::ZExt: |
798 | Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); |
799 | break; |
800 | case CCValAssign::AExt: |
801 | Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); |
802 | break; |
803 | } |
804 | |
805 | if (VA.isMemLoc()) { |
806 | unsigned LocMemOffset = VA.getLocMemOffset(); |
807 | SDValue MemAddr = DAG.getConstant(LocMemOffset, dl, |
808 | StackPtr.getValueType()); |
809 | MemAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, MemAddr); |
810 | if (ArgAlign) |
811 | LargestAlignSeen = std::max(LargestAlignSeen, |
812 | VA.getLocVT().getStoreSizeInBits() >> 3); |
813 | if (Flags.isByVal()) { |
814 | // The argument is a struct passed by value. According to LLVM, "Arg" |
815 | // is is pointer. |
816 | MemOpChains.push_back(CreateCopyOfByValArgument(Arg, MemAddr, Chain, |
817 | Flags, DAG, dl)); |
818 | } else { |
819 | MachinePointerInfo LocPI = MachinePointerInfo::getStack( |
820 | DAG.getMachineFunction(), LocMemOffset); |
821 | SDValue S = DAG.getStore(Chain, dl, Arg, MemAddr, LocPI); |
822 | MemOpChains.push_back(S); |
823 | } |
824 | continue; |
825 | } |
826 | |
827 | // Arguments that can be passed on register must be kept at RegsToPass |
828 | // vector. |
829 | if (VA.isRegLoc()) |
830 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); |
831 | } |
832 | |
833 | if (NeedsArgAlign && Subtarget.hasV60TOps()) { |
834 | DEBUG(dbgs() << "Function needs byte stack align due to call args\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { dbgs() << "Function needs byte stack align due to call args\n" ; } } while (false); |
835 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
836 | // V6 vectors passed by value have 64 or 128 byte alignment depending |
837 | // on whether we are 64 byte vector mode or 128 byte. |
838 | bool UseHVXDbl = Subtarget.useHVXDblOps(); |
839 | assert(Subtarget.useHVXOps())((Subtarget.useHVXOps()) ? static_cast<void> (0) : __assert_fail ("Subtarget.useHVXOps()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 839, __PRETTY_FUNCTION__)); |
840 | const unsigned ObjAlign = UseHVXDbl ? 128 : 64; |
841 | LargestAlignSeen = std::max(LargestAlignSeen, ObjAlign); |
842 | MFI.ensureMaxAlignment(LargestAlignSeen); |
843 | } |
844 | // Transform all store nodes into one single node because all store |
845 | // nodes are independent of each other. |
846 | if (!MemOpChains.empty()) |
847 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains); |
848 | |
849 | SDValue Glue; |
850 | if (!IsTailCall) { |
851 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, dl); |
852 | Glue = Chain.getValue(1); |
853 | } |
854 | |
855 | // Build a sequence of copy-to-reg nodes chained together with token |
856 | // chain and flag operands which copy the outgoing args into registers. |
857 | // The Glue is necessary since all emitted instructions must be |
858 | // stuck together. |
859 | if (!IsTailCall) { |
860 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
861 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
862 | RegsToPass[i].second, Glue); |
863 | Glue = Chain.getValue(1); |
864 | } |
865 | } else { |
866 | // For tail calls lower the arguments to the 'real' stack slot. |
867 | // |
868 | // Force all the incoming stack arguments to be loaded from the stack |
869 | // before any new outgoing arguments are stored to the stack, because the |
870 | // outgoing stack slots may alias the incoming argument stack slots, and |
871 | // the alias isn't otherwise explicit. This is slightly more conservative |
872 | // than necessary, because it means that each store effectively depends |
873 | // on every argument instead of just those arguments it would clobber. |
874 | // |
875 | // Do not flag preceding copytoreg stuff together with the following stuff. |
876 | Glue = SDValue(); |
877 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
878 | Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, |
879 | RegsToPass[i].second, Glue); |
880 | Glue = Chain.getValue(1); |
881 | } |
882 | Glue = SDValue(); |
883 | } |
884 | |
885 | bool LongCalls = MF.getSubtarget<HexagonSubtarget>().useLongCalls(); |
886 | unsigned Flags = LongCalls ? HexagonII::HMOTF_ConstExtended : 0; |
887 | |
888 | // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every |
889 | // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol |
890 | // node so that legalize doesn't hack it. |
891 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { |
892 | Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, PtrVT, 0, Flags); |
893 | } else if (ExternalSymbolSDNode *S = |
894 | dyn_cast<ExternalSymbolSDNode>(Callee)) { |
895 | Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, Flags); |
896 | } |
897 | |
898 | // Returns a chain & a flag for retval copy to use. |
899 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
900 | SmallVector<SDValue, 8> Ops; |
901 | Ops.push_back(Chain); |
902 | Ops.push_back(Callee); |
903 | |
904 | // Add argument registers to the end of the list so that they are |
905 | // known live into the call. |
906 | for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { |
907 | Ops.push_back(DAG.getRegister(RegsToPass[i].first, |
908 | RegsToPass[i].second.getValueType())); |
909 | } |
910 | |
911 | const uint32_t *Mask = HRI.getCallPreservedMask(MF, CallConv); |
912 | 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\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 912, __PRETTY_FUNCTION__)); |
913 | Ops.push_back(DAG.getRegisterMask(Mask)); |
914 | |
915 | if (Glue.getNode()) |
916 | Ops.push_back(Glue); |
917 | |
918 | if (IsTailCall) { |
919 | MF.getFrameInfo().setHasTailCall(); |
920 | return DAG.getNode(HexagonISD::TC_RETURN, dl, NodeTys, Ops); |
921 | } |
922 | |
923 | unsigned OpCode = DoesNotReturn ? HexagonISD::CALLnr : HexagonISD::CALL; |
924 | Chain = DAG.getNode(OpCode, dl, NodeTys, Ops); |
925 | Glue = Chain.getValue(1); |
926 | |
927 | // Create the CALLSEQ_END node. |
928 | Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true), |
929 | DAG.getIntPtrConstant(0, dl, true), Glue, dl); |
930 | Glue = Chain.getValue(1); |
931 | |
932 | // Handle result values, copying them out of physregs into vregs that we |
933 | // return. |
934 | return LowerCallResult(Chain, Glue, CallConv, IsVarArg, Ins, dl, DAG, |
935 | InVals, OutVals, Callee); |
936 | } |
937 | |
938 | static bool getIndexedAddressParts(SDNode *Ptr, EVT VT, |
939 | SDValue &Base, SDValue &Offset, |
940 | bool &IsInc, SelectionDAG &DAG) { |
941 | if (Ptr->getOpcode() != ISD::ADD) |
942 | return false; |
943 | |
944 | auto &HST = static_cast<const HexagonSubtarget&>(DAG.getSubtarget()); |
945 | bool UseHVX = HST.useHVXOps(); |
946 | bool UseHVXDbl = HST.useHVXDblOps(); |
947 | |
948 | bool ValidHVXDblType = |
949 | (UseHVX && UseHVXDbl) && (VT == MVT::v32i32 || VT == MVT::v16i64 || |
950 | VT == MVT::v64i16 || VT == MVT::v128i8); |
951 | bool ValidHVXType = |
952 | UseHVX && !UseHVXDbl && (VT == MVT::v16i32 || VT == MVT::v8i64 || |
953 | VT == MVT::v32i16 || VT == MVT::v64i8); |
954 | |
955 | if (ValidHVXDblType || ValidHVXType || |
956 | VT == MVT::i64 || VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) { |
957 | IsInc = (Ptr->getOpcode() == ISD::ADD); |
958 | Base = Ptr->getOperand(0); |
959 | Offset = Ptr->getOperand(1); |
960 | // Ensure that Offset is a constant. |
961 | return isa<ConstantSDNode>(Offset); |
962 | } |
963 | |
964 | return false; |
965 | } |
966 | |
967 | /// getPostIndexedAddressParts - returns true by value, base pointer and |
968 | /// offset pointer and addressing mode by reference if this node can be |
969 | /// combined with a load / store to form a post-indexed load / store. |
970 | bool HexagonTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, |
971 | SDValue &Base, |
972 | SDValue &Offset, |
973 | ISD::MemIndexedMode &AM, |
974 | SelectionDAG &DAG) const |
975 | { |
976 | EVT VT; |
977 | SDValue Ptr; |
978 | |
979 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { |
980 | VT = LD->getMemoryVT(); |
981 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { |
982 | VT = ST->getMemoryVT(); |
983 | if (ST->getValue().getValueType() == MVT::i64 && ST->isTruncatingStore()) |
984 | return false; |
985 | } else { |
986 | return false; |
987 | } |
988 | |
989 | bool IsInc = false; |
990 | bool isLegal = getIndexedAddressParts(Op, VT, Base, Offset, IsInc, DAG); |
991 | if (isLegal) { |
992 | auto &HII = *Subtarget.getInstrInfo(); |
993 | int32_t OffsetVal = cast<ConstantSDNode>(Offset.getNode())->getSExtValue(); |
994 | if (HII.isValidAutoIncImm(VT, OffsetVal)) { |
995 | AM = IsInc ? ISD::POST_INC : ISD::POST_DEC; |
996 | return true; |
997 | } |
998 | } |
999 | |
1000 | return false; |
1001 | } |
1002 | |
1003 | SDValue |
1004 | HexagonTargetLowering::LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const { |
1005 | SDNode *Node = Op.getNode(); |
1006 | MachineFunction &MF = DAG.getMachineFunction(); |
1007 | auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>(); |
1008 | switch (Node->getOpcode()) { |
1009 | case ISD::INLINEASM: { |
1010 | unsigned NumOps = Node->getNumOperands(); |
1011 | if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue) |
1012 | --NumOps; // Ignore the flag operand. |
1013 | |
1014 | for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) { |
1015 | if (FuncInfo.hasClobberLR()) |
1016 | break; |
1017 | unsigned Flags = |
1018 | cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue(); |
1019 | unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags); |
1020 | ++i; // Skip the ID value. |
1021 | |
1022 | switch (InlineAsm::getKind(Flags)) { |
1023 | default: llvm_unreachable("Bad flags!")::llvm::llvm_unreachable_internal("Bad flags!", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1023); |
1024 | case InlineAsm::Kind_RegDef: |
1025 | case InlineAsm::Kind_RegUse: |
1026 | case InlineAsm::Kind_Imm: |
1027 | case InlineAsm::Kind_Clobber: |
1028 | case InlineAsm::Kind_Mem: { |
1029 | for (; NumVals; --NumVals, ++i) {} |
1030 | break; |
1031 | } |
1032 | case InlineAsm::Kind_RegDefEarlyClobber: { |
1033 | for (; NumVals; --NumVals, ++i) { |
1034 | unsigned Reg = |
1035 | cast<RegisterSDNode>(Node->getOperand(i))->getReg(); |
1036 | |
1037 | // Check it to be lr |
1038 | const HexagonRegisterInfo *QRI = Subtarget.getRegisterInfo(); |
1039 | if (Reg == QRI->getRARegister()) { |
1040 | FuncInfo.setHasClobberLR(true); |
1041 | break; |
1042 | } |
1043 | } |
1044 | break; |
1045 | } |
1046 | } |
1047 | } |
1048 | } |
1049 | } // Node->getOpcode |
1050 | return Op; |
1051 | } |
1052 | |
1053 | // Need to transform ISD::PREFETCH into something that doesn't inherit |
1054 | // all of the properties of ISD::PREFETCH, specifically SDNPMayLoad and |
1055 | // SDNPMayStore. |
1056 | SDValue HexagonTargetLowering::LowerPREFETCH(SDValue Op, |
1057 | SelectionDAG &DAG) const { |
1058 | SDValue Chain = Op.getOperand(0); |
1059 | SDValue Addr = Op.getOperand(1); |
1060 | // Lower it to DCFETCH($reg, #0). A "pat" will try to merge the offset in, |
1061 | // if the "reg" is fed by an "add". |
1062 | SDLoc DL(Op); |
1063 | SDValue Zero = DAG.getConstant(0, DL, MVT::i32); |
1064 | return DAG.getNode(HexagonISD::DCFETCH, DL, MVT::Other, Chain, Addr, Zero); |
1065 | } |
1066 | |
1067 | // Custom-handle ISD::READCYCLECOUNTER because the target-independent SDNode |
1068 | // is marked as having side-effects, while the register read on Hexagon does |
1069 | // not have any. TableGen refuses to accept the direct pattern from that node |
1070 | // to the A4_tfrcpp. |
1071 | SDValue HexagonTargetLowering::LowerREADCYCLECOUNTER(SDValue Op, |
1072 | SelectionDAG &DAG) const { |
1073 | SDValue Chain = Op.getOperand(0); |
1074 | SDLoc dl(Op); |
1075 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); |
1076 | return DAG.getNode(HexagonISD::READCYCLE, dl, VTs, Chain); |
1077 | } |
1078 | |
1079 | SDValue HexagonTargetLowering::LowerINTRINSIC_VOID(SDValue Op, |
1080 | SelectionDAG &DAG) const { |
1081 | SDValue Chain = Op.getOperand(0); |
1082 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); |
1083 | // Lower the hexagon_prefetch builtin to DCFETCH, as above. |
1084 | if (IntNo == Intrinsic::hexagon_prefetch) { |
1085 | SDValue Addr = Op.getOperand(2); |
1086 | SDLoc DL(Op); |
1087 | SDValue Zero = DAG.getConstant(0, DL, MVT::i32); |
1088 | return DAG.getNode(HexagonISD::DCFETCH, DL, MVT::Other, Chain, Addr, Zero); |
1089 | } |
1090 | return SDValue(); |
1091 | } |
1092 | |
1093 | SDValue |
1094 | HexagonTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, |
1095 | SelectionDAG &DAG) const { |
1096 | SDValue Chain = Op.getOperand(0); |
1097 | SDValue Size = Op.getOperand(1); |
1098 | SDValue Align = Op.getOperand(2); |
1099 | SDLoc dl(Op); |
1100 | |
1101 | ConstantSDNode *AlignConst = dyn_cast<ConstantSDNode>(Align); |
1102 | assert(AlignConst && "Non-constant Align in LowerDYNAMIC_STACKALLOC")((AlignConst && "Non-constant Align in LowerDYNAMIC_STACKALLOC" ) ? static_cast<void> (0) : __assert_fail ("AlignConst && \"Non-constant Align in LowerDYNAMIC_STACKALLOC\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1102, __PRETTY_FUNCTION__)); |
1103 | |
1104 | unsigned A = AlignConst->getSExtValue(); |
1105 | auto &HFI = *Subtarget.getFrameLowering(); |
1106 | // "Zero" means natural stack alignment. |
1107 | if (A == 0) |
1108 | A = HFI.getStackAlignment(); |
1109 | |
1110 | DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { { dbgs () << __func__ << " Align: " << A << " Size: "; Size.getNode()->dump(& DAG); dbgs() << "\n"; }; } } while (false) |
1111 | dbgs () << __func__ << " Align: " << A << " Size: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { { dbgs () << __func__ << " Align: " << A << " Size: "; Size.getNode()->dump(& DAG); dbgs() << "\n"; }; } } while (false) |
1112 | Size.getNode()->dump(&DAG);do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { { dbgs () << __func__ << " Align: " << A << " Size: "; Size.getNode()->dump(& DAG); dbgs() << "\n"; }; } } while (false) |
1113 | dbgs() << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { { dbgs () << __func__ << " Align: " << A << " Size: "; Size.getNode()->dump(& DAG); dbgs() << "\n"; }; } } while (false) |
1114 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("hexagon-lowering")) { { dbgs () << __func__ << " Align: " << A << " Size: "; Size.getNode()->dump(& DAG); dbgs() << "\n"; }; } } while (false); |
1115 | |
1116 | SDValue AC = DAG.getConstant(A, dl, MVT::i32); |
1117 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other); |
1118 | SDValue AA = DAG.getNode(HexagonISD::ALLOCA, dl, VTs, Chain, Size, AC); |
1119 | |
1120 | DAG.ReplaceAllUsesOfValueWith(Op, AA); |
1121 | return AA; |
1122 | } |
1123 | |
1124 | SDValue HexagonTargetLowering::LowerFormalArguments( |
1125 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1126 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1127 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1128 | MachineFunction &MF = DAG.getMachineFunction(); |
1129 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1130 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
1131 | auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>(); |
1132 | |
1133 | // Assign locations to all of the incoming arguments. |
1134 | SmallVector<CCValAssign, 16> ArgLocs; |
1135 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
1136 | *DAG.getContext()); |
1137 | |
1138 | CCInfo.AnalyzeFormalArguments(Ins, CC_Hexagon); |
1139 | |
1140 | // For LLVM, in the case when returning a struct by value (>8byte), |
1141 | // the first argument is a pointer that points to the location on caller's |
1142 | // stack where the return value will be stored. For Hexagon, the location on |
1143 | // caller's stack is passed only when the struct size is smaller than (and |
1144 | // equal to) 8 bytes. If not, no address will be passed into callee and |
1145 | // callee return the result direclty through R0/R1. |
1146 | |
1147 | SmallVector<SDValue, 8> MemOps; |
1148 | bool UseHVX = Subtarget.useHVXOps(), UseHVXDbl = Subtarget.useHVXDblOps(); |
1149 | |
1150 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1151 | CCValAssign &VA = ArgLocs[i]; |
1152 | ISD::ArgFlagsTy Flags = Ins[i].Flags; |
1153 | unsigned ObjSize; |
1154 | unsigned StackLocation; |
1155 | int FI; |
1156 | |
1157 | if ( (VA.isRegLoc() && !Flags.isByVal()) |
1158 | || (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() > 8)) { |
1159 | // Arguments passed in registers |
1160 | // 1. int, long long, ptr args that get allocated in register. |
1161 | // 2. Large struct that gets an register to put its address in. |
1162 | EVT RegVT = VA.getLocVT(); |
1163 | if (RegVT == MVT::i8 || RegVT == MVT::i16 || |
1164 | RegVT == MVT::i32 || RegVT == MVT::f32) { |
1165 | unsigned VReg = |
1166 | RegInfo.createVirtualRegister(&Hexagon::IntRegsRegClass); |
1167 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1168 | SDValue Copy = DAG.getCopyFromReg(Chain, dl, VReg, RegVT); |
1169 | // Treat values of type MVT::i1 specially: they are passed in |
1170 | // registers of type i32, but they need to remain as values of |
1171 | // type i1 for consistency of the argument lowering. |
1172 | if (VA.getValVT() == MVT::i1) { |
1173 | // Generate a copy into a predicate register and use the value |
1174 | // of the register as the "InVal". |
1175 | unsigned PReg = |
1176 | RegInfo.createVirtualRegister(&Hexagon::PredRegsRegClass); |
1177 | SDNode *T = DAG.getMachineNode(Hexagon::C2_tfrrp, dl, MVT::i1, |
1178 | Copy.getValue(0)); |
1179 | Copy = DAG.getCopyToReg(Copy.getValue(1), dl, PReg, SDValue(T, 0)); |
1180 | Copy = DAG.getCopyFromReg(Copy, dl, PReg, MVT::i1); |
1181 | } |
1182 | InVals.push_back(Copy); |
1183 | Chain = Copy.getValue(1); |
1184 | } else if (RegVT == MVT::i64 || RegVT == MVT::f64) { |
1185 | unsigned VReg = |
1186 | RegInfo.createVirtualRegister(&Hexagon::DoubleRegsRegClass); |
1187 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1188 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1189 | |
1190 | // Single Vector |
1191 | } else if ((RegVT == MVT::v8i64 || RegVT == MVT::v16i32 || |
1192 | RegVT == MVT::v32i16 || RegVT == MVT::v64i8)) { |
1193 | unsigned VReg = |
1194 | RegInfo.createVirtualRegister(&Hexagon::VectorRegsRegClass); |
1195 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1196 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1197 | } else if (UseHVX && UseHVXDbl && |
1198 | ((RegVT == MVT::v16i64 || RegVT == MVT::v32i32 || |
1199 | RegVT == MVT::v64i16 || RegVT == MVT::v128i8))) { |
1200 | unsigned VReg = |
1201 | RegInfo.createVirtualRegister(&Hexagon::VectorRegs128BRegClass); |
1202 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1203 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1204 | |
1205 | // Double Vector |
1206 | } else if ((RegVT == MVT::v16i64 || RegVT == MVT::v32i32 || |
1207 | RegVT == MVT::v64i16 || RegVT == MVT::v128i8)) { |
1208 | unsigned VReg = |
1209 | RegInfo.createVirtualRegister(&Hexagon::VecDblRegsRegClass); |
1210 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1211 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1212 | } else if (UseHVX && UseHVXDbl && |
1213 | ((RegVT == MVT::v32i64 || RegVT == MVT::v64i32 || |
1214 | RegVT == MVT::v128i16 || RegVT == MVT::v256i8))) { |
1215 | unsigned VReg = |
1216 | RegInfo.createVirtualRegister(&Hexagon::VecDblRegs128BRegClass); |
1217 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1218 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1219 | } else if (RegVT == MVT::v512i1 || RegVT == MVT::v1024i1) { |
1220 | assert(0 && "need to support VecPred regs")((0 && "need to support VecPred regs") ? static_cast< void> (0) : __assert_fail ("0 && \"need to support VecPred regs\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1220, __PRETTY_FUNCTION__)); |
1221 | unsigned VReg = |
1222 | RegInfo.createVirtualRegister(&Hexagon::VecPredRegsRegClass); |
1223 | RegInfo.addLiveIn(VA.getLocReg(), VReg); |
1224 | InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT)); |
1225 | } else { |
1226 | assert (0)((0) ? static_cast<void> (0) : __assert_fail ("0", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1226, __PRETTY_FUNCTION__)); |
1227 | } |
1228 | } else if (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() <= 8) { |
1229 | assert (0 && "ByValSize must be bigger than 8 bytes")((0 && "ByValSize must be bigger than 8 bytes") ? static_cast <void> (0) : __assert_fail ("0 && \"ByValSize must be bigger than 8 bytes\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1229, __PRETTY_FUNCTION__)); |
1230 | } else { |
1231 | // Sanity check. |
1232 | assert(VA.isMemLoc())((VA.isMemLoc()) ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc()", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1232, __PRETTY_FUNCTION__)); |
1233 | |
1234 | if (Flags.isByVal()) { |
1235 | // If it's a byval parameter, then we need to compute the |
1236 | // "real" size, not the size of the pointer. |
1237 | ObjSize = Flags.getByValSize(); |
1238 | } else { |
1239 | ObjSize = VA.getLocVT().getStoreSizeInBits() >> 3; |
1240 | } |
1241 | |
1242 | StackLocation = HEXAGON_LRFP_SIZE8 + VA.getLocMemOffset(); |
1243 | // Create the frame index object for this incoming parameter... |
1244 | FI = MFI.CreateFixedObject(ObjSize, StackLocation, true); |
1245 | |
1246 | // Create the SelectionDAG nodes cordl, responding to a load |
1247 | // from this parameter. |
1248 | SDValue FIN = DAG.getFrameIndex(FI, MVT::i32); |
1249 | |
1250 | if (Flags.isByVal()) { |
1251 | // If it's a pass-by-value aggregate, then do not dereference the stack |
1252 | // location. Instead, we should generate a reference to the stack |
1253 | // location. |
1254 | InVals.push_back(FIN); |
1255 | } else { |
1256 | InVals.push_back( |
1257 | DAG.getLoad(VA.getValVT(), dl, Chain, FIN, MachinePointerInfo())); |
1258 | } |
1259 | } |
1260 | } |
1261 | |
1262 | if (!MemOps.empty()) |
1263 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps); |
1264 | |
1265 | if (isVarArg) { |
1266 | // This will point to the next argument passed via stack. |
1267 | int FrameIndex = MFI.CreateFixedObject(Hexagon_PointerSize(4), |
1268 | HEXAGON_LRFP_SIZE8 + |
1269 | CCInfo.getNextStackOffset(), |
1270 | true); |
1271 | FuncInfo.setVarArgsFrameIndex(FrameIndex); |
1272 | } |
1273 | |
1274 | return Chain; |
1275 | } |
1276 | |
1277 | SDValue |
1278 | HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
1279 | // VASTART stores the address of the VarArgsFrameIndex slot into the |
1280 | // memory location argument. |
1281 | MachineFunction &MF = DAG.getMachineFunction(); |
1282 | HexagonMachineFunctionInfo *QFI = MF.getInfo<HexagonMachineFunctionInfo>(); |
1283 | SDValue Addr = DAG.getFrameIndex(QFI->getVarArgsFrameIndex(), MVT::i32); |
1284 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); |
1285 | return DAG.getStore(Op.getOperand(0), SDLoc(Op), Addr, Op.getOperand(1), |
1286 | MachinePointerInfo(SV)); |
1287 | } |
1288 | |
1289 | // Creates a SPLAT instruction for a constant value VAL. |
1290 | static SDValue createSplat(SelectionDAG &DAG, const SDLoc &dl, EVT VT, |
1291 | SDValue Val) { |
1292 | if (VT.getSimpleVT() == MVT::v4i8) |
1293 | return DAG.getNode(HexagonISD::VSPLATB, dl, VT, Val); |
1294 | |
1295 | if (VT.getSimpleVT() == MVT::v4i16) |
1296 | return DAG.getNode(HexagonISD::VSPLATH, dl, VT, Val); |
1297 | |
1298 | return SDValue(); |
1299 | } |
1300 | |
1301 | static bool isSExtFree(SDValue N) { |
1302 | // A sign-extend of a truncate of a sign-extend is free. |
1303 | if (N.getOpcode() == ISD::TRUNCATE && |
1304 | N.getOperand(0).getOpcode() == ISD::AssertSext) |
1305 | return true; |
1306 | // We have sign-extended loads. |
1307 | if (N.getOpcode() == ISD::LOAD) |
1308 | return true; |
1309 | return false; |
1310 | } |
1311 | |
1312 | SDValue HexagonTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
1313 | SDLoc dl(Op); |
1314 | |
1315 | SDValue LHS = Op.getOperand(0); |
1316 | SDValue RHS = Op.getOperand(1); |
1317 | SDValue Cmp = Op.getOperand(2); |
1318 | ISD::CondCode CC = cast<CondCodeSDNode>(Cmp)->get(); |
1319 | |
1320 | EVT VT = Op.getValueType(); |
1321 | EVT LHSVT = LHS.getValueType(); |
1322 | EVT RHSVT = RHS.getValueType(); |
1323 | |
1324 | if (LHSVT == MVT::v2i16) { |
1325 | assert(ISD::isSignedIntSetCC(CC) || ISD::isUnsignedIntSetCC(CC))((ISD::isSignedIntSetCC(CC) || ISD::isUnsignedIntSetCC(CC)) ? static_cast<void> (0) : __assert_fail ("ISD::isSignedIntSetCC(CC) || ISD::isUnsignedIntSetCC(CC)" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1325, __PRETTY_FUNCTION__)); |
1326 | unsigned ExtOpc = ISD::isSignedIntSetCC(CC) ? ISD::SIGN_EXTEND |
1327 | : ISD::ZERO_EXTEND; |
1328 | SDValue LX = DAG.getNode(ExtOpc, dl, MVT::v2i32, LHS); |
1329 | SDValue RX = DAG.getNode(ExtOpc, dl, MVT::v2i32, RHS); |
1330 | SDValue SC = DAG.getNode(ISD::SETCC, dl, MVT::v2i1, LX, RX, Cmp); |
1331 | return SC; |
1332 | } |
1333 | |
1334 | // Treat all other vector types as legal. |
1335 | if (VT.isVector()) |
1336 | return Op; |
1337 | |
1338 | // Equals and not equals should use sign-extend, not zero-extend, since |
1339 | // we can represent small negative values in the compare instructions. |
1340 | // The LLVM default is to use zero-extend arbitrarily in these cases. |
1341 | if ((CC == ISD::SETEQ || CC == ISD::SETNE) && |
1342 | (RHSVT == MVT::i8 || RHSVT == MVT::i16) && |
1343 | (LHSVT == MVT::i8 || LHSVT == MVT::i16)) { |
1344 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS); |
1345 | if (C && C->getAPIntValue().isNegative()) { |
1346 | LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS); |
1347 | RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS); |
1348 | return DAG.getNode(ISD::SETCC, dl, Op.getValueType(), |
1349 | LHS, RHS, Op.getOperand(2)); |
1350 | } |
1351 | if (isSExtFree(LHS) || isSExtFree(RHS)) { |
1352 | LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS); |
1353 | RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS); |
1354 | return DAG.getNode(ISD::SETCC, dl, Op.getValueType(), |
1355 | LHS, RHS, Op.getOperand(2)); |
1356 | } |
1357 | } |
1358 | return SDValue(); |
1359 | } |
1360 | |
1361 | SDValue |
1362 | HexagonTargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const { |
1363 | SDValue PredOp = Op.getOperand(0); |
1364 | SDValue Op1 = Op.getOperand(1), Op2 = Op.getOperand(2); |
1365 | EVT OpVT = Op1.getValueType(); |
1366 | SDLoc DL(Op); |
1367 | |
1368 | if (OpVT == MVT::v2i16) { |
1369 | SDValue X1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op1); |
1370 | SDValue X2 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op2); |
1371 | SDValue SL = DAG.getNode(ISD::VSELECT, DL, MVT::v2i32, PredOp, X1, X2); |
1372 | SDValue TR = DAG.getNode(ISD::TRUNCATE, DL, MVT::v2i16, SL); |
1373 | return TR; |
1374 | } |
1375 | |
1376 | return SDValue(); |
1377 | } |
1378 | |
1379 | // Handle only specific vector loads. |
1380 | SDValue HexagonTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
1381 | EVT VT = Op.getValueType(); |
1382 | SDLoc DL(Op); |
1383 | LoadSDNode *LoadNode = cast<LoadSDNode>(Op); |
1384 | SDValue Chain = LoadNode->getChain(); |
1385 | SDValue Ptr = Op.getOperand(1); |
1386 | SDValue LoweredLoad; |
1387 | SDValue Result; |
1388 | SDValue Base = LoadNode->getBasePtr(); |
1389 | ISD::LoadExtType Ext = LoadNode->getExtensionType(); |
1390 | unsigned Alignment = LoadNode->getAlignment(); |
1391 | SDValue LoadChain; |
1392 | |
1393 | if(Ext == ISD::NON_EXTLOAD) |
1394 | Ext = ISD::ZEXTLOAD; |
1395 | |
1396 | if (VT == MVT::v4i16) { |
1397 | if (Alignment == 2) { |
1398 | SDValue Loads[4]; |
1399 | // Base load. |
1400 | Loads[0] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Base, |
1401 | LoadNode->getPointerInfo(), MVT::i16, Alignment, |
1402 | LoadNode->getMemOperand()->getFlags()); |
1403 | // Base+2 load. |
1404 | SDValue Increment = DAG.getConstant(2, DL, MVT::i32); |
1405 | Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment); |
1406 | Loads[1] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr, |
1407 | LoadNode->getPointerInfo(), MVT::i16, Alignment, |
1408 | LoadNode->getMemOperand()->getFlags()); |
1409 | // SHL 16, then OR base and base+2. |
1410 | SDValue ShiftAmount = DAG.getConstant(16, DL, MVT::i32); |
1411 | SDValue Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[1], ShiftAmount); |
1412 | SDValue Tmp2 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[0]); |
1413 | // Base + 4. |
1414 | Increment = DAG.getConstant(4, DL, MVT::i32); |
1415 | Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment); |
1416 | Loads[2] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr, |
1417 | LoadNode->getPointerInfo(), MVT::i16, Alignment, |
1418 | LoadNode->getMemOperand()->getFlags()); |
1419 | // Base + 6. |
1420 | Increment = DAG.getConstant(6, DL, MVT::i32); |
1421 | Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment); |
1422 | Loads[3] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr, |
1423 | LoadNode->getPointerInfo(), MVT::i16, Alignment, |
1424 | LoadNode->getMemOperand()->getFlags()); |
1425 | // SHL 16, then OR base+4 and base+6. |
1426 | Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[3], ShiftAmount); |
1427 | SDValue Tmp4 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[2]); |
1428 | // Combine to i64. This could be optimised out later if we can |
1429 | // affect reg allocation of this code. |
1430 | Result = DAG.getNode(HexagonISD::COMBINE, DL, MVT::i64, Tmp4, Tmp2); |
1431 | LoadChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, |
1432 | Loads[0].getValue(1), Loads[1].getValue(1), |
1433 | Loads[2].getValue(1), Loads[3].getValue(1)); |
1434 | } else { |
1435 | // Perform default type expansion. |
1436 | Result = DAG.getLoad(MVT::i64, DL, Chain, Ptr, LoadNode->getPointerInfo(), |
1437 | LoadNode->getAlignment(), |
1438 | LoadNode->getMemOperand()->getFlags()); |
1439 | LoadChain = Result.getValue(1); |
1440 | } |
1441 | } else |
1442 | llvm_unreachable("Custom lowering unsupported load")::llvm::llvm_unreachable_internal("Custom lowering unsupported load" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1442); |
1443 | |
1444 | Result = DAG.getNode(ISD::BITCAST, DL, VT, Result); |
1445 | // Since we pretend to lower a load, we need the original chain |
1446 | // info attached to the result. |
1447 | SDValue Ops[] = { Result, LoadChain }; |
1448 | |
1449 | return DAG.getMergeValues(Ops, DL); |
1450 | } |
1451 | |
1452 | SDValue |
1453 | HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { |
1454 | EVT ValTy = Op.getValueType(); |
1455 | ConstantPoolSDNode *CPN = cast<ConstantPoolSDNode>(Op); |
1456 | unsigned Align = CPN->getAlignment(); |
1457 | bool IsPositionIndependent = isPositionIndependent(); |
1458 | unsigned char TF = IsPositionIndependent ? HexagonII::MO_PCREL : 0; |
1459 | |
1460 | unsigned Offset = 0; |
1461 | SDValue T; |
1462 | if (CPN->isMachineConstantPoolEntry()) |
1463 | T = DAG.getTargetConstantPool(CPN->getMachineCPVal(), ValTy, Align, Offset, |
1464 | TF); |
1465 | else |
1466 | T = DAG.getTargetConstantPool(CPN->getConstVal(), ValTy, Align, Offset, |
1467 | TF); |
1468 | |
1469 | assert(cast<ConstantPoolSDNode>(T)->getTargetFlags() == TF &&((cast<ConstantPoolSDNode>(T)->getTargetFlags() == TF && "Inconsistent target flag encountered") ? static_cast <void> (0) : __assert_fail ("cast<ConstantPoolSDNode>(T)->getTargetFlags() == TF && \"Inconsistent target flag encountered\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1470, __PRETTY_FUNCTION__)) |
1470 | "Inconsistent target flag encountered")((cast<ConstantPoolSDNode>(T)->getTargetFlags() == TF && "Inconsistent target flag encountered") ? static_cast <void> (0) : __assert_fail ("cast<ConstantPoolSDNode>(T)->getTargetFlags() == TF && \"Inconsistent target flag encountered\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1470, __PRETTY_FUNCTION__)); |
1471 | |
1472 | if (IsPositionIndependent) |
1473 | return DAG.getNode(HexagonISD::AT_PCREL, SDLoc(Op), ValTy, T); |
1474 | return DAG.getNode(HexagonISD::CP, SDLoc(Op), ValTy, T); |
1475 | } |
1476 | |
1477 | SDValue |
1478 | HexagonTargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const { |
1479 | EVT VT = Op.getValueType(); |
1480 | int Idx = cast<JumpTableSDNode>(Op)->getIndex(); |
1481 | if (isPositionIndependent()) { |
1482 | SDValue T = DAG.getTargetJumpTable(Idx, VT, HexagonII::MO_PCREL); |
1483 | return DAG.getNode(HexagonISD::AT_PCREL, SDLoc(Op), VT, T); |
1484 | } |
1485 | |
1486 | SDValue T = DAG.getTargetJumpTable(Idx, VT); |
1487 | return DAG.getNode(HexagonISD::JT, SDLoc(Op), VT, T); |
1488 | } |
1489 | |
1490 | SDValue |
1491 | HexagonTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const { |
1492 | const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); |
1493 | MachineFunction &MF = DAG.getMachineFunction(); |
1494 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1495 | MFI.setReturnAddressIsTaken(true); |
1496 | |
1497 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
1498 | return SDValue(); |
1499 | |
1500 | EVT VT = Op.getValueType(); |
1501 | SDLoc dl(Op); |
1502 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
1503 | if (Depth) { |
1504 | SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); |
1505 | SDValue Offset = DAG.getConstant(4, dl, MVT::i32); |
1506 | return DAG.getLoad(VT, dl, DAG.getEntryNode(), |
1507 | DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset), |
1508 | MachinePointerInfo()); |
1509 | } |
1510 | |
1511 | // Return LR, which contains the return address. Mark it an implicit live-in. |
1512 | unsigned Reg = MF.addLiveIn(HRI.getRARegister(), getRegClassFor(MVT::i32)); |
1513 | return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT); |
1514 | } |
1515 | |
1516 | SDValue |
1517 | HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { |
1518 | const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo(); |
1519 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
1520 | MFI.setFrameAddressIsTaken(true); |
1521 | |
1522 | EVT VT = Op.getValueType(); |
1523 | SDLoc dl(Op); |
1524 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); |
1525 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, |
1526 | HRI.getFrameRegister(), VT); |
1527 | while (Depth--) |
1528 | FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, |
1529 | MachinePointerInfo()); |
1530 | return FrameAddr; |
1531 | } |
1532 | |
1533 | SDValue |
1534 | HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const { |
1535 | SDLoc dl(Op); |
1536 | return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0)); |
1537 | } |
1538 | |
1539 | SDValue |
1540 | HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op, SelectionDAG &DAG) const { |
1541 | SDLoc dl(Op); |
1542 | auto *GAN = cast<GlobalAddressSDNode>(Op); |
1543 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
1544 | auto *GV = GAN->getGlobal(); |
1545 | int64_t Offset = GAN->getOffset(); |
1546 | |
1547 | auto &HLOF = *HTM.getObjFileLowering(); |
1548 | Reloc::Model RM = HTM.getRelocationModel(); |
1549 | |
1550 | if (RM == Reloc::Static) { |
1551 | SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset); |
1552 | const GlobalObject *GO = GV->getBaseObject(); |
1553 | if (GO && HLOF.isGlobalInSmallSection(GO, HTM)) |
1554 | return DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, GA); |
1555 | return DAG.getNode(HexagonISD::CONST32, dl, PtrVT, GA); |
1556 | } |
1557 | |
1558 | bool UsePCRel = getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV); |
1559 | if (UsePCRel) { |
1560 | SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset, |
1561 | HexagonII::MO_PCREL); |
1562 | return DAG.getNode(HexagonISD::AT_PCREL, dl, PtrVT, GA); |
1563 | } |
1564 | |
1565 | // Use GOT index. |
1566 | SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); |
1567 | SDValue GA = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, HexagonII::MO_GOT); |
1568 | SDValue Off = DAG.getConstant(Offset, dl, MVT::i32); |
1569 | return DAG.getNode(HexagonISD::AT_GOT, dl, PtrVT, GOT, GA, Off); |
1570 | } |
1571 | |
1572 | // Specifies that for loads and stores VT can be promoted to PromotedLdStVT. |
1573 | SDValue |
1574 | HexagonTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { |
1575 | const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); |
1576 | SDLoc dl(Op); |
1577 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
1578 | |
1579 | Reloc::Model RM = HTM.getRelocationModel(); |
1580 | if (RM == Reloc::Static) { |
1581 | SDValue A = DAG.getTargetBlockAddress(BA, PtrVT); |
1582 | return DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, A); |
1583 | } |
1584 | |
1585 | SDValue A = DAG.getTargetBlockAddress(BA, PtrVT, 0, HexagonII::MO_PCREL); |
1586 | return DAG.getNode(HexagonISD::AT_PCREL, dl, PtrVT, A); |
1587 | } |
1588 | |
1589 | SDValue |
1590 | HexagonTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, SelectionDAG &DAG) |
1591 | const { |
1592 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); |
1593 | SDValue GOTSym = DAG.getTargetExternalSymbol(HEXAGON_GOT_SYM_NAME"_GLOBAL_OFFSET_TABLE_", PtrVT, |
1594 | HexagonII::MO_PCREL); |
1595 | return DAG.getNode(HexagonISD::AT_PCREL, SDLoc(Op), PtrVT, GOTSym); |
1596 | } |
1597 | |
1598 | SDValue |
1599 | HexagonTargetLowering::GetDynamicTLSAddr(SelectionDAG &DAG, SDValue Chain, |
1600 | GlobalAddressSDNode *GA, SDValue Glue, EVT PtrVT, unsigned ReturnReg, |
1601 | unsigned char OperandFlags) const { |
1602 | MachineFunction &MF = DAG.getMachineFunction(); |
1603 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1604 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
1605 | SDLoc dl(GA); |
1606 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, |
1607 | GA->getValueType(0), |
1608 | GA->getOffset(), |
1609 | OperandFlags); |
1610 | // Create Operands for the call.The Operands should have the following: |
1611 | // 1. Chain SDValue |
1612 | // 2. Callee which in this case is the Global address value. |
1613 | // 3. Registers live into the call.In this case its R0, as we |
1614 | // have just one argument to be passed. |
1615 | // 4. Glue. |
1616 | // Note: The order is important. |
1617 | |
1618 | const auto &HRI = *Subtarget.getRegisterInfo(); |
1619 | const uint32_t *Mask = HRI.getCallPreservedMask(MF, CallingConv::C); |
1620 | 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\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1620, __PRETTY_FUNCTION__)); |
1621 | SDValue Ops[] = { Chain, TGA, DAG.getRegister(Hexagon::R0, PtrVT), |
1622 | DAG.getRegisterMask(Mask), Glue }; |
1623 | Chain = DAG.getNode(HexagonISD::CALL, dl, NodeTys, Ops); |
1624 | |
1625 | // Inform MFI that function has calls. |
1626 | MFI.setAdjustsStack(true); |
1627 | |
1628 | Glue = Chain.getValue(1); |
1629 | return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Glue); |
1630 | } |
1631 | |
1632 | // |
1633 | // Lower using the intial executable model for TLS addresses |
1634 | // |
1635 | SDValue |
1636 | HexagonTargetLowering::LowerToTLSInitialExecModel(GlobalAddressSDNode *GA, |
1637 | SelectionDAG &DAG) const { |
1638 | SDLoc dl(GA); |
1639 | int64_t Offset = GA->getOffset(); |
1640 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
1641 | |
1642 | // Get the thread pointer. |
1643 | SDValue TP = DAG.getCopyFromReg(DAG.getEntryNode(), dl, Hexagon::UGP, PtrVT); |
1644 | |
1645 | bool IsPositionIndependent = isPositionIndependent(); |
1646 | unsigned char TF = |
1647 | IsPositionIndependent ? HexagonII::MO_IEGOT : HexagonII::MO_IE; |
1648 | |
1649 | // First generate the TLS symbol address |
1650 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, PtrVT, |
1651 | Offset, TF); |
1652 | |
1653 | SDValue Sym = DAG.getNode(HexagonISD::CONST32, dl, PtrVT, TGA); |
1654 | |
1655 | if (IsPositionIndependent) { |
1656 | // Generate the GOT pointer in case of position independent code |
1657 | SDValue GOT = LowerGLOBAL_OFFSET_TABLE(Sym, DAG); |
1658 | |
1659 | // Add the TLS Symbol address to GOT pointer.This gives |
1660 | // GOT relative relocation for the symbol. |
1661 | Sym = DAG.getNode(ISD::ADD, dl, PtrVT, GOT, Sym); |
1662 | } |
1663 | |
1664 | // Load the offset value for TLS symbol.This offset is relative to |
1665 | // thread pointer. |
1666 | SDValue LoadOffset = |
1667 | DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Sym, MachinePointerInfo()); |
1668 | |
1669 | // Address of the thread local variable is the add of thread |
1670 | // pointer and the offset of the variable. |
1671 | return DAG.getNode(ISD::ADD, dl, PtrVT, TP, LoadOffset); |
1672 | } |
1673 | |
1674 | // |
1675 | // Lower using the local executable model for TLS addresses |
1676 | // |
1677 | SDValue |
1678 | HexagonTargetLowering::LowerToTLSLocalExecModel(GlobalAddressSDNode *GA, |
1679 | SelectionDAG &DAG) const { |
1680 | SDLoc dl(GA); |
1681 | int64_t Offset = GA->getOffset(); |
1682 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
1683 | |
1684 | // Get the thread pointer. |
1685 | SDValue TP = DAG.getCopyFromReg(DAG.getEntryNode(), dl, Hexagon::UGP, PtrVT); |
1686 | // Generate the TLS symbol address |
1687 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, PtrVT, Offset, |
1688 | HexagonII::MO_TPREL); |
1689 | SDValue Sym = DAG.getNode(HexagonISD::CONST32, dl, PtrVT, TGA); |
1690 | |
1691 | // Address of the thread local variable is the add of thread |
1692 | // pointer and the offset of the variable. |
1693 | return DAG.getNode(ISD::ADD, dl, PtrVT, TP, Sym); |
1694 | } |
1695 | |
1696 | // |
1697 | // Lower using the general dynamic model for TLS addresses |
1698 | // |
1699 | SDValue |
1700 | HexagonTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, |
1701 | SelectionDAG &DAG) const { |
1702 | SDLoc dl(GA); |
1703 | int64_t Offset = GA->getOffset(); |
1704 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
1705 | |
1706 | // First generate the TLS symbol address |
1707 | SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, PtrVT, Offset, |
1708 | HexagonII::MO_GDGOT); |
1709 | |
1710 | // Then, generate the GOT pointer |
1711 | SDValue GOT = LowerGLOBAL_OFFSET_TABLE(TGA, DAG); |
1712 | |
1713 | // Add the TLS symbol and the GOT pointer |
1714 | SDValue Sym = DAG.getNode(HexagonISD::CONST32, dl, PtrVT, TGA); |
1715 | SDValue Chain = DAG.getNode(ISD::ADD, dl, PtrVT, GOT, Sym); |
1716 | |
1717 | // Copy over the argument to R0 |
1718 | SDValue InFlag; |
1719 | Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, Hexagon::R0, Chain, InFlag); |
1720 | InFlag = Chain.getValue(1); |
1721 | |
1722 | unsigned Flags = |
1723 | static_cast<const HexagonSubtarget &>(DAG.getSubtarget()).useLongCalls() |
1724 | ? HexagonII::MO_GDPLT | HexagonII::HMOTF_ConstExtended |
1725 | : HexagonII::MO_GDPLT; |
1726 | |
1727 | return GetDynamicTLSAddr(DAG, Chain, GA, InFlag, PtrVT, |
1728 | Hexagon::R0, Flags); |
1729 | } |
1730 | |
1731 | // |
1732 | // Lower TLS addresses. |
1733 | // |
1734 | // For now for dynamic models, we only support the general dynamic model. |
1735 | // |
1736 | SDValue |
1737 | HexagonTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
1738 | SelectionDAG &DAG) const { |
1739 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); |
1740 | |
1741 | switch (HTM.getTLSModel(GA->getGlobal())) { |
1742 | case TLSModel::GeneralDynamic: |
1743 | case TLSModel::LocalDynamic: |
1744 | return LowerToTLSGeneralDynamicModel(GA, DAG); |
1745 | case TLSModel::InitialExec: |
1746 | return LowerToTLSInitialExecModel(GA, DAG); |
1747 | case TLSModel::LocalExec: |
1748 | return LowerToTLSLocalExecModel(GA, DAG); |
1749 | } |
1750 | llvm_unreachable("Bogus TLS model")::llvm::llvm_unreachable_internal("Bogus TLS model", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 1750); |
1751 | } |
1752 | |
1753 | //===----------------------------------------------------------------------===// |
1754 | // TargetLowering Implementation |
1755 | //===----------------------------------------------------------------------===// |
1756 | |
1757 | HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM, |
1758 | const HexagonSubtarget &ST) |
1759 | : TargetLowering(TM), HTM(static_cast<const HexagonTargetMachine&>(TM)), |
1760 | Subtarget(ST) { |
1761 | bool IsV4 = !Subtarget.hasV5TOps(); |
1762 | auto &HRI = *Subtarget.getRegisterInfo(); |
1763 | bool UseHVX = Subtarget.useHVXOps(); |
1764 | bool UseHVXSgl = Subtarget.useHVXSglOps(); |
1765 | bool UseHVXDbl = Subtarget.useHVXDblOps(); |
1766 | |
1767 | setPrefLoopAlignment(4); |
1768 | setPrefFunctionAlignment(4); |
1769 | setMinFunctionAlignment(2); |
1770 | setStackPointerRegisterToSaveRestore(HRI.getStackRegister()); |
1771 | |
1772 | setMaxAtomicSizeInBitsSupported(64); |
1773 | setMinCmpXchgSizeInBits(32); |
1774 | |
1775 | if (EnableHexSDNodeSched) |
1776 | setSchedulingPreference(Sched::VLIW); |
1777 | else |
1778 | setSchedulingPreference(Sched::Source); |
1779 | |
1780 | // Limits for inline expansion of memcpy/memmove |
1781 | MaxStoresPerMemcpy = MaxStoresPerMemcpyCL; |
1782 | MaxStoresPerMemcpyOptSize = MaxStoresPerMemcpyOptSizeCL; |
1783 | MaxStoresPerMemmove = MaxStoresPerMemmoveCL; |
1784 | MaxStoresPerMemmoveOptSize = MaxStoresPerMemmoveOptSizeCL; |
1785 | MaxStoresPerMemset = MaxStoresPerMemsetCL; |
1786 | MaxStoresPerMemsetOptSize = MaxStoresPerMemsetOptSizeCL; |
1787 | |
1788 | // |
1789 | // Set up register classes. |
1790 | // |
1791 | |
1792 | addRegisterClass(MVT::i1, &Hexagon::PredRegsRegClass); |
1793 | addRegisterClass(MVT::v2i1, &Hexagon::PredRegsRegClass); // bbbbaaaa |
1794 | addRegisterClass(MVT::v4i1, &Hexagon::PredRegsRegClass); // ddccbbaa |
1795 | addRegisterClass(MVT::v8i1, &Hexagon::PredRegsRegClass); // hgfedcba |
1796 | addRegisterClass(MVT::i32, &Hexagon::IntRegsRegClass); |
1797 | addRegisterClass(MVT::v4i8, &Hexagon::IntRegsRegClass); |
1798 | addRegisterClass(MVT::v2i16, &Hexagon::IntRegsRegClass); |
1799 | addRegisterClass(MVT::i64, &Hexagon::DoubleRegsRegClass); |
1800 | addRegisterClass(MVT::v8i8, &Hexagon::DoubleRegsRegClass); |
1801 | addRegisterClass(MVT::v4i16, &Hexagon::DoubleRegsRegClass); |
1802 | addRegisterClass(MVT::v2i32, &Hexagon::DoubleRegsRegClass); |
1803 | |
1804 | if (Subtarget.hasV5TOps()) { |
1805 | addRegisterClass(MVT::f32, &Hexagon::IntRegsRegClass); |
1806 | addRegisterClass(MVT::f64, &Hexagon::DoubleRegsRegClass); |
1807 | } |
1808 | |
1809 | if (Subtarget.hasV60TOps()) { |
1810 | if (Subtarget.useHVXSglOps()) { |
1811 | addRegisterClass(MVT::v64i8, &Hexagon::VectorRegsRegClass); |
1812 | addRegisterClass(MVT::v32i16, &Hexagon::VectorRegsRegClass); |
1813 | addRegisterClass(MVT::v16i32, &Hexagon::VectorRegsRegClass); |
1814 | addRegisterClass(MVT::v8i64, &Hexagon::VectorRegsRegClass); |
1815 | addRegisterClass(MVT::v128i8, &Hexagon::VecDblRegsRegClass); |
1816 | addRegisterClass(MVT::v64i16, &Hexagon::VecDblRegsRegClass); |
1817 | addRegisterClass(MVT::v32i32, &Hexagon::VecDblRegsRegClass); |
1818 | addRegisterClass(MVT::v16i64, &Hexagon::VecDblRegsRegClass); |
1819 | addRegisterClass(MVT::v512i1, &Hexagon::VecPredRegsRegClass); |
1820 | } else if (Subtarget.useHVXDblOps()) { |
1821 | addRegisterClass(MVT::v128i8, &Hexagon::VectorRegs128BRegClass); |
1822 | addRegisterClass(MVT::v64i16, &Hexagon::VectorRegs128BRegClass); |
1823 | addRegisterClass(MVT::v32i32, &Hexagon::VectorRegs128BRegClass); |
1824 | addRegisterClass(MVT::v16i64, &Hexagon::VectorRegs128BRegClass); |
1825 | addRegisterClass(MVT::v256i8, &Hexagon::VecDblRegs128BRegClass); |
1826 | addRegisterClass(MVT::v128i16, &Hexagon::VecDblRegs128BRegClass); |
1827 | addRegisterClass(MVT::v64i32, &Hexagon::VecDblRegs128BRegClass); |
1828 | addRegisterClass(MVT::v32i64, &Hexagon::VecDblRegs128BRegClass); |
1829 | addRegisterClass(MVT::v1024i1, &Hexagon::VecPredRegs128BRegClass); |
1830 | } |
1831 | } |
1832 | |
1833 | // |
1834 | // Handling of scalar operations. |
1835 | // |
1836 | // All operations default to "legal", except: |
1837 | // - indexed loads and stores (pre-/post-incremented), |
1838 | // - ANY_EXTEND_VECTOR_INREG, ATOMIC_CMP_SWAP_WITH_SUCCESS, CONCAT_VECTORS, |
1839 | // ConstantFP, DEBUGTRAP, FCEIL, FCOPYSIGN, FEXP, FEXP2, FFLOOR, FGETSIGN, |
1840 | // FLOG, FLOG2, FLOG10, FMAXNUM, FMINNUM, FNEARBYINT, FRINT, FROUND, TRAP, |
1841 | // FTRUNC, PREFETCH, SIGN_EXTEND_VECTOR_INREG, ZERO_EXTEND_VECTOR_INREG, |
1842 | // which default to "expand" for at least one type. |
1843 | |
1844 | // Misc operations. |
1845 | setOperationAction(ISD::ConstantFP, MVT::f32, Legal); // Default: expand |
1846 | setOperationAction(ISD::ConstantFP, MVT::f64, Legal); // Default: expand |
1847 | |
1848 | setOperationAction(ISD::ConstantPool, MVT::i32, Custom); |
1849 | setOperationAction(ISD::JumpTable, MVT::i32, Custom); |
1850 | setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand); |
1851 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
1852 | setOperationAction(ISD::INLINEASM, MVT::Other, Custom); |
1853 | setOperationAction(ISD::PREFETCH, MVT::Other, Custom); |
1854 | setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom); |
1855 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
1856 | setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); |
1857 | setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); |
1858 | setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); |
1859 | setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom); |
1860 | |
1861 | // Custom legalize GlobalAddress nodes into CONST32. |
1862 | setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); |
1863 | setOperationAction(ISD::GlobalAddress, MVT::i8, Custom); |
1864 | setOperationAction(ISD::BlockAddress, MVT::i32, Custom); |
1865 | |
1866 | // Hexagon needs to optimize cases with negative constants. |
1867 | setOperationAction(ISD::SETCC, MVT::i8, Custom); |
1868 | setOperationAction(ISD::SETCC, MVT::i16, Custom); |
1869 | |
1870 | // VASTART needs to be custom lowered to use the VarArgsFrameIndex. |
1871 | setOperationAction(ISD::VASTART, MVT::Other, Custom); |
1872 | setOperationAction(ISD::VAEND, MVT::Other, Expand); |
1873 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
1874 | |
1875 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); |
1876 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); |
1877 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom); |
1878 | |
1879 | if (EmitJumpTables) |
1880 | setMinimumJumpTableEntries(MinimumJumpTables); |
1881 | else |
1882 | setMinimumJumpTableEntries(std::numeric_limits<int>::max()); |
1883 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); |
1884 | |
1885 | // Hexagon has instructions for add/sub with carry. The problem with |
1886 | // modeling these instructions is that they produce 2 results: Rdd and Px. |
1887 | // To model the update of Px, we will have to use Defs[p0..p3] which will |
1888 | // cause any predicate live range to spill. So, we pretend we dont't have |
1889 | // these instructions. |
1890 | setOperationAction(ISD::ADDE, MVT::i8, Expand); |
1891 | setOperationAction(ISD::ADDE, MVT::i16, Expand); |
1892 | setOperationAction(ISD::ADDE, MVT::i32, Expand); |
1893 | setOperationAction(ISD::ADDE, MVT::i64, Expand); |
1894 | setOperationAction(ISD::SUBE, MVT::i8, Expand); |
1895 | setOperationAction(ISD::SUBE, MVT::i16, Expand); |
1896 | setOperationAction(ISD::SUBE, MVT::i32, Expand); |
1897 | setOperationAction(ISD::SUBE, MVT::i64, Expand); |
1898 | setOperationAction(ISD::ADDC, MVT::i8, Expand); |
1899 | setOperationAction(ISD::ADDC, MVT::i16, Expand); |
1900 | setOperationAction(ISD::ADDC, MVT::i32, Expand); |
1901 | setOperationAction(ISD::ADDC, MVT::i64, Expand); |
1902 | setOperationAction(ISD::SUBC, MVT::i8, Expand); |
1903 | setOperationAction(ISD::SUBC, MVT::i16, Expand); |
1904 | setOperationAction(ISD::SUBC, MVT::i32, Expand); |
1905 | setOperationAction(ISD::SUBC, MVT::i64, Expand); |
1906 | |
1907 | // Only add and sub that detect overflow are the saturating ones. |
1908 | for (MVT VT : MVT::integer_valuetypes()) { |
1909 | setOperationAction(ISD::UADDO, VT, Expand); |
1910 | setOperationAction(ISD::SADDO, VT, Expand); |
1911 | setOperationAction(ISD::USUBO, VT, Expand); |
1912 | setOperationAction(ISD::SSUBO, VT, Expand); |
1913 | } |
1914 | |
1915 | setOperationAction(ISD::CTLZ, MVT::i8, Promote); |
1916 | setOperationAction(ISD::CTLZ, MVT::i16, Promote); |
1917 | setOperationAction(ISD::CTTZ, MVT::i8, Promote); |
1918 | setOperationAction(ISD::CTTZ, MVT::i16, Promote); |
1919 | |
1920 | // In V5, popcount can count # of 1s in i64 but returns i32. |
1921 | // On V4 it will be expanded (set later). |
1922 | setOperationAction(ISD::CTPOP, MVT::i8, Promote); |
1923 | setOperationAction(ISD::CTPOP, MVT::i16, Promote); |
1924 | setOperationAction(ISD::CTPOP, MVT::i32, Promote); |
1925 | setOperationAction(ISD::CTPOP, MVT::i64, Legal); |
1926 | |
1927 | setOperationAction(ISD::BITREVERSE, MVT::i32, Legal); |
1928 | setOperationAction(ISD::BITREVERSE, MVT::i64, Legal); |
1929 | setOperationAction(ISD::BSWAP, MVT::i32, Legal); |
1930 | setOperationAction(ISD::BSWAP, MVT::i64, Legal); |
1931 | setOperationAction(ISD::MUL, MVT::i64, Legal); |
1932 | |
1933 | for (unsigned IntExpOp : |
1934 | { ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, |
1935 | ISD::SDIVREM, ISD::UDIVREM, ISD::ROTL, ISD::ROTR, |
1936 | ISD::SHL_PARTS, ISD::SRA_PARTS, ISD::SRL_PARTS, |
1937 | ISD::SMUL_LOHI, ISD::UMUL_LOHI }) { |
1938 | setOperationAction(IntExpOp, MVT::i32, Expand); |
1939 | setOperationAction(IntExpOp, MVT::i64, Expand); |
1940 | } |
1941 | |
1942 | for (unsigned FPExpOp : |
1943 | {ISD::FDIV, ISD::FREM, ISD::FSQRT, ISD::FSIN, ISD::FCOS, ISD::FSINCOS, |
1944 | ISD::FPOW, ISD::FCOPYSIGN}) { |
1945 | setOperationAction(FPExpOp, MVT::f32, Expand); |
1946 | setOperationAction(FPExpOp, MVT::f64, Expand); |
1947 | } |
1948 | |
1949 | // No extending loads from i32. |
1950 | for (MVT VT : MVT::integer_valuetypes()) { |
1951 | setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand); |
1952 | setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand); |
1953 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand); |
1954 | } |
1955 | // Turn FP truncstore into trunc + store. |
1956 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
1957 | // Turn FP extload into load/fpextend. |
1958 | for (MVT VT : MVT::fp_valuetypes()) |
1959 | setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand); |
1960 | |
1961 | // Expand BR_CC and SELECT_CC for all integer and fp types. |
1962 | for (MVT VT : MVT::integer_valuetypes()) { |
1963 | setOperationAction(ISD::BR_CC, VT, Expand); |
1964 | setOperationAction(ISD::SELECT_CC, VT, Expand); |
1965 | } |
1966 | for (MVT VT : MVT::fp_valuetypes()) { |
1967 | setOperationAction(ISD::BR_CC, VT, Expand); |
1968 | setOperationAction(ISD::SELECT_CC, VT, Expand); |
1969 | } |
1970 | setOperationAction(ISD::BR_CC, MVT::Other, Expand); |
1971 | |
1972 | // |
1973 | // Handling of vector operations. |
1974 | // |
1975 | |
1976 | // Custom lower v4i16 load only. Let v4i16 store to be |
1977 | // promoted for now. |
1978 | promoteLdStType(MVT::v4i8, MVT::i32); |
1979 | promoteLdStType(MVT::v2i16, MVT::i32); |
1980 | promoteLdStType(MVT::v8i8, MVT::i64); |
1981 | promoteLdStType(MVT::v2i32, MVT::i64); |
1982 | |
1983 | setOperationAction(ISD::LOAD, MVT::v4i16, Custom); |
1984 | setOperationAction(ISD::STORE, MVT::v4i16, Promote); |
1985 | AddPromotedToType(ISD::LOAD, MVT::v4i16, MVT::i64); |
1986 | AddPromotedToType(ISD::STORE, MVT::v4i16, MVT::i64); |
1987 | |
1988 | // Set the action for vector operations to "expand", then override it with |
1989 | // either "custom" or "legal" for specific cases. |
1990 | static const unsigned VectExpOps[] = { |
1991 | // Integer arithmetic: |
1992 | ISD::ADD, ISD::SUB, ISD::MUL, ISD::SDIV, ISD::UDIV, |
1993 | ISD::SREM, ISD::UREM, ISD::SDIVREM, ISD::UDIVREM, ISD::ADDC, |
1994 | ISD::SUBC, ISD::SADDO, ISD::UADDO, ISD::SSUBO, ISD::USUBO, |
1995 | ISD::SMUL_LOHI, ISD::UMUL_LOHI, |
1996 | // Logical/bit: |
1997 | ISD::AND, ISD::OR, ISD::XOR, ISD::ROTL, ISD::ROTR, |
1998 | ISD::CTPOP, ISD::CTLZ, ISD::CTTZ, |
1999 | // Floating point arithmetic/math functions: |
2000 | ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FMA, ISD::FDIV, |
2001 | ISD::FREM, ISD::FNEG, ISD::FABS, ISD::FSQRT, ISD::FSIN, |
2002 | ISD::FCOS, ISD::FPOW, ISD::FLOG, ISD::FLOG2, |
2003 | ISD::FLOG10, ISD::FEXP, ISD::FEXP2, ISD::FCEIL, ISD::FTRUNC, |
2004 | ISD::FRINT, ISD::FNEARBYINT, ISD::FROUND, ISD::FFLOOR, |
2005 | ISD::FMINNUM, ISD::FMAXNUM, ISD::FSINCOS, |
2006 | // Misc: |
2007 | ISD::BR_CC, ISD::SELECT_CC, ISD::ConstantPool, |
2008 | // Vector: |
2009 | ISD::BUILD_VECTOR, ISD::SCALAR_TO_VECTOR, |
2010 | ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT, |
2011 | ISD::EXTRACT_SUBVECTOR, ISD::INSERT_SUBVECTOR, |
2012 | ISD::CONCAT_VECTORS, ISD::VECTOR_SHUFFLE |
2013 | }; |
2014 | |
2015 | for (MVT VT : MVT::vector_valuetypes()) { |
2016 | for (unsigned VectExpOp : VectExpOps) |
2017 | setOperationAction(VectExpOp, VT, Expand); |
2018 | |
2019 | // Expand all extending loads and truncating stores: |
2020 | for (MVT TargetVT : MVT::vector_valuetypes()) { |
2021 | if (TargetVT == VT) |
2022 | continue; |
2023 | setLoadExtAction(ISD::EXTLOAD, TargetVT, VT, Expand); |
2024 | setLoadExtAction(ISD::ZEXTLOAD, TargetVT, VT, Expand); |
2025 | setLoadExtAction(ISD::SEXTLOAD, TargetVT, VT, Expand); |
2026 | setTruncStoreAction(VT, TargetVT, Expand); |
2027 | } |
2028 | |
2029 | // Normalize all inputs to SELECT to be vectors of i32. |
2030 | if (VT.getVectorElementType() != MVT::i32) { |
2031 | MVT VT32 = MVT::getVectorVT(MVT::i32, VT.getSizeInBits()/32); |
2032 | setOperationAction(ISD::SELECT, VT, Promote); |
2033 | AddPromotedToType(ISD::SELECT, VT, VT32); |
2034 | } |
2035 | setOperationAction(ISD::SRA, VT, Custom); |
2036 | setOperationAction(ISD::SHL, VT, Custom); |
2037 | setOperationAction(ISD::SRL, VT, Custom); |
2038 | } |
2039 | |
2040 | // Types natively supported: |
2041 | for (MVT NativeVT : {MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v32i1, MVT::v64i1, |
2042 | MVT::v4i8, MVT::v8i8, MVT::v2i16, MVT::v4i16, MVT::v1i32, |
2043 | MVT::v2i32, MVT::v1i64}) { |
2044 | setOperationAction(ISD::BUILD_VECTOR, NativeVT, Custom); |
2045 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, NativeVT, Custom); |
2046 | setOperationAction(ISD::INSERT_VECTOR_ELT, NativeVT, Custom); |
2047 | setOperationAction(ISD::EXTRACT_SUBVECTOR, NativeVT, Custom); |
2048 | setOperationAction(ISD::INSERT_SUBVECTOR, NativeVT, Custom); |
2049 | setOperationAction(ISD::CONCAT_VECTORS, NativeVT, Custom); |
2050 | |
2051 | setOperationAction(ISD::ADD, NativeVT, Legal); |
2052 | setOperationAction(ISD::SUB, NativeVT, Legal); |
2053 | setOperationAction(ISD::MUL, NativeVT, Legal); |
2054 | setOperationAction(ISD::AND, NativeVT, Legal); |
2055 | setOperationAction(ISD::OR, NativeVT, Legal); |
2056 | setOperationAction(ISD::XOR, NativeVT, Legal); |
2057 | } |
2058 | |
2059 | setOperationAction(ISD::SETCC, MVT::v2i16, Custom); |
2060 | setOperationAction(ISD::VSELECT, MVT::v2i16, Custom); |
2061 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom); |
2062 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom); |
2063 | |
2064 | if (UseHVX) { |
2065 | if (UseHVXSgl) { |
2066 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v128i8, Custom); |
2067 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i16, Custom); |
2068 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i32, Custom); |
2069 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i64, Custom); |
2070 | // We try to generate the vpack{e/o} instructions. If we fail |
2071 | // we fall back upon ExpandOp. |
2072 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom); |
2073 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom); |
2074 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v64i8, Custom); |
2075 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32i16, Custom); |
2076 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v16i32, Custom); |
2077 | } else if (UseHVXDbl) { |
2078 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v256i8, Custom); |
2079 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v128i16, Custom); |
2080 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i32, Custom); |
2081 | setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i64, Custom); |
2082 | // We try to generate the vpack{e/o} instructions. If we fail |
2083 | // we fall back upon ExpandOp. |
2084 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v128i8, Custom); |
2085 | setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i16, Custom); |
2086 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom); |
2087 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v128i8, Custom); |
2088 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v64i16, Custom); |
2089 | setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v32i32, Custom); |
2090 | } else { |
2091 | llvm_unreachable("Unrecognized HVX mode")::llvm::llvm_unreachable_internal("Unrecognized HVX mode", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2091); |
2092 | } |
2093 | } |
2094 | // Subtarget-specific operation actions. |
2095 | // |
2096 | if (Subtarget.hasV5TOps()) { |
2097 | setOperationAction(ISD::FMA, MVT::f64, Expand); |
2098 | setOperationAction(ISD::FADD, MVT::f64, Expand); |
2099 | setOperationAction(ISD::FSUB, MVT::f64, Expand); |
2100 | setOperationAction(ISD::FMUL, MVT::f64, Expand); |
2101 | |
2102 | setOperationAction(ISD::FMINNUM, MVT::f32, Legal); |
2103 | setOperationAction(ISD::FMAXNUM, MVT::f32, Legal); |
2104 | |
2105 | setOperationAction(ISD::FP_TO_UINT, MVT::i1, Promote); |
2106 | setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote); |
2107 | setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote); |
2108 | setOperationAction(ISD::FP_TO_SINT, MVT::i1, Promote); |
2109 | setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote); |
2110 | setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote); |
2111 | setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote); |
2112 | setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote); |
2113 | setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote); |
2114 | setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote); |
2115 | setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote); |
2116 | setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote); |
2117 | } else { // V4 |
2118 | setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand); |
2119 | setOperationAction(ISD::SINT_TO_FP, MVT::i64, Expand); |
2120 | setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand); |
2121 | setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand); |
2122 | setOperationAction(ISD::FP_TO_SINT, MVT::f64, Expand); |
2123 | setOperationAction(ISD::FP_TO_SINT, MVT::f32, Expand); |
2124 | setOperationAction(ISD::FP_EXTEND, MVT::f32, Expand); |
2125 | setOperationAction(ISD::FP_ROUND, MVT::f64, Expand); |
2126 | setCondCodeAction(ISD::SETUNE, MVT::f64, Expand); |
2127 | |
2128 | setOperationAction(ISD::CTPOP, MVT::i8, Expand); |
2129 | setOperationAction(ISD::CTPOP, MVT::i16, Expand); |
2130 | setOperationAction(ISD::CTPOP, MVT::i32, Expand); |
2131 | setOperationAction(ISD::CTPOP, MVT::i64, Expand); |
2132 | |
2133 | // Expand these operations for both f32 and f64: |
2134 | for (unsigned FPExpOpV4 : |
2135 | {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FABS, ISD::FNEG, ISD::FMA}) { |
2136 | setOperationAction(FPExpOpV4, MVT::f32, Expand); |
2137 | setOperationAction(FPExpOpV4, MVT::f64, Expand); |
2138 | } |
2139 | |
2140 | for (ISD::CondCode FPExpCCV4 : |
2141 | {ISD::SETOEQ, ISD::SETOGT, ISD::SETOLT, ISD::SETOGE, ISD::SETOLE, |
2142 | ISD::SETUO, ISD::SETO}) { |
2143 | setCondCodeAction(FPExpCCV4, MVT::f32, Expand); |
2144 | setCondCodeAction(FPExpCCV4, MVT::f64, Expand); |
2145 | } |
2146 | } |
2147 | |
2148 | // Handling of indexed loads/stores: default is "expand". |
2149 | // |
2150 | for (MVT VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) { |
2151 | setIndexedLoadAction(ISD::POST_INC, VT, Legal); |
2152 | setIndexedStoreAction(ISD::POST_INC, VT, Legal); |
2153 | } |
2154 | |
2155 | if (UseHVXSgl) { |
2156 | for (MVT VT : {MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64, |
2157 | MVT::v128i8, MVT::v64i16, MVT::v32i32, MVT::v16i64}) { |
2158 | setIndexedLoadAction(ISD::POST_INC, VT, Legal); |
2159 | setIndexedStoreAction(ISD::POST_INC, VT, Legal); |
2160 | } |
2161 | } else if (UseHVXDbl) { |
2162 | for (MVT VT : {MVT::v128i8, MVT::v64i16, MVT::v32i32, MVT::v16i64, |
2163 | MVT::v256i8, MVT::v128i16, MVT::v64i32, MVT::v32i64}) { |
2164 | setIndexedLoadAction(ISD::POST_INC, VT, Legal); |
2165 | setIndexedStoreAction(ISD::POST_INC, VT, Legal); |
2166 | } |
2167 | } |
2168 | |
2169 | computeRegisterProperties(&HRI); |
2170 | |
2171 | // |
2172 | // Library calls for unsupported operations |
2173 | // |
2174 | bool FastMath = EnableFastMath; |
2175 | |
2176 | setLibcallName(RTLIB::SDIV_I32, "__hexagon_divsi3"); |
2177 | setLibcallName(RTLIB::SDIV_I64, "__hexagon_divdi3"); |
2178 | setLibcallName(RTLIB::UDIV_I32, "__hexagon_udivsi3"); |
2179 | setLibcallName(RTLIB::UDIV_I64, "__hexagon_udivdi3"); |
2180 | setLibcallName(RTLIB::SREM_I32, "__hexagon_modsi3"); |
2181 | setLibcallName(RTLIB::SREM_I64, "__hexagon_moddi3"); |
2182 | setLibcallName(RTLIB::UREM_I32, "__hexagon_umodsi3"); |
2183 | setLibcallName(RTLIB::UREM_I64, "__hexagon_umoddi3"); |
2184 | |
2185 | setLibcallName(RTLIB::SINTTOFP_I128_F64, "__hexagon_floattidf"); |
2186 | setLibcallName(RTLIB::SINTTOFP_I128_F32, "__hexagon_floattisf"); |
2187 | setLibcallName(RTLIB::FPTOUINT_F32_I128, "__hexagon_fixunssfti"); |
2188 | setLibcallName(RTLIB::FPTOUINT_F64_I128, "__hexagon_fixunsdfti"); |
2189 | setLibcallName(RTLIB::FPTOSINT_F32_I128, "__hexagon_fixsfti"); |
2190 | setLibcallName(RTLIB::FPTOSINT_F64_I128, "__hexagon_fixdfti"); |
2191 | |
2192 | if (IsV4) { |
2193 | // Handle single-precision floating point operations on V4. |
2194 | if (FastMath) { |
2195 | setLibcallName(RTLIB::ADD_F32, "__hexagon_fast_addsf3"); |
2196 | setLibcallName(RTLIB::SUB_F32, "__hexagon_fast_subsf3"); |
2197 | setLibcallName(RTLIB::MUL_F32, "__hexagon_fast_mulsf3"); |
2198 | setLibcallName(RTLIB::OGT_F32, "__hexagon_fast_gtsf2"); |
2199 | setLibcallName(RTLIB::OLT_F32, "__hexagon_fast_ltsf2"); |
2200 | // Double-precision compares. |
2201 | setLibcallName(RTLIB::OGT_F64, "__hexagon_fast_gtdf2"); |
2202 | setLibcallName(RTLIB::OLT_F64, "__hexagon_fast_ltdf2"); |
2203 | } else { |
2204 | setLibcallName(RTLIB::ADD_F32, "__hexagon_addsf3"); |
2205 | setLibcallName(RTLIB::SUB_F32, "__hexagon_subsf3"); |
2206 | setLibcallName(RTLIB::MUL_F32, "__hexagon_mulsf3"); |
2207 | setLibcallName(RTLIB::OGT_F32, "__hexagon_gtsf2"); |
2208 | setLibcallName(RTLIB::OLT_F32, "__hexagon_ltsf2"); |
2209 | // Double-precision compares. |
2210 | setLibcallName(RTLIB::OGT_F64, "__hexagon_gtdf2"); |
2211 | setLibcallName(RTLIB::OLT_F64, "__hexagon_ltdf2"); |
2212 | } |
2213 | } |
2214 | |
2215 | // This is the only fast library function for sqrtd. |
2216 | if (FastMath) |
2217 | setLibcallName(RTLIB::SQRT_F64, "__hexagon_fast2_sqrtdf2"); |
2218 | |
2219 | // Prefix is: nothing for "slow-math", |
2220 | // "fast2_" for V4 fast-math and V5+ fast-math double-precision |
2221 | // (actually, keep fast-math and fast-math2 separate for now) |
2222 | if (FastMath) { |
2223 | setLibcallName(RTLIB::ADD_F64, "__hexagon_fast_adddf3"); |
2224 | setLibcallName(RTLIB::SUB_F64, "__hexagon_fast_subdf3"); |
2225 | setLibcallName(RTLIB::MUL_F64, "__hexagon_fast_muldf3"); |
2226 | setLibcallName(RTLIB::DIV_F64, "__hexagon_fast_divdf3"); |
2227 | // Calling __hexagon_fast2_divsf3 with fast-math on V5 (ok). |
2228 | setLibcallName(RTLIB::DIV_F32, "__hexagon_fast_divsf3"); |
2229 | } else { |
2230 | setLibcallName(RTLIB::ADD_F64, "__hexagon_adddf3"); |
2231 | setLibcallName(RTLIB::SUB_F64, "__hexagon_subdf3"); |
2232 | setLibcallName(RTLIB::MUL_F64, "__hexagon_muldf3"); |
2233 | setLibcallName(RTLIB::DIV_F64, "__hexagon_divdf3"); |
2234 | setLibcallName(RTLIB::DIV_F32, "__hexagon_divsf3"); |
2235 | } |
2236 | |
2237 | if (Subtarget.hasV5TOps()) { |
2238 | if (FastMath) |
2239 | setLibcallName(RTLIB::SQRT_F32, "__hexagon_fast2_sqrtf"); |
2240 | else |
2241 | setLibcallName(RTLIB::SQRT_F32, "__hexagon_sqrtf"); |
2242 | } else { |
2243 | // V4 |
2244 | setLibcallName(RTLIB::SINTTOFP_I32_F32, "__hexagon_floatsisf"); |
2245 | setLibcallName(RTLIB::SINTTOFP_I32_F64, "__hexagon_floatsidf"); |
2246 | setLibcallName(RTLIB::SINTTOFP_I64_F32, "__hexagon_floatdisf"); |
2247 | setLibcallName(RTLIB::SINTTOFP_I64_F64, "__hexagon_floatdidf"); |
2248 | setLibcallName(RTLIB::UINTTOFP_I32_F32, "__hexagon_floatunsisf"); |
2249 | setLibcallName(RTLIB::UINTTOFP_I32_F64, "__hexagon_floatunsidf"); |
2250 | setLibcallName(RTLIB::UINTTOFP_I64_F32, "__hexagon_floatundisf"); |
2251 | setLibcallName(RTLIB::UINTTOFP_I64_F64, "__hexagon_floatundidf"); |
2252 | setLibcallName(RTLIB::FPTOUINT_F32_I32, "__hexagon_fixunssfsi"); |
2253 | setLibcallName(RTLIB::FPTOUINT_F32_I64, "__hexagon_fixunssfdi"); |
2254 | setLibcallName(RTLIB::FPTOUINT_F64_I32, "__hexagon_fixunsdfsi"); |
2255 | setLibcallName(RTLIB::FPTOUINT_F64_I64, "__hexagon_fixunsdfdi"); |
2256 | setLibcallName(RTLIB::FPTOSINT_F32_I32, "__hexagon_fixsfsi"); |
2257 | setLibcallName(RTLIB::FPTOSINT_F32_I64, "__hexagon_fixsfdi"); |
2258 | setLibcallName(RTLIB::FPTOSINT_F64_I32, "__hexagon_fixdfsi"); |
2259 | setLibcallName(RTLIB::FPTOSINT_F64_I64, "__hexagon_fixdfdi"); |
2260 | setLibcallName(RTLIB::FPEXT_F32_F64, "__hexagon_extendsfdf2"); |
2261 | setLibcallName(RTLIB::FPROUND_F64_F32, "__hexagon_truncdfsf2"); |
2262 | setLibcallName(RTLIB::OEQ_F32, "__hexagon_eqsf2"); |
2263 | setLibcallName(RTLIB::OEQ_F64, "__hexagon_eqdf2"); |
2264 | setLibcallName(RTLIB::OGE_F32, "__hexagon_gesf2"); |
2265 | setLibcallName(RTLIB::OGE_F64, "__hexagon_gedf2"); |
2266 | setLibcallName(RTLIB::OLE_F32, "__hexagon_lesf2"); |
2267 | setLibcallName(RTLIB::OLE_F64, "__hexagon_ledf2"); |
2268 | setLibcallName(RTLIB::UNE_F32, "__hexagon_nesf2"); |
2269 | setLibcallName(RTLIB::UNE_F64, "__hexagon_nedf2"); |
2270 | setLibcallName(RTLIB::UO_F32, "__hexagon_unordsf2"); |
2271 | setLibcallName(RTLIB::UO_F64, "__hexagon_unorddf2"); |
2272 | setLibcallName(RTLIB::O_F32, "__hexagon_unordsf2"); |
2273 | setLibcallName(RTLIB::O_F64, "__hexagon_unorddf2"); |
2274 | } |
2275 | |
2276 | // These cause problems when the shift amount is non-constant. |
2277 | setLibcallName(RTLIB::SHL_I128, nullptr); |
2278 | setLibcallName(RTLIB::SRL_I128, nullptr); |
2279 | setLibcallName(RTLIB::SRA_I128, nullptr); |
2280 | } |
2281 | |
2282 | const char* HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const { |
2283 | switch ((HexagonISD::NodeType)Opcode) { |
2284 | case HexagonISD::ALLOCA: return "HexagonISD::ALLOCA"; |
2285 | case HexagonISD::AT_GOT: return "HexagonISD::AT_GOT"; |
2286 | case HexagonISD::AT_PCREL: return "HexagonISD::AT_PCREL"; |
2287 | case HexagonISD::BARRIER: return "HexagonISD::BARRIER"; |
2288 | case HexagonISD::CALL: return "HexagonISD::CALL"; |
2289 | case HexagonISD::CALLnr: return "HexagonISD::CALLnr"; |
2290 | case HexagonISD::CALLR: return "HexagonISD::CALLR"; |
2291 | case HexagonISD::COMBINE: return "HexagonISD::COMBINE"; |
2292 | case HexagonISD::CONST32_GP: return "HexagonISD::CONST32_GP"; |
2293 | case HexagonISD::CONST32: return "HexagonISD::CONST32"; |
2294 | case HexagonISD::CP: return "HexagonISD::CP"; |
2295 | case HexagonISD::DCFETCH: return "HexagonISD::DCFETCH"; |
2296 | case HexagonISD::EH_RETURN: return "HexagonISD::EH_RETURN"; |
2297 | case HexagonISD::EXTRACTU: return "HexagonISD::EXTRACTU"; |
2298 | case HexagonISD::EXTRACTURP: return "HexagonISD::EXTRACTURP"; |
2299 | case HexagonISD::INSERT: return "HexagonISD::INSERT"; |
2300 | case HexagonISD::INSERTRP: return "HexagonISD::INSERTRP"; |
2301 | case HexagonISD::JT: return "HexagonISD::JT"; |
2302 | case HexagonISD::PACKHL: return "HexagonISD::PACKHL"; |
2303 | case HexagonISD::RET_FLAG: return "HexagonISD::RET_FLAG"; |
2304 | case HexagonISD::SHUFFEB: return "HexagonISD::SHUFFEB"; |
2305 | case HexagonISD::SHUFFEH: return "HexagonISD::SHUFFEH"; |
2306 | case HexagonISD::SHUFFOB: return "HexagonISD::SHUFFOB"; |
2307 | case HexagonISD::SHUFFOH: return "HexagonISD::SHUFFOH"; |
2308 | case HexagonISD::TC_RETURN: return "HexagonISD::TC_RETURN"; |
2309 | case HexagonISD::VCMPBEQ: return "HexagonISD::VCMPBEQ"; |
2310 | case HexagonISD::VCMPBGT: return "HexagonISD::VCMPBGT"; |
2311 | case HexagonISD::VCMPBGTU: return "HexagonISD::VCMPBGTU"; |
2312 | case HexagonISD::VCMPHEQ: return "HexagonISD::VCMPHEQ"; |
2313 | case HexagonISD::VCMPHGT: return "HexagonISD::VCMPHGT"; |
2314 | case HexagonISD::VCMPHGTU: return "HexagonISD::VCMPHGTU"; |
2315 | case HexagonISD::VCMPWEQ: return "HexagonISD::VCMPWEQ"; |
2316 | case HexagonISD::VCMPWGT: return "HexagonISD::VCMPWGT"; |
2317 | case HexagonISD::VCMPWGTU: return "HexagonISD::VCMPWGTU"; |
2318 | case HexagonISD::VCOMBINE: return "HexagonISD::VCOMBINE"; |
2319 | case HexagonISD::VPACK: return "HexagonISD::VPACK"; |
2320 | case HexagonISD::VSHLH: return "HexagonISD::VSHLH"; |
2321 | case HexagonISD::VSHLW: return "HexagonISD::VSHLW"; |
2322 | case HexagonISD::VSPLATB: return "HexagonISD::VSPLTB"; |
2323 | case HexagonISD::VSPLATH: return "HexagonISD::VSPLATH"; |
2324 | case HexagonISD::VSRAH: return "HexagonISD::VSRAH"; |
2325 | case HexagonISD::VSRAW: return "HexagonISD::VSRAW"; |
2326 | case HexagonISD::VSRLH: return "HexagonISD::VSRLH"; |
2327 | case HexagonISD::VSRLW: return "HexagonISD::VSRLW"; |
2328 | case HexagonISD::VSXTBH: return "HexagonISD::VSXTBH"; |
2329 | case HexagonISD::VSXTBW: return "HexagonISD::VSXTBW"; |
2330 | case HexagonISD::READCYCLE: return "HexagonISD::READCYCLE"; |
2331 | case HexagonISD::OP_END: break; |
2332 | } |
2333 | return nullptr; |
2334 | } |
2335 | |
2336 | bool HexagonTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { |
2337 | EVT MTy1 = EVT::getEVT(Ty1); |
2338 | EVT MTy2 = EVT::getEVT(Ty2); |
2339 | if (!MTy1.isSimple() || !MTy2.isSimple()) |
2340 | return false; |
2341 | return (MTy1.getSimpleVT() == MVT::i64) && (MTy2.getSimpleVT() == MVT::i32); |
2342 | } |
2343 | |
2344 | bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { |
2345 | if (!VT1.isSimple() || !VT2.isSimple()) |
2346 | return false; |
2347 | return (VT1.getSimpleVT() == MVT::i64) && (VT2.getSimpleVT() == MVT::i32); |
2348 | } |
2349 | |
2350 | bool HexagonTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const { |
2351 | return isOperationLegalOrCustom(ISD::FMA, VT); |
2352 | } |
2353 | |
2354 | // Should we expand the build vector with shuffles? |
2355 | bool HexagonTargetLowering::shouldExpandBuildVectorWithShuffles(EVT VT, |
2356 | unsigned DefinedValues) const { |
2357 | // Hexagon vector shuffle operates on element sizes of bytes or halfwords |
2358 | EVT EltVT = VT.getVectorElementType(); |
2359 | int EltBits = EltVT.getSizeInBits(); |
2360 | if ((EltBits != 8) && (EltBits != 16)) |
2361 | return false; |
2362 | |
2363 | return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues); |
2364 | } |
2365 | |
2366 | static StridedLoadKind isStridedLoad(const ArrayRef<int> &Mask) { |
2367 | int even_start = -2; |
2368 | int odd_start = -1; |
2369 | size_t mask_len = Mask.size(); |
2370 | for (auto idx : Mask) { |
2371 | if ((idx - even_start) == 2) |
2372 | even_start = idx; |
2373 | else |
2374 | break; |
2375 | } |
2376 | if (even_start == (int)(mask_len * 2) - 2) |
2377 | return StridedLoadKind::Even; |
2378 | for (auto idx : Mask) { |
2379 | if ((idx - odd_start) == 2) |
2380 | odd_start = idx; |
2381 | else |
2382 | break; |
2383 | } |
2384 | if (odd_start == (int)(mask_len * 2) - 1) |
2385 | return StridedLoadKind::Odd; |
2386 | |
2387 | return StridedLoadKind::NoPattern; |
2388 | } |
2389 | |
2390 | bool HexagonTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &Mask, |
2391 | EVT VT) const { |
2392 | if (Subtarget.useHVXOps()) |
2393 | return isStridedLoad(Mask) != StridedLoadKind::NoPattern; |
2394 | return true; |
2395 | } |
2396 | |
2397 | // Lower a vector shuffle (V1, V2, V3). V1 and V2 are the two vectors |
2398 | // to select data from, V3 is the permutation. |
2399 | SDValue |
2400 | HexagonTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) |
2401 | const { |
2402 | const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op); |
2403 | SDValue V1 = Op.getOperand(0); |
2404 | SDValue V2 = Op.getOperand(1); |
2405 | SDLoc dl(Op); |
2406 | EVT VT = Op.getValueType(); |
2407 | bool UseHVX = Subtarget.useHVXOps(); |
2408 | |
2409 | if (V2.isUndef()) |
2410 | V2 = V1; |
2411 | |
2412 | if (SVN->isSplat()) { |
2413 | int Lane = SVN->getSplatIndex(); |
2414 | if (Lane == -1) Lane = 0; |
2415 | |
2416 | // Test if V1 is a SCALAR_TO_VECTOR. |
2417 | if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) |
2418 | return createSplat(DAG, dl, VT, V1.getOperand(0)); |
2419 | |
2420 | // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR |
2421 | // (and probably will turn into a SCALAR_TO_VECTOR once legalization |
2422 | // reaches it). |
2423 | if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR && |
2424 | !isa<ConstantSDNode>(V1.getOperand(0))) { |
2425 | bool IsScalarToVector = true; |
2426 | for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i) { |
2427 | if (!V1.getOperand(i).isUndef()) { |
2428 | IsScalarToVector = false; |
2429 | break; |
2430 | } |
2431 | } |
2432 | if (IsScalarToVector) |
2433 | return createSplat(DAG, dl, VT, V1.getOperand(0)); |
2434 | } |
2435 | return createSplat(DAG, dl, VT, DAG.getConstant(Lane, dl, MVT::i32)); |
2436 | } |
2437 | |
2438 | if (UseHVX) { |
2439 | ArrayRef<int> Mask = SVN->getMask(); |
2440 | size_t MaskLen = Mask.size(); |
2441 | int ElemSizeInBits = VT.getScalarSizeInBits(); |
2442 | if ((Subtarget.useHVXSglOps() && (ElemSizeInBits * MaskLen) == 64 * 8) || |
2443 | (Subtarget.useHVXDblOps() && (ElemSizeInBits * MaskLen) == 128 * 8)) { |
2444 | // Return 1 for odd and 2 of even |
2445 | StridedLoadKind Pattern = isStridedLoad(Mask); |
2446 | |
2447 | if (Pattern == StridedLoadKind::NoPattern) |
2448 | return SDValue(); |
2449 | |
2450 | SDValue Vec0 = Op.getOperand(0); |
2451 | SDValue Vec1 = Op.getOperand(1); |
2452 | SDValue StridePattern = DAG.getConstant(Pattern, dl, MVT::i32); |
2453 | SDValue Ops[] = { Vec1, Vec0, StridePattern }; |
2454 | return DAG.getNode(HexagonISD::VPACK, dl, VT, Ops); |
2455 | } |
2456 | // We used to assert in the "else" part here, but that is bad for Halide |
2457 | // Halide creates intermediate double registers by interleaving two |
2458 | // concatenated vector registers. The interleaving requires vector_shuffle |
2459 | // nodes and we shouldn't barf on a double register result of a |
2460 | // vector_shuffle because it is most likely an intermediate result. |
2461 | } |
2462 | // FIXME: We need to support more general vector shuffles. See |
2463 | // below the comment from the ARM backend that deals in the general |
2464 | // case with the vector shuffles. For now, let expand handle these. |
2465 | return SDValue(); |
2466 | |
2467 | // If the shuffle is not directly supported and it has 4 elements, use |
2468 | // the PerfectShuffle-generated table to synthesize it from other shuffles. |
2469 | } |
2470 | |
2471 | // If BUILD_VECTOR has same base element repeated several times, |
2472 | // report true. |
2473 | static bool isCommonSplatElement(BuildVectorSDNode *BVN) { |
2474 | unsigned NElts = BVN->getNumOperands(); |
2475 | SDValue V0 = BVN->getOperand(0); |
2476 | |
2477 | for (unsigned i = 1, e = NElts; i != e; ++i) { |
2478 | if (BVN->getOperand(i) != V0) |
2479 | return false; |
2480 | } |
2481 | return true; |
2482 | } |
2483 | |
2484 | // Lower a vector shift. Try to convert |
2485 | // <VT> = SHL/SRA/SRL <VT> by <VT> to Hexagon specific |
2486 | // <VT> = SHL/SRA/SRL <VT> by <IT/i32>. |
2487 | SDValue |
2488 | HexagonTargetLowering::LowerVECTOR_SHIFT(SDValue Op, SelectionDAG &DAG) const { |
2489 | BuildVectorSDNode *BVN = nullptr; |
2490 | SDValue V1 = Op.getOperand(0); |
2491 | SDValue V2 = Op.getOperand(1); |
2492 | SDValue V3; |
2493 | SDLoc dl(Op); |
2494 | EVT VT = Op.getValueType(); |
2495 | |
2496 | if ((BVN = dyn_cast<BuildVectorSDNode>(V1.getNode())) && |
2497 | isCommonSplatElement(BVN)) |
2498 | V3 = V2; |
2499 | else if ((BVN = dyn_cast<BuildVectorSDNode>(V2.getNode())) && |
2500 | isCommonSplatElement(BVN)) |
2501 | V3 = V1; |
2502 | else |
2503 | return SDValue(); |
2504 | |
2505 | SDValue CommonSplat = BVN->getOperand(0); |
2506 | SDValue Result; |
2507 | |
2508 | if (VT.getSimpleVT() == MVT::v4i16) { |
2509 | switch (Op.getOpcode()) { |
2510 | case ISD::SRA: |
2511 | Result = DAG.getNode(HexagonISD::VSRAH, dl, VT, V3, CommonSplat); |
2512 | break; |
2513 | case ISD::SHL: |
2514 | Result = DAG.getNode(HexagonISD::VSHLH, dl, VT, V3, CommonSplat); |
2515 | break; |
2516 | case ISD::SRL: |
2517 | Result = DAG.getNode(HexagonISD::VSRLH, dl, VT, V3, CommonSplat); |
2518 | break; |
2519 | default: |
2520 | return SDValue(); |
2521 | } |
2522 | } else if (VT.getSimpleVT() == MVT::v2i32) { |
2523 | switch (Op.getOpcode()) { |
2524 | case ISD::SRA: |
2525 | Result = DAG.getNode(HexagonISD::VSRAW, dl, VT, V3, CommonSplat); |
2526 | break; |
2527 | case ISD::SHL: |
2528 | Result = DAG.getNode(HexagonISD::VSHLW, dl, VT, V3, CommonSplat); |
2529 | break; |
2530 | case ISD::SRL: |
2531 | Result = DAG.getNode(HexagonISD::VSRLW, dl, VT, V3, CommonSplat); |
2532 | break; |
2533 | default: |
2534 | return SDValue(); |
2535 | } |
2536 | } else { |
2537 | return SDValue(); |
2538 | } |
2539 | |
2540 | return DAG.getNode(ISD::BITCAST, dl, VT, Result); |
2541 | } |
2542 | |
2543 | SDValue |
2544 | HexagonTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { |
2545 | BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); |
2546 | SDLoc dl(Op); |
2547 | EVT VT = Op.getValueType(); |
2548 | |
2549 | unsigned Size = VT.getSizeInBits(); |
2550 | |
2551 | // Only handle vectors of 64 bits or shorter. |
2552 | if (Size > 64) |
2553 | return SDValue(); |
2554 | |
2555 | APInt APSplatBits, APSplatUndef; |
2556 | unsigned SplatBitSize; |
2557 | bool HasAnyUndefs; |
2558 | unsigned NElts = BVN->getNumOperands(); |
2559 | |
2560 | // Try to generate a SPLAT instruction. |
2561 | if ((VT.getSimpleVT() == MVT::v4i8 || VT.getSimpleVT() == MVT::v4i16) && |
2562 | (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize, |
2563 | HasAnyUndefs, 0, true) && SplatBitSize <= 16)) { |
2564 | unsigned SplatBits = APSplatBits.getZExtValue(); |
2565 | int32_t SextVal = ((int32_t) (SplatBits << (32 - SplatBitSize)) >> |
2566 | (32 - SplatBitSize)); |
2567 | return createSplat(DAG, dl, VT, DAG.getConstant(SextVal, dl, MVT::i32)); |
2568 | } |
2569 | |
2570 | // Try to generate COMBINE to build v2i32 vectors. |
2571 | if (VT.getSimpleVT() == MVT::v2i32) { |
2572 | SDValue V0 = BVN->getOperand(0); |
2573 | SDValue V1 = BVN->getOperand(1); |
2574 | |
2575 | if (V0.isUndef()) |
2576 | V0 = DAG.getConstant(0, dl, MVT::i32); |
2577 | if (V1.isUndef()) |
2578 | V1 = DAG.getConstant(0, dl, MVT::i32); |
2579 | |
2580 | ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(V0); |
2581 | ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(V1); |
2582 | // If the element isn't a constant, it is in a register: |
2583 | // generate a COMBINE Register Register instruction. |
2584 | if (!C0 || !C1) |
2585 | return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0); |
2586 | |
2587 | // If one of the operands is an 8 bit integer constant, generate |
2588 | // a COMBINE Immediate Immediate instruction. |
2589 | if (isInt<8>(C0->getSExtValue()) || |
2590 | isInt<8>(C1->getSExtValue())) |
2591 | return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0); |
2592 | } |
2593 | |
2594 | // Try to generate a S2_packhl to build v2i16 vectors. |
2595 | if (VT.getSimpleVT() == MVT::v2i16) { |
2596 | for (unsigned i = 0, e = NElts; i != e; ++i) { |
2597 | if (BVN->getOperand(i).isUndef()) |
2598 | continue; |
2599 | ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(BVN->getOperand(i)); |
2600 | // If the element isn't a constant, it is in a register: |
2601 | // generate a S2_packhl instruction. |
2602 | if (!Cst) { |
2603 | SDValue pack = DAG.getNode(HexagonISD::PACKHL, dl, MVT::v4i16, |
2604 | BVN->getOperand(1), BVN->getOperand(0)); |
2605 | |
2606 | return DAG.getTargetExtractSubreg(Hexagon::isub_lo, dl, MVT::v2i16, |
2607 | pack); |
2608 | } |
2609 | } |
2610 | } |
2611 | |
2612 | // In the general case, generate a CONST32 or a CONST64 for constant vectors, |
2613 | // and insert_vector_elt for all the other cases. |
2614 | uint64_t Res = 0; |
2615 | unsigned EltSize = Size / NElts; |
2616 | SDValue ConstVal; |
2617 | uint64_t Mask = ~uint64_t(0ULL) >> (64 - EltSize); |
2618 | bool HasNonConstantElements = false; |
2619 | |
2620 | for (unsigned i = 0, e = NElts; i != e; ++i) { |
2621 | // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon's |
2622 | // combine, const64, etc. are Big Endian. |
2623 | unsigned OpIdx = NElts - i - 1; |
2624 | SDValue Operand = BVN->getOperand(OpIdx); |
2625 | if (Operand.isUndef()) |
2626 | continue; |
2627 | |
2628 | int64_t Val = 0; |
2629 | if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Operand)) |
2630 | Val = Cst->getSExtValue(); |
2631 | else |
2632 | HasNonConstantElements = true; |
2633 | |
2634 | Val &= Mask; |
2635 | Res = (Res << EltSize) | Val; |
2636 | } |
2637 | |
2638 | if (Size > 64) |
2639 | return SDValue(); |
2640 | |
2641 | if (Size == 64) |
2642 | ConstVal = DAG.getConstant(Res, dl, MVT::i64); |
2643 | else |
2644 | ConstVal = DAG.getConstant(Res, dl, MVT::i32); |
2645 | |
2646 | // When there are non constant operands, add them with INSERT_VECTOR_ELT to |
2647 | // ConstVal, the constant part of the vector. |
2648 | if (HasNonConstantElements) { |
2649 | EVT EltVT = VT.getVectorElementType(); |
2650 | SDValue Width = DAG.getConstant(EltVT.getSizeInBits(), dl, MVT::i64); |
2651 | SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width, |
2652 | DAG.getConstant(32, dl, MVT::i64)); |
2653 | |
2654 | for (unsigned i = 0, e = NElts; i != e; ++i) { |
2655 | // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon |
2656 | // is Big Endian. |
2657 | unsigned OpIdx = NElts - i - 1; |
2658 | SDValue Operand = BVN->getOperand(OpIdx); |
2659 | if (isa<ConstantSDNode>(Operand)) |
2660 | // This operand is already in ConstVal. |
2661 | continue; |
2662 | |
2663 | if (VT.getSizeInBits() == 64 && |
2664 | Operand.getValueSizeInBits() == 32) { |
2665 | SDValue C = DAG.getConstant(0, dl, MVT::i32); |
2666 | Operand = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Operand); |
2667 | } |
2668 | |
2669 | SDValue Idx = DAG.getConstant(OpIdx, dl, MVT::i64); |
2670 | SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, Width); |
2671 | SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset); |
2672 | const SDValue Ops[] = {ConstVal, Operand, Combined}; |
2673 | |
2674 | if (VT.getSizeInBits() == 32) |
2675 | ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops); |
2676 | else |
2677 | ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops); |
2678 | } |
2679 | } |
2680 | |
2681 | return DAG.getNode(ISD::BITCAST, dl, VT, ConstVal); |
2682 | } |
2683 | |
2684 | SDValue |
2685 | HexagonTargetLowering::LowerCONCAT_VECTORS(SDValue Op, |
2686 | SelectionDAG &DAG) const { |
2687 | SDLoc dl(Op); |
2688 | bool UseHVX = Subtarget.useHVXOps(); |
2689 | EVT VT = Op.getValueType(); |
2690 | unsigned NElts = Op.getNumOperands(); |
2691 | SDValue Vec0 = Op.getOperand(0); |
2692 | EVT VecVT = Vec0.getValueType(); |
2693 | unsigned Width = VecVT.getSizeInBits(); |
2694 | |
2695 | if (NElts == 2) { |
2696 | MVT ST = VecVT.getSimpleVT(); |
2697 | // We are trying to concat two v2i16 to a single v4i16, or two v4i8 |
2698 | // into a single v8i8. |
2699 | if (ST == MVT::v2i16 || ST == MVT::v4i8) |
2700 | return DAG.getNode(HexagonISD::COMBINE, dl, VT, Op.getOperand(1), Vec0); |
2701 | |
2702 | if (UseHVX) { |
2703 | assert((Width == 64*8 && Subtarget.useHVXSglOps()) ||(((Width == 64*8 && Subtarget.useHVXSglOps()) || (Width == 128*8 && Subtarget.useHVXDblOps())) ? static_cast <void> (0) : __assert_fail ("(Width == 64*8 && Subtarget.useHVXSglOps()) || (Width == 128*8 && Subtarget.useHVXDblOps())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2704, __PRETTY_FUNCTION__)) |
2704 | (Width == 128*8 && Subtarget.useHVXDblOps()))(((Width == 64*8 && Subtarget.useHVXSglOps()) || (Width == 128*8 && Subtarget.useHVXDblOps())) ? static_cast <void> (0) : __assert_fail ("(Width == 64*8 && Subtarget.useHVXSglOps()) || (Width == 128*8 && Subtarget.useHVXDblOps())" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2704, __PRETTY_FUNCTION__)); |
2705 | SDValue Vec1 = Op.getOperand(1); |
2706 | MVT OpTy = Subtarget.useHVXSglOps() ? MVT::v16i32 : MVT::v32i32; |
2707 | MVT ReTy = Subtarget.useHVXSglOps() ? MVT::v32i32 : MVT::v64i32; |
2708 | SDValue B0 = DAG.getNode(ISD::BITCAST, dl, OpTy, Vec0); |
2709 | SDValue B1 = DAG.getNode(ISD::BITCAST, dl, OpTy, Vec1); |
2710 | SDValue VC = DAG.getNode(HexagonISD::VCOMBINE, dl, ReTy, B1, B0); |
2711 | return DAG.getNode(ISD::BITCAST, dl, VT, VC); |
2712 | } |
2713 | } |
2714 | |
2715 | if (VT.getSizeInBits() != 32 && VT.getSizeInBits() != 64) |
2716 | return SDValue(); |
2717 | |
2718 | SDValue C0 = DAG.getConstant(0, dl, MVT::i64); |
2719 | SDValue C32 = DAG.getConstant(32, dl, MVT::i64); |
2720 | SDValue W = DAG.getConstant(Width, dl, MVT::i64); |
2721 | // Create the "width" part of the argument to insert_rp/insertp_rp. |
2722 | SDValue S = DAG.getNode(ISD::SHL, dl, MVT::i64, W, C32); |
2723 | SDValue V = C0; |
2724 | |
2725 | for (unsigned i = 0, e = NElts; i != e; ++i) { |
2726 | unsigned N = NElts-i-1; |
2727 | SDValue OpN = Op.getOperand(N); |
2728 | |
2729 | if (VT.getSizeInBits() == 64 && OpN.getValueSizeInBits() == 32) { |
2730 | SDValue C = DAG.getConstant(0, dl, MVT::i32); |
2731 | OpN = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, OpN); |
2732 | } |
2733 | SDValue Idx = DAG.getConstant(N, dl, MVT::i64); |
2734 | SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, W); |
2735 | SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, S, Offset); |
2736 | if (VT.getSizeInBits() == 32) |
2737 | V = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, {V, OpN, Or}); |
2738 | else if (VT.getSizeInBits() == 64) |
2739 | V = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, {V, OpN, Or}); |
2740 | else |
2741 | return SDValue(); |
2742 | } |
2743 | |
2744 | return DAG.getNode(ISD::BITCAST, dl, VT, V); |
2745 | } |
2746 | |
2747 | SDValue |
2748 | HexagonTargetLowering::LowerEXTRACT_SUBVECTOR_HVX(SDValue Op, |
2749 | SelectionDAG &DAG) const { |
2750 | EVT VT = Op.getOperand(0).getValueType(); |
2751 | SDLoc dl(Op); |
2752 | bool UseHVX = Subtarget.useHVXOps(); |
2753 | bool UseHVXSgl = Subtarget.useHVXSglOps(); |
2754 | // Just in case... |
2755 | |
2756 | if (!VT.isVector() || !UseHVX) |
2757 | return SDValue(); |
2758 | |
2759 | EVT ResVT = Op.getValueType(); |
2760 | unsigned ResSize = ResVT.getSizeInBits(); |
2761 | unsigned VectorSizeInBits = UseHVXSgl ? (64 * 8) : (128 * 8); |
2762 | unsigned OpSize = VT.getSizeInBits(); |
2763 | |
2764 | // We deal only with cases where the result is the vector size |
2765 | // and the vector operand is a double register. |
2766 | if (!(ResVT.isByteSized() && ResSize == VectorSizeInBits) || |
2767 | !(VT.isByteSized() && OpSize == 2 * VectorSizeInBits)) |
2768 | return SDValue(); |
2769 | |
2770 | ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op.getOperand(1)); |
2771 | if (!Cst) |
2772 | return SDValue(); |
2773 | unsigned Val = Cst->getZExtValue(); |
2774 | |
2775 | // These two will get lowered to an appropriate EXTRACT_SUBREG in ISel. |
2776 | if (Val == 0) { |
2777 | SDValue Vec = Op.getOperand(0); |
2778 | return DAG.getTargetExtractSubreg(Hexagon::vsub_lo, dl, ResVT, Vec); |
2779 | } |
2780 | |
2781 | if (ResVT.getVectorNumElements() == Val) { |
2782 | SDValue Vec = Op.getOperand(0); |
2783 | return DAG.getTargetExtractSubreg(Hexagon::vsub_hi, dl, ResVT, Vec); |
2784 | } |
2785 | |
2786 | return SDValue(); |
2787 | } |
2788 | |
2789 | SDValue |
2790 | HexagonTargetLowering::LowerEXTRACT_VECTOR(SDValue Op, |
2791 | SelectionDAG &DAG) const { |
2792 | // If we are dealing with EXTRACT_SUBVECTOR on a HVX type, we may |
2793 | // be able to simplify it to an EXTRACT_SUBREG. |
2794 | if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR && Subtarget.useHVXOps() && |
2795 | isHvxVectorType(Op.getValueType().getSimpleVT())) |
2796 | return LowerEXTRACT_SUBVECTOR_HVX(Op, DAG); |
2797 | |
2798 | EVT VT = Op.getValueType(); |
2799 | int VTN = VT.isVector() ? VT.getVectorNumElements() : 1; |
2800 | SDLoc dl(Op); |
2801 | SDValue Idx = Op.getOperand(1); |
2802 | SDValue Vec = Op.getOperand(0); |
2803 | EVT VecVT = Vec.getValueType(); |
2804 | EVT EltVT = VecVT.getVectorElementType(); |
2805 | int EltSize = EltVT.getSizeInBits(); |
2806 | SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT ? |
2807 | EltSize : VTN * EltSize, dl, MVT::i64); |
2808 | |
2809 | // Constant element number. |
2810 | if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Idx)) { |
2811 | uint64_t X = CI->getZExtValue(); |
2812 | SDValue Offset = DAG.getConstant(X * EltSize, dl, MVT::i32); |
2813 | const SDValue Ops[] = {Vec, Width, Offset}; |
2814 | |
2815 | ConstantSDNode *CW = dyn_cast<ConstantSDNode>(Width); |
2816 | assert(CW && "Non constant width in LowerEXTRACT_VECTOR")((CW && "Non constant width in LowerEXTRACT_VECTOR") ? static_cast<void> (0) : __assert_fail ("CW && \"Non constant width in LowerEXTRACT_VECTOR\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2816, __PRETTY_FUNCTION__)); |
2817 | |
2818 | SDValue N; |
2819 | MVT SVT = VecVT.getSimpleVT(); |
2820 | uint64_t W = CW->getZExtValue(); |
2821 | |
2822 | if (W == 32) { |
2823 | // Translate this node into EXTRACT_SUBREG. |
2824 | unsigned Subreg = (X == 0) ? Hexagon::isub_lo : 0; |
Value stored to 'Subreg' during its initialization is never read | |
2825 | |
2826 | if (X == 0) |
2827 | Subreg = Hexagon::isub_lo; |
2828 | else if (SVT == MVT::v2i32 && X == 1) |
2829 | Subreg = Hexagon::isub_hi; |
2830 | else if (SVT == MVT::v4i16 && X == 2) |
2831 | Subreg = Hexagon::isub_hi; |
2832 | else if (SVT == MVT::v8i8 && X == 4) |
2833 | Subreg = Hexagon::isub_hi; |
2834 | else |
2835 | llvm_unreachable("Bad offset")::llvm::llvm_unreachable_internal("Bad offset", "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2835); |
2836 | N = DAG.getTargetExtractSubreg(Subreg, dl, MVT::i32, Vec); |
2837 | |
2838 | } else if (SVT.getSizeInBits() == 32) { |
2839 | N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i32, Ops); |
2840 | } else if (SVT.getSizeInBits() == 64) { |
2841 | N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i64, Ops); |
2842 | if (VT.getSizeInBits() == 32) |
2843 | N = DAG.getTargetExtractSubreg(Hexagon::isub_lo, dl, MVT::i32, N); |
2844 | } else |
2845 | return SDValue(); |
2846 | |
2847 | return DAG.getNode(ISD::BITCAST, dl, VT, N); |
2848 | } |
2849 | |
2850 | // Variable element number. |
2851 | SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx, |
2852 | DAG.getConstant(EltSize, dl, MVT::i32)); |
2853 | SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width, |
2854 | DAG.getConstant(32, dl, MVT::i64)); |
2855 | SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset); |
2856 | |
2857 | const SDValue Ops[] = {Vec, Combined}; |
2858 | |
2859 | SDValue N; |
2860 | if (VecVT.getSizeInBits() == 32) { |
2861 | N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i32, Ops); |
2862 | } else { |
2863 | N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i64, Ops); |
2864 | if (VT.getSizeInBits() == 32) |
2865 | N = DAG.getTargetExtractSubreg(Hexagon::isub_lo, dl, MVT::i32, N); |
2866 | } |
2867 | return DAG.getNode(ISD::BITCAST, dl, VT, N); |
2868 | } |
2869 | |
2870 | SDValue |
2871 | HexagonTargetLowering::LowerINSERT_VECTOR(SDValue Op, |
2872 | SelectionDAG &DAG) const { |
2873 | EVT VT = Op.getValueType(); |
2874 | int VTN = VT.isVector() ? VT.getVectorNumElements() : 1; |
2875 | SDLoc dl(Op); |
2876 | SDValue Vec = Op.getOperand(0); |
2877 | SDValue Val = Op.getOperand(1); |
2878 | SDValue Idx = Op.getOperand(2); |
2879 | EVT VecVT = Vec.getValueType(); |
2880 | EVT EltVT = VecVT.getVectorElementType(); |
2881 | int EltSize = EltVT.getSizeInBits(); |
2882 | SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::INSERT_VECTOR_ELT ? |
2883 | EltSize : VTN * EltSize, dl, MVT::i64); |
2884 | |
2885 | if (ConstantSDNode *C = cast<ConstantSDNode>(Idx)) { |
2886 | SDValue Offset = DAG.getConstant(C->getSExtValue() * EltSize, dl, MVT::i32); |
2887 | const SDValue Ops[] = {Vec, Val, Width, Offset}; |
2888 | |
2889 | SDValue N; |
2890 | if (VT.getSizeInBits() == 32) |
2891 | N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i32, Ops); |
2892 | else if (VT.getSizeInBits() == 64) |
2893 | N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i64, Ops); |
2894 | else |
2895 | return SDValue(); |
2896 | |
2897 | return DAG.getNode(ISD::BITCAST, dl, VT, N); |
2898 | } |
2899 | |
2900 | // Variable element number. |
2901 | SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx, |
2902 | DAG.getConstant(EltSize, dl, MVT::i32)); |
2903 | SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width, |
2904 | DAG.getConstant(32, dl, MVT::i64)); |
2905 | SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset); |
2906 | |
2907 | if (VT.getSizeInBits() == 64 && Val.getValueSizeInBits() == 32) { |
2908 | SDValue C = DAG.getConstant(0, dl, MVT::i32); |
2909 | Val = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Val); |
2910 | } |
2911 | |
2912 | const SDValue Ops[] = {Vec, Val, Combined}; |
2913 | |
2914 | SDValue N; |
2915 | if (VT.getSizeInBits() == 32) |
2916 | N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops); |
2917 | else if (VT.getSizeInBits() == 64) |
2918 | N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops); |
2919 | else |
2920 | return SDValue(); |
2921 | |
2922 | return DAG.getNode(ISD::BITCAST, dl, VT, N); |
2923 | } |
2924 | |
2925 | bool |
2926 | HexagonTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { |
2927 | // Assuming the caller does not have either a signext or zeroext modifier, and |
2928 | // only one value is accepted, any reasonable truncation is allowed. |
2929 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
2930 | return false; |
2931 | |
2932 | // FIXME: in principle up to 64-bit could be made safe, but it would be very |
2933 | // fragile at the moment: any support for multiple value returns would be |
2934 | // liable to disallow tail calls involving i64 -> iN truncation in many cases. |
2935 | return Ty1->getPrimitiveSizeInBits() <= 32; |
2936 | } |
2937 | |
2938 | SDValue |
2939 | HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { |
2940 | SDValue Chain = Op.getOperand(0); |
2941 | SDValue Offset = Op.getOperand(1); |
2942 | SDValue Handler = Op.getOperand(2); |
2943 | SDLoc dl(Op); |
2944 | auto PtrVT = getPointerTy(DAG.getDataLayout()); |
2945 | |
2946 | // Mark function as containing a call to EH_RETURN. |
2947 | HexagonMachineFunctionInfo *FuncInfo = |
2948 | DAG.getMachineFunction().getInfo<HexagonMachineFunctionInfo>(); |
2949 | FuncInfo->setHasEHReturn(); |
2950 | |
2951 | unsigned OffsetReg = Hexagon::R28; |
2952 | |
2953 | SDValue StoreAddr = |
2954 | DAG.getNode(ISD::ADD, dl, PtrVT, DAG.getRegister(Hexagon::R30, PtrVT), |
2955 | DAG.getIntPtrConstant(4, dl)); |
2956 | Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo()); |
2957 | Chain = DAG.getCopyToReg(Chain, dl, OffsetReg, Offset); |
2958 | |
2959 | // Not needed we already use it as explict input to EH_RETURN. |
2960 | // MF.getRegInfo().addLiveOut(OffsetReg); |
2961 | |
2962 | return DAG.getNode(HexagonISD::EH_RETURN, dl, MVT::Other, Chain); |
2963 | } |
2964 | |
2965 | SDValue |
2966 | HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
2967 | unsigned Opc = Op.getOpcode(); |
2968 | switch (Opc) { |
2969 | default: |
2970 | #ifndef NDEBUG |
2971 | Op.getNode()->dumpr(&DAG); |
2972 | if (Opc > HexagonISD::OP_BEGIN && Opc < HexagonISD::OP_END) |
2973 | errs() << "Check for a non-legal type in this operation\n"; |
2974 | #endif |
2975 | llvm_unreachable("Should not custom lower this!")::llvm::llvm_unreachable_internal("Should not custom lower this!" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 2975); |
2976 | case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); |
2977 | case ISD::INSERT_SUBVECTOR: return LowerINSERT_VECTOR(Op, DAG); |
2978 | case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR(Op, DAG); |
2979 | case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_VECTOR(Op, DAG); |
2980 | case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR(Op, DAG); |
2981 | case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); |
2982 | case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); |
2983 | case ISD::SRA: |
2984 | case ISD::SHL: |
2985 | case ISD::SRL: return LowerVECTOR_SHIFT(Op, DAG); |
2986 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
2987 | case ISD::JumpTable: return LowerJumpTable(Op, DAG); |
2988 | case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); |
2989 | // Frame & Return address. Currently unimplemented. |
2990 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); |
2991 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); |
2992 | case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); |
2993 | case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG); |
2994 | case ISD::GlobalAddress: return LowerGLOBALADDRESS(Op, DAG); |
2995 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
2996 | case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); |
2997 | case ISD::VASTART: return LowerVASTART(Op, DAG); |
2998 | // Custom lower some vector loads. |
2999 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
3000 | case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); |
3001 | case ISD::SETCC: return LowerSETCC(Op, DAG); |
3002 | case ISD::VSELECT: return LowerVSELECT(Op, DAG); |
3003 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
3004 | case ISD::INTRINSIC_VOID: return LowerINTRINSIC_VOID(Op, DAG); |
3005 | case ISD::INLINEASM: return LowerINLINEASM(Op, DAG); |
3006 | case ISD::PREFETCH: return LowerPREFETCH(Op, DAG); |
3007 | case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG); |
3008 | } |
3009 | } |
3010 | |
3011 | /// Returns relocation base for the given PIC jumptable. |
3012 | SDValue |
3013 | HexagonTargetLowering::getPICJumpTableRelocBase(SDValue Table, |
3014 | SelectionDAG &DAG) const { |
3015 | int Idx = cast<JumpTableSDNode>(Table)->getIndex(); |
3016 | EVT VT = Table.getValueType(); |
3017 | SDValue T = DAG.getTargetJumpTable(Idx, VT, HexagonII::MO_PCREL); |
3018 | return DAG.getNode(HexagonISD::AT_PCREL, SDLoc(Table), VT, T); |
3019 | } |
3020 | |
3021 | //===----------------------------------------------------------------------===// |
3022 | // Inline Assembly Support |
3023 | //===----------------------------------------------------------------------===// |
3024 | |
3025 | TargetLowering::ConstraintType |
3026 | HexagonTargetLowering::getConstraintType(StringRef Constraint) const { |
3027 | if (Constraint.size() == 1) { |
3028 | switch (Constraint[0]) { |
3029 | case 'q': |
3030 | case 'v': |
3031 | if (Subtarget.useHVXOps()) |
3032 | return C_Register; |
3033 | break; |
3034 | } |
3035 | } |
3036 | return TargetLowering::getConstraintType(Constraint); |
3037 | } |
3038 | |
3039 | std::pair<unsigned, const TargetRegisterClass*> |
3040 | HexagonTargetLowering::getRegForInlineAsmConstraint( |
3041 | const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { |
3042 | bool UseHVX = Subtarget.useHVXOps(), UseHVXDbl = Subtarget.useHVXDblOps(); |
3043 | |
3044 | if (Constraint.size() == 1) { |
3045 | switch (Constraint[0]) { |
3046 | case 'r': // R0-R31 |
3047 | switch (VT.SimpleTy) { |
3048 | default: |
3049 | llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type")::llvm::llvm_unreachable_internal("getRegForInlineAsmConstraint Unhandled data type" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3049); |
3050 | case MVT::i1: |
3051 | case MVT::i8: |
3052 | case MVT::i16: |
3053 | case MVT::i32: |
3054 | case MVT::f32: |
3055 | return std::make_pair(0U, &Hexagon::IntRegsRegClass); |
3056 | case MVT::i64: |
3057 | case MVT::f64: |
3058 | return std::make_pair(0U, &Hexagon::DoubleRegsRegClass); |
3059 | } |
3060 | case 'q': // q0-q3 |
3061 | switch (VT.getSizeInBits()) { |
3062 | default: |
3063 | llvm_unreachable("getRegForInlineAsmConstraint Unhandled vector size")::llvm::llvm_unreachable_internal("getRegForInlineAsmConstraint Unhandled vector size" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3063); |
3064 | case 512: |
3065 | return std::make_pair(0U, &Hexagon::VecPredRegsRegClass); |
3066 | case 1024: |
3067 | return std::make_pair(0U, &Hexagon::VecPredRegs128BRegClass); |
3068 | } |
3069 | case 'v': // V0-V31 |
3070 | switch (VT.getSizeInBits()) { |
3071 | default: |
3072 | llvm_unreachable("getRegForInlineAsmConstraint Unhandled vector size")::llvm::llvm_unreachable_internal("getRegForInlineAsmConstraint Unhandled vector size" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3072); |
3073 | case 512: |
3074 | return std::make_pair(0U, &Hexagon::VectorRegsRegClass); |
3075 | case 1024: |
3076 | if (Subtarget.hasV60TOps() && UseHVX && UseHVXDbl) |
3077 | return std::make_pair(0U, &Hexagon::VectorRegs128BRegClass); |
3078 | return std::make_pair(0U, &Hexagon::VecDblRegsRegClass); |
3079 | case 2048: |
3080 | return std::make_pair(0U, &Hexagon::VecDblRegs128BRegClass); |
3081 | } |
3082 | |
3083 | default: |
3084 | llvm_unreachable("Unknown asm register class")::llvm::llvm_unreachable_internal("Unknown asm register class" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3084); |
3085 | } |
3086 | } |
3087 | |
3088 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
3089 | } |
3090 | |
3091 | /// isFPImmLegal - Returns true if the target can instruction select the |
3092 | /// specified FP immediate natively. If false, the legalizer will |
3093 | /// materialize the FP immediate as a load from a constant pool. |
3094 | bool HexagonTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const { |
3095 | return Subtarget.hasV5TOps(); |
3096 | } |
3097 | |
3098 | /// isLegalAddressingMode - Return true if the addressing mode represented by |
3099 | /// AM is legal for this target, for a load/store of the specified type. |
3100 | bool HexagonTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
3101 | const AddrMode &AM, Type *Ty, |
3102 | unsigned AS) const { |
3103 | if (Ty->isSized()) { |
3104 | // When LSR detects uses of the same base address to access different |
3105 | // types (e.g. unions), it will assume a conservative type for these |
3106 | // uses: |
3107 | // LSR Use: Kind=Address of void in addrspace(4294967295), ... |
3108 | // The type Ty passed here would then be "void". Skip the alignment |
3109 | // checks, but do not return false right away, since that confuses |
3110 | // LSR into crashing. |
3111 | unsigned A = DL.getABITypeAlignment(Ty); |
3112 | // The base offset must be a multiple of the alignment. |
3113 | if ((AM.BaseOffs % A) != 0) |
3114 | return false; |
3115 | // The shifted offset must fit in 11 bits. |
3116 | if (!isInt<11>(AM.BaseOffs >> Log2_32(A))) |
3117 | return false; |
3118 | } |
3119 | |
3120 | // No global is ever allowed as a base. |
3121 | if (AM.BaseGV) |
3122 | return false; |
3123 | |
3124 | int Scale = AM.Scale; |
3125 | if (Scale < 0) |
3126 | Scale = -Scale; |
3127 | switch (Scale) { |
3128 | case 0: // No scale reg, "r+i", "r", or just "i". |
3129 | break; |
3130 | default: // No scaled addressing mode. |
3131 | return false; |
3132 | } |
3133 | return true; |
3134 | } |
3135 | |
3136 | /// Return true if folding a constant offset with the given GlobalAddress is |
3137 | /// legal. It is frequently not legal in PIC relocation models. |
3138 | bool HexagonTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) |
3139 | const { |
3140 | return HTM.getRelocationModel() == Reloc::Static; |
3141 | } |
3142 | |
3143 | /// isLegalICmpImmediate - Return true if the specified immediate is legal |
3144 | /// icmp immediate, that is the target has icmp instructions which can compare |
3145 | /// a register against the immediate without having to materialize the |
3146 | /// immediate into a register. |
3147 | bool HexagonTargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
3148 | return Imm >= -512 && Imm <= 511; |
3149 | } |
3150 | |
3151 | /// IsEligibleForTailCallOptimization - Check whether the call is eligible |
3152 | /// for tail call optimization. Targets which want to do tail call |
3153 | /// optimization should implement this function. |
3154 | bool HexagonTargetLowering::IsEligibleForTailCallOptimization( |
3155 | SDValue Callee, |
3156 | CallingConv::ID CalleeCC, |
3157 | bool isVarArg, |
3158 | bool isCalleeStructRet, |
3159 | bool isCallerStructRet, |
3160 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
3161 | const SmallVectorImpl<SDValue> &OutVals, |
3162 | const SmallVectorImpl<ISD::InputArg> &Ins, |
3163 | SelectionDAG& DAG) const { |
3164 | const Function *CallerF = DAG.getMachineFunction().getFunction(); |
3165 | CallingConv::ID CallerCC = CallerF->getCallingConv(); |
3166 | bool CCMatch = CallerCC == CalleeCC; |
3167 | |
3168 | // *************************************************************************** |
3169 | // Look for obvious safe cases to perform tail call optimization that do not |
3170 | // require ABI changes. |
3171 | // *************************************************************************** |
3172 | |
3173 | // If this is a tail call via a function pointer, then don't do it! |
3174 | if (!isa<GlobalAddressSDNode>(Callee) && |
3175 | !isa<ExternalSymbolSDNode>(Callee)) { |
3176 | return false; |
3177 | } |
3178 | |
3179 | // Do not optimize if the calling conventions do not match and the conventions |
3180 | // used are not C or Fast. |
3181 | if (!CCMatch) { |
3182 | bool R = (CallerCC == CallingConv::C || CallerCC == CallingConv::Fast); |
3183 | bool E = (CalleeCC == CallingConv::C || CalleeCC == CallingConv::Fast); |
3184 | // If R & E, then ok. |
3185 | if (!R || !E) |
3186 | return false; |
3187 | } |
3188 | |
3189 | // Do not tail call optimize vararg calls. |
3190 | if (isVarArg) |
3191 | return false; |
3192 | |
3193 | // Also avoid tail call optimization if either caller or callee uses struct |
3194 | // return semantics. |
3195 | if (isCalleeStructRet || isCallerStructRet) |
3196 | return false; |
3197 | |
3198 | // In addition to the cases above, we also disable Tail Call Optimization if |
3199 | // the calling convention code that at least one outgoing argument needs to |
3200 | // go on the stack. We cannot check that here because at this point that |
3201 | // information is not available. |
3202 | return true; |
3203 | } |
3204 | |
3205 | /// Returns the target specific optimal type for load and store operations as |
3206 | /// a result of memset, memcpy, and memmove lowering. |
3207 | /// |
3208 | /// If DstAlign is zero that means it's safe to destination alignment can |
3209 | /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't |
3210 | /// a need to check it against alignment requirement, probably because the |
3211 | /// source does not need to be loaded. If 'IsMemset' is true, that means it's |
3212 | /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of |
3213 | /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it |
3214 | /// does not need to be loaded. It returns EVT::Other if the type should be |
3215 | /// determined using generic target-independent logic. |
3216 | EVT HexagonTargetLowering::getOptimalMemOpType(uint64_t Size, |
3217 | unsigned DstAlign, unsigned SrcAlign, bool IsMemset, bool ZeroMemset, |
3218 | bool MemcpyStrSrc, MachineFunction &MF) const { |
3219 | |
3220 | auto Aligned = [](unsigned GivenA, unsigned MinA) -> bool { |
3221 | return (GivenA % MinA) == 0; |
3222 | }; |
3223 | |
3224 | if (Size >= 8 && Aligned(DstAlign, 8) && (IsMemset || Aligned(SrcAlign, 8))) |
3225 | return MVT::i64; |
3226 | if (Size >= 4 && Aligned(DstAlign, 4) && (IsMemset || Aligned(SrcAlign, 4))) |
3227 | return MVT::i32; |
3228 | if (Size >= 2 && Aligned(DstAlign, 2) && (IsMemset || Aligned(SrcAlign, 2))) |
3229 | return MVT::i16; |
3230 | |
3231 | return MVT::Other; |
3232 | } |
3233 | |
3234 | bool HexagonTargetLowering::allowsMisalignedMemoryAccesses(EVT VT, |
3235 | unsigned AS, unsigned Align, bool *Fast) const { |
3236 | if (Fast) |
3237 | *Fast = false; |
3238 | |
3239 | switch (VT.getSimpleVT().SimpleTy) { |
3240 | default: |
3241 | return false; |
3242 | case MVT::v64i8: |
3243 | case MVT::v128i8: |
3244 | case MVT::v256i8: |
3245 | case MVT::v32i16: |
3246 | case MVT::v64i16: |
3247 | case MVT::v128i16: |
3248 | case MVT::v16i32: |
3249 | case MVT::v32i32: |
3250 | case MVT::v64i32: |
3251 | case MVT::v8i64: |
3252 | case MVT::v16i64: |
3253 | case MVT::v32i64: |
3254 | return true; |
3255 | } |
3256 | return false; |
3257 | } |
3258 | |
3259 | std::pair<const TargetRegisterClass*, uint8_t> |
3260 | HexagonTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI, |
3261 | MVT VT) const { |
3262 | const TargetRegisterClass *RRC = nullptr; |
3263 | |
3264 | uint8_t Cost = 1; |
3265 | switch (VT.SimpleTy) { |
3266 | default: |
3267 | return TargetLowering::findRepresentativeClass(TRI, VT); |
3268 | case MVT::v64i8: |
3269 | case MVT::v32i16: |
3270 | case MVT::v16i32: |
3271 | case MVT::v8i64: |
3272 | RRC = &Hexagon::VectorRegsRegClass; |
3273 | break; |
3274 | case MVT::v128i8: |
3275 | case MVT::v64i16: |
3276 | case MVT::v32i32: |
3277 | case MVT::v16i64: |
3278 | if (Subtarget.hasV60TOps() && Subtarget.useHVXOps() && |
3279 | Subtarget.useHVXDblOps()) |
3280 | RRC = &Hexagon::VectorRegs128BRegClass; |
3281 | else |
3282 | RRC = &Hexagon::VecDblRegsRegClass; |
3283 | break; |
3284 | case MVT::v256i8: |
3285 | case MVT::v128i16: |
3286 | case MVT::v64i32: |
3287 | case MVT::v32i64: |
3288 | RRC = &Hexagon::VecDblRegs128BRegClass; |
3289 | break; |
3290 | } |
3291 | return std::make_pair(RRC, Cost); |
3292 | } |
3293 | |
3294 | Value *HexagonTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr, |
3295 | AtomicOrdering Ord) const { |
3296 | BasicBlock *BB = Builder.GetInsertBlock(); |
3297 | Module *M = BB->getParent()->getParent(); |
3298 | Type *Ty = cast<PointerType>(Addr->getType())->getElementType(); |
3299 | unsigned SZ = Ty->getPrimitiveSizeInBits(); |
3300 | assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic loads supported")(((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic loads supported" ) ? static_cast<void> (0) : __assert_fail ("(SZ == 32 || SZ == 64) && \"Only 32/64-bit atomic loads supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3300, __PRETTY_FUNCTION__)); |
3301 | Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_L2_loadw_locked |
3302 | : Intrinsic::hexagon_L4_loadd_locked; |
3303 | Value *Fn = Intrinsic::getDeclaration(M, IntID); |
3304 | return Builder.CreateCall(Fn, Addr, "larx"); |
3305 | } |
3306 | |
3307 | /// Perform a store-conditional operation to Addr. Return the status of the |
3308 | /// store. This should be 0 if the store succeeded, non-zero otherwise. |
3309 | Value *HexagonTargetLowering::emitStoreConditional(IRBuilder<> &Builder, |
3310 | Value *Val, Value *Addr, AtomicOrdering Ord) const { |
3311 | BasicBlock *BB = Builder.GetInsertBlock(); |
3312 | Module *M = BB->getParent()->getParent(); |
3313 | Type *Ty = Val->getType(); |
3314 | unsigned SZ = Ty->getPrimitiveSizeInBits(); |
3315 | assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic stores supported")(((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic stores supported" ) ? static_cast<void> (0) : __assert_fail ("(SZ == 32 || SZ == 64) && \"Only 32/64-bit atomic stores supported\"" , "/tmp/buildd/llvm-toolchain-snapshot-5.0~svn306458/lib/Target/Hexagon/HexagonISelLowering.cpp" , 3315, __PRETTY_FUNCTION__)); |
3316 | Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_S2_storew_locked |
3317 | : Intrinsic::hexagon_S4_stored_locked; |
3318 | Value *Fn = Intrinsic::getDeclaration(M, IntID); |
3319 | Value *Call = Builder.CreateCall(Fn, {Addr, Val}, "stcx"); |
3320 | Value *Cmp = Builder.CreateICmpEQ(Call, Builder.getInt32(0), ""); |
3321 | Value *Ext = Builder.CreateZExt(Cmp, Type::getInt32Ty(M->getContext())); |
3322 | return Ext; |
3323 | } |
3324 | |
3325 | TargetLowering::AtomicExpansionKind |
3326 | HexagonTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const { |
3327 | // Do not expand loads and stores that don't exceed 64 bits. |
3328 | return LI->getType()->getPrimitiveSizeInBits() > 64 |
3329 | ? AtomicExpansionKind::LLOnly |
3330 | : AtomicExpansionKind::None; |
3331 | } |
3332 | |
3333 | bool HexagonTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const { |
3334 | // Do not expand loads and stores that don't exceed 64 bits. |
3335 | return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() > 64; |
3336 | } |
3337 | |
3338 | bool HexagonTargetLowering::shouldExpandAtomicCmpXchgInIR( |
3339 | AtomicCmpXchgInst *AI) const { |
3340 | const DataLayout &DL = AI->getModule()->getDataLayout(); |
3341 | unsigned Size = DL.getTypeStoreSize(AI->getCompareOperand()->getType()); |
3342 | return Size >= 4 && Size <= 8; |
3343 | } |