File: | llvm/include/llvm/CodeGen/SelectionDAGNodes.h |
Warning: | line 1110, column 10 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- RISCVISelLowering.cpp - RISCV DAG Lowering Implementation --------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This file defines the interfaces that RISCV uses to lower LLVM code into a | |||
10 | // selection DAG. | |||
11 | // | |||
12 | //===----------------------------------------------------------------------===// | |||
13 | ||||
14 | #include "RISCVISelLowering.h" | |||
15 | #include "MCTargetDesc/RISCVMatInt.h" | |||
16 | #include "RISCV.h" | |||
17 | #include "RISCVMachineFunctionInfo.h" | |||
18 | #include "RISCVRegisterInfo.h" | |||
19 | #include "RISCVSubtarget.h" | |||
20 | #include "RISCVTargetMachine.h" | |||
21 | #include "llvm/ADT/SmallSet.h" | |||
22 | #include "llvm/ADT/Statistic.h" | |||
23 | #include "llvm/CodeGen/CallingConvLower.h" | |||
24 | #include "llvm/CodeGen/MachineFrameInfo.h" | |||
25 | #include "llvm/CodeGen/MachineFunction.h" | |||
26 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
27 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
28 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" | |||
29 | #include "llvm/CodeGen/ValueTypes.h" | |||
30 | #include "llvm/IR/DiagnosticInfo.h" | |||
31 | #include "llvm/IR/DiagnosticPrinter.h" | |||
32 | #include "llvm/IR/IntrinsicsRISCV.h" | |||
33 | #include "llvm/Support/Debug.h" | |||
34 | #include "llvm/Support/ErrorHandling.h" | |||
35 | #include "llvm/Support/KnownBits.h" | |||
36 | #include "llvm/Support/MathExtras.h" | |||
37 | #include "llvm/Support/raw_ostream.h" | |||
38 | ||||
39 | using namespace llvm; | |||
40 | ||||
41 | #define DEBUG_TYPE"riscv-lower" "riscv-lower" | |||
42 | ||||
43 | STATISTIC(NumTailCalls, "Number of tail calls")static llvm::Statistic NumTailCalls = {"riscv-lower", "NumTailCalls" , "Number of tail calls"}; | |||
44 | ||||
45 | RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM, | |||
46 | const RISCVSubtarget &STI) | |||
47 | : TargetLowering(TM), Subtarget(STI) { | |||
48 | ||||
49 | if (Subtarget.isRV32E()) | |||
50 | report_fatal_error("Codegen not yet implemented for RV32E"); | |||
51 | ||||
52 | RISCVABI::ABI ABI = Subtarget.getTargetABI(); | |||
53 | assert(ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI")((ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI" ) ? static_cast<void> (0) : __assert_fail ("ABI != RISCVABI::ABI_Unknown && \"Improperly initialised target ABI\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 53, __PRETTY_FUNCTION__)); | |||
54 | ||||
55 | if ((ABI == RISCVABI::ABI_ILP32F || ABI == RISCVABI::ABI_LP64F) && | |||
56 | !Subtarget.hasStdExtF()) { | |||
57 | errs() << "Hard-float 'f' ABI can't be used for a target that " | |||
58 | "doesn't support the F instruction set extension (ignoring " | |||
59 | "target-abi)\n"; | |||
60 | ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32; | |||
61 | } else if ((ABI == RISCVABI::ABI_ILP32D || ABI == RISCVABI::ABI_LP64D) && | |||
62 | !Subtarget.hasStdExtD()) { | |||
63 | errs() << "Hard-float 'd' ABI can't be used for a target that " | |||
64 | "doesn't support the D instruction set extension (ignoring " | |||
65 | "target-abi)\n"; | |||
66 | ABI = Subtarget.is64Bit() ? RISCVABI::ABI_LP64 : RISCVABI::ABI_ILP32; | |||
67 | } | |||
68 | ||||
69 | switch (ABI) { | |||
70 | default: | |||
71 | report_fatal_error("Don't know how to lower this ABI"); | |||
72 | case RISCVABI::ABI_ILP32: | |||
73 | case RISCVABI::ABI_ILP32F: | |||
74 | case RISCVABI::ABI_ILP32D: | |||
75 | case RISCVABI::ABI_LP64: | |||
76 | case RISCVABI::ABI_LP64F: | |||
77 | case RISCVABI::ABI_LP64D: | |||
78 | break; | |||
79 | } | |||
80 | ||||
81 | MVT XLenVT = Subtarget.getXLenVT(); | |||
82 | ||||
83 | // Set up the register classes. | |||
84 | addRegisterClass(XLenVT, &RISCV::GPRRegClass); | |||
85 | ||||
86 | if (Subtarget.hasStdExtZfh()) | |||
87 | addRegisterClass(MVT::f16, &RISCV::FPR16RegClass); | |||
88 | if (Subtarget.hasStdExtF()) | |||
89 | addRegisterClass(MVT::f32, &RISCV::FPR32RegClass); | |||
90 | if (Subtarget.hasStdExtD()) | |||
91 | addRegisterClass(MVT::f64, &RISCV::FPR64RegClass); | |||
92 | ||||
93 | static const MVT::SimpleValueType BoolVecVTs[] = { | |||
94 | MVT::nxv1i1, MVT::nxv2i1, MVT::nxv4i1, MVT::nxv8i1, | |||
95 | MVT::nxv16i1, MVT::nxv32i1, MVT::nxv64i1}; | |||
96 | static const MVT::SimpleValueType IntVecVTs[] = { | |||
97 | MVT::nxv1i8, MVT::nxv2i8, MVT::nxv4i8, MVT::nxv8i8, MVT::nxv16i8, | |||
98 | MVT::nxv32i8, MVT::nxv64i8, MVT::nxv1i16, MVT::nxv2i16, MVT::nxv4i16, | |||
99 | MVT::nxv8i16, MVT::nxv16i16, MVT::nxv32i16, MVT::nxv1i32, MVT::nxv2i32, | |||
100 | MVT::nxv4i32, MVT::nxv8i32, MVT::nxv16i32, MVT::nxv1i64, MVT::nxv2i64, | |||
101 | MVT::nxv4i64, MVT::nxv8i64}; | |||
102 | static const MVT::SimpleValueType F16VecVTs[] = { | |||
103 | MVT::nxv1f16, MVT::nxv2f16, MVT::nxv4f16, | |||
104 | MVT::nxv8f16, MVT::nxv16f16, MVT::nxv32f16}; | |||
105 | static const MVT::SimpleValueType F32VecVTs[] = { | |||
106 | MVT::nxv1f32, MVT::nxv2f32, MVT::nxv4f32, MVT::nxv8f32, MVT::nxv16f32}; | |||
107 | static const MVT::SimpleValueType F64VecVTs[] = { | |||
108 | MVT::nxv1f64, MVT::nxv2f64, MVT::nxv4f64, MVT::nxv8f64}; | |||
109 | ||||
110 | if (Subtarget.hasStdExtV()) { | |||
111 | auto addRegClassForRVV = [this](MVT VT) { | |||
112 | unsigned Size = VT.getSizeInBits().getKnownMinValue(); | |||
113 | assert(Size <= 512 && isPowerOf2_32(Size))((Size <= 512 && isPowerOf2_32(Size)) ? static_cast <void> (0) : __assert_fail ("Size <= 512 && isPowerOf2_32(Size)" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 113, __PRETTY_FUNCTION__)); | |||
114 | const TargetRegisterClass *RC; | |||
115 | if (Size <= 64) | |||
116 | RC = &RISCV::VRRegClass; | |||
117 | else if (Size == 128) | |||
118 | RC = &RISCV::VRM2RegClass; | |||
119 | else if (Size == 256) | |||
120 | RC = &RISCV::VRM4RegClass; | |||
121 | else | |||
122 | RC = &RISCV::VRM8RegClass; | |||
123 | ||||
124 | addRegisterClass(VT, RC); | |||
125 | }; | |||
126 | ||||
127 | for (MVT VT : BoolVecVTs) | |||
128 | addRegClassForRVV(VT); | |||
129 | for (MVT VT : IntVecVTs) | |||
130 | addRegClassForRVV(VT); | |||
131 | ||||
132 | if (Subtarget.hasStdExtZfh()) | |||
133 | for (MVT VT : F16VecVTs) | |||
134 | addRegClassForRVV(VT); | |||
135 | ||||
136 | if (Subtarget.hasStdExtF()) | |||
137 | for (MVT VT : F32VecVTs) | |||
138 | addRegClassForRVV(VT); | |||
139 | ||||
140 | if (Subtarget.hasStdExtD()) | |||
141 | for (MVT VT : F64VecVTs) | |||
142 | addRegClassForRVV(VT); | |||
143 | ||||
144 | if (Subtarget.useRVVForFixedLengthVectors()) { | |||
145 | auto addRegClassForFixedVectors = [this](MVT VT) { | |||
146 | unsigned LMul = Subtarget.getLMULForFixedLengthVector(VT); | |||
147 | const TargetRegisterClass *RC; | |||
148 | if (LMul == 1) | |||
149 | RC = &RISCV::VRRegClass; | |||
150 | else if (LMul == 2) | |||
151 | RC = &RISCV::VRM2RegClass; | |||
152 | else if (LMul == 4) | |||
153 | RC = &RISCV::VRM4RegClass; | |||
154 | else if (LMul == 8) | |||
155 | RC = &RISCV::VRM8RegClass; | |||
156 | else | |||
157 | llvm_unreachable("Unexpected LMul!")::llvm::llvm_unreachable_internal("Unexpected LMul!", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 157); | |||
158 | ||||
159 | addRegisterClass(VT, RC); | |||
160 | }; | |||
161 | for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) | |||
162 | if (useRVVForFixedLengthVectorVT(VT)) | |||
163 | addRegClassForFixedVectors(VT); | |||
164 | ||||
165 | for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) | |||
166 | if (useRVVForFixedLengthVectorVT(VT)) | |||
167 | addRegClassForFixedVectors(VT); | |||
168 | } | |||
169 | } | |||
170 | ||||
171 | // Compute derived properties from the register classes. | |||
172 | computeRegisterProperties(STI.getRegisterInfo()); | |||
173 | ||||
174 | setStackPointerRegisterToSaveRestore(RISCV::X2); | |||
175 | ||||
176 | for (auto N : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) | |||
177 | setLoadExtAction(N, XLenVT, MVT::i1, Promote); | |||
178 | ||||
179 | // TODO: add all necessary setOperationAction calls. | |||
180 | setOperationAction(ISD::DYNAMIC_STACKALLOC, XLenVT, Expand); | |||
181 | ||||
182 | setOperationAction(ISD::BR_JT, MVT::Other, Expand); | |||
183 | setOperationAction(ISD::BR_CC, XLenVT, Expand); | |||
184 | setOperationAction(ISD::SELECT_CC, XLenVT, Expand); | |||
185 | ||||
186 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); | |||
187 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); | |||
188 | ||||
189 | setOperationAction(ISD::VASTART, MVT::Other, Custom); | |||
190 | setOperationAction(ISD::VAARG, MVT::Other, Expand); | |||
191 | setOperationAction(ISD::VACOPY, MVT::Other, Expand); | |||
192 | setOperationAction(ISD::VAEND, MVT::Other, Expand); | |||
193 | ||||
194 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); | |||
195 | if (!Subtarget.hasStdExtZbb()) { | |||
196 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); | |||
197 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); | |||
198 | } | |||
199 | ||||
200 | if (Subtarget.is64Bit()) { | |||
201 | setOperationAction(ISD::ADD, MVT::i32, Custom); | |||
202 | setOperationAction(ISD::SUB, MVT::i32, Custom); | |||
203 | setOperationAction(ISD::SHL, MVT::i32, Custom); | |||
204 | setOperationAction(ISD::SRA, MVT::i32, Custom); | |||
205 | setOperationAction(ISD::SRL, MVT::i32, Custom); | |||
206 | } | |||
207 | ||||
208 | if (!Subtarget.hasStdExtM()) { | |||
209 | setOperationAction(ISD::MUL, XLenVT, Expand); | |||
210 | setOperationAction(ISD::MULHS, XLenVT, Expand); | |||
211 | setOperationAction(ISD::MULHU, XLenVT, Expand); | |||
212 | setOperationAction(ISD::SDIV, XLenVT, Expand); | |||
213 | setOperationAction(ISD::UDIV, XLenVT, Expand); | |||
214 | setOperationAction(ISD::SREM, XLenVT, Expand); | |||
215 | setOperationAction(ISD::UREM, XLenVT, Expand); | |||
216 | } | |||
217 | ||||
218 | if (Subtarget.is64Bit() && Subtarget.hasStdExtM()) { | |||
219 | setOperationAction(ISD::MUL, MVT::i32, Custom); | |||
220 | ||||
221 | setOperationAction(ISD::SDIV, MVT::i8, Custom); | |||
222 | setOperationAction(ISD::UDIV, MVT::i8, Custom); | |||
223 | setOperationAction(ISD::UREM, MVT::i8, Custom); | |||
224 | setOperationAction(ISD::SDIV, MVT::i16, Custom); | |||
225 | setOperationAction(ISD::UDIV, MVT::i16, Custom); | |||
226 | setOperationAction(ISD::UREM, MVT::i16, Custom); | |||
227 | setOperationAction(ISD::SDIV, MVT::i32, Custom); | |||
228 | setOperationAction(ISD::UDIV, MVT::i32, Custom); | |||
229 | setOperationAction(ISD::UREM, MVT::i32, Custom); | |||
230 | } | |||
231 | ||||
232 | setOperationAction(ISD::SDIVREM, XLenVT, Expand); | |||
233 | setOperationAction(ISD::UDIVREM, XLenVT, Expand); | |||
234 | setOperationAction(ISD::SMUL_LOHI, XLenVT, Expand); | |||
235 | setOperationAction(ISD::UMUL_LOHI, XLenVT, Expand); | |||
236 | ||||
237 | setOperationAction(ISD::SHL_PARTS, XLenVT, Custom); | |||
238 | setOperationAction(ISD::SRL_PARTS, XLenVT, Custom); | |||
239 | setOperationAction(ISD::SRA_PARTS, XLenVT, Custom); | |||
240 | ||||
241 | if (Subtarget.hasStdExtZbb() || Subtarget.hasStdExtZbp()) { | |||
242 | if (Subtarget.is64Bit()) { | |||
243 | setOperationAction(ISD::ROTL, MVT::i32, Custom); | |||
244 | setOperationAction(ISD::ROTR, MVT::i32, Custom); | |||
245 | } | |||
246 | } else { | |||
247 | setOperationAction(ISD::ROTL, XLenVT, Expand); | |||
248 | setOperationAction(ISD::ROTR, XLenVT, Expand); | |||
249 | } | |||
250 | ||||
251 | if (Subtarget.hasStdExtZbp()) { | |||
252 | // Custom lower bswap/bitreverse so we can convert them to GREVI to enable | |||
253 | // more combining. | |||
254 | setOperationAction(ISD::BITREVERSE, XLenVT, Custom); | |||
255 | setOperationAction(ISD::BSWAP, XLenVT, Custom); | |||
256 | ||||
257 | if (Subtarget.is64Bit()) { | |||
258 | setOperationAction(ISD::BITREVERSE, MVT::i32, Custom); | |||
259 | setOperationAction(ISD::BSWAP, MVT::i32, Custom); | |||
260 | } | |||
261 | } else { | |||
262 | // With Zbb we have an XLen rev8 instruction, but not GREVI. So we'll | |||
263 | // pattern match it directly in isel. | |||
264 | setOperationAction(ISD::BSWAP, XLenVT, | |||
265 | Subtarget.hasStdExtZbb() ? Legal : Expand); | |||
266 | } | |||
267 | ||||
268 | if (Subtarget.hasStdExtZbb()) { | |||
269 | setOperationAction(ISD::SMIN, XLenVT, Legal); | |||
270 | setOperationAction(ISD::SMAX, XLenVT, Legal); | |||
271 | setOperationAction(ISD::UMIN, XLenVT, Legal); | |||
272 | setOperationAction(ISD::UMAX, XLenVT, Legal); | |||
273 | } else { | |||
274 | setOperationAction(ISD::CTTZ, XLenVT, Expand); | |||
275 | setOperationAction(ISD::CTLZ, XLenVT, Expand); | |||
276 | setOperationAction(ISD::CTPOP, XLenVT, Expand); | |||
277 | } | |||
278 | ||||
279 | if (Subtarget.hasStdExtZbt()) { | |||
280 | setOperationAction(ISD::FSHL, XLenVT, Custom); | |||
281 | setOperationAction(ISD::FSHR, XLenVT, Custom); | |||
282 | setOperationAction(ISD::SELECT, XLenVT, Legal); | |||
283 | ||||
284 | if (Subtarget.is64Bit()) { | |||
285 | setOperationAction(ISD::FSHL, MVT::i32, Custom); | |||
286 | setOperationAction(ISD::FSHR, MVT::i32, Custom); | |||
287 | } | |||
288 | } else { | |||
289 | setOperationAction(ISD::SELECT, XLenVT, Custom); | |||
290 | } | |||
291 | ||||
292 | ISD::CondCode FPCCToExpand[] = { | |||
293 | ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT, | |||
294 | ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUNE, ISD::SETGT, | |||
295 | ISD::SETGE, ISD::SETNE, ISD::SETO, ISD::SETUO}; | |||
296 | ||||
297 | ISD::NodeType FPOpToExpand[] = { | |||
298 | ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM, ISD::FP16_TO_FP, | |||
299 | ISD::FP_TO_FP16}; | |||
300 | ||||
301 | if (Subtarget.hasStdExtZfh()) | |||
302 | setOperationAction(ISD::BITCAST, MVT::i16, Custom); | |||
303 | ||||
304 | if (Subtarget.hasStdExtZfh()) { | |||
305 | setOperationAction(ISD::FMINNUM, MVT::f16, Legal); | |||
306 | setOperationAction(ISD::FMAXNUM, MVT::f16, Legal); | |||
307 | for (auto CC : FPCCToExpand) | |||
308 | setCondCodeAction(CC, MVT::f16, Expand); | |||
309 | setOperationAction(ISD::SELECT_CC, MVT::f16, Expand); | |||
310 | setOperationAction(ISD::SELECT, MVT::f16, Custom); | |||
311 | setOperationAction(ISD::BR_CC, MVT::f16, Expand); | |||
312 | for (auto Op : FPOpToExpand) | |||
313 | setOperationAction(Op, MVT::f16, Expand); | |||
314 | } | |||
315 | ||||
316 | if (Subtarget.hasStdExtF()) { | |||
317 | setOperationAction(ISD::FMINNUM, MVT::f32, Legal); | |||
318 | setOperationAction(ISD::FMAXNUM, MVT::f32, Legal); | |||
319 | for (auto CC : FPCCToExpand) | |||
320 | setCondCodeAction(CC, MVT::f32, Expand); | |||
321 | setOperationAction(ISD::SELECT_CC, MVT::f32, Expand); | |||
322 | setOperationAction(ISD::SELECT, MVT::f32, Custom); | |||
323 | setOperationAction(ISD::BR_CC, MVT::f32, Expand); | |||
324 | for (auto Op : FPOpToExpand) | |||
325 | setOperationAction(Op, MVT::f32, Expand); | |||
326 | setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand); | |||
327 | setTruncStoreAction(MVT::f32, MVT::f16, Expand); | |||
328 | } | |||
329 | ||||
330 | if (Subtarget.hasStdExtF() && Subtarget.is64Bit()) | |||
331 | setOperationAction(ISD::BITCAST, MVT::i32, Custom); | |||
332 | ||||
333 | if (Subtarget.hasStdExtD()) { | |||
334 | setOperationAction(ISD::FMINNUM, MVT::f64, Legal); | |||
335 | setOperationAction(ISD::FMAXNUM, MVT::f64, Legal); | |||
336 | for (auto CC : FPCCToExpand) | |||
337 | setCondCodeAction(CC, MVT::f64, Expand); | |||
338 | setOperationAction(ISD::SELECT_CC, MVT::f64, Expand); | |||
339 | setOperationAction(ISD::SELECT, MVT::f64, Custom); | |||
340 | setOperationAction(ISD::BR_CC, MVT::f64, Expand); | |||
341 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); | |||
342 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); | |||
343 | for (auto Op : FPOpToExpand) | |||
344 | setOperationAction(Op, MVT::f64, Expand); | |||
345 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand); | |||
346 | setTruncStoreAction(MVT::f64, MVT::f16, Expand); | |||
347 | } | |||
348 | ||||
349 | if (Subtarget.is64Bit()) { | |||
350 | setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); | |||
351 | setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); | |||
352 | setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom); | |||
353 | setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom); | |||
354 | } | |||
355 | ||||
356 | setOperationAction(ISD::GlobalAddress, XLenVT, Custom); | |||
357 | setOperationAction(ISD::BlockAddress, XLenVT, Custom); | |||
358 | setOperationAction(ISD::ConstantPool, XLenVT, Custom); | |||
359 | setOperationAction(ISD::JumpTable, XLenVT, Custom); | |||
360 | ||||
361 | setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom); | |||
362 | ||||
363 | // TODO: On M-mode only targets, the cycle[h] CSR may not be present. | |||
364 | // Unfortunately this can't be determined just from the ISA naming string. | |||
365 | setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, | |||
366 | Subtarget.is64Bit() ? Legal : Custom); | |||
367 | ||||
368 | setOperationAction(ISD::TRAP, MVT::Other, Legal); | |||
369 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); | |||
370 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); | |||
371 | ||||
372 | if (Subtarget.hasStdExtA()) { | |||
373 | setMaxAtomicSizeInBitsSupported(Subtarget.getXLen()); | |||
374 | setMinCmpXchgSizeInBits(32); | |||
375 | } else { | |||
376 | setMaxAtomicSizeInBitsSupported(0); | |||
377 | } | |||
378 | ||||
379 | setBooleanContents(ZeroOrOneBooleanContent); | |||
380 | ||||
381 | if (Subtarget.hasStdExtV()) { | |||
382 | setBooleanVectorContents(ZeroOrOneBooleanContent); | |||
383 | ||||
384 | setOperationAction(ISD::VSCALE, XLenVT, Custom); | |||
385 | ||||
386 | // RVV intrinsics may have illegal operands. | |||
387 | // We also need to custom legalize vmv.x.s. | |||
388 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom); | |||
389 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom); | |||
390 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom); | |||
391 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i16, Custom); | |||
392 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i32, Custom); | |||
393 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i32, Custom); | |||
394 | ||||
395 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); | |||
396 | ||||
397 | if (Subtarget.is64Bit()) { | |||
398 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom); | |||
399 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom); | |||
400 | } else { | |||
401 | // We must custom-lower certain vXi64 operations on RV32 due to the vector | |||
402 | // element type being illegal. | |||
403 | setOperationAction(ISD::SPLAT_VECTOR, MVT::i64, Custom); | |||
404 | setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::i64, Custom); | |||
405 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::i64, Custom); | |||
406 | ||||
407 | setOperationAction(ISD::VECREDUCE_ADD, MVT::i64, Custom); | |||
408 | setOperationAction(ISD::VECREDUCE_AND, MVT::i64, Custom); | |||
409 | setOperationAction(ISD::VECREDUCE_OR, MVT::i64, Custom); | |||
410 | setOperationAction(ISD::VECREDUCE_XOR, MVT::i64, Custom); | |||
411 | setOperationAction(ISD::VECREDUCE_SMAX, MVT::i64, Custom); | |||
412 | setOperationAction(ISD::VECREDUCE_SMIN, MVT::i64, Custom); | |||
413 | setOperationAction(ISD::VECREDUCE_UMAX, MVT::i64, Custom); | |||
414 | setOperationAction(ISD::VECREDUCE_UMIN, MVT::i64, Custom); | |||
415 | } | |||
416 | ||||
417 | for (MVT VT : BoolVecVTs) { | |||
418 | setOperationAction(ISD::SPLAT_VECTOR, VT, Legal); | |||
419 | ||||
420 | // Mask VTs are custom-expanded into a series of standard nodes | |||
421 | setOperationAction(ISD::TRUNCATE, VT, Custom); | |||
422 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
423 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
424 | } | |||
425 | ||||
426 | for (MVT VT : IntVecVTs) { | |||
427 | setOperationAction(ISD::SPLAT_VECTOR, VT, Legal); | |||
428 | ||||
429 | setOperationAction(ISD::SMIN, VT, Legal); | |||
430 | setOperationAction(ISD::SMAX, VT, Legal); | |||
431 | setOperationAction(ISD::UMIN, VT, Legal); | |||
432 | setOperationAction(ISD::UMAX, VT, Legal); | |||
433 | ||||
434 | if (!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64) | |||
435 | setOperationAction(ISD::ABS, VT, Custom); | |||
436 | ||||
437 | setOperationAction(ISD::ROTL, VT, Expand); | |||
438 | setOperationAction(ISD::ROTR, VT, Expand); | |||
439 | ||||
440 | // Custom-lower extensions and truncations from/to mask types. | |||
441 | setOperationAction(ISD::ANY_EXTEND, VT, Custom); | |||
442 | setOperationAction(ISD::SIGN_EXTEND, VT, Custom); | |||
443 | setOperationAction(ISD::ZERO_EXTEND, VT, Custom); | |||
444 | ||||
445 | // RVV has native int->float & float->int conversions where the | |||
446 | // element type sizes are within one power-of-two of each other. Any | |||
447 | // wider distances between type sizes have to be lowered as sequences | |||
448 | // which progressively narrow the gap in stages. | |||
449 | setOperationAction(ISD::SINT_TO_FP, VT, Custom); | |||
450 | setOperationAction(ISD::UINT_TO_FP, VT, Custom); | |||
451 | setOperationAction(ISD::FP_TO_SINT, VT, Custom); | |||
452 | setOperationAction(ISD::FP_TO_UINT, VT, Custom); | |||
453 | ||||
454 | // Integer VTs are lowered as a series of "RISCVISD::TRUNCATE_VECTOR_VL" | |||
455 | // nodes which truncate by one power of two at a time. | |||
456 | setOperationAction(ISD::TRUNCATE, VT, Custom); | |||
457 | ||||
458 | // Custom-lower insert/extract operations to simplify patterns. | |||
459 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
460 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
461 | ||||
462 | // Custom-lower reduction operations to set up the corresponding custom | |||
463 | // nodes' operands. | |||
464 | setOperationAction(ISD::VECREDUCE_ADD, VT, Custom); | |||
465 | setOperationAction(ISD::VECREDUCE_AND, VT, Custom); | |||
466 | setOperationAction(ISD::VECREDUCE_OR, VT, Custom); | |||
467 | setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); | |||
468 | setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom); | |||
469 | setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom); | |||
470 | setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom); | |||
471 | setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom); | |||
472 | ||||
473 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
474 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
475 | ||||
476 | setOperationAction(ISD::VECTOR_REVERSE, VT, Custom); | |||
477 | } | |||
478 | ||||
479 | // Expand various CCs to best match the RVV ISA, which natively supports UNE | |||
480 | // but no other unordered comparisons, and supports all ordered comparisons | |||
481 | // except ONE. Additionally, we expand GT,OGT,GE,OGE for optimization | |||
482 | // purposes; they are expanded to their swapped-operand CCs (LT,OLT,LE,OLE), | |||
483 | // and we pattern-match those back to the "original", swapping operands once | |||
484 | // more. This way we catch both operations and both "vf" and "fv" forms with | |||
485 | // fewer patterns. | |||
486 | ISD::CondCode VFPCCToExpand[] = { | |||
487 | ISD::SETO, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT, | |||
488 | ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUO, | |||
489 | ISD::SETGT, ISD::SETOGT, ISD::SETGE, ISD::SETOGE, | |||
490 | }; | |||
491 | ||||
492 | // Sets common operation actions on RVV floating-point vector types. | |||
493 | const auto SetCommonVFPActions = [&](MVT VT) { | |||
494 | setOperationAction(ISD::SPLAT_VECTOR, VT, Legal); | |||
495 | // RVV has native FP_ROUND & FP_EXTEND conversions where the element type | |||
496 | // sizes are within one power-of-two of each other. Therefore conversions | |||
497 | // between vXf16 and vXf64 must be lowered as sequences which convert via | |||
498 | // vXf32. | |||
499 | setOperationAction(ISD::FP_ROUND, VT, Custom); | |||
500 | setOperationAction(ISD::FP_EXTEND, VT, Custom); | |||
501 | // Custom-lower insert/extract operations to simplify patterns. | |||
502 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
503 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
504 | // Expand various condition codes (explained above). | |||
505 | for (auto CC : VFPCCToExpand) | |||
506 | setCondCodeAction(CC, VT, Expand); | |||
507 | ||||
508 | setOperationAction(ISD::VECREDUCE_FADD, VT, Custom); | |||
509 | setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom); | |||
510 | setOperationAction(ISD::FCOPYSIGN, VT, Legal); | |||
511 | ||||
512 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
513 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
514 | ||||
515 | setOperationAction(ISD::VECTOR_REVERSE, VT, Custom); | |||
516 | }; | |||
517 | ||||
518 | if (Subtarget.hasStdExtZfh()) | |||
519 | for (MVT VT : F16VecVTs) | |||
520 | SetCommonVFPActions(VT); | |||
521 | ||||
522 | if (Subtarget.hasStdExtF()) | |||
523 | for (MVT VT : F32VecVTs) | |||
524 | SetCommonVFPActions(VT); | |||
525 | ||||
526 | if (Subtarget.hasStdExtD()) | |||
527 | for (MVT VT : F64VecVTs) | |||
528 | SetCommonVFPActions(VT); | |||
529 | ||||
530 | if (Subtarget.useRVVForFixedLengthVectors()) { | |||
531 | for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) { | |||
532 | if (!useRVVForFixedLengthVectorVT(VT)) | |||
533 | continue; | |||
534 | ||||
535 | // By default everything must be expanded. | |||
536 | for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) | |||
537 | setOperationAction(Op, VT, Expand); | |||
538 | for (MVT OtherVT : MVT::fixedlen_vector_valuetypes()) | |||
539 | setTruncStoreAction(VT, OtherVT, Expand); | |||
540 | ||||
541 | // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed. | |||
542 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
543 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
544 | ||||
545 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
546 | setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); | |||
547 | ||||
548 | setOperationAction(ISD::LOAD, VT, Custom); | |||
549 | setOperationAction(ISD::STORE, VT, Custom); | |||
550 | ||||
551 | setOperationAction(ISD::SETCC, VT, Custom); | |||
552 | ||||
553 | setOperationAction(ISD::TRUNCATE, VT, Custom); | |||
554 | ||||
555 | // Operations below are different for between masks and other vectors. | |||
556 | if (VT.getVectorElementType() == MVT::i1) { | |||
557 | setOperationAction(ISD::AND, VT, Custom); | |||
558 | setOperationAction(ISD::OR, VT, Custom); | |||
559 | setOperationAction(ISD::XOR, VT, Custom); | |||
560 | continue; | |||
561 | } | |||
562 | ||||
563 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
564 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
565 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
566 | ||||
567 | setOperationAction(ISD::ADD, VT, Custom); | |||
568 | setOperationAction(ISD::MUL, VT, Custom); | |||
569 | setOperationAction(ISD::SUB, VT, Custom); | |||
570 | setOperationAction(ISD::AND, VT, Custom); | |||
571 | setOperationAction(ISD::OR, VT, Custom); | |||
572 | setOperationAction(ISD::XOR, VT, Custom); | |||
573 | setOperationAction(ISD::SDIV, VT, Custom); | |||
574 | setOperationAction(ISD::SREM, VT, Custom); | |||
575 | setOperationAction(ISD::UDIV, VT, Custom); | |||
576 | setOperationAction(ISD::UREM, VT, Custom); | |||
577 | setOperationAction(ISD::SHL, VT, Custom); | |||
578 | setOperationAction(ISD::SRA, VT, Custom); | |||
579 | setOperationAction(ISD::SRL, VT, Custom); | |||
580 | ||||
581 | setOperationAction(ISD::SMIN, VT, Custom); | |||
582 | setOperationAction(ISD::SMAX, VT, Custom); | |||
583 | setOperationAction(ISD::UMIN, VT, Custom); | |||
584 | setOperationAction(ISD::UMAX, VT, Custom); | |||
585 | setOperationAction(ISD::ABS, VT, Custom); | |||
586 | ||||
587 | setOperationAction(ISD::MULHS, VT, Custom); | |||
588 | setOperationAction(ISD::MULHU, VT, Custom); | |||
589 | ||||
590 | setOperationAction(ISD::SINT_TO_FP, VT, Custom); | |||
591 | setOperationAction(ISD::UINT_TO_FP, VT, Custom); | |||
592 | setOperationAction(ISD::FP_TO_SINT, VT, Custom); | |||
593 | setOperationAction(ISD::FP_TO_UINT, VT, Custom); | |||
594 | ||||
595 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
596 | ||||
597 | setOperationAction(ISD::ANY_EXTEND, VT, Custom); | |||
598 | setOperationAction(ISD::SIGN_EXTEND, VT, Custom); | |||
599 | setOperationAction(ISD::ZERO_EXTEND, VT, Custom); | |||
600 | ||||
601 | setOperationAction(ISD::BITCAST, VT, Custom); | |||
602 | ||||
603 | // Custom-lower reduction operations to set up the corresponding custom | |||
604 | // nodes' operands. | |||
605 | setOperationAction(ISD::VECREDUCE_ADD, VT, Custom); | |||
606 | setOperationAction(ISD::VECREDUCE_AND, VT, Custom); | |||
607 | setOperationAction(ISD::VECREDUCE_OR, VT, Custom); | |||
608 | setOperationAction(ISD::VECREDUCE_XOR, VT, Custom); | |||
609 | setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom); | |||
610 | setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom); | |||
611 | setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom); | |||
612 | setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom); | |||
613 | } | |||
614 | ||||
615 | for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) { | |||
616 | if (!useRVVForFixedLengthVectorVT(VT)) | |||
617 | continue; | |||
618 | ||||
619 | // By default everything must be expanded. | |||
620 | for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) | |||
621 | setOperationAction(Op, VT, Expand); | |||
622 | for (MVT OtherVT : MVT::fp_fixedlen_vector_valuetypes()) { | |||
623 | setLoadExtAction(ISD::EXTLOAD, OtherVT, VT, Expand); | |||
624 | setTruncStoreAction(VT, OtherVT, Expand); | |||
625 | } | |||
626 | ||||
627 | // We use EXTRACT_SUBVECTOR as a "cast" from scalable to fixed. | |||
628 | setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom); | |||
629 | setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom); | |||
630 | ||||
631 | setOperationAction(ISD::BUILD_VECTOR, VT, Custom); | |||
632 | setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); | |||
633 | setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom); | |||
634 | setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); | |||
635 | ||||
636 | setOperationAction(ISD::LOAD, VT, Custom); | |||
637 | setOperationAction(ISD::STORE, VT, Custom); | |||
638 | setOperationAction(ISD::FADD, VT, Custom); | |||
639 | setOperationAction(ISD::FSUB, VT, Custom); | |||
640 | setOperationAction(ISD::FMUL, VT, Custom); | |||
641 | setOperationAction(ISD::FDIV, VT, Custom); | |||
642 | setOperationAction(ISD::FNEG, VT, Custom); | |||
643 | setOperationAction(ISD::FABS, VT, Custom); | |||
644 | setOperationAction(ISD::FSQRT, VT, Custom); | |||
645 | setOperationAction(ISD::FMA, VT, Custom); | |||
646 | ||||
647 | setOperationAction(ISD::FP_ROUND, VT, Custom); | |||
648 | setOperationAction(ISD::FP_EXTEND, VT, Custom); | |||
649 | ||||
650 | for (auto CC : VFPCCToExpand) | |||
651 | setCondCodeAction(CC, VT, Expand); | |||
652 | ||||
653 | setOperationAction(ISD::VSELECT, VT, Custom); | |||
654 | ||||
655 | setOperationAction(ISD::BITCAST, VT, Custom); | |||
656 | ||||
657 | setOperationAction(ISD::VECREDUCE_FADD, VT, Custom); | |||
658 | setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom); | |||
659 | } | |||
660 | } | |||
661 | } | |||
662 | ||||
663 | // Function alignments. | |||
664 | const Align FunctionAlignment(Subtarget.hasStdExtC() ? 2 : 4); | |||
665 | setMinFunctionAlignment(FunctionAlignment); | |||
666 | setPrefFunctionAlignment(FunctionAlignment); | |||
667 | ||||
668 | setMinimumJumpTableEntries(5); | |||
669 | ||||
670 | // Jumps are expensive, compared to logic | |||
671 | setJumpIsExpensive(); | |||
672 | ||||
673 | // We can use any register for comparisons | |||
674 | setHasMultipleConditionRegisters(); | |||
675 | ||||
676 | setTargetDAGCombine(ISD::SETCC); | |||
677 | if (Subtarget.hasStdExtZbp()) { | |||
678 | setTargetDAGCombine(ISD::OR); | |||
679 | } | |||
680 | if (Subtarget.hasStdExtV()) | |||
681 | setTargetDAGCombine(ISD::FCOPYSIGN); | |||
682 | } | |||
683 | ||||
684 | EVT RISCVTargetLowering::getSetCCResultType(const DataLayout &DL, | |||
685 | LLVMContext &Context, | |||
686 | EVT VT) const { | |||
687 | if (!VT.isVector()) | |||
688 | return getPointerTy(DL); | |||
689 | if (Subtarget.hasStdExtV() && | |||
690 | (VT.isScalableVector() || Subtarget.useRVVForFixedLengthVectors())) | |||
691 | return EVT::getVectorVT(Context, MVT::i1, VT.getVectorElementCount()); | |||
692 | return VT.changeVectorElementTypeToInteger(); | |||
693 | } | |||
694 | ||||
695 | bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, | |||
696 | const CallInst &I, | |||
697 | MachineFunction &MF, | |||
698 | unsigned Intrinsic) const { | |||
699 | switch (Intrinsic) { | |||
700 | default: | |||
701 | return false; | |||
702 | case Intrinsic::riscv_masked_atomicrmw_xchg_i32: | |||
703 | case Intrinsic::riscv_masked_atomicrmw_add_i32: | |||
704 | case Intrinsic::riscv_masked_atomicrmw_sub_i32: | |||
705 | case Intrinsic::riscv_masked_atomicrmw_nand_i32: | |||
706 | case Intrinsic::riscv_masked_atomicrmw_max_i32: | |||
707 | case Intrinsic::riscv_masked_atomicrmw_min_i32: | |||
708 | case Intrinsic::riscv_masked_atomicrmw_umax_i32: | |||
709 | case Intrinsic::riscv_masked_atomicrmw_umin_i32: | |||
710 | case Intrinsic::riscv_masked_cmpxchg_i32: | |||
711 | PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType()); | |||
712 | Info.opc = ISD::INTRINSIC_W_CHAIN; | |||
713 | Info.memVT = MVT::getVT(PtrTy->getElementType()); | |||
714 | Info.ptrVal = I.getArgOperand(0); | |||
715 | Info.offset = 0; | |||
716 | Info.align = Align(4); | |||
717 | Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore | | |||
718 | MachineMemOperand::MOVolatile; | |||
719 | return true; | |||
720 | } | |||
721 | } | |||
722 | ||||
723 | bool RISCVTargetLowering::isLegalAddressingMode(const DataLayout &DL, | |||
724 | const AddrMode &AM, Type *Ty, | |||
725 | unsigned AS, | |||
726 | Instruction *I) const { | |||
727 | // No global is ever allowed as a base. | |||
728 | if (AM.BaseGV) | |||
729 | return false; | |||
730 | ||||
731 | // Require a 12-bit signed offset. | |||
732 | if (!isInt<12>(AM.BaseOffs)) | |||
733 | return false; | |||
734 | ||||
735 | switch (AM.Scale) { | |||
736 | case 0: // "r+i" or just "i", depending on HasBaseReg. | |||
737 | break; | |||
738 | case 1: | |||
739 | if (!AM.HasBaseReg) // allow "r+i". | |||
740 | break; | |||
741 | return false; // disallow "r+r" or "r+r+i". | |||
742 | default: | |||
743 | return false; | |||
744 | } | |||
745 | ||||
746 | return true; | |||
747 | } | |||
748 | ||||
749 | bool RISCVTargetLowering::isLegalICmpImmediate(int64_t Imm) const { | |||
750 | return isInt<12>(Imm); | |||
751 | } | |||
752 | ||||
753 | bool RISCVTargetLowering::isLegalAddImmediate(int64_t Imm) const { | |||
754 | return isInt<12>(Imm); | |||
755 | } | |||
756 | ||||
757 | // On RV32, 64-bit integers are split into their high and low parts and held | |||
758 | // in two different registers, so the trunc is free since the low register can | |||
759 | // just be used. | |||
760 | bool RISCVTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const { | |||
761 | if (Subtarget.is64Bit() || !SrcTy->isIntegerTy() || !DstTy->isIntegerTy()) | |||
762 | return false; | |||
763 | unsigned SrcBits = SrcTy->getPrimitiveSizeInBits(); | |||
764 | unsigned DestBits = DstTy->getPrimitiveSizeInBits(); | |||
765 | return (SrcBits == 64 && DestBits == 32); | |||
766 | } | |||
767 | ||||
768 | bool RISCVTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const { | |||
769 | if (Subtarget.is64Bit() || SrcVT.isVector() || DstVT.isVector() || | |||
770 | !SrcVT.isInteger() || !DstVT.isInteger()) | |||
771 | return false; | |||
772 | unsigned SrcBits = SrcVT.getSizeInBits(); | |||
773 | unsigned DestBits = DstVT.getSizeInBits(); | |||
774 | return (SrcBits == 64 && DestBits == 32); | |||
775 | } | |||
776 | ||||
777 | bool RISCVTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { | |||
778 | // Zexts are free if they can be combined with a load. | |||
779 | if (auto *LD = dyn_cast<LoadSDNode>(Val)) { | |||
780 | EVT MemVT = LD->getMemoryVT(); | |||
781 | if ((MemVT == MVT::i8 || MemVT == MVT::i16 || | |||
782 | (Subtarget.is64Bit() && MemVT == MVT::i32)) && | |||
783 | (LD->getExtensionType() == ISD::NON_EXTLOAD || | |||
784 | LD->getExtensionType() == ISD::ZEXTLOAD)) | |||
785 | return true; | |||
786 | } | |||
787 | ||||
788 | return TargetLowering::isZExtFree(Val, VT2); | |||
789 | } | |||
790 | ||||
791 | bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const { | |||
792 | return Subtarget.is64Bit() && SrcVT == MVT::i32 && DstVT == MVT::i64; | |||
793 | } | |||
794 | ||||
795 | bool RISCVTargetLowering::isCheapToSpeculateCttz() const { | |||
796 | return Subtarget.hasStdExtZbb(); | |||
797 | } | |||
798 | ||||
799 | bool RISCVTargetLowering::isCheapToSpeculateCtlz() const { | |||
800 | return Subtarget.hasStdExtZbb(); | |||
801 | } | |||
802 | ||||
803 | bool RISCVTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, | |||
804 | bool ForCodeSize) const { | |||
805 | if (VT == MVT::f16 && !Subtarget.hasStdExtZfh()) | |||
806 | return false; | |||
807 | if (VT == MVT::f32 && !Subtarget.hasStdExtF()) | |||
808 | return false; | |||
809 | if (VT == MVT::f64 && !Subtarget.hasStdExtD()) | |||
810 | return false; | |||
811 | if (Imm.isNegZero()) | |||
812 | return false; | |||
813 | return Imm.isZero(); | |||
814 | } | |||
815 | ||||
816 | bool RISCVTargetLowering::hasBitPreservingFPLogic(EVT VT) const { | |||
817 | return (VT == MVT::f16 && Subtarget.hasStdExtZfh()) || | |||
818 | (VT == MVT::f32 && Subtarget.hasStdExtF()) || | |||
819 | (VT == MVT::f64 && Subtarget.hasStdExtD()); | |||
820 | } | |||
821 | ||||
822 | // Changes the condition code and swaps operands if necessary, so the SetCC | |||
823 | // operation matches one of the comparisons supported directly in the RISC-V | |||
824 | // ISA. | |||
825 | static void normaliseSetCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) { | |||
826 | switch (CC) { | |||
827 | default: | |||
828 | break; | |||
829 | case ISD::SETGT: | |||
830 | case ISD::SETLE: | |||
831 | case ISD::SETUGT: | |||
832 | case ISD::SETULE: | |||
833 | CC = ISD::getSetCCSwappedOperands(CC); | |||
834 | std::swap(LHS, RHS); | |||
835 | break; | |||
836 | } | |||
837 | } | |||
838 | ||||
839 | // Return the RISC-V branch opcode that matches the given DAG integer | |||
840 | // condition code. The CondCode must be one of those supported by the RISC-V | |||
841 | // ISA (see normaliseSetCC). | |||
842 | static unsigned getBranchOpcodeForIntCondCode(ISD::CondCode CC) { | |||
843 | switch (CC) { | |||
844 | default: | |||
845 | llvm_unreachable("Unsupported CondCode")::llvm::llvm_unreachable_internal("Unsupported CondCode", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 845); | |||
846 | case ISD::SETEQ: | |||
847 | return RISCV::BEQ; | |||
848 | case ISD::SETNE: | |||
849 | return RISCV::BNE; | |||
850 | case ISD::SETLT: | |||
851 | return RISCV::BLT; | |||
852 | case ISD::SETGE: | |||
853 | return RISCV::BGE; | |||
854 | case ISD::SETULT: | |||
855 | return RISCV::BLTU; | |||
856 | case ISD::SETUGE: | |||
857 | return RISCV::BGEU; | |||
858 | } | |||
859 | } | |||
860 | ||||
861 | RISCVVLMUL RISCVTargetLowering::getLMUL(MVT VT) { | |||
862 | assert(VT.isScalableVector() && "Expecting a scalable vector type")((VT.isScalableVector() && "Expecting a scalable vector type" ) ? static_cast<void> (0) : __assert_fail ("VT.isScalableVector() && \"Expecting a scalable vector type\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 862, __PRETTY_FUNCTION__)); | |||
863 | unsigned KnownSize = VT.getSizeInBits().getKnownMinValue(); | |||
864 | if (VT.getVectorElementType() == MVT::i1) | |||
865 | KnownSize *= 8; | |||
866 | ||||
867 | switch (KnownSize) { | |||
868 | default: | |||
869 | llvm_unreachable("Invalid LMUL.")::llvm::llvm_unreachable_internal("Invalid LMUL.", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 869); | |||
870 | case 8: | |||
871 | return RISCVVLMUL::LMUL_F8; | |||
872 | case 16: | |||
873 | return RISCVVLMUL::LMUL_F4; | |||
874 | case 32: | |||
875 | return RISCVVLMUL::LMUL_F2; | |||
876 | case 64: | |||
877 | return RISCVVLMUL::LMUL_1; | |||
878 | case 128: | |||
879 | return RISCVVLMUL::LMUL_2; | |||
880 | case 256: | |||
881 | return RISCVVLMUL::LMUL_4; | |||
882 | case 512: | |||
883 | return RISCVVLMUL::LMUL_8; | |||
884 | } | |||
885 | } | |||
886 | ||||
887 | unsigned RISCVTargetLowering::getRegClassIDForLMUL(RISCVVLMUL LMul) { | |||
888 | switch (LMul) { | |||
889 | default: | |||
890 | llvm_unreachable("Invalid LMUL.")::llvm::llvm_unreachable_internal("Invalid LMUL.", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 890); | |||
891 | case RISCVVLMUL::LMUL_F8: | |||
892 | case RISCVVLMUL::LMUL_F4: | |||
893 | case RISCVVLMUL::LMUL_F2: | |||
894 | case RISCVVLMUL::LMUL_1: | |||
895 | return RISCV::VRRegClassID; | |||
896 | case RISCVVLMUL::LMUL_2: | |||
897 | return RISCV::VRM2RegClassID; | |||
898 | case RISCVVLMUL::LMUL_4: | |||
899 | return RISCV::VRM4RegClassID; | |||
900 | case RISCVVLMUL::LMUL_8: | |||
901 | return RISCV::VRM8RegClassID; | |||
902 | } | |||
903 | } | |||
904 | ||||
905 | unsigned RISCVTargetLowering::getSubregIndexByMVT(MVT VT, unsigned Index) { | |||
906 | RISCVVLMUL LMUL = getLMUL(VT); | |||
907 | if (LMUL == RISCVVLMUL::LMUL_F8 || LMUL == RISCVVLMUL::LMUL_F4 || | |||
908 | LMUL == RISCVVLMUL::LMUL_F2 || LMUL == RISCVVLMUL::LMUL_1) { | |||
909 | static_assert(RISCV::sub_vrm1_7 == RISCV::sub_vrm1_0 + 7, | |||
910 | "Unexpected subreg numbering"); | |||
911 | return RISCV::sub_vrm1_0 + Index; | |||
912 | } | |||
913 | if (LMUL == RISCVVLMUL::LMUL_2) { | |||
914 | static_assert(RISCV::sub_vrm2_3 == RISCV::sub_vrm2_0 + 3, | |||
915 | "Unexpected subreg numbering"); | |||
916 | return RISCV::sub_vrm2_0 + Index; | |||
917 | } | |||
918 | if (LMUL == RISCVVLMUL::LMUL_4) { | |||
919 | static_assert(RISCV::sub_vrm4_1 == RISCV::sub_vrm4_0 + 1, | |||
920 | "Unexpected subreg numbering"); | |||
921 | return RISCV::sub_vrm4_0 + Index; | |||
922 | } | |||
923 | llvm_unreachable("Invalid vector type.")::llvm::llvm_unreachable_internal("Invalid vector type.", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 923); | |||
924 | } | |||
925 | ||||
926 | unsigned RISCVTargetLowering::getRegClassIDForVecVT(MVT VT) { | |||
927 | if (VT.getVectorElementType() == MVT::i1) | |||
928 | return RISCV::VRRegClassID; | |||
929 | return getRegClassIDForLMUL(getLMUL(VT)); | |||
930 | } | |||
931 | ||||
932 | // Attempt to decompose a subvector insert/extract between VecVT and | |||
933 | // SubVecVT via subregister indices. Returns the subregister index that | |||
934 | // can perform the subvector insert/extract with the given element index, as | |||
935 | // well as the index corresponding to any leftover subvectors that must be | |||
936 | // further inserted/extracted within the register class for SubVecVT. | |||
937 | std::pair<unsigned, unsigned> | |||
938 | RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs( | |||
939 | MVT VecVT, MVT SubVecVT, unsigned InsertExtractIdx, | |||
940 | const RISCVRegisterInfo *TRI) { | |||
941 | static_assert((RISCV::VRM8RegClassID > RISCV::VRM4RegClassID && | |||
942 | RISCV::VRM4RegClassID > RISCV::VRM2RegClassID && | |||
943 | RISCV::VRM2RegClassID > RISCV::VRRegClassID), | |||
944 | "Register classes not ordered"); | |||
945 | unsigned VecRegClassID = getRegClassIDForVecVT(VecVT); | |||
946 | unsigned SubRegClassID = getRegClassIDForVecVT(SubVecVT); | |||
947 | // Try to compose a subregister index that takes us from the incoming | |||
948 | // LMUL>1 register class down to the outgoing one. At each step we half | |||
949 | // the LMUL: | |||
950 | // nxv16i32@12 -> nxv2i32: sub_vrm4_1_then_sub_vrm2_1_then_sub_vrm1_0 | |||
951 | // Note that this is not guaranteed to find a subregister index, such as | |||
952 | // when we are extracting from one VR type to another. | |||
953 | unsigned SubRegIdx = RISCV::NoSubRegister; | |||
954 | for (const unsigned RCID : | |||
955 | {RISCV::VRM4RegClassID, RISCV::VRM2RegClassID, RISCV::VRRegClassID}) | |||
956 | if (VecRegClassID > RCID && SubRegClassID <= RCID) { | |||
957 | VecVT = VecVT.getHalfNumVectorElementsVT(); | |||
958 | bool IsHi = | |||
959 | InsertExtractIdx >= VecVT.getVectorElementCount().getKnownMinValue(); | |||
960 | SubRegIdx = TRI->composeSubRegIndices(SubRegIdx, | |||
961 | getSubregIndexByMVT(VecVT, IsHi)); | |||
962 | if (IsHi) | |||
963 | InsertExtractIdx -= VecVT.getVectorElementCount().getKnownMinValue(); | |||
964 | } | |||
965 | return {SubRegIdx, InsertExtractIdx}; | |||
966 | } | |||
967 | ||||
968 | // Return the largest legal scalable vector type that matches VT's element type. | |||
969 | MVT RISCVTargetLowering::getContainerForFixedLengthVector( | |||
970 | const TargetLowering &TLI, MVT VT, const RISCVSubtarget &Subtarget) { | |||
971 | assert(VT.isFixedLengthVector() && TLI.isTypeLegal(VT) &&((VT.isFixedLengthVector() && TLI.isTypeLegal(VT) && "Expected legal fixed length vector!") ? static_cast<void > (0) : __assert_fail ("VT.isFixedLengthVector() && TLI.isTypeLegal(VT) && \"Expected legal fixed length vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 972, __PRETTY_FUNCTION__)) | |||
972 | "Expected legal fixed length vector!")((VT.isFixedLengthVector() && TLI.isTypeLegal(VT) && "Expected legal fixed length vector!") ? static_cast<void > (0) : __assert_fail ("VT.isFixedLengthVector() && TLI.isTypeLegal(VT) && \"Expected legal fixed length vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 972, __PRETTY_FUNCTION__)); | |||
973 | ||||
974 | unsigned LMul = Subtarget.getLMULForFixedLengthVector(VT); | |||
975 | assert(LMul <= 8 && isPowerOf2_32(LMul) && "Unexpected LMUL!")((LMul <= 8 && isPowerOf2_32(LMul) && "Unexpected LMUL!" ) ? static_cast<void> (0) : __assert_fail ("LMul <= 8 && isPowerOf2_32(LMul) && \"Unexpected LMUL!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 975, __PRETTY_FUNCTION__)); | |||
976 | ||||
977 | MVT EltVT = VT.getVectorElementType(); | |||
978 | switch (EltVT.SimpleTy) { | |||
979 | default: | |||
980 | llvm_unreachable("unexpected element type for RVV container")::llvm::llvm_unreachable_internal("unexpected element type for RVV container" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 980); | |||
981 | case MVT::i1: { | |||
982 | // Masks are calculated assuming 8-bit elements since that's when we need | |||
983 | // the most elements. | |||
984 | unsigned EltsPerBlock = RISCV::RVVBitsPerBlock / 8; | |||
985 | return MVT::getScalableVectorVT(MVT::i1, LMul * EltsPerBlock); | |||
986 | } | |||
987 | case MVT::i8: | |||
988 | case MVT::i16: | |||
989 | case MVT::i32: | |||
990 | case MVT::i64: | |||
991 | case MVT::f16: | |||
992 | case MVT::f32: | |||
993 | case MVT::f64: { | |||
994 | unsigned EltsPerBlock = RISCV::RVVBitsPerBlock / EltVT.getSizeInBits(); | |||
995 | return MVT::getScalableVectorVT(EltVT, LMul * EltsPerBlock); | |||
996 | } | |||
997 | } | |||
998 | } | |||
999 | ||||
1000 | MVT RISCVTargetLowering::getContainerForFixedLengthVector( | |||
1001 | SelectionDAG &DAG, MVT VT, const RISCVSubtarget &Subtarget) { | |||
1002 | return getContainerForFixedLengthVector(DAG.getTargetLoweringInfo(), VT, | |||
1003 | Subtarget); | |||
1004 | } | |||
1005 | ||||
1006 | MVT RISCVTargetLowering::getContainerForFixedLengthVector(MVT VT) const { | |||
1007 | return getContainerForFixedLengthVector(*this, VT, getSubtarget()); | |||
1008 | } | |||
1009 | ||||
1010 | // Grow V to consume an entire RVV register. | |||
1011 | static SDValue convertToScalableVector(EVT VT, SDValue V, SelectionDAG &DAG, | |||
1012 | const RISCVSubtarget &Subtarget) { | |||
1013 | assert(VT.isScalableVector() &&((VT.isScalableVector() && "Expected to convert into a scalable vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isScalableVector() && \"Expected to convert into a scalable vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1014, __PRETTY_FUNCTION__)) | |||
1014 | "Expected to convert into a scalable vector!")((VT.isScalableVector() && "Expected to convert into a scalable vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isScalableVector() && \"Expected to convert into a scalable vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1014, __PRETTY_FUNCTION__)); | |||
1015 | assert(V.getValueType().isFixedLengthVector() &&((V.getValueType().isFixedLengthVector() && "Expected a fixed length vector operand!" ) ? static_cast<void> (0) : __assert_fail ("V.getValueType().isFixedLengthVector() && \"Expected a fixed length vector operand!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1016, __PRETTY_FUNCTION__)) | |||
1016 | "Expected a fixed length vector operand!")((V.getValueType().isFixedLengthVector() && "Expected a fixed length vector operand!" ) ? static_cast<void> (0) : __assert_fail ("V.getValueType().isFixedLengthVector() && \"Expected a fixed length vector operand!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1016, __PRETTY_FUNCTION__)); | |||
1017 | SDLoc DL(V); | |||
1018 | SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT()); | |||
1019 | return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero); | |||
1020 | } | |||
1021 | ||||
1022 | // Shrink V so it's just big enough to maintain a VT's worth of data. | |||
1023 | static SDValue convertFromScalableVector(EVT VT, SDValue V, SelectionDAG &DAG, | |||
1024 | const RISCVSubtarget &Subtarget) { | |||
1025 | assert(VT.isFixedLengthVector() &&((VT.isFixedLengthVector() && "Expected to convert into a fixed length vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Expected to convert into a fixed length vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1026, __PRETTY_FUNCTION__)) | |||
1026 | "Expected to convert into a fixed length vector!")((VT.isFixedLengthVector() && "Expected to convert into a fixed length vector!" ) ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Expected to convert into a fixed length vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1026, __PRETTY_FUNCTION__)); | |||
1027 | assert(V.getValueType().isScalableVector() &&((V.getValueType().isScalableVector() && "Expected a scalable vector operand!" ) ? static_cast<void> (0) : __assert_fail ("V.getValueType().isScalableVector() && \"Expected a scalable vector operand!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1028, __PRETTY_FUNCTION__)) | |||
1028 | "Expected a scalable vector operand!")((V.getValueType().isScalableVector() && "Expected a scalable vector operand!" ) ? static_cast<void> (0) : __assert_fail ("V.getValueType().isScalableVector() && \"Expected a scalable vector operand!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1028, __PRETTY_FUNCTION__)); | |||
1029 | SDLoc DL(V); | |||
1030 | SDValue Zero = DAG.getConstant(0, DL, Subtarget.getXLenVT()); | |||
1031 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero); | |||
1032 | } | |||
1033 | ||||
1034 | // Gets the two common "VL" operands: an all-ones mask and the vector length. | |||
1035 | // VecVT is a vector type, either fixed-length or scalable, and ContainerVT is | |||
1036 | // the vector type that it is contained in. | |||
1037 | static std::pair<SDValue, SDValue> | |||
1038 | getDefaultVLOps(MVT VecVT, MVT ContainerVT, SDLoc DL, SelectionDAG &DAG, | |||
1039 | const RISCVSubtarget &Subtarget) { | |||
1040 | assert(ContainerVT.isScalableVector() && "Expecting scalable container type")((ContainerVT.isScalableVector() && "Expecting scalable container type" ) ? static_cast<void> (0) : __assert_fail ("ContainerVT.isScalableVector() && \"Expecting scalable container type\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1040, __PRETTY_FUNCTION__)); | |||
1041 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1042 | SDValue VL = VecVT.isFixedLengthVector() | |||
1043 | ? DAG.getConstant(VecVT.getVectorNumElements(), DL, XLenVT) | |||
1044 | : DAG.getRegister(RISCV::X0, XLenVT); | |||
1045 | MVT MaskVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount()); | |||
1046 | SDValue Mask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL); | |||
1047 | return {Mask, VL}; | |||
1048 | } | |||
1049 | ||||
1050 | // As above but assuming the given type is a scalable vector type. | |||
1051 | static std::pair<SDValue, SDValue> | |||
1052 | getDefaultScalableVLOps(MVT VecVT, SDLoc DL, SelectionDAG &DAG, | |||
1053 | const RISCVSubtarget &Subtarget) { | |||
1054 | assert(VecVT.isScalableVector() && "Expecting a scalable vector")((VecVT.isScalableVector() && "Expecting a scalable vector" ) ? static_cast<void> (0) : __assert_fail ("VecVT.isScalableVector() && \"Expecting a scalable vector\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1054, __PRETTY_FUNCTION__)); | |||
1055 | return getDefaultVLOps(VecVT, VecVT, DL, DAG, Subtarget); | |||
1056 | } | |||
1057 | ||||
1058 | // The state of RVV BUILD_VECTOR and VECTOR_SHUFFLE lowering is that very few | |||
1059 | // of either is (currently) supported. This can get us into an infinite loop | |||
1060 | // where we try to lower a BUILD_VECTOR as a VECTOR_SHUFFLE as a BUILD_VECTOR | |||
1061 | // as a ..., etc. | |||
1062 | // Until either (or both) of these can reliably lower any node, reporting that | |||
1063 | // we don't want to expand BUILD_VECTORs via VECTOR_SHUFFLEs at least breaks | |||
1064 | // the infinite loop. Note that this lowers BUILD_VECTOR through the stack, | |||
1065 | // which is not desirable. | |||
1066 | bool RISCVTargetLowering::shouldExpandBuildVectorWithShuffles( | |||
1067 | EVT VT, unsigned DefinedValues) const { | |||
1068 | return false; | |||
1069 | } | |||
1070 | ||||
1071 | static SDValue lowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG, | |||
1072 | const RISCVSubtarget &Subtarget) { | |||
1073 | MVT VT = Op.getSimpleValueType(); | |||
1074 | assert(VT.isFixedLengthVector() && "Unexpected vector!")((VT.isFixedLengthVector() && "Unexpected vector!") ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Unexpected vector!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1074, __PRETTY_FUNCTION__)); | |||
1075 | ||||
1076 | MVT ContainerVT = | |||
1077 | RISCVTargetLowering::getContainerForFixedLengthVector(DAG, VT, Subtarget); | |||
1078 | ||||
1079 | SDLoc DL(Op); | |||
1080 | SDValue Mask, VL; | |||
1081 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
1082 | ||||
1083 | if (VT.getVectorElementType() == MVT::i1) { | |||
1084 | if (ISD::isBuildVectorAllZeros(Op.getNode())) { | |||
1085 | SDValue VMClr = DAG.getNode(RISCVISD::VMCLR_VL, DL, ContainerVT, VL); | |||
1086 | return convertFromScalableVector(VT, VMClr, DAG, Subtarget); | |||
1087 | } | |||
1088 | ||||
1089 | if (ISD::isBuildVectorAllOnes(Op.getNode())) { | |||
1090 | SDValue VMSet = DAG.getNode(RISCVISD::VMSET_VL, DL, ContainerVT, VL); | |||
1091 | return convertFromScalableVector(VT, VMSet, DAG, Subtarget); | |||
1092 | } | |||
1093 | ||||
1094 | return SDValue(); | |||
1095 | } | |||
1096 | ||||
1097 | if (SDValue Splat = cast<BuildVectorSDNode>(Op)->getSplatValue()) { | |||
1098 | unsigned Opc = VT.isFloatingPoint() ? RISCVISD::VFMV_V_F_VL | |||
1099 | : RISCVISD::VMV_V_X_VL; | |||
1100 | Splat = DAG.getNode(Opc, DL, ContainerVT, Splat, VL); | |||
1101 | return convertFromScalableVector(VT, Splat, DAG, Subtarget); | |||
1102 | } | |||
1103 | ||||
1104 | // Try and match an index sequence, which we can lower directly to the vid | |||
1105 | // instruction. An all-undef vector is matched by getSplatValue, above. | |||
1106 | if (VT.isInteger()) { | |||
1107 | bool IsVID = true; | |||
1108 | for (unsigned i = 0, e = Op.getNumOperands(); i < e && IsVID; i++) | |||
1109 | IsVID &= Op.getOperand(i).isUndef() || | |||
1110 | (isa<ConstantSDNode>(Op.getOperand(i)) && | |||
1111 | Op.getConstantOperandVal(i) == i); | |||
1112 | ||||
1113 | if (IsVID) { | |||
1114 | SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, ContainerVT, Mask, VL); | |||
1115 | return convertFromScalableVector(VT, VID, DAG, Subtarget); | |||
1116 | } | |||
1117 | } | |||
1118 | ||||
1119 | return SDValue(); | |||
1120 | } | |||
1121 | ||||
1122 | static SDValue lowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG, | |||
1123 | const RISCVSubtarget &Subtarget) { | |||
1124 | SDValue V1 = Op.getOperand(0); | |||
1125 | SDLoc DL(Op); | |||
1126 | MVT VT = Op.getSimpleValueType(); | |||
1127 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); | |||
1128 | ||||
1129 | if (SVN->isSplat()) { | |||
1130 | int Lane = SVN->getSplatIndex(); | |||
1131 | if (Lane >= 0) { | |||
1132 | MVT ContainerVT = RISCVTargetLowering::getContainerForFixedLengthVector( | |||
1133 | DAG, VT, Subtarget); | |||
1134 | ||||
1135 | V1 = convertToScalableVector(ContainerVT, V1, DAG, Subtarget); | |||
1136 | assert(Lane < (int)VT.getVectorNumElements() && "Unexpected lane!")((Lane < (int)VT.getVectorNumElements() && "Unexpected lane!" ) ? static_cast<void> (0) : __assert_fail ("Lane < (int)VT.getVectorNumElements() && \"Unexpected lane!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1136, __PRETTY_FUNCTION__)); | |||
1137 | ||||
1138 | SDValue Mask, VL; | |||
1139 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
1140 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1141 | SDValue Gather = | |||
1142 | DAG.getNode(RISCVISD::VRGATHER_VX_VL, DL, ContainerVT, V1, | |||
1143 | DAG.getConstant(Lane, DL, XLenVT), Mask, VL); | |||
1144 | return convertFromScalableVector(VT, Gather, DAG, Subtarget); | |||
1145 | } | |||
1146 | } | |||
1147 | ||||
1148 | return SDValue(); | |||
1149 | } | |||
1150 | ||||
1151 | static SDValue getRVVFPExtendOrRound(SDValue Op, MVT VT, MVT ContainerVT, | |||
1152 | SDLoc DL, SelectionDAG &DAG, | |||
1153 | const RISCVSubtarget &Subtarget) { | |||
1154 | if (VT.isScalableVector()) | |||
1155 | return DAG.getFPExtendOrRound(Op, DL, VT); | |||
1156 | assert(VT.isFixedLengthVector() &&((VT.isFixedLengthVector() && "Unexpected value type for RVV FP extend/round lowering" ) ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Unexpected value type for RVV FP extend/round lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1157, __PRETTY_FUNCTION__)) | |||
1157 | "Unexpected value type for RVV FP extend/round lowering")((VT.isFixedLengthVector() && "Unexpected value type for RVV FP extend/round lowering" ) ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Unexpected value type for RVV FP extend/round lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1157, __PRETTY_FUNCTION__)); | |||
1158 | SDValue Mask, VL; | |||
1159 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
1160 | unsigned RVVOpc = ContainerVT.bitsGT(Op.getSimpleValueType()) | |||
1161 | ? RISCVISD::FP_EXTEND_VL | |||
1162 | : RISCVISD::FP_ROUND_VL; | |||
1163 | return DAG.getNode(RVVOpc, DL, ContainerVT, Op, Mask, VL); | |||
1164 | } | |||
1165 | ||||
1166 | SDValue RISCVTargetLowering::LowerOperation(SDValue Op, | |||
1167 | SelectionDAG &DAG) const { | |||
1168 | switch (Op.getOpcode()) { | |||
| ||||
1169 | default: | |||
1170 | report_fatal_error("unimplemented operand"); | |||
1171 | case ISD::GlobalAddress: | |||
1172 | return lowerGlobalAddress(Op, DAG); | |||
1173 | case ISD::BlockAddress: | |||
1174 | return lowerBlockAddress(Op, DAG); | |||
1175 | case ISD::ConstantPool: | |||
1176 | return lowerConstantPool(Op, DAG); | |||
1177 | case ISD::JumpTable: | |||
1178 | return lowerJumpTable(Op, DAG); | |||
1179 | case ISD::GlobalTLSAddress: | |||
1180 | return lowerGlobalTLSAddress(Op, DAG); | |||
1181 | case ISD::SELECT: | |||
1182 | return lowerSELECT(Op, DAG); | |||
1183 | case ISD::VASTART: | |||
1184 | return lowerVASTART(Op, DAG); | |||
1185 | case ISD::FRAMEADDR: | |||
1186 | return lowerFRAMEADDR(Op, DAG); | |||
1187 | case ISD::RETURNADDR: | |||
1188 | return lowerRETURNADDR(Op, DAG); | |||
1189 | case ISD::SHL_PARTS: | |||
1190 | return lowerShiftLeftParts(Op, DAG); | |||
1191 | case ISD::SRA_PARTS: | |||
1192 | return lowerShiftRightParts(Op, DAG, true); | |||
1193 | case ISD::SRL_PARTS: | |||
1194 | return lowerShiftRightParts(Op, DAG, false); | |||
1195 | case ISD::BITCAST: { | |||
1196 | SDValue Op0 = Op.getOperand(0); | |||
1197 | // We can handle fixed length vector bitcasts with a simple replacement | |||
1198 | // in isel. | |||
1199 | if (Op.getValueType().isFixedLengthVector()) { | |||
1200 | if (Op0.getValueType().isFixedLengthVector()) | |||
1201 | return Op; | |||
1202 | return SDValue(); | |||
1203 | } | |||
1204 | assert(((Subtarget.is64Bit() && Subtarget.hasStdExtF()) ||((((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1206, __PRETTY_FUNCTION__)) | |||
1205 | Subtarget.hasStdExtZfh()) &&((((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1206, __PRETTY_FUNCTION__)) | |||
1206 | "Unexpected custom legalisation")((((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("((Subtarget.is64Bit() && Subtarget.hasStdExtF()) || Subtarget.hasStdExtZfh()) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1206, __PRETTY_FUNCTION__)); | |||
1207 | SDLoc DL(Op); | |||
1208 | if (Op.getValueType() == MVT::f16 && Subtarget.hasStdExtZfh()) { | |||
1209 | if (Op0.getValueType() != MVT::i16) | |||
1210 | return SDValue(); | |||
1211 | SDValue NewOp0 = | |||
1212 | DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), Op0); | |||
1213 | SDValue FPConv = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, NewOp0); | |||
1214 | return FPConv; | |||
1215 | } else if (Op.getValueType() == MVT::f32 && Subtarget.is64Bit() && | |||
1216 | Subtarget.hasStdExtF()) { | |||
1217 | if (Op0.getValueType() != MVT::i32) | |||
1218 | return SDValue(); | |||
1219 | SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0); | |||
1220 | SDValue FPConv = | |||
1221 | DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, NewOp0); | |||
1222 | return FPConv; | |||
1223 | } | |||
1224 | return SDValue(); | |||
1225 | } | |||
1226 | case ISD::INTRINSIC_WO_CHAIN: | |||
1227 | return LowerINTRINSIC_WO_CHAIN(Op, DAG); | |||
1228 | case ISD::INTRINSIC_W_CHAIN: | |||
1229 | return LowerINTRINSIC_W_CHAIN(Op, DAG); | |||
1230 | case ISD::BSWAP: | |||
1231 | case ISD::BITREVERSE: { | |||
1232 | // Convert BSWAP/BITREVERSE to GREVI to enable GREVI combinining. | |||
1233 | assert(Subtarget.hasStdExtZbp() && "Unexpected custom legalisation")((Subtarget.hasStdExtZbp() && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasStdExtZbp() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1233, __PRETTY_FUNCTION__)); | |||
1234 | MVT VT = Op.getSimpleValueType(); | |||
1235 | SDLoc DL(Op); | |||
1236 | // Start with the maximum immediate value which is the bitwidth - 1. | |||
1237 | unsigned Imm = VT.getSizeInBits() - 1; | |||
1238 | // If this is BSWAP rather than BITREVERSE, clear the lower 3 bits. | |||
1239 | if (Op.getOpcode() == ISD::BSWAP) | |||
1240 | Imm &= ~0x7U; | |||
1241 | return DAG.getNode(RISCVISD::GREVI, DL, VT, Op.getOperand(0), | |||
1242 | DAG.getTargetConstant(Imm, DL, Subtarget.getXLenVT())); | |||
1243 | } | |||
1244 | case ISD::FSHL: | |||
1245 | case ISD::FSHR: { | |||
1246 | MVT VT = Op.getSimpleValueType(); | |||
1247 | assert(VT == Subtarget.getXLenVT() && "Unexpected custom legalization")((VT == Subtarget.getXLenVT() && "Unexpected custom legalization" ) ? static_cast<void> (0) : __assert_fail ("VT == Subtarget.getXLenVT() && \"Unexpected custom legalization\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1247, __PRETTY_FUNCTION__)); | |||
1248 | SDLoc DL(Op); | |||
1249 | // FSL/FSR take a log2(XLen)+1 bit shift amount but XLenVT FSHL/FSHR only | |||
1250 | // use log(XLen) bits. Mask the shift amount accordingly. | |||
1251 | unsigned ShAmtWidth = Subtarget.getXLen() - 1; | |||
1252 | SDValue ShAmt = DAG.getNode(ISD::AND, DL, VT, Op.getOperand(2), | |||
1253 | DAG.getConstant(ShAmtWidth, DL, VT)); | |||
1254 | unsigned Opc = Op.getOpcode() == ISD::FSHL ? RISCVISD::FSL : RISCVISD::FSR; | |||
1255 | return DAG.getNode(Opc, DL, VT, Op.getOperand(0), Op.getOperand(1), ShAmt); | |||
1256 | } | |||
1257 | case ISD::TRUNCATE: { | |||
1258 | SDLoc DL(Op); | |||
1259 | MVT VT = Op.getSimpleValueType(); | |||
1260 | // Only custom-lower vector truncates | |||
1261 | if (!VT.isVector()) | |||
1262 | return Op; | |||
1263 | ||||
1264 | // Truncates to mask types are handled differently | |||
1265 | if (VT.getVectorElementType() == MVT::i1) | |||
1266 | return lowerVectorMaskTrunc(Op, DAG); | |||
1267 | ||||
1268 | // RVV only has truncates which operate from SEW*2->SEW, so lower arbitrary | |||
1269 | // truncates as a series of "RISCVISD::TRUNCATE_VECTOR_VL" nodes which | |||
1270 | // truncate by one power of two at a time. | |||
1271 | MVT DstEltVT = VT.getVectorElementType(); | |||
1272 | ||||
1273 | SDValue Src = Op.getOperand(0); | |||
1274 | MVT SrcVT = Src.getSimpleValueType(); | |||
1275 | MVT SrcEltVT = SrcVT.getVectorElementType(); | |||
1276 | ||||
1277 | assert(DstEltVT.bitsLT(SrcEltVT) &&((DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT .getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits ()) && "Unexpected vector truncate lowering") ? static_cast <void> (0) : __assert_fail ("DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits()) && \"Unexpected vector truncate lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1280, __PRETTY_FUNCTION__)) | |||
1278 | isPowerOf2_64(DstEltVT.getSizeInBits()) &&((DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT .getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits ()) && "Unexpected vector truncate lowering") ? static_cast <void> (0) : __assert_fail ("DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits()) && \"Unexpected vector truncate lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1280, __PRETTY_FUNCTION__)) | |||
1279 | isPowerOf2_64(SrcEltVT.getSizeInBits()) &&((DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT .getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits ()) && "Unexpected vector truncate lowering") ? static_cast <void> (0) : __assert_fail ("DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits()) && \"Unexpected vector truncate lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1280, __PRETTY_FUNCTION__)) | |||
1280 | "Unexpected vector truncate lowering")((DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT .getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits ()) && "Unexpected vector truncate lowering") ? static_cast <void> (0) : __assert_fail ("DstEltVT.bitsLT(SrcEltVT) && isPowerOf2_64(DstEltVT.getSizeInBits()) && isPowerOf2_64(SrcEltVT.getSizeInBits()) && \"Unexpected vector truncate lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1280, __PRETTY_FUNCTION__)); | |||
1281 | ||||
1282 | MVT ContainerVT = SrcVT; | |||
1283 | if (SrcVT.isFixedLengthVector()) { | |||
1284 | ContainerVT = getContainerForFixedLengthVector(SrcVT); | |||
1285 | Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget); | |||
1286 | } | |||
1287 | ||||
1288 | SDValue Result = Src; | |||
1289 | SDValue Mask, VL; | |||
1290 | std::tie(Mask, VL) = | |||
1291 | getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget); | |||
1292 | LLVMContext &Context = *DAG.getContext(); | |||
1293 | const ElementCount Count = ContainerVT.getVectorElementCount(); | |||
1294 | do { | |||
1295 | SrcEltVT = MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2); | |||
1296 | EVT ResultVT = EVT::getVectorVT(Context, SrcEltVT, Count); | |||
1297 | Result = DAG.getNode(RISCVISD::TRUNCATE_VECTOR_VL, DL, ResultVT, Result, | |||
1298 | Mask, VL); | |||
1299 | } while (SrcEltVT != DstEltVT); | |||
1300 | ||||
1301 | if (SrcVT.isFixedLengthVector()) | |||
1302 | Result = convertFromScalableVector(VT, Result, DAG, Subtarget); | |||
1303 | ||||
1304 | return Result; | |||
1305 | } | |||
1306 | case ISD::ANY_EXTEND: | |||
1307 | case ISD::ZERO_EXTEND: | |||
1308 | if (Op.getOperand(0).getValueType().isVector() && | |||
1309 | Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1) | |||
1310 | return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ 1); | |||
1311 | return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VZEXT_VL); | |||
1312 | case ISD::SIGN_EXTEND: | |||
1313 | if (Op.getOperand(0).getValueType().isVector() && | |||
1314 | Op.getOperand(0).getValueType().getVectorElementType() == MVT::i1) | |||
1315 | return lowerVectorMaskExt(Op, DAG, /*ExtVal*/ -1); | |||
1316 | return lowerFixedLengthVectorExtendToRVV(Op, DAG, RISCVISD::VSEXT_VL); | |||
1317 | case ISD::SPLAT_VECTOR: | |||
1318 | return lowerSPLATVECTOR(Op, DAG); | |||
1319 | case ISD::INSERT_VECTOR_ELT: | |||
1320 | return lowerINSERT_VECTOR_ELT(Op, DAG); | |||
1321 | case ISD::EXTRACT_VECTOR_ELT: | |||
1322 | return lowerEXTRACT_VECTOR_ELT(Op, DAG); | |||
1323 | case ISD::VSCALE: { | |||
1324 | MVT VT = Op.getSimpleValueType(); | |||
1325 | SDLoc DL(Op); | |||
1326 | SDValue VLENB = DAG.getNode(RISCVISD::READ_VLENB, DL, VT); | |||
1327 | // We define our scalable vector types for lmul=1 to use a 64 bit known | |||
1328 | // minimum size. e.g. <vscale x 2 x i32>. VLENB is in bytes so we calculate | |||
1329 | // vscale as VLENB / 8. | |||
1330 | SDValue VScale = DAG.getNode(ISD::SRL, DL, VT, VLENB, | |||
1331 | DAG.getConstant(3, DL, VT)); | |||
1332 | return DAG.getNode(ISD::MUL, DL, VT, VScale, Op.getOperand(0)); | |||
1333 | } | |||
1334 | case ISD::FP_EXTEND: { | |||
1335 | // RVV can only do fp_extend to types double the size as the source. We | |||
1336 | // custom-lower f16->f64 extensions to two hops of ISD::FP_EXTEND, going | |||
1337 | // via f32. | |||
1338 | SDLoc DL(Op); | |||
1339 | MVT VT = Op.getSimpleValueType(); | |||
1340 | SDValue Src = Op.getOperand(0); | |||
1341 | MVT SrcVT = Src.getSimpleValueType(); | |||
1342 | ||||
1343 | // Prepare any fixed-length vector operands. | |||
1344 | MVT ContainerVT = VT; | |||
1345 | if (SrcVT.isFixedLengthVector()) { | |||
1346 | ContainerVT = getContainerForFixedLengthVector(VT); | |||
1347 | MVT SrcContainerVT = | |||
1348 | ContainerVT.changeVectorElementType(SrcVT.getVectorElementType()); | |||
1349 | Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget); | |||
1350 | } | |||
1351 | ||||
1352 | if (!VT.isVector() || VT.getVectorElementType() != MVT::f64 || | |||
1353 | SrcVT.getVectorElementType() != MVT::f16) { | |||
1354 | // For scalable vectors, we only need to close the gap between | |||
1355 | // vXf16->vXf64. | |||
1356 | if (!VT.isFixedLengthVector()) | |||
1357 | return Op; | |||
1358 | // For fixed-length vectors, lower the FP_EXTEND to a custom "VL" version. | |||
1359 | Src = getRVVFPExtendOrRound(Src, VT, ContainerVT, DL, DAG, Subtarget); | |||
1360 | return convertFromScalableVector(VT, Src, DAG, Subtarget); | |||
1361 | } | |||
1362 | ||||
1363 | MVT InterVT = VT.changeVectorElementType(MVT::f32); | |||
1364 | MVT InterContainerVT = ContainerVT.changeVectorElementType(MVT::f32); | |||
1365 | SDValue IntermediateExtend = getRVVFPExtendOrRound( | |||
1366 | Src, InterVT, InterContainerVT, DL, DAG, Subtarget); | |||
1367 | ||||
1368 | SDValue Extend = getRVVFPExtendOrRound(IntermediateExtend, VT, ContainerVT, | |||
1369 | DL, DAG, Subtarget); | |||
1370 | if (VT.isFixedLengthVector()) | |||
1371 | return convertFromScalableVector(VT, Extend, DAG, Subtarget); | |||
1372 | return Extend; | |||
1373 | } | |||
1374 | case ISD::FP_ROUND: { | |||
1375 | // RVV can only do fp_round to types half the size as the source. We | |||
1376 | // custom-lower f64->f16 rounds via RVV's round-to-odd float | |||
1377 | // conversion instruction. | |||
1378 | SDLoc DL(Op); | |||
1379 | MVT VT = Op.getSimpleValueType(); | |||
1380 | SDValue Src = Op.getOperand(0); | |||
1381 | MVT SrcVT = Src.getSimpleValueType(); | |||
1382 | ||||
1383 | // Prepare any fixed-length vector operands. | |||
1384 | MVT ContainerVT = VT; | |||
1385 | if (VT.isFixedLengthVector()) { | |||
1386 | MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT); | |||
1387 | ContainerVT = | |||
1388 | SrcContainerVT.changeVectorElementType(VT.getVectorElementType()); | |||
1389 | Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget); | |||
1390 | } | |||
1391 | ||||
1392 | if (!VT.isVector() || VT.getVectorElementType() != MVT::f16 || | |||
1393 | SrcVT.getVectorElementType() != MVT::f64) { | |||
1394 | // For scalable vectors, we only need to close the gap between | |||
1395 | // vXf64<->vXf16. | |||
1396 | if (!VT.isFixedLengthVector()) | |||
1397 | return Op; | |||
1398 | // For fixed-length vectors, lower the FP_ROUND to a custom "VL" version. | |||
1399 | Src = getRVVFPExtendOrRound(Src, VT, ContainerVT, DL, DAG, Subtarget); | |||
1400 | return convertFromScalableVector(VT, Src, DAG, Subtarget); | |||
1401 | } | |||
1402 | ||||
1403 | SDValue Mask, VL; | |||
1404 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
1405 | ||||
1406 | MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32); | |||
1407 | SDValue IntermediateRound = | |||
1408 | DAG.getNode(RISCVISD::VFNCVT_ROD_VL, DL, InterVT, Src, Mask, VL); | |||
1409 | SDValue Round = getRVVFPExtendOrRound(IntermediateRound, VT, ContainerVT, | |||
1410 | DL, DAG, Subtarget); | |||
1411 | ||||
1412 | if (VT.isFixedLengthVector()) | |||
1413 | return convertFromScalableVector(VT, Round, DAG, Subtarget); | |||
1414 | return Round; | |||
1415 | } | |||
1416 | case ISD::FP_TO_SINT: | |||
1417 | case ISD::FP_TO_UINT: | |||
1418 | case ISD::SINT_TO_FP: | |||
1419 | case ISD::UINT_TO_FP: { | |||
1420 | // RVV can only do fp<->int conversions to types half/double the size as | |||
1421 | // the source. We custom-lower any conversions that do two hops into | |||
1422 | // sequences. | |||
1423 | MVT VT = Op.getSimpleValueType(); | |||
1424 | if (!VT.isVector()) | |||
1425 | return Op; | |||
1426 | SDLoc DL(Op); | |||
1427 | SDValue Src = Op.getOperand(0); | |||
1428 | MVT EltVT = VT.getVectorElementType(); | |||
1429 | MVT SrcVT = Src.getSimpleValueType(); | |||
1430 | MVT SrcEltVT = SrcVT.getVectorElementType(); | |||
1431 | unsigned EltSize = EltVT.getSizeInBits(); | |||
1432 | unsigned SrcEltSize = SrcEltVT.getSizeInBits(); | |||
1433 | assert(isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) &&((isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) && "Unexpected vector element types") ? static_cast< void> (0) : __assert_fail ("isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) && \"Unexpected vector element types\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1434, __PRETTY_FUNCTION__)) | |||
1434 | "Unexpected vector element types")((isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) && "Unexpected vector element types") ? static_cast< void> (0) : __assert_fail ("isPowerOf2_32(EltSize) && isPowerOf2_32(SrcEltSize) && \"Unexpected vector element types\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1434, __PRETTY_FUNCTION__)); | |||
1435 | ||||
1436 | bool IsInt2FP = SrcEltVT.isInteger(); | |||
1437 | // Widening conversions | |||
1438 | if (EltSize > SrcEltSize && (EltSize / SrcEltSize >= 4)) { | |||
1439 | if (IsInt2FP) { | |||
1440 | // Do a regular integer sign/zero extension then convert to float. | |||
1441 | MVT IVecVT = MVT::getVectorVT(MVT::getIntegerVT(EltVT.getSizeInBits()), | |||
1442 | VT.getVectorElementCount()); | |||
1443 | unsigned ExtOpcode = Op.getOpcode() == ISD::UINT_TO_FP | |||
1444 | ? ISD::ZERO_EXTEND | |||
1445 | : ISD::SIGN_EXTEND; | |||
1446 | SDValue Ext = DAG.getNode(ExtOpcode, DL, IVecVT, Src); | |||
1447 | return DAG.getNode(Op.getOpcode(), DL, VT, Ext); | |||
1448 | } | |||
1449 | // FP2Int | |||
1450 | assert(SrcEltVT == MVT::f16 && "Unexpected FP_TO_[US]INT lowering")((SrcEltVT == MVT::f16 && "Unexpected FP_TO_[US]INT lowering" ) ? static_cast<void> (0) : __assert_fail ("SrcEltVT == MVT::f16 && \"Unexpected FP_TO_[US]INT lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1450, __PRETTY_FUNCTION__)); | |||
1451 | // Do one doubling fp_extend then complete the operation by converting | |||
1452 | // to int. | |||
1453 | MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount()); | |||
1454 | SDValue FExt = DAG.getFPExtendOrRound(Src, DL, InterimFVT); | |||
1455 | return DAG.getNode(Op.getOpcode(), DL, VT, FExt); | |||
1456 | } | |||
1457 | ||||
1458 | // Narrowing conversions | |||
1459 | if (SrcEltSize > EltSize && (SrcEltSize / EltSize >= 4)) { | |||
1460 | if (IsInt2FP) { | |||
1461 | // One narrowing int_to_fp, then an fp_round. | |||
1462 | assert(EltVT == MVT::f16 && "Unexpected [US]_TO_FP lowering")((EltVT == MVT::f16 && "Unexpected [US]_TO_FP lowering" ) ? static_cast<void> (0) : __assert_fail ("EltVT == MVT::f16 && \"Unexpected [US]_TO_FP lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1462, __PRETTY_FUNCTION__)); | |||
1463 | MVT InterimFVT = MVT::getVectorVT(MVT::f32, VT.getVectorElementCount()); | |||
1464 | SDValue Int2FP = DAG.getNode(Op.getOpcode(), DL, InterimFVT, Src); | |||
1465 | return DAG.getFPExtendOrRound(Int2FP, DL, VT); | |||
1466 | } | |||
1467 | // FP2Int | |||
1468 | // One narrowing fp_to_int, then truncate the integer. If the float isn't | |||
1469 | // representable by the integer, the result is poison. | |||
1470 | MVT IVecVT = | |||
1471 | MVT::getVectorVT(MVT::getIntegerVT(SrcEltVT.getSizeInBits() / 2), | |||
1472 | VT.getVectorElementCount()); | |||
1473 | SDValue FP2Int = DAG.getNode(Op.getOpcode(), DL, IVecVT, Src); | |||
1474 | return DAG.getNode(ISD::TRUNCATE, DL, VT, FP2Int); | |||
1475 | } | |||
1476 | ||||
1477 | // Scalable vectors can exit here. Patterns will handle equally-sized | |||
1478 | // conversions halving/doubling ones. | |||
1479 | if (!VT.isFixedLengthVector()) | |||
1480 | return Op; | |||
1481 | ||||
1482 | // For fixed-length vectors we lower to a custom "VL" node. | |||
1483 | unsigned RVVOpc = 0; | |||
1484 | switch (Op.getOpcode()) { | |||
1485 | default: | |||
1486 | llvm_unreachable("Impossible opcode")::llvm::llvm_unreachable_internal("Impossible opcode", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1486); | |||
1487 | case ISD::FP_TO_SINT: | |||
1488 | RVVOpc = RISCVISD::FP_TO_SINT_VL; | |||
1489 | break; | |||
1490 | case ISD::FP_TO_UINT: | |||
1491 | RVVOpc = RISCVISD::FP_TO_UINT_VL; | |||
1492 | break; | |||
1493 | case ISD::SINT_TO_FP: | |||
1494 | RVVOpc = RISCVISD::SINT_TO_FP_VL; | |||
1495 | break; | |||
1496 | case ISD::UINT_TO_FP: | |||
1497 | RVVOpc = RISCVISD::UINT_TO_FP_VL; | |||
1498 | break; | |||
1499 | } | |||
1500 | ||||
1501 | MVT ContainerVT, SrcContainerVT; | |||
1502 | // Derive the reference container type from the larger vector type. | |||
1503 | if (SrcEltSize > EltSize) { | |||
1504 | SrcContainerVT = getContainerForFixedLengthVector(SrcVT); | |||
1505 | ContainerVT = | |||
1506 | SrcContainerVT.changeVectorElementType(VT.getVectorElementType()); | |||
1507 | } else { | |||
1508 | ContainerVT = getContainerForFixedLengthVector(VT); | |||
1509 | SrcContainerVT = ContainerVT.changeVectorElementType(SrcEltVT); | |||
1510 | } | |||
1511 | ||||
1512 | SDValue Mask, VL; | |||
1513 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
1514 | ||||
1515 | Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget); | |||
1516 | Src = DAG.getNode(RVVOpc, DL, ContainerVT, Src, Mask, VL); | |||
1517 | return convertFromScalableVector(VT, Src, DAG, Subtarget); | |||
1518 | } | |||
1519 | case ISD::VECREDUCE_ADD: | |||
1520 | case ISD::VECREDUCE_UMAX: | |||
1521 | case ISD::VECREDUCE_SMAX: | |||
1522 | case ISD::VECREDUCE_UMIN: | |||
1523 | case ISD::VECREDUCE_SMIN: | |||
1524 | case ISD::VECREDUCE_AND: | |||
1525 | case ISD::VECREDUCE_OR: | |||
1526 | case ISD::VECREDUCE_XOR: | |||
1527 | return lowerVECREDUCE(Op, DAG); | |||
1528 | case ISD::VECREDUCE_FADD: | |||
1529 | case ISD::VECREDUCE_SEQ_FADD: | |||
1530 | return lowerFPVECREDUCE(Op, DAG); | |||
1531 | case ISD::INSERT_SUBVECTOR: | |||
1532 | return lowerINSERT_SUBVECTOR(Op, DAG); | |||
1533 | case ISD::EXTRACT_SUBVECTOR: | |||
1534 | return lowerEXTRACT_SUBVECTOR(Op, DAG); | |||
1535 | case ISD::VECTOR_REVERSE: | |||
1536 | return lowerVECTOR_REVERSE(Op, DAG); | |||
1537 | case ISD::BUILD_VECTOR: | |||
1538 | return lowerBUILD_VECTOR(Op, DAG, Subtarget); | |||
1539 | case ISD::VECTOR_SHUFFLE: | |||
1540 | return lowerVECTOR_SHUFFLE(Op, DAG, Subtarget); | |||
1541 | case ISD::CONCAT_VECTORS: { | |||
1542 | // Split CONCAT_VECTORS into a series of INSERT_SUBVECTOR nodes. This is | |||
1543 | // better than going through the stack, as the default expansion does. | |||
1544 | SDLoc DL(Op); | |||
1545 | MVT VT = Op.getSimpleValueType(); | |||
1546 | assert(VT.isFixedLengthVector() && "Unexpected CONCAT_VECTORS lowering")((VT.isFixedLengthVector() && "Unexpected CONCAT_VECTORS lowering" ) ? static_cast<void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Unexpected CONCAT_VECTORS lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 1546, __PRETTY_FUNCTION__)); | |||
1547 | unsigned NumOpElts = | |||
1548 | Op.getOperand(0).getSimpleValueType().getVectorNumElements(); | |||
1549 | SDValue Vec = DAG.getUNDEF(VT); | |||
1550 | for (const auto &OpIdx : enumerate(Op->ops())) | |||
1551 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, Vec, OpIdx.value(), | |||
1552 | DAG.getIntPtrConstant(OpIdx.index() * NumOpElts, DL)); | |||
1553 | return Vec; | |||
1554 | } | |||
1555 | case ISD::LOAD: | |||
1556 | return lowerFixedLengthVectorLoadToRVV(Op, DAG); | |||
1557 | case ISD::STORE: | |||
1558 | return lowerFixedLengthVectorStoreToRVV(Op, DAG); | |||
1559 | case ISD::SETCC: | |||
1560 | return lowerFixedLengthVectorSetccToRVV(Op, DAG); | |||
1561 | case ISD::ADD: | |||
1562 | return lowerToScalableOp(Op, DAG, RISCVISD::ADD_VL); | |||
1563 | case ISD::SUB: | |||
1564 | return lowerToScalableOp(Op, DAG, RISCVISD::SUB_VL); | |||
1565 | case ISD::MUL: | |||
1566 | return lowerToScalableOp(Op, DAG, RISCVISD::MUL_VL); | |||
1567 | case ISD::MULHS: | |||
1568 | return lowerToScalableOp(Op, DAG, RISCVISD::MULHS_VL); | |||
1569 | case ISD::MULHU: | |||
1570 | return lowerToScalableOp(Op, DAG, RISCVISD::MULHU_VL); | |||
1571 | case ISD::AND: | |||
1572 | return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMAND_VL, | |||
1573 | RISCVISD::AND_VL); | |||
1574 | case ISD::OR: | |||
1575 | return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMOR_VL, | |||
1576 | RISCVISD::OR_VL); | |||
1577 | case ISD::XOR: | |||
1578 | return lowerFixedLengthVectorLogicOpToRVV(Op, DAG, RISCVISD::VMXOR_VL, | |||
1579 | RISCVISD::XOR_VL); | |||
1580 | case ISD::SDIV: | |||
1581 | return lowerToScalableOp(Op, DAG, RISCVISD::SDIV_VL); | |||
1582 | case ISD::SREM: | |||
1583 | return lowerToScalableOp(Op, DAG, RISCVISD::SREM_VL); | |||
1584 | case ISD::UDIV: | |||
1585 | return lowerToScalableOp(Op, DAG, RISCVISD::UDIV_VL); | |||
1586 | case ISD::UREM: | |||
1587 | return lowerToScalableOp(Op, DAG, RISCVISD::UREM_VL); | |||
1588 | case ISD::SHL: | |||
1589 | return lowerToScalableOp(Op, DAG, RISCVISD::SHL_VL); | |||
1590 | case ISD::SRA: | |||
1591 | return lowerToScalableOp(Op, DAG, RISCVISD::SRA_VL); | |||
1592 | case ISD::SRL: | |||
1593 | return lowerToScalableOp(Op, DAG, RISCVISD::SRL_VL); | |||
1594 | case ISD::FADD: | |||
1595 | return lowerToScalableOp(Op, DAG, RISCVISD::FADD_VL); | |||
1596 | case ISD::FSUB: | |||
1597 | return lowerToScalableOp(Op, DAG, RISCVISD::FSUB_VL); | |||
1598 | case ISD::FMUL: | |||
1599 | return lowerToScalableOp(Op, DAG, RISCVISD::FMUL_VL); | |||
1600 | case ISD::FDIV: | |||
1601 | return lowerToScalableOp(Op, DAG, RISCVISD::FDIV_VL); | |||
1602 | case ISD::FNEG: | |||
1603 | return lowerToScalableOp(Op, DAG, RISCVISD::FNEG_VL); | |||
1604 | case ISD::FABS: | |||
1605 | return lowerToScalableOp(Op, DAG, RISCVISD::FABS_VL); | |||
1606 | case ISD::FSQRT: | |||
1607 | return lowerToScalableOp(Op, DAG, RISCVISD::FSQRT_VL); | |||
1608 | case ISD::FMA: | |||
1609 | return lowerToScalableOp(Op, DAG, RISCVISD::FMA_VL); | |||
1610 | case ISD::SMIN: | |||
1611 | return lowerToScalableOp(Op, DAG, RISCVISD::SMIN_VL); | |||
1612 | case ISD::SMAX: | |||
1613 | return lowerToScalableOp(Op, DAG, RISCVISD::SMAX_VL); | |||
1614 | case ISD::UMIN: | |||
1615 | return lowerToScalableOp(Op, DAG, RISCVISD::UMIN_VL); | |||
1616 | case ISD::UMAX: | |||
1617 | return lowerToScalableOp(Op, DAG, RISCVISD::UMAX_VL); | |||
1618 | case ISD::ABS: | |||
1619 | return lowerABS(Op, DAG); | |||
1620 | case ISD::VSELECT: | |||
1621 | return lowerFixedLengthVectorSelectToRVV(Op, DAG); | |||
1622 | } | |||
1623 | } | |||
1624 | ||||
1625 | static SDValue getTargetNode(GlobalAddressSDNode *N, SDLoc DL, EVT Ty, | |||
1626 | SelectionDAG &DAG, unsigned Flags) { | |||
1627 | return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags); | |||
1628 | } | |||
1629 | ||||
1630 | static SDValue getTargetNode(BlockAddressSDNode *N, SDLoc DL, EVT Ty, | |||
1631 | SelectionDAG &DAG, unsigned Flags) { | |||
1632 | return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(), | |||
1633 | Flags); | |||
1634 | } | |||
1635 | ||||
1636 | static SDValue getTargetNode(ConstantPoolSDNode *N, SDLoc DL, EVT Ty, | |||
1637 | SelectionDAG &DAG, unsigned Flags) { | |||
1638 | return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(), | |||
1639 | N->getOffset(), Flags); | |||
1640 | } | |||
1641 | ||||
1642 | static SDValue getTargetNode(JumpTableSDNode *N, SDLoc DL, EVT Ty, | |||
1643 | SelectionDAG &DAG, unsigned Flags) { | |||
1644 | return DAG.getTargetJumpTable(N->getIndex(), Ty, Flags); | |||
1645 | } | |||
1646 | ||||
1647 | template <class NodeTy> | |||
1648 | SDValue RISCVTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG, | |||
1649 | bool IsLocal) const { | |||
1650 | SDLoc DL(N); | |||
1651 | EVT Ty = getPointerTy(DAG.getDataLayout()); | |||
1652 | ||||
1653 | if (isPositionIndependent()) { | |||
1654 | SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0); | |||
1655 | if (IsLocal) | |||
1656 | // Use PC-relative addressing to access the symbol. This generates the | |||
1657 | // pattern (PseudoLLA sym), which expands to (addi (auipc %pcrel_hi(sym)) | |||
1658 | // %pcrel_lo(auipc)). | |||
1659 | return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0); | |||
1660 | ||||
1661 | // Use PC-relative addressing to access the GOT for this symbol, then load | |||
1662 | // the address from the GOT. This generates the pattern (PseudoLA sym), | |||
1663 | // which expands to (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))). | |||
1664 | return SDValue(DAG.getMachineNode(RISCV::PseudoLA, DL, Ty, Addr), 0); | |||
1665 | } | |||
1666 | ||||
1667 | switch (getTargetMachine().getCodeModel()) { | |||
1668 | default: | |||
1669 | report_fatal_error("Unsupported code model for lowering"); | |||
1670 | case CodeModel::Small: { | |||
1671 | // Generate a sequence for accessing addresses within the first 2 GiB of | |||
1672 | // address space. This generates the pattern (addi (lui %hi(sym)) %lo(sym)). | |||
1673 | SDValue AddrHi = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_HI); | |||
1674 | SDValue AddrLo = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_LO); | |||
1675 | SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0); | |||
1676 | return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNHi, AddrLo), 0); | |||
1677 | } | |||
1678 | case CodeModel::Medium: { | |||
1679 | // Generate a sequence for accessing addresses within any 2GiB range within | |||
1680 | // the address space. This generates the pattern (PseudoLLA sym), which | |||
1681 | // expands to (addi (auipc %pcrel_hi(sym)) %pcrel_lo(auipc)). | |||
1682 | SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0); | |||
1683 | return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0); | |||
1684 | } | |||
1685 | } | |||
1686 | } | |||
1687 | ||||
1688 | SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op, | |||
1689 | SelectionDAG &DAG) const { | |||
1690 | SDLoc DL(Op); | |||
1691 | EVT Ty = Op.getValueType(); | |||
1692 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); | |||
1693 | int64_t Offset = N->getOffset(); | |||
1694 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1695 | ||||
1696 | const GlobalValue *GV = N->getGlobal(); | |||
1697 | bool IsLocal = getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV); | |||
1698 | SDValue Addr = getAddr(N, DAG, IsLocal); | |||
1699 | ||||
1700 | // In order to maximise the opportunity for common subexpression elimination, | |||
1701 | // emit a separate ADD node for the global address offset instead of folding | |||
1702 | // it in the global address node. Later peephole optimisations may choose to | |||
1703 | // fold it back in when profitable. | |||
1704 | if (Offset != 0) | |||
1705 | return DAG.getNode(ISD::ADD, DL, Ty, Addr, | |||
1706 | DAG.getConstant(Offset, DL, XLenVT)); | |||
1707 | return Addr; | |||
1708 | } | |||
1709 | ||||
1710 | SDValue RISCVTargetLowering::lowerBlockAddress(SDValue Op, | |||
1711 | SelectionDAG &DAG) const { | |||
1712 | BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op); | |||
1713 | ||||
1714 | return getAddr(N, DAG); | |||
1715 | } | |||
1716 | ||||
1717 | SDValue RISCVTargetLowering::lowerConstantPool(SDValue Op, | |||
1718 | SelectionDAG &DAG) const { | |||
1719 | ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op); | |||
1720 | ||||
1721 | return getAddr(N, DAG); | |||
1722 | } | |||
1723 | ||||
1724 | SDValue RISCVTargetLowering::lowerJumpTable(SDValue Op, | |||
1725 | SelectionDAG &DAG) const { | |||
1726 | JumpTableSDNode *N = cast<JumpTableSDNode>(Op); | |||
1727 | ||||
1728 | return getAddr(N, DAG); | |||
1729 | } | |||
1730 | ||||
1731 | SDValue RISCVTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N, | |||
1732 | SelectionDAG &DAG, | |||
1733 | bool UseGOT) const { | |||
1734 | SDLoc DL(N); | |||
1735 | EVT Ty = getPointerTy(DAG.getDataLayout()); | |||
1736 | const GlobalValue *GV = N->getGlobal(); | |||
1737 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1738 | ||||
1739 | if (UseGOT) { | |||
1740 | // Use PC-relative addressing to access the GOT for this TLS symbol, then | |||
1741 | // load the address from the GOT and add the thread pointer. This generates | |||
1742 | // the pattern (PseudoLA_TLS_IE sym), which expands to | |||
1743 | // (ld (auipc %tls_ie_pcrel_hi(sym)) %pcrel_lo(auipc)). | |||
1744 | SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0); | |||
1745 | SDValue Load = | |||
1746 | SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_IE, DL, Ty, Addr), 0); | |||
1747 | ||||
1748 | // Add the thread pointer. | |||
1749 | SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT); | |||
1750 | return DAG.getNode(ISD::ADD, DL, Ty, Load, TPReg); | |||
1751 | } | |||
1752 | ||||
1753 | // Generate a sequence for accessing the address relative to the thread | |||
1754 | // pointer, with the appropriate adjustment for the thread pointer offset. | |||
1755 | // This generates the pattern | |||
1756 | // (add (add_tprel (lui %tprel_hi(sym)) tp %tprel_add(sym)) %tprel_lo(sym)) | |||
1757 | SDValue AddrHi = | |||
1758 | DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_HI); | |||
1759 | SDValue AddrAdd = | |||
1760 | DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_ADD); | |||
1761 | SDValue AddrLo = | |||
1762 | DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_LO); | |||
1763 | ||||
1764 | SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0); | |||
1765 | SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT); | |||
1766 | SDValue MNAdd = SDValue( | |||
1767 | DAG.getMachineNode(RISCV::PseudoAddTPRel, DL, Ty, MNHi, TPReg, AddrAdd), | |||
1768 | 0); | |||
1769 | return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNAdd, AddrLo), 0); | |||
1770 | } | |||
1771 | ||||
1772 | SDValue RISCVTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N, | |||
1773 | SelectionDAG &DAG) const { | |||
1774 | SDLoc DL(N); | |||
1775 | EVT Ty = getPointerTy(DAG.getDataLayout()); | |||
1776 | IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits()); | |||
1777 | const GlobalValue *GV = N->getGlobal(); | |||
1778 | ||||
1779 | // Use a PC-relative addressing mode to access the global dynamic GOT address. | |||
1780 | // This generates the pattern (PseudoLA_TLS_GD sym), which expands to | |||
1781 | // (addi (auipc %tls_gd_pcrel_hi(sym)) %pcrel_lo(auipc)). | |||
1782 | SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0); | |||
1783 | SDValue Load = | |||
1784 | SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_GD, DL, Ty, Addr), 0); | |||
1785 | ||||
1786 | // Prepare argument list to generate call. | |||
1787 | ArgListTy Args; | |||
1788 | ArgListEntry Entry; | |||
1789 | Entry.Node = Load; | |||
1790 | Entry.Ty = CallTy; | |||
1791 | Args.push_back(Entry); | |||
1792 | ||||
1793 | // Setup call to __tls_get_addr. | |||
1794 | TargetLowering::CallLoweringInfo CLI(DAG); | |||
1795 | CLI.setDebugLoc(DL) | |||
1796 | .setChain(DAG.getEntryNode()) | |||
1797 | .setLibCallee(CallingConv::C, CallTy, | |||
1798 | DAG.getExternalSymbol("__tls_get_addr", Ty), | |||
1799 | std::move(Args)); | |||
1800 | ||||
1801 | return LowerCallTo(CLI).first; | |||
1802 | } | |||
1803 | ||||
1804 | SDValue RISCVTargetLowering::lowerGlobalTLSAddress(SDValue Op, | |||
1805 | SelectionDAG &DAG) const { | |||
1806 | SDLoc DL(Op); | |||
1807 | EVT Ty = Op.getValueType(); | |||
1808 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op); | |||
1809 | int64_t Offset = N->getOffset(); | |||
1810 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1811 | ||||
1812 | TLSModel::Model Model = getTargetMachine().getTLSModel(N->getGlobal()); | |||
1813 | ||||
1814 | if (DAG.getMachineFunction().getFunction().getCallingConv() == | |||
1815 | CallingConv::GHC) | |||
1816 | report_fatal_error("In GHC calling convention TLS is not supported"); | |||
1817 | ||||
1818 | SDValue Addr; | |||
1819 | switch (Model) { | |||
1820 | case TLSModel::LocalExec: | |||
1821 | Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/false); | |||
1822 | break; | |||
1823 | case TLSModel::InitialExec: | |||
1824 | Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/true); | |||
1825 | break; | |||
1826 | case TLSModel::LocalDynamic: | |||
1827 | case TLSModel::GeneralDynamic: | |||
1828 | Addr = getDynamicTLSAddr(N, DAG); | |||
1829 | break; | |||
1830 | } | |||
1831 | ||||
1832 | // In order to maximise the opportunity for common subexpression elimination, | |||
1833 | // emit a separate ADD node for the global address offset instead of folding | |||
1834 | // it in the global address node. Later peephole optimisations may choose to | |||
1835 | // fold it back in when profitable. | |||
1836 | if (Offset != 0) | |||
1837 | return DAG.getNode(ISD::ADD, DL, Ty, Addr, | |||
1838 | DAG.getConstant(Offset, DL, XLenVT)); | |||
1839 | return Addr; | |||
1840 | } | |||
1841 | ||||
1842 | SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const { | |||
1843 | SDValue CondV = Op.getOperand(0); | |||
1844 | SDValue TrueV = Op.getOperand(1); | |||
1845 | SDValue FalseV = Op.getOperand(2); | |||
1846 | SDLoc DL(Op); | |||
1847 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1848 | ||||
1849 | // If the result type is XLenVT and CondV is the output of a SETCC node | |||
1850 | // which also operated on XLenVT inputs, then merge the SETCC node into the | |||
1851 | // lowered RISCVISD::SELECT_CC to take advantage of the integer | |||
1852 | // compare+branch instructions. i.e.: | |||
1853 | // (select (setcc lhs, rhs, cc), truev, falsev) | |||
1854 | // -> (riscvisd::select_cc lhs, rhs, cc, truev, falsev) | |||
1855 | if (Op.getSimpleValueType() == XLenVT && CondV.getOpcode() == ISD::SETCC && | |||
1856 | CondV.getOperand(0).getSimpleValueType() == XLenVT) { | |||
1857 | SDValue LHS = CondV.getOperand(0); | |||
1858 | SDValue RHS = CondV.getOperand(1); | |||
1859 | auto CC = cast<CondCodeSDNode>(CondV.getOperand(2)); | |||
1860 | ISD::CondCode CCVal = CC->get(); | |||
1861 | ||||
1862 | normaliseSetCC(LHS, RHS, CCVal); | |||
1863 | ||||
1864 | SDValue TargetCC = DAG.getConstant(CCVal, DL, XLenVT); | |||
1865 | SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV}; | |||
1866 | return DAG.getNode(RISCVISD::SELECT_CC, DL, Op.getValueType(), Ops); | |||
1867 | } | |||
1868 | ||||
1869 | // Otherwise: | |||
1870 | // (select condv, truev, falsev) | |||
1871 | // -> (riscvisd::select_cc condv, zero, setne, truev, falsev) | |||
1872 | SDValue Zero = DAG.getConstant(0, DL, XLenVT); | |||
1873 | SDValue SetNE = DAG.getConstant(ISD::SETNE, DL, XLenVT); | |||
1874 | ||||
1875 | SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV}; | |||
1876 | ||||
1877 | return DAG.getNode(RISCVISD::SELECT_CC, DL, Op.getValueType(), Ops); | |||
1878 | } | |||
1879 | ||||
1880 | SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { | |||
1881 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1882 | RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>(); | |||
1883 | ||||
1884 | SDLoc DL(Op); | |||
1885 | SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), | |||
1886 | getPointerTy(MF.getDataLayout())); | |||
1887 | ||||
1888 | // vastart just stores the address of the VarArgsFrameIndex slot into the | |||
1889 | // memory location argument. | |||
1890 | const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); | |||
1891 | return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1), | |||
1892 | MachinePointerInfo(SV)); | |||
1893 | } | |||
1894 | ||||
1895 | SDValue RISCVTargetLowering::lowerFRAMEADDR(SDValue Op, | |||
1896 | SelectionDAG &DAG) const { | |||
1897 | const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo(); | |||
1898 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1899 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
1900 | MFI.setFrameAddressIsTaken(true); | |||
1901 | Register FrameReg = RI.getFrameRegister(MF); | |||
1902 | int XLenInBytes = Subtarget.getXLen() / 8; | |||
1903 | ||||
1904 | EVT VT = Op.getValueType(); | |||
1905 | SDLoc DL(Op); | |||
1906 | SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT); | |||
1907 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
1908 | while (Depth--) { | |||
1909 | int Offset = -(XLenInBytes * 2); | |||
1910 | SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr, | |||
1911 | DAG.getIntPtrConstant(Offset, DL)); | |||
1912 | FrameAddr = | |||
1913 | DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo()); | |||
1914 | } | |||
1915 | return FrameAddr; | |||
1916 | } | |||
1917 | ||||
1918 | SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op, | |||
1919 | SelectionDAG &DAG) const { | |||
1920 | const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo(); | |||
1921 | MachineFunction &MF = DAG.getMachineFunction(); | |||
1922 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
1923 | MFI.setReturnAddressIsTaken(true); | |||
1924 | MVT XLenVT = Subtarget.getXLenVT(); | |||
1925 | int XLenInBytes = Subtarget.getXLen() / 8; | |||
1926 | ||||
1927 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) | |||
1928 | return SDValue(); | |||
1929 | ||||
1930 | EVT VT = Op.getValueType(); | |||
1931 | SDLoc DL(Op); | |||
1932 | unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
1933 | if (Depth) { | |||
1934 | int Off = -XLenInBytes; | |||
1935 | SDValue FrameAddr = lowerFRAMEADDR(Op, DAG); | |||
1936 | SDValue Offset = DAG.getConstant(Off, DL, VT); | |||
1937 | return DAG.getLoad(VT, DL, DAG.getEntryNode(), | |||
1938 | DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset), | |||
1939 | MachinePointerInfo()); | |||
1940 | } | |||
1941 | ||||
1942 | // Return the value of the return address register, marking it an implicit | |||
1943 | // live-in. | |||
1944 | Register Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT)); | |||
1945 | return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT); | |||
1946 | } | |||
1947 | ||||
1948 | SDValue RISCVTargetLowering::lowerShiftLeftParts(SDValue Op, | |||
1949 | SelectionDAG &DAG) const { | |||
1950 | SDLoc DL(Op); | |||
1951 | SDValue Lo = Op.getOperand(0); | |||
1952 | SDValue Hi = Op.getOperand(1); | |||
1953 | SDValue Shamt = Op.getOperand(2); | |||
1954 | EVT VT = Lo.getValueType(); | |||
1955 | ||||
1956 | // if Shamt-XLEN < 0: // Shamt < XLEN | |||
1957 | // Lo = Lo << Shamt | |||
1958 | // Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 - Shamt)) | |||
1959 | // else: | |||
1960 | // Lo = 0 | |||
1961 | // Hi = Lo << (Shamt-XLEN) | |||
1962 | ||||
1963 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
1964 | SDValue One = DAG.getConstant(1, DL, VT); | |||
1965 | SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT); | |||
1966 | SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT); | |||
1967 | SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen); | |||
1968 | SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt); | |||
1969 | ||||
1970 | SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt); | |||
1971 | SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One); | |||
1972 | SDValue ShiftRightLo = | |||
1973 | DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, XLenMinus1Shamt); | |||
1974 | SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt); | |||
1975 | SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo); | |||
1976 | SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusXLen); | |||
1977 | ||||
1978 | SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT); | |||
1979 | ||||
1980 | Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero); | |||
1981 | Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse); | |||
1982 | ||||
1983 | SDValue Parts[2] = {Lo, Hi}; | |||
1984 | return DAG.getMergeValues(Parts, DL); | |||
1985 | } | |||
1986 | ||||
1987 | SDValue RISCVTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, | |||
1988 | bool IsSRA) const { | |||
1989 | SDLoc DL(Op); | |||
1990 | SDValue Lo = Op.getOperand(0); | |||
1991 | SDValue Hi = Op.getOperand(1); | |||
1992 | SDValue Shamt = Op.getOperand(2); | |||
1993 | EVT VT = Lo.getValueType(); | |||
1994 | ||||
1995 | // SRA expansion: | |||
1996 | // if Shamt-XLEN < 0: // Shamt < XLEN | |||
1997 | // Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt)) | |||
1998 | // Hi = Hi >>s Shamt | |||
1999 | // else: | |||
2000 | // Lo = Hi >>s (Shamt-XLEN); | |||
2001 | // Hi = Hi >>s (XLEN-1) | |||
2002 | // | |||
2003 | // SRL expansion: | |||
2004 | // if Shamt-XLEN < 0: // Shamt < XLEN | |||
2005 | // Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt)) | |||
2006 | // Hi = Hi >>u Shamt | |||
2007 | // else: | |||
2008 | // Lo = Hi >>u (Shamt-XLEN); | |||
2009 | // Hi = 0; | |||
2010 | ||||
2011 | unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL; | |||
2012 | ||||
2013 | SDValue Zero = DAG.getConstant(0, DL, VT); | |||
2014 | SDValue One = DAG.getConstant(1, DL, VT); | |||
2015 | SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT); | |||
2016 | SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT); | |||
2017 | SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen); | |||
2018 | SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt); | |||
2019 | ||||
2020 | SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt); | |||
2021 | SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One); | |||
2022 | SDValue ShiftLeftHi = | |||
2023 | DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, XLenMinus1Shamt); | |||
2024 | SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi); | |||
2025 | SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt); | |||
2026 | SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusXLen); | |||
2027 | SDValue HiFalse = | |||
2028 | IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, XLenMinus1) : Zero; | |||
2029 | ||||
2030 | SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT); | |||
2031 | ||||
2032 | Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse); | |||
2033 | Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse); | |||
2034 | ||||
2035 | SDValue Parts[2] = {Lo, Hi}; | |||
2036 | return DAG.getMergeValues(Parts, DL); | |||
2037 | } | |||
2038 | ||||
2039 | // Custom-lower a SPLAT_VECTOR where XLEN<SEW, as the SEW element type is | |||
2040 | // illegal (currently only vXi64 RV32). | |||
2041 | // FIXME: We could also catch non-constant sign-extended i32 values and lower | |||
2042 | // them to SPLAT_VECTOR_I64 | |||
2043 | SDValue RISCVTargetLowering::lowerSPLATVECTOR(SDValue Op, | |||
2044 | SelectionDAG &DAG) const { | |||
2045 | SDLoc DL(Op); | |||
2046 | EVT VecVT = Op.getValueType(); | |||
2047 | assert(!Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64 &&((!Subtarget.is64Bit() && VecVT.getVectorElementType( ) == MVT::i64 && "Unexpected SPLAT_VECTOR lowering") ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64 && \"Unexpected SPLAT_VECTOR lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2048, __PRETTY_FUNCTION__)) | |||
2048 | "Unexpected SPLAT_VECTOR lowering")((!Subtarget.is64Bit() && VecVT.getVectorElementType( ) == MVT::i64 && "Unexpected SPLAT_VECTOR lowering") ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64 && \"Unexpected SPLAT_VECTOR lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2048, __PRETTY_FUNCTION__)); | |||
2049 | SDValue SplatVal = Op.getOperand(0); | |||
2050 | ||||
2051 | // If we can prove that the value is a sign-extended 32-bit value, lower this | |||
2052 | // as a custom node in order to try and match RVV vector/scalar instructions. | |||
2053 | if (auto *CVal = dyn_cast<ConstantSDNode>(SplatVal)) { | |||
2054 | if (isInt<32>(CVal->getSExtValue())) | |||
2055 | return DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, | |||
2056 | DAG.getConstant(CVal->getSExtValue(), DL, MVT::i32)); | |||
2057 | } | |||
2058 | ||||
2059 | if (SplatVal.getOpcode() == ISD::SIGN_EXTEND && | |||
2060 | SplatVal.getOperand(0).getValueType() == MVT::i32) { | |||
2061 | return DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, | |||
2062 | SplatVal.getOperand(0)); | |||
2063 | } | |||
2064 | ||||
2065 | // Else, on RV32 we lower an i64-element SPLAT_VECTOR thus, being careful not | |||
2066 | // to accidentally sign-extend the 32-bit halves to the e64 SEW: | |||
2067 | // vmv.v.x vX, hi | |||
2068 | // vsll.vx vX, vX, /*32*/ | |||
2069 | // vmv.v.x vY, lo | |||
2070 | // vsll.vx vY, vY, /*32*/ | |||
2071 | // vsrl.vx vY, vY, /*32*/ | |||
2072 | // vor.vv vX, vX, vY | |||
2073 | SDValue One = DAG.getConstant(1, DL, MVT::i32); | |||
2074 | SDValue Zero = DAG.getConstant(0, DL, MVT::i32); | |||
2075 | SDValue ThirtyTwoV = DAG.getConstant(32, DL, VecVT); | |||
2076 | SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, SplatVal, Zero); | |||
2077 | SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, SplatVal, One); | |||
2078 | ||||
2079 | Lo = DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, Lo); | |||
2080 | Lo = DAG.getNode(ISD::SHL, DL, VecVT, Lo, ThirtyTwoV); | |||
2081 | Lo = DAG.getNode(ISD::SRL, DL, VecVT, Lo, ThirtyTwoV); | |||
2082 | ||||
2083 | if (isNullConstant(Hi)) | |||
2084 | return Lo; | |||
2085 | ||||
2086 | Hi = DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, Hi); | |||
2087 | Hi = DAG.getNode(ISD::SHL, DL, VecVT, Hi, ThirtyTwoV); | |||
2088 | ||||
2089 | return DAG.getNode(ISD::OR, DL, VecVT, Lo, Hi); | |||
2090 | } | |||
2091 | ||||
2092 | // Custom-lower extensions from mask vectors by using a vselect either with 1 | |||
2093 | // for zero/any-extension or -1 for sign-extension: | |||
2094 | // (vXiN = (s|z)ext vXi1:vmask) -> (vXiN = vselect vmask, (-1 or 1), 0) | |||
2095 | // Note that any-extension is lowered identically to zero-extension. | |||
2096 | SDValue RISCVTargetLowering::lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG, | |||
2097 | int64_t ExtTrueVal) const { | |||
2098 | SDLoc DL(Op); | |||
2099 | MVT VecVT = Op.getSimpleValueType(); | |||
2100 | SDValue Src = Op.getOperand(0); | |||
2101 | // Only custom-lower extensions from mask types | |||
2102 | assert(Src.getValueType().isVector() &&((Src.getValueType().isVector() && Src.getValueType() .getVectorElementType() == MVT::i1) ? static_cast<void> (0) : __assert_fail ("Src.getValueType().isVector() && Src.getValueType().getVectorElementType() == MVT::i1" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2103, __PRETTY_FUNCTION__)) | |||
2103 | Src.getValueType().getVectorElementType() == MVT::i1)((Src.getValueType().isVector() && Src.getValueType() .getVectorElementType() == MVT::i1) ? static_cast<void> (0) : __assert_fail ("Src.getValueType().isVector() && Src.getValueType().getVectorElementType() == MVT::i1" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2103, __PRETTY_FUNCTION__)); | |||
2104 | ||||
2105 | MVT XLenVT = Subtarget.getXLenVT(); | |||
2106 | SDValue SplatZero = DAG.getConstant(0, DL, XLenVT); | |||
2107 | SDValue SplatTrueVal = DAG.getConstant(ExtTrueVal, DL, XLenVT); | |||
2108 | ||||
2109 | if (VecVT.isScalableVector()) { | |||
2110 | // Be careful not to introduce illegal scalar types at this stage, and be | |||
2111 | // careful also about splatting constants as on RV32, vXi64 SPLAT_VECTOR is | |||
2112 | // illegal and must be expanded. Since we know that the constants are | |||
2113 | // sign-extended 32-bit values, we use SPLAT_VECTOR_I64 directly. | |||
2114 | bool IsRV32E64 = | |||
2115 | !Subtarget.is64Bit() && VecVT.getVectorElementType() == MVT::i64; | |||
2116 | ||||
2117 | if (!IsRV32E64) { | |||
2118 | SplatZero = DAG.getSplatVector(VecVT, DL, SplatZero); | |||
2119 | SplatTrueVal = DAG.getSplatVector(VecVT, DL, SplatTrueVal); | |||
2120 | } else { | |||
2121 | SplatZero = DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, SplatZero); | |||
2122 | SplatTrueVal = | |||
2123 | DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VecVT, SplatTrueVal); | |||
2124 | } | |||
2125 | ||||
2126 | return DAG.getNode(ISD::VSELECT, DL, VecVT, Src, SplatTrueVal, SplatZero); | |||
2127 | } | |||
2128 | ||||
2129 | MVT ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2130 | MVT I1ContainerVT = | |||
2131 | MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount()); | |||
2132 | ||||
2133 | SDValue CC = convertToScalableVector(I1ContainerVT, Src, DAG, Subtarget); | |||
2134 | ||||
2135 | SDValue Mask, VL; | |||
2136 | std::tie(Mask, VL) = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget); | |||
2137 | ||||
2138 | SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, SplatZero, VL); | |||
2139 | SplatTrueVal = | |||
2140 | DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, SplatTrueVal, VL); | |||
2141 | SDValue Select = DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, CC, | |||
2142 | SplatTrueVal, SplatZero, VL); | |||
2143 | ||||
2144 | return convertFromScalableVector(VecVT, Select, DAG, Subtarget); | |||
2145 | } | |||
2146 | ||||
2147 | SDValue RISCVTargetLowering::lowerFixedLengthVectorExtendToRVV( | |||
2148 | SDValue Op, SelectionDAG &DAG, unsigned ExtendOpc) const { | |||
2149 | MVT ExtVT = Op.getSimpleValueType(); | |||
2150 | // Only custom-lower extensions from fixed-length vector types. | |||
2151 | if (!ExtVT.isFixedLengthVector()) | |||
2152 | return Op; | |||
2153 | MVT VT = Op.getOperand(0).getSimpleValueType(); | |||
2154 | // Grab the canonical container type for the extended type. Infer the smaller | |||
2155 | // type from that to ensure the same number of vector elements, as we know | |||
2156 | // the LMUL will be sufficient to hold the smaller type. | |||
2157 | MVT ContainerExtVT = getContainerForFixedLengthVector(ExtVT); | |||
2158 | // Get the extended container type manually to ensure the same number of | |||
2159 | // vector elements between source and dest. | |||
2160 | MVT ContainerVT = MVT::getVectorVT(VT.getVectorElementType(), | |||
2161 | ContainerExtVT.getVectorElementCount()); | |||
2162 | ||||
2163 | SDValue Op1 = | |||
2164 | convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget); | |||
2165 | ||||
2166 | SDLoc DL(Op); | |||
2167 | SDValue Mask, VL; | |||
2168 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
2169 | ||||
2170 | SDValue Ext = DAG.getNode(ExtendOpc, DL, ContainerExtVT, Op1, Mask, VL); | |||
2171 | ||||
2172 | return convertFromScalableVector(ExtVT, Ext, DAG, Subtarget); | |||
2173 | } | |||
2174 | ||||
2175 | // Custom-lower truncations from vectors to mask vectors by using a mask and a | |||
2176 | // setcc operation: | |||
2177 | // (vXi1 = trunc vXiN vec) -> (vXi1 = setcc (and vec, 1), 0, ne) | |||
2178 | SDValue RISCVTargetLowering::lowerVectorMaskTrunc(SDValue Op, | |||
2179 | SelectionDAG &DAG) const { | |||
2180 | SDLoc DL(Op); | |||
2181 | EVT MaskVT = Op.getValueType(); | |||
2182 | // Only expect to custom-lower truncations to mask types | |||
2183 | assert(MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 &&((MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 && "Unexpected type for vector mask lowering" ) ? static_cast<void> (0) : __assert_fail ("MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 && \"Unexpected type for vector mask lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2184, __PRETTY_FUNCTION__)) | |||
2184 | "Unexpected type for vector mask lowering")((MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 && "Unexpected type for vector mask lowering" ) ? static_cast<void> (0) : __assert_fail ("MaskVT.isVector() && MaskVT.getVectorElementType() == MVT::i1 && \"Unexpected type for vector mask lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2184, __PRETTY_FUNCTION__)); | |||
2185 | SDValue Src = Op.getOperand(0); | |||
2186 | MVT VecVT = Src.getSimpleValueType(); | |||
2187 | ||||
2188 | // If this is a fixed vector, we need to convert it to a scalable vector. | |||
2189 | MVT ContainerVT = VecVT; | |||
2190 | if (VecVT.isFixedLengthVector()) { | |||
2191 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2192 | Src = convertToScalableVector(ContainerVT, Src, DAG, Subtarget); | |||
2193 | } | |||
2194 | ||||
2195 | SDValue SplatOne = DAG.getConstant(1, DL, Subtarget.getXLenVT()); | |||
2196 | SDValue SplatZero = DAG.getConstant(0, DL, Subtarget.getXLenVT()); | |||
2197 | ||||
2198 | SplatOne = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, SplatOne); | |||
2199 | SplatZero = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, SplatZero); | |||
2200 | ||||
2201 | if (VecVT.isScalableVector()) { | |||
2202 | SDValue Trunc = DAG.getNode(ISD::AND, DL, VecVT, Src, SplatOne); | |||
2203 | return DAG.getSetCC(DL, MaskVT, Trunc, SplatZero, ISD::SETNE); | |||
2204 | } | |||
2205 | ||||
2206 | SDValue Mask, VL; | |||
2207 | std::tie(Mask, VL) = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget); | |||
2208 | ||||
2209 | MVT MaskContainerVT = ContainerVT.changeVectorElementType(MVT::i1); | |||
2210 | SDValue Trunc = | |||
2211 | DAG.getNode(RISCVISD::AND_VL, DL, ContainerVT, Src, SplatOne, Mask, VL); | |||
2212 | Trunc = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskContainerVT, Trunc, SplatZero, | |||
2213 | DAG.getCondCode(ISD::SETNE), Mask, VL); | |||
2214 | return convertFromScalableVector(MaskVT, Trunc, DAG, Subtarget); | |||
2215 | } | |||
2216 | ||||
2217 | SDValue RISCVTargetLowering::lowerINSERT_VECTOR_ELT(SDValue Op, | |||
2218 | SelectionDAG &DAG) const { | |||
2219 | SDLoc DL(Op); | |||
2220 | MVT VecVT = Op.getSimpleValueType(); | |||
2221 | SDValue Vec = Op.getOperand(0); | |||
2222 | SDValue Val = Op.getOperand(1); | |||
2223 | SDValue Idx = Op.getOperand(2); | |||
2224 | ||||
2225 | MVT ContainerVT = VecVT; | |||
2226 | // If the operand is a fixed-length vector, convert to a scalable one. | |||
2227 | if (VecVT.isFixedLengthVector()) { | |||
2228 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2229 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
2230 | } | |||
2231 | ||||
2232 | SDValue Mask, VL; | |||
2233 | std::tie(Mask, VL) = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget); | |||
2234 | ||||
2235 | // Custom-legalize INSERT_VECTOR_ELT where XLEN>=SEW, so that the vector is | |||
2236 | // first slid down into position, the value is inserted into the first | |||
2237 | // position, and the vector is slid back up. We do this to simplify patterns. | |||
2238 | // (slideup vec, (insertelt (slidedown impdef, vec, idx), val, 0), idx), | |||
2239 | if (Subtarget.is64Bit() || Val.getValueType() != MVT::i64) { | |||
2240 | if (isNullConstant(Idx)) | |||
2241 | return DAG.getNode(RISCVISD::VMV_S_XF_VL, DL, ContainerVT, Vec, Val, VL); | |||
2242 | SDValue Slidedown = | |||
2243 | DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, ContainerVT, | |||
2244 | DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL); | |||
2245 | SDValue InsertElt0 = | |||
2246 | DAG.getNode(RISCVISD::VMV_S_XF_VL, DL, ContainerVT, Slidedown, Val, VL); | |||
2247 | return DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, ContainerVT, Vec, InsertElt0, | |||
2248 | Idx, Mask, VL); | |||
2249 | } | |||
2250 | ||||
2251 | // Custom-legalize INSERT_VECTOR_ELT where XLEN<SEW, as the SEW element type | |||
2252 | // is illegal (currently only vXi64 RV32). | |||
2253 | // Since there is no easy way of getting a single element into a vector when | |||
2254 | // XLEN<SEW, we lower the operation to the following sequence: | |||
2255 | // splat vVal, rVal | |||
2256 | // vid.v vVid | |||
2257 | // vmseq.vx mMask, vVid, rIdx | |||
2258 | // vmerge.vvm vDest, vSrc, vVal, mMask | |||
2259 | // This essentially merges the original vector with the inserted element by | |||
2260 | // using a mask whose only set bit is that corresponding to the insert | |||
2261 | // index. | |||
2262 | SDValue SplattedVal = DAG.getSplatVector(ContainerVT, DL, Val); | |||
2263 | SDValue SplattedIdx = | |||
2264 | DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, Idx, VL); | |||
2265 | ||||
2266 | SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, ContainerVT, Mask, VL); | |||
2267 | auto SetCCVT = | |||
2268 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), ContainerVT); | |||
2269 | SDValue SelectCond = | |||
2270 | DAG.getNode(RISCVISD::SETCC_VL, DL, SetCCVT, VID, SplattedIdx, | |||
2271 | DAG.getCondCode(ISD::SETEQ), Mask, VL); | |||
2272 | SDValue Select = DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, | |||
2273 | SelectCond, SplattedVal, Vec, VL); | |||
2274 | if (!VecVT.isFixedLengthVector()) | |||
2275 | return Select; | |||
2276 | return convertFromScalableVector(VecVT, Select, DAG, Subtarget); | |||
2277 | } | |||
2278 | ||||
2279 | // Custom-lower EXTRACT_VECTOR_ELT operations to slide the vector down, then | |||
2280 | // extract the first element: (extractelt (slidedown vec, idx), 0). For integer | |||
2281 | // types this is done using VMV_X_S to allow us to glean information about the | |||
2282 | // sign bits of the result. | |||
2283 | SDValue RISCVTargetLowering::lowerEXTRACT_VECTOR_ELT(SDValue Op, | |||
2284 | SelectionDAG &DAG) const { | |||
2285 | SDLoc DL(Op); | |||
2286 | SDValue Idx = Op.getOperand(1); | |||
2287 | SDValue Vec = Op.getOperand(0); | |||
2288 | EVT EltVT = Op.getValueType(); | |||
2289 | MVT VecVT = Vec.getSimpleValueType(); | |||
2290 | MVT XLenVT = Subtarget.getXLenVT(); | |||
2291 | ||||
2292 | // If this is a fixed vector, we need to convert it to a scalable vector. | |||
2293 | MVT ContainerVT = VecVT; | |||
2294 | if (VecVT.isFixedLengthVector()) { | |||
2295 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2296 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
2297 | } | |||
2298 | ||||
2299 | // If the index is 0, the vector is already in the right position. | |||
2300 | if (!isNullConstant(Idx)) { | |||
2301 | // Use a VL of 1 to avoid processing more elements than we need. | |||
2302 | SDValue VL = DAG.getConstant(1, DL, XLenVT); | |||
2303 | MVT MaskVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount()); | |||
2304 | SDValue Mask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL); | |||
2305 | Vec = DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, ContainerVT, | |||
2306 | DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL); | |||
2307 | } | |||
2308 | ||||
2309 | if (!EltVT.isInteger()) { | |||
2310 | // Floating-point extracts are handled in TableGen. | |||
2311 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Vec, | |||
2312 | DAG.getConstant(0, DL, XLenVT)); | |||
2313 | } | |||
2314 | ||||
2315 | SDValue Elt0 = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec); | |||
2316 | return DAG.getNode(ISD::TRUNCATE, DL, EltVT, Elt0); | |||
2317 | } | |||
2318 | ||||
2319 | SDValue RISCVTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, | |||
2320 | SelectionDAG &DAG) const { | |||
2321 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); | |||
2322 | SDLoc DL(Op); | |||
2323 | ||||
2324 | if (Subtarget.hasStdExtV()) { | |||
2325 | // Some RVV intrinsics may claim that they want an integer operand to be | |||
2326 | // extended. | |||
2327 | if (const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II = | |||
2328 | RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo)) { | |||
2329 | if (II->ExtendedOperand) { | |||
2330 | assert(II->ExtendedOperand < Op.getNumOperands())((II->ExtendedOperand < Op.getNumOperands()) ? static_cast <void> (0) : __assert_fail ("II->ExtendedOperand < Op.getNumOperands()" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2330, __PRETTY_FUNCTION__)); | |||
2331 | SmallVector<SDValue, 8> Operands(Op->op_begin(), Op->op_end()); | |||
2332 | SDValue &ScalarOp = Operands[II->ExtendedOperand]; | |||
2333 | EVT OpVT = ScalarOp.getValueType(); | |||
2334 | if (OpVT == MVT::i8 || OpVT == MVT::i16 || | |||
2335 | (OpVT == MVT::i32 && Subtarget.is64Bit())) { | |||
2336 | // If the operand is a constant, sign extend to increase our chances | |||
2337 | // of being able to use a .vi instruction. ANY_EXTEND would become a | |||
2338 | // a zero extend and the simm5 check in isel would fail. | |||
2339 | // FIXME: Should we ignore the upper bits in isel instead? | |||
2340 | unsigned ExtOpc = isa<ConstantSDNode>(ScalarOp) ? ISD::SIGN_EXTEND | |||
2341 | : ISD::ANY_EXTEND; | |||
2342 | ScalarOp = DAG.getNode(ExtOpc, DL, Subtarget.getXLenVT(), ScalarOp); | |||
2343 | return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op.getValueType(), | |||
2344 | Operands); | |||
2345 | } | |||
2346 | } | |||
2347 | } | |||
2348 | } | |||
2349 | ||||
2350 | switch (IntNo) { | |||
2351 | default: | |||
2352 | return SDValue(); // Don't custom lower most intrinsics. | |||
2353 | case Intrinsic::thread_pointer: { | |||
2354 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
2355 | return DAG.getRegister(RISCV::X4, PtrVT); | |||
2356 | } | |||
2357 | case Intrinsic::riscv_vmv_x_s: | |||
2358 | assert(Op.getValueType() == Subtarget.getXLenVT() && "Unexpected VT!")((Op.getValueType() == Subtarget.getXLenVT() && "Unexpected VT!" ) ? static_cast<void> (0) : __assert_fail ("Op.getValueType() == Subtarget.getXLenVT() && \"Unexpected VT!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2358, __PRETTY_FUNCTION__)); | |||
2359 | return DAG.getNode(RISCVISD::VMV_X_S, DL, Op.getValueType(), | |||
2360 | Op.getOperand(1)); | |||
2361 | case Intrinsic::riscv_vmv_v_x: { | |||
2362 | SDValue Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, Subtarget.getXLenVT(), | |||
2363 | Op.getOperand(1)); | |||
2364 | return DAG.getNode(RISCVISD::VMV_V_X_VL, DL, Op.getValueType(), | |||
2365 | Scalar, Op.getOperand(2)); | |||
2366 | } | |||
2367 | case Intrinsic::riscv_vfmv_v_f: | |||
2368 | return DAG.getNode(RISCVISD::VFMV_V_F_VL, DL, Op.getValueType(), | |||
2369 | Op.getOperand(1), Op.getOperand(2)); | |||
2370 | } | |||
2371 | } | |||
2372 | ||||
2373 | SDValue RISCVTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, | |||
2374 | SelectionDAG &DAG) const { | |||
2375 | unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); | |||
2376 | SDLoc DL(Op); | |||
2377 | ||||
2378 | if (Subtarget.hasStdExtV()) { | |||
2379 | // Some RVV intrinsics may claim that they want an integer operand to be | |||
2380 | // extended. | |||
2381 | if (const RISCVVIntrinsicsTable::RISCVVIntrinsicInfo *II = | |||
2382 | RISCVVIntrinsicsTable::getRISCVVIntrinsicInfo(IntNo)) { | |||
2383 | if (II->ExtendedOperand) { | |||
2384 | // The operands start from the second argument in INTRINSIC_W_CHAIN. | |||
2385 | unsigned ExtendOp = II->ExtendedOperand + 1; | |||
2386 | assert(ExtendOp < Op.getNumOperands())((ExtendOp < Op.getNumOperands()) ? static_cast<void> (0) : __assert_fail ("ExtendOp < Op.getNumOperands()", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2386, __PRETTY_FUNCTION__)); | |||
2387 | SmallVector<SDValue, 8> Operands(Op->op_begin(), Op->op_end()); | |||
2388 | SDValue &ScalarOp = Operands[ExtendOp]; | |||
2389 | EVT OpVT = ScalarOp.getValueType(); | |||
2390 | if (OpVT == MVT::i8 || OpVT == MVT::i16 || | |||
2391 | (OpVT == MVT::i32 && Subtarget.is64Bit())) { | |||
2392 | // If the operand is a constant, sign extend to increase our chances | |||
2393 | // of being able to use a .vi instruction. ANY_EXTEND would become a | |||
2394 | // a zero extend and the simm5 check in isel would fail. | |||
2395 | // FIXME: Should we ignore the upper bits in isel instead? | |||
2396 | unsigned ExtOpc = isa<ConstantSDNode>(ScalarOp) ? ISD::SIGN_EXTEND | |||
2397 | : ISD::ANY_EXTEND; | |||
2398 | ScalarOp = DAG.getNode(ExtOpc, DL, Subtarget.getXLenVT(), ScalarOp); | |||
2399 | return DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, Op->getVTList(), | |||
2400 | Operands); | |||
2401 | } | |||
2402 | } | |||
2403 | } | |||
2404 | } | |||
2405 | ||||
2406 | return SDValue(); // Don't custom lower most intrinsics. | |||
2407 | } | |||
2408 | ||||
2409 | static MVT getLMUL1VT(MVT VT) { | |||
2410 | assert(VT.getVectorElementType().getSizeInBits() <= 64 &&((VT.getVectorElementType().getSizeInBits() <= 64 && "Unexpected vector MVT") ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType().getSizeInBits() <= 64 && \"Unexpected vector MVT\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2411, __PRETTY_FUNCTION__)) | |||
2411 | "Unexpected vector MVT")((VT.getVectorElementType().getSizeInBits() <= 64 && "Unexpected vector MVT") ? static_cast<void> (0) : __assert_fail ("VT.getVectorElementType().getSizeInBits() <= 64 && \"Unexpected vector MVT\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2411, __PRETTY_FUNCTION__)); | |||
2412 | return MVT::getScalableVectorVT( | |||
2413 | VT.getVectorElementType(), | |||
2414 | RISCV::RVVBitsPerBlock / VT.getVectorElementType().getSizeInBits()); | |||
2415 | } | |||
2416 | ||||
2417 | static unsigned getRVVReductionOp(unsigned ISDOpcode) { | |||
2418 | switch (ISDOpcode) { | |||
2419 | default: | |||
2420 | llvm_unreachable("Unhandled reduction")::llvm::llvm_unreachable_internal("Unhandled reduction", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2420); | |||
2421 | case ISD::VECREDUCE_ADD: | |||
2422 | return RISCVISD::VECREDUCE_ADD_VL; | |||
2423 | case ISD::VECREDUCE_UMAX: | |||
2424 | return RISCVISD::VECREDUCE_UMAX_VL; | |||
2425 | case ISD::VECREDUCE_SMAX: | |||
2426 | return RISCVISD::VECREDUCE_SMAX_VL; | |||
2427 | case ISD::VECREDUCE_UMIN: | |||
2428 | return RISCVISD::VECREDUCE_UMIN_VL; | |||
2429 | case ISD::VECREDUCE_SMIN: | |||
2430 | return RISCVISD::VECREDUCE_SMIN_VL; | |||
2431 | case ISD::VECREDUCE_AND: | |||
2432 | return RISCVISD::VECREDUCE_AND_VL; | |||
2433 | case ISD::VECREDUCE_OR: | |||
2434 | return RISCVISD::VECREDUCE_OR_VL; | |||
2435 | case ISD::VECREDUCE_XOR: | |||
2436 | return RISCVISD::VECREDUCE_XOR_VL; | |||
2437 | } | |||
2438 | } | |||
2439 | ||||
2440 | // Take a (supported) standard ISD reduction opcode and transform it to a RISCV | |||
2441 | // reduction opcode. Note that this returns a vector type, which must be | |||
2442 | // further processed to access the scalar result in element 0. | |||
2443 | SDValue RISCVTargetLowering::lowerVECREDUCE(SDValue Op, | |||
2444 | SelectionDAG &DAG) const { | |||
2445 | SDLoc DL(Op); | |||
2446 | MVT VecVT = Op.getOperand(0).getSimpleValueType(); | |||
2447 | MVT VecEltVT = VecVT.getVectorElementType(); | |||
2448 | ||||
2449 | // Avoid creating vectors with illegal type. | |||
2450 | if (!isTypeLegal(VecVT)) | |||
2451 | return SDValue(); | |||
2452 | ||||
2453 | unsigned RVVOpcode = getRVVReductionOp(Op.getOpcode()); | |||
2454 | ||||
2455 | SDValue Vec = Op.getOperand(0); | |||
2456 | ||||
2457 | MVT ContainerVT = VecVT; | |||
2458 | if (VecVT.isFixedLengthVector()) { | |||
2459 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2460 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
2461 | } | |||
2462 | ||||
2463 | MVT M1VT = getLMUL1VT(ContainerVT); | |||
2464 | ||||
2465 | SDValue Mask, VL; | |||
2466 | std::tie(Mask, VL) = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget); | |||
2467 | ||||
2468 | // FIXME: This is a VLMAX splat which might be too large and can prevent | |||
2469 | // vsetvli removal. | |||
2470 | SDValue NeutralElem = DAG.getNeutralElement( | |||
2471 | ISD::getVecReduceBaseOpcode(Op.getOpcode()), DL, VecEltVT, SDNodeFlags()); | |||
2472 | SDValue IdentitySplat = DAG.getSplatVector(M1VT, DL, NeutralElem); | |||
2473 | SDValue Reduction = | |||
2474 | DAG.getNode(RVVOpcode, DL, M1VT, Vec, IdentitySplat, Mask, VL); | |||
2475 | SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VecEltVT, Reduction, | |||
2476 | DAG.getConstant(0, DL, Subtarget.getXLenVT())); | |||
2477 | return DAG.getSExtOrTrunc(Elt0, DL, Op.getValueType()); | |||
2478 | } | |||
2479 | ||||
2480 | // Given a reduction op, this function returns the matching reduction opcode, | |||
2481 | // the vector SDValue and the scalar SDValue required to lower this to a | |||
2482 | // RISCVISD node. | |||
2483 | static std::tuple<unsigned, SDValue, SDValue> | |||
2484 | getRVVFPReductionOpAndOperands(SDValue Op, SelectionDAG &DAG, EVT EltVT) { | |||
2485 | SDLoc DL(Op); | |||
2486 | switch (Op.getOpcode()) { | |||
2487 | default: | |||
2488 | llvm_unreachable("Unhandled reduction")::llvm::llvm_unreachable_internal("Unhandled reduction", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2488); | |||
2489 | case ISD::VECREDUCE_FADD: | |||
2490 | return std::make_tuple(RISCVISD::VECREDUCE_FADD_VL, Op.getOperand(0), | |||
2491 | DAG.getConstantFP(0.0, DL, EltVT)); | |||
2492 | case ISD::VECREDUCE_SEQ_FADD: | |||
2493 | return std::make_tuple(RISCVISD::VECREDUCE_SEQ_FADD_VL, Op.getOperand(1), | |||
2494 | Op.getOperand(0)); | |||
2495 | } | |||
2496 | } | |||
2497 | ||||
2498 | SDValue RISCVTargetLowering::lowerFPVECREDUCE(SDValue Op, | |||
2499 | SelectionDAG &DAG) const { | |||
2500 | SDLoc DL(Op); | |||
2501 | MVT VecEltVT = Op.getSimpleValueType(); | |||
2502 | ||||
2503 | unsigned RVVOpcode; | |||
2504 | SDValue VectorVal, ScalarVal; | |||
2505 | std::tie(RVVOpcode, VectorVal, ScalarVal) = | |||
2506 | getRVVFPReductionOpAndOperands(Op, DAG, VecEltVT); | |||
2507 | MVT VecVT = VectorVal.getSimpleValueType(); | |||
2508 | ||||
2509 | MVT ContainerVT = VecVT; | |||
2510 | if (VecVT.isFixedLengthVector()) { | |||
2511 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2512 | VectorVal = convertToScalableVector(ContainerVT, VectorVal, DAG, Subtarget); | |||
2513 | } | |||
2514 | ||||
2515 | MVT M1VT = getLMUL1VT(VectorVal.getSimpleValueType()); | |||
2516 | ||||
2517 | SDValue Mask, VL; | |||
2518 | std::tie(Mask, VL) = getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget); | |||
2519 | ||||
2520 | // FIXME: This is a VLMAX splat which might be too large and can prevent | |||
2521 | // vsetvli removal. | |||
2522 | SDValue ScalarSplat = DAG.getSplatVector(M1VT, DL, ScalarVal); | |||
2523 | SDValue Reduction = | |||
2524 | DAG.getNode(RVVOpcode, DL, M1VT, VectorVal, ScalarSplat, Mask, VL); | |||
2525 | return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VecEltVT, Reduction, | |||
2526 | DAG.getConstant(0, DL, Subtarget.getXLenVT())); | |||
2527 | } | |||
2528 | ||||
2529 | SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op, | |||
2530 | SelectionDAG &DAG) const { | |||
2531 | SDValue Vec = Op.getOperand(0); | |||
2532 | SDValue SubVec = Op.getOperand(1); | |||
2533 | MVT VecVT = Vec.getSimpleValueType(); | |||
2534 | MVT SubVecVT = SubVec.getSimpleValueType(); | |||
2535 | ||||
2536 | SDLoc DL(Op); | |||
2537 | MVT XLenVT = Subtarget.getXLenVT(); | |||
2538 | unsigned OrigIdx = Op.getConstantOperandVal(2); | |||
2539 | const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
2540 | ||||
2541 | // We don't have the ability to slide mask vectors up indexed by their i1 | |||
2542 | // elements; the smallest we can do is i8. Often we are able to bitcast to | |||
2543 | // equivalent i8 vectors. Note that when inserting a fixed-length vector | |||
2544 | // into a scalable one, we might not necessarily have enough scalable | |||
2545 | // elements to safely divide by 8: nxv1i1 = insert nxv1i1, v4i1 is valid. | |||
2546 | if (SubVecVT.getVectorElementType() == MVT::i1 && | |||
2547 | (OrigIdx != 0 || !Vec.isUndef())) { | |||
2548 | if (VecVT.getVectorMinNumElements() >= 8 && | |||
2549 | SubVecVT.getVectorMinNumElements() >= 8) { | |||
2550 | assert(OrigIdx % 8 == 0 && "Invalid index")((OrigIdx % 8 == 0 && "Invalid index") ? static_cast< void> (0) : __assert_fail ("OrigIdx % 8 == 0 && \"Invalid index\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2550, __PRETTY_FUNCTION__)); | |||
2551 | assert(VecVT.getVectorMinNumElements() % 8 == 0 &&((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2553, __PRETTY_FUNCTION__)) | |||
2552 | SubVecVT.getVectorMinNumElements() % 8 == 0 &&((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2553, __PRETTY_FUNCTION__)) | |||
2553 | "Unexpected mask vector lowering")((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2553, __PRETTY_FUNCTION__)); | |||
2554 | OrigIdx /= 8; | |||
2555 | SubVecVT = | |||
2556 | MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8, | |||
2557 | SubVecVT.isScalableVector()); | |||
2558 | VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8, | |||
2559 | VecVT.isScalableVector()); | |||
2560 | Vec = DAG.getBitcast(VecVT, Vec); | |||
2561 | SubVec = DAG.getBitcast(SubVecVT, SubVec); | |||
2562 | } else { | |||
2563 | // We can't slide this mask vector up indexed by its i1 elements. | |||
2564 | // This poses a problem when we wish to insert a scalable vector which | |||
2565 | // can't be re-expressed as a larger type. Just choose the slow path and | |||
2566 | // extend to a larger type, then truncate back down. | |||
2567 | MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8); | |||
2568 | MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8); | |||
2569 | Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec); | |||
2570 | SubVec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtSubVecVT, SubVec); | |||
2571 | Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ExtVecVT, Vec, SubVec, | |||
2572 | Op.getOperand(2)); | |||
2573 | SDValue SplatZero = DAG.getConstant(0, DL, ExtVecVT); | |||
2574 | return DAG.getSetCC(DL, VecVT, Vec, SplatZero, ISD::SETNE); | |||
2575 | } | |||
2576 | } | |||
2577 | ||||
2578 | // If the subvector vector is a fixed-length type, we cannot use subregister | |||
2579 | // manipulation to simplify the codegen; we don't know which register of a | |||
2580 | // LMUL group contains the specific subvector as we only know the minimum | |||
2581 | // register size. Therefore we must slide the vector group up the full | |||
2582 | // amount. | |||
2583 | if (SubVecVT.isFixedLengthVector()) { | |||
2584 | if (OrigIdx == 0 && Vec.isUndef()) | |||
2585 | return Op; | |||
2586 | MVT ContainerVT = VecVT; | |||
2587 | if (VecVT.isFixedLengthVector()) { | |||
2588 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2589 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
2590 | } | |||
2591 | SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ContainerVT, | |||
2592 | DAG.getUNDEF(ContainerVT), SubVec, | |||
2593 | DAG.getConstant(0, DL, XLenVT)); | |||
2594 | SDValue Mask = | |||
2595 | getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first; | |||
2596 | // Set the vector length to only the number of elements we care about. Note | |||
2597 | // that for slideup this includes the offset. | |||
2598 | SDValue VL = | |||
2599 | DAG.getConstant(OrigIdx + SubVecVT.getVectorNumElements(), DL, XLenVT); | |||
2600 | SDValue SlideupAmt = DAG.getConstant(OrigIdx, DL, XLenVT); | |||
2601 | SDValue Slideup = DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, ContainerVT, Vec, | |||
2602 | SubVec, SlideupAmt, Mask, VL); | |||
2603 | if (!VecVT.isFixedLengthVector()) | |||
2604 | return Slideup; | |||
2605 | return convertFromScalableVector(VecVT, Slideup, DAG, Subtarget); | |||
2606 | } | |||
2607 | ||||
2608 | unsigned SubRegIdx, RemIdx; | |||
2609 | std::tie(SubRegIdx, RemIdx) = | |||
2610 | RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs( | |||
2611 | VecVT, SubVecVT, OrigIdx, TRI); | |||
2612 | ||||
2613 | RISCVVLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecVT); | |||
2614 | bool IsSubVecPartReg = SubVecLMUL == RISCVVLMUL::LMUL_F2 || | |||
2615 | SubVecLMUL == RISCVVLMUL::LMUL_F4 || | |||
2616 | SubVecLMUL == RISCVVLMUL::LMUL_F8; | |||
2617 | ||||
2618 | // 1. If the Idx has been completely eliminated and this subvector's size is | |||
2619 | // a vector register or a multiple thereof, or the surrounding elements are | |||
2620 | // undef, then this is a subvector insert which naturally aligns to a vector | |||
2621 | // register. These can easily be handled using subregister manipulation. | |||
2622 | // 2. If the subvector is smaller than a vector register, then the insertion | |||
2623 | // must preserve the undisturbed elements of the register. We do this by | |||
2624 | // lowering to an EXTRACT_SUBVECTOR grabbing the nearest LMUL=1 vector type | |||
2625 | // (which resolves to a subregister copy), performing a VSLIDEUP to place the | |||
2626 | // subvector within the vector register, and an INSERT_SUBVECTOR of that | |||
2627 | // LMUL=1 type back into the larger vector (resolving to another subregister | |||
2628 | // operation). See below for how our VSLIDEUP works. We go via a LMUL=1 type | |||
2629 | // to avoid allocating a large register group to hold our subvector. | |||
2630 | if (RemIdx == 0 && (!IsSubVecPartReg || Vec.isUndef())) | |||
2631 | return Op; | |||
2632 | ||||
2633 | // VSLIDEUP works by leaving elements 0<i<OFFSET undisturbed, elements | |||
2634 | // OFFSET<=i<VL set to the "subvector" and vl<=i<VLMAX set to the tail policy | |||
2635 | // (in our case undisturbed). This means we can set up a subvector insertion | |||
2636 | // where OFFSET is the insertion offset, and the VL is the OFFSET plus the | |||
2637 | // size of the subvector. | |||
2638 | MVT InterSubVT = VecVT; | |||
2639 | SDValue AlignedExtract = Vec; | |||
2640 | unsigned AlignedIdx = OrigIdx - RemIdx; | |||
2641 | if (VecVT.bitsGT(getLMUL1VT(VecVT))) { | |||
2642 | InterSubVT = getLMUL1VT(VecVT); | |||
2643 | // Extract a subvector equal to the nearest full vector register type. This | |||
2644 | // should resolve to a EXTRACT_SUBREG instruction. | |||
2645 | AlignedExtract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec, | |||
2646 | DAG.getConstant(AlignedIdx, DL, XLenVT)); | |||
2647 | } | |||
2648 | ||||
2649 | SDValue SlideupAmt = DAG.getConstant(RemIdx, DL, XLenVT); | |||
2650 | // For scalable vectors this must be further multiplied by vscale. | |||
2651 | SlideupAmt = DAG.getNode(ISD::VSCALE, DL, XLenVT, SlideupAmt); | |||
2652 | ||||
2653 | SDValue Mask, VL; | |||
2654 | std::tie(Mask, VL) = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget); | |||
2655 | ||||
2656 | // Construct the vector length corresponding to RemIdx + length(SubVecVT). | |||
2657 | VL = DAG.getConstant(SubVecVT.getVectorMinNumElements(), DL, XLenVT); | |||
2658 | VL = DAG.getNode(ISD::VSCALE, DL, XLenVT, VL); | |||
2659 | VL = DAG.getNode(ISD::ADD, DL, XLenVT, SlideupAmt, VL); | |||
2660 | ||||
2661 | SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InterSubVT, | |||
2662 | DAG.getUNDEF(InterSubVT), SubVec, | |||
2663 | DAG.getConstant(0, DL, XLenVT)); | |||
2664 | ||||
2665 | SDValue Slideup = DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, InterSubVT, | |||
2666 | AlignedExtract, SubVec, SlideupAmt, Mask, VL); | |||
2667 | ||||
2668 | // If required, insert this subvector back into the correct vector register. | |||
2669 | // This should resolve to an INSERT_SUBREG instruction. | |||
2670 | if (VecVT.bitsGT(InterSubVT)) | |||
2671 | Slideup = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, Vec, Slideup, | |||
2672 | DAG.getConstant(AlignedIdx, DL, XLenVT)); | |||
2673 | ||||
2674 | // We might have bitcast from a mask type: cast back to the original type if | |||
2675 | // required. | |||
2676 | return DAG.getBitcast(Op.getSimpleValueType(), Slideup); | |||
2677 | } | |||
2678 | ||||
2679 | SDValue RISCVTargetLowering::lowerEXTRACT_SUBVECTOR(SDValue Op, | |||
2680 | SelectionDAG &DAG) const { | |||
2681 | SDValue Vec = Op.getOperand(0); | |||
2682 | MVT SubVecVT = Op.getSimpleValueType(); | |||
2683 | MVT VecVT = Vec.getSimpleValueType(); | |||
2684 | ||||
2685 | SDLoc DL(Op); | |||
2686 | MVT XLenVT = Subtarget.getXLenVT(); | |||
2687 | unsigned OrigIdx = Op.getConstantOperandVal(1); | |||
2688 | const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
2689 | ||||
2690 | // We don't have the ability to slide mask vectors down indexed by their i1 | |||
2691 | // elements; the smallest we can do is i8. Often we are able to bitcast to | |||
2692 | // equivalent i8 vectors. Note that when extracting a fixed-length vector | |||
2693 | // from a scalable one, we might not necessarily have enough scalable | |||
2694 | // elements to safely divide by 8: v8i1 = extract nxv1i1 is valid. | |||
2695 | if (SubVecVT.getVectorElementType() == MVT::i1 && OrigIdx != 0) { | |||
2696 | if (VecVT.getVectorMinNumElements() >= 8 && | |||
2697 | SubVecVT.getVectorMinNumElements() >= 8) { | |||
2698 | assert(OrigIdx % 8 == 0 && "Invalid index")((OrigIdx % 8 == 0 && "Invalid index") ? static_cast< void> (0) : __assert_fail ("OrigIdx % 8 == 0 && \"Invalid index\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2698, __PRETTY_FUNCTION__)); | |||
2699 | assert(VecVT.getVectorMinNumElements() % 8 == 0 &&((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2701, __PRETTY_FUNCTION__)) | |||
2700 | SubVecVT.getVectorMinNumElements() % 8 == 0 &&((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2701, __PRETTY_FUNCTION__)) | |||
2701 | "Unexpected mask vector lowering")((VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT .getVectorMinNumElements() % 8 == 0 && "Unexpected mask vector lowering" ) ? static_cast<void> (0) : __assert_fail ("VecVT.getVectorMinNumElements() % 8 == 0 && SubVecVT.getVectorMinNumElements() % 8 == 0 && \"Unexpected mask vector lowering\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 2701, __PRETTY_FUNCTION__)); | |||
2702 | OrigIdx /= 8; | |||
2703 | SubVecVT = | |||
2704 | MVT::getVectorVT(MVT::i8, SubVecVT.getVectorMinNumElements() / 8, | |||
2705 | SubVecVT.isScalableVector()); | |||
2706 | VecVT = MVT::getVectorVT(MVT::i8, VecVT.getVectorMinNumElements() / 8, | |||
2707 | VecVT.isScalableVector()); | |||
2708 | Vec = DAG.getBitcast(VecVT, Vec); | |||
2709 | } else { | |||
2710 | // We can't slide this mask vector down, indexed by its i1 elements. | |||
2711 | // This poses a problem when we wish to extract a scalable vector which | |||
2712 | // can't be re-expressed as a larger type. Just choose the slow path and | |||
2713 | // extend to a larger type, then truncate back down. | |||
2714 | // TODO: We could probably improve this when extracting certain fixed | |||
2715 | // from fixed, where we can extract as i8 and shift the correct element | |||
2716 | // right to reach the desired subvector? | |||
2717 | MVT ExtVecVT = VecVT.changeVectorElementType(MVT::i8); | |||
2718 | MVT ExtSubVecVT = SubVecVT.changeVectorElementType(MVT::i8); | |||
2719 | Vec = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVecVT, Vec); | |||
2720 | Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtSubVecVT, Vec, | |||
2721 | Op.getOperand(1)); | |||
2722 | SDValue SplatZero = DAG.getConstant(0, DL, ExtSubVecVT); | |||
2723 | return DAG.getSetCC(DL, SubVecVT, Vec, SplatZero, ISD::SETNE); | |||
2724 | } | |||
2725 | } | |||
2726 | ||||
2727 | // If the subvector vector is a fixed-length type, we cannot use subregister | |||
2728 | // manipulation to simplify the codegen; we don't know which register of a | |||
2729 | // LMUL group contains the specific subvector as we only know the minimum | |||
2730 | // register size. Therefore we must slide the vector group down the full | |||
2731 | // amount. | |||
2732 | if (SubVecVT.isFixedLengthVector()) { | |||
2733 | // With an index of 0 this is a cast-like subvector, which can be performed | |||
2734 | // with subregister operations. | |||
2735 | if (OrigIdx == 0) | |||
2736 | return Op; | |||
2737 | MVT ContainerVT = VecVT; | |||
2738 | if (VecVT.isFixedLengthVector()) { | |||
2739 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
2740 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
2741 | } | |||
2742 | SDValue Mask = | |||
2743 | getDefaultVLOps(VecVT, ContainerVT, DL, DAG, Subtarget).first; | |||
2744 | // Set the vector length to only the number of elements we care about. This | |||
2745 | // avoids sliding down elements we're going to discard straight away. | |||
2746 | SDValue VL = DAG.getConstant(SubVecVT.getVectorNumElements(), DL, XLenVT); | |||
2747 | SDValue SlidedownAmt = DAG.getConstant(OrigIdx, DL, XLenVT); | |||
2748 | SDValue Slidedown = | |||
2749 | DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, ContainerVT, | |||
2750 | DAG.getUNDEF(ContainerVT), Vec, SlidedownAmt, Mask, VL); | |||
2751 | // Now we can use a cast-like subvector extract to get the result. | |||
2752 | return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown, | |||
2753 | DAG.getConstant(0, DL, XLenVT)); | |||
2754 | } | |||
2755 | ||||
2756 | unsigned SubRegIdx, RemIdx; | |||
2757 | std::tie(SubRegIdx, RemIdx) = | |||
2758 | RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs( | |||
2759 | VecVT, SubVecVT, OrigIdx, TRI); | |||
2760 | ||||
2761 | // If the Idx has been completely eliminated then this is a subvector extract | |||
2762 | // which naturally aligns to a vector register. These can easily be handled | |||
2763 | // using subregister manipulation. | |||
2764 | if (RemIdx == 0) | |||
2765 | return Op; | |||
2766 | ||||
2767 | // Else we must shift our vector register directly to extract the subvector. | |||
2768 | // Do this using VSLIDEDOWN. | |||
2769 | ||||
2770 | // If the vector type is an LMUL-group type, extract a subvector equal to the | |||
2771 | // nearest full vector register type. This should resolve to a EXTRACT_SUBREG | |||
2772 | // instruction. | |||
2773 | MVT InterSubVT = VecVT; | |||
2774 | if (VecVT.bitsGT(getLMUL1VT(VecVT))) { | |||
2775 | InterSubVT = getLMUL1VT(VecVT); | |||
2776 | Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec, | |||
2777 | DAG.getConstant(OrigIdx - RemIdx, DL, XLenVT)); | |||
2778 | } | |||
2779 | ||||
2780 | // Slide this vector register down by the desired number of elements in order | |||
2781 | // to place the desired subvector starting at element 0. | |||
2782 | SDValue SlidedownAmt = DAG.getConstant(RemIdx, DL, XLenVT); | |||
2783 | // For scalable vectors this must be further multiplied by vscale. | |||
2784 | SlidedownAmt = DAG.getNode(ISD::VSCALE, DL, XLenVT, SlidedownAmt); | |||
2785 | ||||
2786 | SDValue Mask, VL; | |||
2787 | std::tie(Mask, VL) = getDefaultScalableVLOps(InterSubVT, DL, DAG, Subtarget); | |||
2788 | SDValue Slidedown = | |||
2789 | DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, InterSubVT, | |||
2790 | DAG.getUNDEF(InterSubVT), Vec, SlidedownAmt, Mask, VL); | |||
2791 | ||||
2792 | // Now the vector is in the right position, extract our final subvector. This | |||
2793 | // should resolve to a COPY. | |||
2794 | Slidedown = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVecVT, Slidedown, | |||
2795 | DAG.getConstant(0, DL, XLenVT)); | |||
2796 | ||||
2797 | // We might have bitcast from a mask type: cast back to the original type if | |||
2798 | // required. | |||
2799 | return DAG.getBitcast(Op.getSimpleValueType(), Slidedown); | |||
2800 | } | |||
2801 | ||||
2802 | // Implement vector_reverse using vrgather.vv with indices determined by | |||
2803 | // subtracting the id of each element from (VLMAX-1). This will convert | |||
2804 | // the indices like so: | |||
2805 | // (0, 1,..., VLMAX-2, VLMAX-1) -> (VLMAX-1, VLMAX-2,..., 1, 0). | |||
2806 | // TODO: This code assumes VLMAX <= 65536 for LMUL=8 SEW=16. | |||
2807 | SDValue RISCVTargetLowering::lowerVECTOR_REVERSE(SDValue Op, | |||
2808 | SelectionDAG &DAG) const { | |||
2809 | SDLoc DL(Op); | |||
2810 | MVT VecVT = Op.getSimpleValueType(); | |||
2811 | unsigned EltSize = VecVT.getScalarSizeInBits(); | |||
2812 | unsigned MinSize = VecVT.getSizeInBits().getKnownMinValue(); | |||
2813 | ||||
2814 | unsigned MaxVLMAX = 0; | |||
2815 | unsigned VectorBitsMax = Subtarget.getMaxRVVVectorSizeInBits(); | |||
2816 | if (VectorBitsMax != 0) | |||
2817 | MaxVLMAX = ((VectorBitsMax / EltSize) * MinSize) / RISCV::RVVBitsPerBlock; | |||
2818 | ||||
2819 | unsigned GatherOpc = RISCVISD::VRGATHER_VV_VL; | |||
2820 | MVT IntVT = VecVT.changeVectorElementTypeToInteger(); | |||
2821 | ||||
2822 | // If this is SEW=8 and VLMAX is unknown or more than 256, we need | |||
2823 | // to use vrgatherei16.vv. | |||
2824 | // TODO: It's also possible to use vrgatherei16.vv for other types to | |||
2825 | // decrease register width for the index calculation. | |||
2826 | if ((MaxVLMAX == 0 || MaxVLMAX > 256) && EltSize == 8) { | |||
2827 | // If this is LMUL=8, we have to split before can use vrgatherei16.vv. | |||
2828 | // Reverse each half, then reassemble them in reverse order. | |||
2829 | // NOTE: It's also possible that after splitting that VLMAX no longer | |||
2830 | // requires vrgatherei16.vv. | |||
2831 | if (MinSize == (8 * RISCV::RVVBitsPerBlock)) { | |||
2832 | SDValue Lo, Hi; | |||
2833 | std::tie(Lo, Hi) = DAG.SplitVectorOperand(Op.getNode(), 0); | |||
2834 | EVT LoVT, HiVT; | |||
2835 | std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VecVT); | |||
2836 | Lo = DAG.getNode(ISD::VECTOR_REVERSE, DL, LoVT, Lo); | |||
2837 | Hi = DAG.getNode(ISD::VECTOR_REVERSE, DL, HiVT, Hi); | |||
2838 | // Reassemble the low and high pieces reversed. | |||
2839 | // FIXME: This is a CONCAT_VECTORS. | |||
2840 | SDValue Res = | |||
2841 | DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, DAG.getUNDEF(VecVT), Hi, | |||
2842 | DAG.getIntPtrConstant(0, DL)); | |||
2843 | return DAG.getNode( | |||
2844 | ISD::INSERT_SUBVECTOR, DL, VecVT, Res, Lo, | |||
2845 | DAG.getIntPtrConstant(LoVT.getVectorMinNumElements(), DL)); | |||
2846 | } | |||
2847 | ||||
2848 | // Just promote the int type to i16 which will double the LMUL. | |||
2849 | IntVT = MVT::getVectorVT(MVT::i16, VecVT.getVectorElementCount()); | |||
2850 | GatherOpc = RISCVISD::VRGATHEREI16_VV_VL; | |||
2851 | } | |||
2852 | ||||
2853 | MVT XLenVT = Subtarget.getXLenVT(); | |||
2854 | SDValue Mask, VL; | |||
2855 | std::tie(Mask, VL) = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget); | |||
2856 | ||||
2857 | // Calculate VLMAX-1 for the desired SEW. | |||
2858 | unsigned MinElts = VecVT.getVectorMinNumElements(); | |||
2859 | SDValue VLMax = DAG.getNode(ISD::VSCALE, DL, XLenVT, | |||
2860 | DAG.getConstant(MinElts, DL, XLenVT)); | |||
2861 | SDValue VLMinus1 = | |||
2862 | DAG.getNode(ISD::SUB, DL, XLenVT, VLMax, DAG.getConstant(1, DL, XLenVT)); | |||
2863 | ||||
2864 | // Splat VLMAX-1 taking care to handle SEW==64 on RV32. | |||
2865 | bool IsRV32E64 = | |||
2866 | !Subtarget.is64Bit() && IntVT.getVectorElementType() == MVT::i64; | |||
2867 | SDValue SplatVL; | |||
2868 | if (!IsRV32E64) | |||
2869 | SplatVL = DAG.getSplatVector(IntVT, DL, VLMinus1); | |||
2870 | else | |||
2871 | SplatVL = DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, IntVT, VLMinus1); | |||
2872 | ||||
2873 | SDValue VID = DAG.getNode(RISCVISD::VID_VL, DL, IntVT, Mask, VL); | |||
2874 | SDValue Indices = | |||
2875 | DAG.getNode(RISCVISD::SUB_VL, DL, IntVT, SplatVL, VID, Mask, VL); | |||
2876 | ||||
2877 | return DAG.getNode(GatherOpc, DL, VecVT, Op.getOperand(0), Indices, Mask, VL); | |||
2878 | } | |||
2879 | ||||
2880 | SDValue | |||
2881 | RISCVTargetLowering::lowerFixedLengthVectorLoadToRVV(SDValue Op, | |||
2882 | SelectionDAG &DAG) const { | |||
2883 | auto *Load = cast<LoadSDNode>(Op); | |||
2884 | ||||
2885 | SDLoc DL(Op); | |||
2886 | MVT VT = Op.getSimpleValueType(); | |||
2887 | MVT ContainerVT = getContainerForFixedLengthVector(VT); | |||
2888 | ||||
2889 | SDValue VL = | |||
2890 | DAG.getConstant(VT.getVectorNumElements(), DL, Subtarget.getXLenVT()); | |||
2891 | ||||
2892 | SDVTList VTs = DAG.getVTList({ContainerVT, MVT::Other}); | |||
2893 | SDValue NewLoad = DAG.getMemIntrinsicNode( | |||
2894 | RISCVISD::VLE_VL, DL, VTs, {Load->getChain(), Load->getBasePtr(), VL}, | |||
2895 | Load->getMemoryVT(), Load->getMemOperand()); | |||
2896 | ||||
2897 | SDValue Result = convertFromScalableVector(VT, NewLoad, DAG, Subtarget); | |||
2898 | return DAG.getMergeValues({Result, Load->getChain()}, DL); | |||
2899 | } | |||
2900 | ||||
2901 | SDValue | |||
2902 | RISCVTargetLowering::lowerFixedLengthVectorStoreToRVV(SDValue Op, | |||
2903 | SelectionDAG &DAG) const { | |||
2904 | auto *Store = cast<StoreSDNode>(Op); | |||
2905 | ||||
2906 | SDLoc DL(Op); | |||
2907 | MVT VT = Store->getValue().getSimpleValueType(); | |||
2908 | ||||
2909 | // FIXME: We probably need to zero any extra bits in a byte for mask stores. | |||
2910 | // This is tricky to do. | |||
2911 | ||||
2912 | MVT ContainerVT = getContainerForFixedLengthVector(VT); | |||
2913 | ||||
2914 | SDValue VL = | |||
2915 | DAG.getConstant(VT.getVectorNumElements(), DL, Subtarget.getXLenVT()); | |||
2916 | ||||
2917 | SDValue NewValue = | |||
2918 | convertToScalableVector(ContainerVT, Store->getValue(), DAG, Subtarget); | |||
2919 | return DAG.getMemIntrinsicNode( | |||
2920 | RISCVISD::VSE_VL, DL, DAG.getVTList(MVT::Other), | |||
2921 | {Store->getChain(), NewValue, Store->getBasePtr(), VL}, | |||
2922 | Store->getMemoryVT(), Store->getMemOperand()); | |||
2923 | } | |||
2924 | ||||
2925 | SDValue | |||
2926 | RISCVTargetLowering::lowerFixedLengthVectorSetccToRVV(SDValue Op, | |||
2927 | SelectionDAG &DAG) const { | |||
2928 | MVT InVT = Op.getOperand(0).getSimpleValueType(); | |||
2929 | MVT ContainerVT = getContainerForFixedLengthVector(InVT); | |||
2930 | ||||
2931 | MVT VT = Op.getSimpleValueType(); | |||
2932 | ||||
2933 | SDValue Op1 = | |||
2934 | convertToScalableVector(ContainerVT, Op.getOperand(0), DAG, Subtarget); | |||
2935 | SDValue Op2 = | |||
2936 | convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget); | |||
2937 | ||||
2938 | SDLoc DL(Op); | |||
2939 | SDValue VL = | |||
2940 | DAG.getConstant(VT.getVectorNumElements(), DL, Subtarget.getXLenVT()); | |||
2941 | ||||
2942 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); | |||
2943 | ||||
2944 | bool Invert = false; | |||
2945 | Optional<unsigned> LogicOpc; | |||
2946 | if (ContainerVT.isFloatingPoint()) { | |||
2947 | bool Swap = false; | |||
2948 | switch (CC) { | |||
2949 | default: | |||
2950 | break; | |||
2951 | case ISD::SETULE: | |||
2952 | case ISD::SETULT: | |||
2953 | Swap = true; | |||
2954 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | |||
2955 | case ISD::SETUGE: | |||
2956 | case ISD::SETUGT: | |||
2957 | CC = getSetCCInverse(CC, ContainerVT); | |||
2958 | Invert = true; | |||
2959 | break; | |||
2960 | case ISD::SETOGE: | |||
2961 | case ISD::SETOGT: | |||
2962 | case ISD::SETGE: | |||
2963 | case ISD::SETGT: | |||
2964 | Swap = true; | |||
2965 | break; | |||
2966 | case ISD::SETUEQ: | |||
2967 | // Use !((OLT Op1, Op2) || (OLT Op2, Op1)) | |||
2968 | Invert = true; | |||
2969 | LogicOpc = RISCVISD::VMOR_VL; | |||
2970 | CC = ISD::SETOLT; | |||
2971 | break; | |||
2972 | case ISD::SETONE: | |||
2973 | // Use ((OLT Op1, Op2) || (OLT Op2, Op1)) | |||
2974 | LogicOpc = RISCVISD::VMOR_VL; | |||
2975 | CC = ISD::SETOLT; | |||
2976 | break; | |||
2977 | case ISD::SETO: | |||
2978 | // Use (OEQ Op1, Op1) && (OEQ Op2, Op2) | |||
2979 | LogicOpc = RISCVISD::VMAND_VL; | |||
2980 | CC = ISD::SETOEQ; | |||
2981 | break; | |||
2982 | case ISD::SETUO: | |||
2983 | // Use (UNE Op1, Op1) || (UNE Op2, Op2) | |||
2984 | LogicOpc = RISCVISD::VMOR_VL; | |||
2985 | CC = ISD::SETUNE; | |||
2986 | break; | |||
2987 | } | |||
2988 | ||||
2989 | if (Swap) { | |||
2990 | CC = getSetCCSwappedOperands(CC); | |||
2991 | std::swap(Op1, Op2); | |||
2992 | } | |||
2993 | } | |||
2994 | ||||
2995 | MVT MaskVT = MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount()); | |||
2996 | SDValue Mask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL); | |||
2997 | ||||
2998 | // There are 3 cases we need to emit. | |||
2999 | // 1. For (OEQ Op1, Op1) && (OEQ Op2, Op2) or (UNE Op1, Op1) || (UNE Op2, Op2) | |||
3000 | // we need to compare each operand with itself. | |||
3001 | // 2. For (OLT Op1, Op2) || (OLT Op2, Op1) we need to compare Op1 and Op2 in | |||
3002 | // both orders. | |||
3003 | // 3. For any other case we just need one compare with Op1 and Op2. | |||
3004 | SDValue Cmp; | |||
3005 | if (LogicOpc && (CC == ISD::SETOEQ || CC == ISD::SETUNE)) { | |||
3006 | Cmp = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT, Op1, Op1, | |||
3007 | DAG.getCondCode(CC), Mask, VL); | |||
3008 | SDValue Cmp2 = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT, Op2, Op2, | |||
3009 | DAG.getCondCode(CC), Mask, VL); | |||
3010 | Cmp = DAG.getNode(*LogicOpc, DL, MaskVT, Cmp, Cmp2, VL); | |||
3011 | } else { | |||
3012 | Cmp = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT, Op1, Op2, | |||
3013 | DAG.getCondCode(CC), Mask, VL); | |||
3014 | if (LogicOpc) { | |||
3015 | SDValue Cmp2 = DAG.getNode(RISCVISD::SETCC_VL, DL, MaskVT, Op2, Op1, | |||
3016 | DAG.getCondCode(CC), Mask, VL); | |||
3017 | Cmp = DAG.getNode(*LogicOpc, DL, MaskVT, Cmp, Cmp2, VL); | |||
3018 | } | |||
3019 | } | |||
3020 | ||||
3021 | if (Invert) { | |||
3022 | SDValue AllOnes = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL); | |||
3023 | Cmp = DAG.getNode(RISCVISD::VMXOR_VL, DL, MaskVT, Cmp, AllOnes, VL); | |||
3024 | } | |||
3025 | ||||
3026 | return convertFromScalableVector(VT, Cmp, DAG, Subtarget); | |||
3027 | } | |||
3028 | ||||
3029 | SDValue RISCVTargetLowering::lowerFixedLengthVectorLogicOpToRVV( | |||
3030 | SDValue Op, SelectionDAG &DAG, unsigned MaskOpc, unsigned VecOpc) const { | |||
3031 | MVT VT = Op.getSimpleValueType(); | |||
3032 | ||||
3033 | if (VT.getVectorElementType() == MVT::i1) | |||
3034 | return lowerToScalableOp(Op, DAG, MaskOpc, /*HasMask*/ false); | |||
3035 | ||||
3036 | return lowerToScalableOp(Op, DAG, VecOpc, /*HasMask*/ true); | |||
3037 | } | |||
3038 | ||||
3039 | // Lower vector ABS to smax(X, sub(0, X)). | |||
3040 | SDValue RISCVTargetLowering::lowerABS(SDValue Op, SelectionDAG &DAG) const { | |||
3041 | SDLoc DL(Op); | |||
3042 | MVT VT = Op.getSimpleValueType(); | |||
3043 | SDValue X = Op.getOperand(0); | |||
3044 | ||||
3045 | // For scalable vectors we just need to deal with i64 on RV32 since the | |||
3046 | // default expansion crashes in getConstant. | |||
3047 | if (VT.isScalableVector()) { | |||
3048 | assert(!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64 &&((!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64 && "Unexpected custom lowering!") ? static_cast <void> (0) : __assert_fail ("!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64 && \"Unexpected custom lowering!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3049, __PRETTY_FUNCTION__)) | |||
3049 | "Unexpected custom lowering!")((!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64 && "Unexpected custom lowering!") ? static_cast <void> (0) : __assert_fail ("!Subtarget.is64Bit() && VT.getVectorElementType() == MVT::i64 && \"Unexpected custom lowering!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3049, __PRETTY_FUNCTION__)); | |||
3050 | SDValue SplatZero = DAG.getNode(RISCVISD::SPLAT_VECTOR_I64, DL, VT, | |||
3051 | DAG.getConstant(0, DL, MVT::i32)); | |||
3052 | SDValue NegX = DAG.getNode(ISD::SUB, DL, VT, SplatZero, X); | |||
3053 | return DAG.getNode(ISD::SMAX, DL, VT, X, NegX); | |||
3054 | } | |||
3055 | ||||
3056 | assert(VT.isFixedLengthVector() && "Unexpected type")((VT.isFixedLengthVector() && "Unexpected type") ? static_cast <void> (0) : __assert_fail ("VT.isFixedLengthVector() && \"Unexpected type\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3056, __PRETTY_FUNCTION__)); | |||
3057 | ||||
3058 | MVT ContainerVT = | |||
3059 | RISCVTargetLowering::getContainerForFixedLengthVector(DAG, VT, Subtarget); | |||
3060 | X = convertToScalableVector(ContainerVT, X, DAG, Subtarget); | |||
3061 | ||||
3062 | SDValue Mask, VL; | |||
3063 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
3064 | ||||
3065 | SDValue SplatZero = | |||
3066 | DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, | |||
3067 | DAG.getConstant(0, DL, Subtarget.getXLenVT())); | |||
3068 | SDValue NegX = | |||
3069 | DAG.getNode(RISCVISD::SUB_VL, DL, ContainerVT, SplatZero, X, Mask, VL); | |||
3070 | SDValue Max = | |||
3071 | DAG.getNode(RISCVISD::SMAX_VL, DL, ContainerVT, X, NegX, Mask, VL); | |||
3072 | ||||
3073 | return convertFromScalableVector(VT, Max, DAG, Subtarget); | |||
3074 | } | |||
3075 | ||||
3076 | SDValue RISCVTargetLowering::lowerFixedLengthVectorSelectToRVV( | |||
3077 | SDValue Op, SelectionDAG &DAG) const { | |||
3078 | MVT VT = Op.getSimpleValueType(); | |||
3079 | MVT ContainerVT = getContainerForFixedLengthVector(VT); | |||
3080 | ||||
3081 | MVT I1ContainerVT = | |||
3082 | MVT::getVectorVT(MVT::i1, ContainerVT.getVectorElementCount()); | |||
3083 | ||||
3084 | SDValue CC = | |||
3085 | convertToScalableVector(I1ContainerVT, Op.getOperand(0), DAG, Subtarget); | |||
3086 | SDValue Op1 = | |||
3087 | convertToScalableVector(ContainerVT, Op.getOperand(1), DAG, Subtarget); | |||
3088 | SDValue Op2 = | |||
3089 | convertToScalableVector(ContainerVT, Op.getOperand(2), DAG, Subtarget); | |||
3090 | ||||
3091 | SDLoc DL(Op); | |||
3092 | SDValue Mask, VL; | |||
3093 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
3094 | ||||
3095 | SDValue Select = | |||
3096 | DAG.getNode(RISCVISD::VSELECT_VL, DL, ContainerVT, CC, Op1, Op2, VL); | |||
3097 | ||||
3098 | return convertFromScalableVector(VT, Select, DAG, Subtarget); | |||
3099 | } | |||
3100 | ||||
3101 | SDValue RISCVTargetLowering::lowerToScalableOp(SDValue Op, SelectionDAG &DAG, | |||
3102 | unsigned NewOpc, | |||
3103 | bool HasMask) const { | |||
3104 | MVT VT = Op.getSimpleValueType(); | |||
3105 | assert(useRVVForFixedLengthVectorVT(VT) &&((useRVVForFixedLengthVectorVT(VT) && "Only expected to lower fixed length vector operation!" ) ? static_cast<void> (0) : __assert_fail ("useRVVForFixedLengthVectorVT(VT) && \"Only expected to lower fixed length vector operation!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3106, __PRETTY_FUNCTION__)) | |||
3106 | "Only expected to lower fixed length vector operation!")((useRVVForFixedLengthVectorVT(VT) && "Only expected to lower fixed length vector operation!" ) ? static_cast<void> (0) : __assert_fail ("useRVVForFixedLengthVectorVT(VT) && \"Only expected to lower fixed length vector operation!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3106, __PRETTY_FUNCTION__)); | |||
3107 | MVT ContainerVT = getContainerForFixedLengthVector(VT); | |||
3108 | ||||
3109 | // Create list of operands by converting existing ones to scalable types. | |||
3110 | SmallVector<SDValue, 6> Ops; | |||
3111 | for (const SDValue &V : Op->op_values()) { | |||
3112 | assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!")((!isa<VTSDNode>(V) && "Unexpected VTSDNode node!" ) ? static_cast<void> (0) : __assert_fail ("!isa<VTSDNode>(V) && \"Unexpected VTSDNode node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3112, __PRETTY_FUNCTION__)); | |||
3113 | ||||
3114 | // Pass through non-vector operands. | |||
3115 | if (!V.getValueType().isVector()) { | |||
3116 | Ops.push_back(V); | |||
3117 | continue; | |||
3118 | } | |||
3119 | ||||
3120 | // "cast" fixed length vector to a scalable vector. | |||
3121 | assert(useRVVForFixedLengthVectorVT(V.getSimpleValueType()) &&((useRVVForFixedLengthVectorVT(V.getSimpleValueType()) && "Only fixed length vectors are supported!") ? static_cast< void> (0) : __assert_fail ("useRVVForFixedLengthVectorVT(V.getSimpleValueType()) && \"Only fixed length vectors are supported!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3122, __PRETTY_FUNCTION__)) | |||
3122 | "Only fixed length vectors are supported!")((useRVVForFixedLengthVectorVT(V.getSimpleValueType()) && "Only fixed length vectors are supported!") ? static_cast< void> (0) : __assert_fail ("useRVVForFixedLengthVectorVT(V.getSimpleValueType()) && \"Only fixed length vectors are supported!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3122, __PRETTY_FUNCTION__)); | |||
3123 | Ops.push_back(convertToScalableVector(ContainerVT, V, DAG, Subtarget)); | |||
3124 | } | |||
3125 | ||||
3126 | SDLoc DL(Op); | |||
3127 | SDValue Mask, VL; | |||
3128 | std::tie(Mask, VL) = getDefaultVLOps(VT, ContainerVT, DL, DAG, Subtarget); | |||
3129 | if (HasMask) | |||
3130 | Ops.push_back(Mask); | |||
3131 | Ops.push_back(VL); | |||
3132 | ||||
3133 | SDValue ScalableRes = DAG.getNode(NewOpc, DL, ContainerVT, Ops); | |||
3134 | return convertFromScalableVector(VT, ScalableRes, DAG, Subtarget); | |||
3135 | } | |||
3136 | ||||
3137 | // Returns the opcode of the target-specific SDNode that implements the 32-bit | |||
3138 | // form of the given Opcode. | |||
3139 | static RISCVISD::NodeType getRISCVWOpcode(unsigned Opcode) { | |||
3140 | switch (Opcode) { | |||
3141 | default: | |||
3142 | llvm_unreachable("Unexpected opcode")::llvm::llvm_unreachable_internal("Unexpected opcode", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3142); | |||
3143 | case ISD::SHL: | |||
3144 | return RISCVISD::SLLW; | |||
3145 | case ISD::SRA: | |||
3146 | return RISCVISD::SRAW; | |||
3147 | case ISD::SRL: | |||
3148 | return RISCVISD::SRLW; | |||
3149 | case ISD::SDIV: | |||
3150 | return RISCVISD::DIVW; | |||
3151 | case ISD::UDIV: | |||
3152 | return RISCVISD::DIVUW; | |||
3153 | case ISD::UREM: | |||
3154 | return RISCVISD::REMUW; | |||
3155 | case ISD::ROTL: | |||
3156 | return RISCVISD::ROLW; | |||
3157 | case ISD::ROTR: | |||
3158 | return RISCVISD::RORW; | |||
3159 | case RISCVISD::GREVI: | |||
3160 | return RISCVISD::GREVIW; | |||
3161 | case RISCVISD::GORCI: | |||
3162 | return RISCVISD::GORCIW; | |||
3163 | } | |||
3164 | } | |||
3165 | ||||
3166 | // Converts the given 32-bit operation to a target-specific SelectionDAG node. | |||
3167 | // Because i32 isn't a legal type for RV64, these operations would otherwise | |||
3168 | // be promoted to i64, making it difficult to select the SLLW/DIVUW/.../*W | |||
3169 | // later one because the fact the operation was originally of type i32 is | |||
3170 | // lost. | |||
3171 | static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG, | |||
3172 | unsigned ExtOpc = ISD::ANY_EXTEND) { | |||
3173 | SDLoc DL(N); | |||
3174 | RISCVISD::NodeType WOpcode = getRISCVWOpcode(N->getOpcode()); | |||
3175 | SDValue NewOp0 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(0)); | |||
3176 | SDValue NewOp1 = DAG.getNode(ExtOpc, DL, MVT::i64, N->getOperand(1)); | |||
3177 | SDValue NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1); | |||
3178 | // ReplaceNodeResults requires we maintain the same type for the return value. | |||
3179 | return DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), NewRes); | |||
3180 | } | |||
3181 | ||||
3182 | // Converts the given 32-bit operation to a i64 operation with signed extension | |||
3183 | // semantic to reduce the signed extension instructions. | |||
3184 | static SDValue customLegalizeToWOpWithSExt(SDNode *N, SelectionDAG &DAG) { | |||
3185 | SDLoc DL(N); | |||
3186 | SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0)); | |||
3187 | SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1)); | |||
3188 | SDValue NewWOp = DAG.getNode(N->getOpcode(), DL, MVT::i64, NewOp0, NewOp1); | |||
3189 | SDValue NewRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, NewWOp, | |||
3190 | DAG.getValueType(MVT::i32)); | |||
3191 | return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes); | |||
3192 | } | |||
3193 | ||||
3194 | void RISCVTargetLowering::ReplaceNodeResults(SDNode *N, | |||
3195 | SmallVectorImpl<SDValue> &Results, | |||
3196 | SelectionDAG &DAG) const { | |||
3197 | SDLoc DL(N); | |||
3198 | switch (N->getOpcode()) { | |||
3199 | default: | |||
3200 | llvm_unreachable("Don't know how to custom type legalize this operation!")::llvm::llvm_unreachable_internal("Don't know how to custom type legalize this operation!" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3200); | |||
3201 | case ISD::STRICT_FP_TO_SINT: | |||
3202 | case ISD::STRICT_FP_TO_UINT: | |||
3203 | case ISD::FP_TO_SINT: | |||
3204 | case ISD::FP_TO_UINT: { | |||
3205 | bool IsStrict = N->isStrictFPOpcode(); | |||
3206 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3207, __PRETTY_FUNCTION__)) | |||
3207 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3207, __PRETTY_FUNCTION__)); | |||
3208 | SDValue Op0 = IsStrict ? N->getOperand(1) : N->getOperand(0); | |||
3209 | // If the FP type needs to be softened, emit a library call using the 'si' | |||
3210 | // version. If we left it to default legalization we'd end up with 'di'. If | |||
3211 | // the FP type doesn't need to be softened just let generic type | |||
3212 | // legalization promote the result type. | |||
3213 | if (getTypeAction(*DAG.getContext(), Op0.getValueType()) != | |||
3214 | TargetLowering::TypeSoftenFloat) | |||
3215 | return; | |||
3216 | RTLIB::Libcall LC; | |||
3217 | if (N->getOpcode() == ISD::FP_TO_SINT || | |||
3218 | N->getOpcode() == ISD::STRICT_FP_TO_SINT) | |||
3219 | LC = RTLIB::getFPTOSINT(Op0.getValueType(), N->getValueType(0)); | |||
3220 | else | |||
3221 | LC = RTLIB::getFPTOUINT(Op0.getValueType(), N->getValueType(0)); | |||
3222 | MakeLibCallOptions CallOptions; | |||
3223 | EVT OpVT = Op0.getValueType(); | |||
3224 | CallOptions.setTypeListBeforeSoften(OpVT, N->getValueType(0), true); | |||
3225 | SDValue Chain = IsStrict ? N->getOperand(0) : SDValue(); | |||
3226 | SDValue Result; | |||
3227 | std::tie(Result, Chain) = | |||
3228 | makeLibCall(DAG, LC, N->getValueType(0), Op0, CallOptions, DL, Chain); | |||
3229 | Results.push_back(Result); | |||
3230 | if (IsStrict) | |||
3231 | Results.push_back(Chain); | |||
3232 | break; | |||
3233 | } | |||
3234 | case ISD::READCYCLECOUNTER: { | |||
3235 | assert(!Subtarget.is64Bit() &&((!Subtarget.is64Bit() && "READCYCLECOUNTER only has custom type legalization on riscv32" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && \"READCYCLECOUNTER only has custom type legalization on riscv32\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3236, __PRETTY_FUNCTION__)) | |||
3236 | "READCYCLECOUNTER only has custom type legalization on riscv32")((!Subtarget.is64Bit() && "READCYCLECOUNTER only has custom type legalization on riscv32" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && \"READCYCLECOUNTER only has custom type legalization on riscv32\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3236, __PRETTY_FUNCTION__)); | |||
3237 | ||||
3238 | SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other); | |||
3239 | SDValue RCW = | |||
3240 | DAG.getNode(RISCVISD::READ_CYCLE_WIDE, DL, VTs, N->getOperand(0)); | |||
3241 | ||||
3242 | Results.push_back( | |||
3243 | DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, RCW, RCW.getValue(1))); | |||
3244 | Results.push_back(RCW.getValue(2)); | |||
3245 | break; | |||
3246 | } | |||
3247 | case ISD::ADD: | |||
3248 | case ISD::SUB: | |||
3249 | case ISD::MUL: | |||
3250 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3251, __PRETTY_FUNCTION__)) | |||
3251 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3251, __PRETTY_FUNCTION__)); | |||
3252 | if (N->getOperand(1).getOpcode() == ISD::Constant) | |||
3253 | return; | |||
3254 | Results.push_back(customLegalizeToWOpWithSExt(N, DAG)); | |||
3255 | break; | |||
3256 | case ISD::SHL: | |||
3257 | case ISD::SRA: | |||
3258 | case ISD::SRL: | |||
3259 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3260, __PRETTY_FUNCTION__)) | |||
3260 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3260, __PRETTY_FUNCTION__)); | |||
3261 | if (N->getOperand(1).getOpcode() == ISD::Constant) | |||
3262 | return; | |||
3263 | Results.push_back(customLegalizeToWOp(N, DAG)); | |||
3264 | break; | |||
3265 | case ISD::ROTL: | |||
3266 | case ISD::ROTR: | |||
3267 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3268, __PRETTY_FUNCTION__)) | |||
3268 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3268, __PRETTY_FUNCTION__)); | |||
3269 | Results.push_back(customLegalizeToWOp(N, DAG)); | |||
3270 | break; | |||
3271 | case ISD::SDIV: | |||
3272 | case ISD::UDIV: | |||
3273 | case ISD::UREM: { | |||
3274 | MVT VT = N->getSimpleValueType(0); | |||
3275 | assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&(((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3277, __PRETTY_FUNCTION__)) | |||
3276 | Subtarget.is64Bit() && Subtarget.hasStdExtM() &&(((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3277, __PRETTY_FUNCTION__)) | |||
3277 | "Unexpected custom legalisation")(((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) && Subtarget.is64Bit() && Subtarget.hasStdExtM() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3277, __PRETTY_FUNCTION__)); | |||
3278 | if (N->getOperand(0).getOpcode() == ISD::Constant || | |||
3279 | N->getOperand(1).getOpcode() == ISD::Constant) | |||
3280 | return; | |||
3281 | ||||
3282 | // If the input is i32, use ANY_EXTEND since the W instructions don't read | |||
3283 | // the upper 32 bits. For other types we need to sign or zero extend | |||
3284 | // based on the opcode. | |||
3285 | unsigned ExtOpc = ISD::ANY_EXTEND; | |||
3286 | if (VT != MVT::i32) | |||
3287 | ExtOpc = N->getOpcode() == ISD::SDIV ? ISD::SIGN_EXTEND | |||
3288 | : ISD::ZERO_EXTEND; | |||
3289 | ||||
3290 | Results.push_back(customLegalizeToWOp(N, DAG, ExtOpc)); | |||
3291 | break; | |||
3292 | } | |||
3293 | case ISD::BITCAST: { | |||
3294 | assert(((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtF()) || (N->getValueType( 0) == MVT::i16 && Subtarget.hasStdExtZfh())) && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtF()) || (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh())) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3297, __PRETTY_FUNCTION__)) | |||
3295 | Subtarget.hasStdExtF()) ||((((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtF()) || (N->getValueType( 0) == MVT::i16 && Subtarget.hasStdExtZfh())) && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtF()) || (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh())) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3297, __PRETTY_FUNCTION__)) | |||
3296 | (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh())) &&((((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtF()) || (N->getValueType( 0) == MVT::i16 && Subtarget.hasStdExtZfh())) && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtF()) || (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh())) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3297, __PRETTY_FUNCTION__)) | |||
3297 | "Unexpected custom legalisation")((((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtF()) || (N->getValueType( 0) == MVT::i16 && Subtarget.hasStdExtZfh())) && "Unexpected custom legalisation") ? static_cast<void> ( 0) : __assert_fail ("((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtF()) || (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh())) && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3297, __PRETTY_FUNCTION__)); | |||
3298 | SDValue Op0 = N->getOperand(0); | |||
3299 | if (N->getValueType(0) == MVT::i16 && Subtarget.hasStdExtZfh()) { | |||
3300 | if (Op0.getValueType() != MVT::f16) | |||
3301 | return; | |||
3302 | SDValue FPConv = | |||
3303 | DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, Subtarget.getXLenVT(), Op0); | |||
3304 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, FPConv)); | |||
3305 | } else if (N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && | |||
3306 | Subtarget.hasStdExtF()) { | |||
3307 | if (Op0.getValueType() != MVT::f32) | |||
3308 | return; | |||
3309 | SDValue FPConv = | |||
3310 | DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Op0); | |||
3311 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPConv)); | |||
3312 | } | |||
3313 | break; | |||
3314 | } | |||
3315 | case RISCVISD::GREVI: | |||
3316 | case RISCVISD::GORCI: { | |||
3317 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3318, __PRETTY_FUNCTION__)) | |||
3318 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3318, __PRETTY_FUNCTION__)); | |||
3319 | // This is similar to customLegalizeToWOp, except that we pass the second | |||
3320 | // operand (a TargetConstant) straight through: it is already of type | |||
3321 | // XLenVT. | |||
3322 | SDLoc DL(N); | |||
3323 | RISCVISD::NodeType WOpcode = getRISCVWOpcode(N->getOpcode()); | |||
3324 | SDValue NewOp0 = | |||
3325 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0)); | |||
3326 | SDValue NewRes = | |||
3327 | DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, N->getOperand(1)); | |||
3328 | // ReplaceNodeResults requires we maintain the same type for the return | |||
3329 | // value. | |||
3330 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes)); | |||
3331 | break; | |||
3332 | } | |||
3333 | case RISCVISD::SHFLI: { | |||
3334 | // There is no SHFLIW instruction, but we can just promote the operation. | |||
3335 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3336, __PRETTY_FUNCTION__)) | |||
3336 | "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && "Unexpected custom legalisation") ? static_cast <void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3336, __PRETTY_FUNCTION__)); | |||
3337 | SDLoc DL(N); | |||
3338 | SDValue NewOp0 = | |||
3339 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0)); | |||
3340 | SDValue NewRes = | |||
3341 | DAG.getNode(RISCVISD::SHFLI, DL, MVT::i64, NewOp0, N->getOperand(1)); | |||
3342 | // ReplaceNodeResults requires we maintain the same type for the return | |||
3343 | // value. | |||
3344 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes)); | |||
3345 | break; | |||
3346 | } | |||
3347 | case ISD::BSWAP: | |||
3348 | case ISD::BITREVERSE: { | |||
3349 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtZbp() && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbp() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3350, __PRETTY_FUNCTION__)) | |||
3350 | Subtarget.hasStdExtZbp() && "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtZbp() && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbp() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3350, __PRETTY_FUNCTION__)); | |||
3351 | SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, | |||
3352 | N->getOperand(0)); | |||
3353 | unsigned Imm = N->getOpcode() == ISD::BITREVERSE ? 31 : 24; | |||
3354 | SDValue GREVIW = DAG.getNode(RISCVISD::GREVIW, DL, MVT::i64, NewOp0, | |||
3355 | DAG.getTargetConstant(Imm, DL, | |||
3356 | Subtarget.getXLenVT())); | |||
3357 | // ReplaceNodeResults requires we maintain the same type for the return | |||
3358 | // value. | |||
3359 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, GREVIW)); | |||
3360 | break; | |||
3361 | } | |||
3362 | case ISD::FSHL: | |||
3363 | case ISD::FSHR: { | |||
3364 | assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtZbt() && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbt() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3365, __PRETTY_FUNCTION__)) | |||
3365 | Subtarget.hasStdExtZbt() && "Unexpected custom legalisation")((N->getValueType(0) == MVT::i32 && Subtarget.is64Bit () && Subtarget.hasStdExtZbt() && "Unexpected custom legalisation" ) ? static_cast<void> (0) : __assert_fail ("N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbt() && \"Unexpected custom legalisation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3365, __PRETTY_FUNCTION__)); | |||
3366 | SDValue NewOp0 = | |||
3367 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0)); | |||
3368 | SDValue NewOp1 = | |||
3369 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1)); | |||
3370 | SDValue NewOp2 = | |||
3371 | DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(2)); | |||
3372 | // FSLW/FSRW take a 6 bit shift amount but i32 FSHL/FSHR only use 5 bits. | |||
3373 | // Mask the shift amount to 5 bits. | |||
3374 | NewOp2 = DAG.getNode(ISD::AND, DL, MVT::i64, NewOp2, | |||
3375 | DAG.getConstant(0x1f, DL, MVT::i64)); | |||
3376 | unsigned Opc = | |||
3377 | N->getOpcode() == ISD::FSHL ? RISCVISD::FSLW : RISCVISD::FSRW; | |||
3378 | SDValue NewOp = DAG.getNode(Opc, DL, MVT::i64, NewOp0, NewOp1, NewOp2); | |||
3379 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewOp)); | |||
3380 | break; | |||
3381 | } | |||
3382 | case ISD::EXTRACT_VECTOR_ELT: { | |||
3383 | // Custom-legalize an EXTRACT_VECTOR_ELT where XLEN<SEW, as the SEW element | |||
3384 | // type is illegal (currently only vXi64 RV32). | |||
3385 | // With vmv.x.s, when SEW > XLEN, only the least-significant XLEN bits are | |||
3386 | // transferred to the destination register. We issue two of these from the | |||
3387 | // upper- and lower- halves of the SEW-bit vector element, slid down to the | |||
3388 | // first element. | |||
3389 | SDLoc DL(N); | |||
3390 | SDValue Vec = N->getOperand(0); | |||
3391 | SDValue Idx = N->getOperand(1); | |||
3392 | ||||
3393 | // The vector type hasn't been legalized yet so we can't issue target | |||
3394 | // specific nodes if it needs legalization. | |||
3395 | // FIXME: We would manually legalize if it's important. | |||
3396 | if (!isTypeLegal(Vec.getValueType())) | |||
3397 | return; | |||
3398 | ||||
3399 | MVT VecVT = Vec.getSimpleValueType(); | |||
3400 | ||||
3401 | assert(!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 &&((!Subtarget.is64Bit() && N->getValueType(0) == MVT ::i64 && VecVT.getVectorElementType() == MVT::i64 && "Unexpected EXTRACT_VECTOR_ELT legalization") ? static_cast< void> (0) : __assert_fail ("!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 && VecVT.getVectorElementType() == MVT::i64 && \"Unexpected EXTRACT_VECTOR_ELT legalization\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3403, __PRETTY_FUNCTION__)) | |||
3402 | VecVT.getVectorElementType() == MVT::i64 &&((!Subtarget.is64Bit() && N->getValueType(0) == MVT ::i64 && VecVT.getVectorElementType() == MVT::i64 && "Unexpected EXTRACT_VECTOR_ELT legalization") ? static_cast< void> (0) : __assert_fail ("!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 && VecVT.getVectorElementType() == MVT::i64 && \"Unexpected EXTRACT_VECTOR_ELT legalization\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3403, __PRETTY_FUNCTION__)) | |||
3403 | "Unexpected EXTRACT_VECTOR_ELT legalization")((!Subtarget.is64Bit() && N->getValueType(0) == MVT ::i64 && VecVT.getVectorElementType() == MVT::i64 && "Unexpected EXTRACT_VECTOR_ELT legalization") ? static_cast< void> (0) : __assert_fail ("!Subtarget.is64Bit() && N->getValueType(0) == MVT::i64 && VecVT.getVectorElementType() == MVT::i64 && \"Unexpected EXTRACT_VECTOR_ELT legalization\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3403, __PRETTY_FUNCTION__)); | |||
3404 | ||||
3405 | // If this is a fixed vector, we need to convert it to a scalable vector. | |||
3406 | MVT ContainerVT = VecVT; | |||
3407 | if (VecVT.isFixedLengthVector()) { | |||
3408 | ContainerVT = getContainerForFixedLengthVector(VecVT); | |||
3409 | Vec = convertToScalableVector(ContainerVT, Vec, DAG, Subtarget); | |||
3410 | } | |||
3411 | ||||
3412 | MVT XLenVT = Subtarget.getXLenVT(); | |||
3413 | ||||
3414 | // Use a VL of 1 to avoid processing more elements than we need. | |||
3415 | MVT MaskVT = MVT::getVectorVT(MVT::i1, VecVT.getVectorElementCount()); | |||
3416 | SDValue VL = DAG.getConstant(1, DL, XLenVT); | |||
3417 | SDValue Mask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL); | |||
3418 | ||||
3419 | // Unless the index is known to be 0, we must slide the vector down to get | |||
3420 | // the desired element into index 0. | |||
3421 | if (!isNullConstant(Idx)) { | |||
3422 | Vec = DAG.getNode(RISCVISD::VSLIDEDOWN_VL, DL, ContainerVT, | |||
3423 | DAG.getUNDEF(ContainerVT), Vec, Idx, Mask, VL); | |||
3424 | } | |||
3425 | ||||
3426 | // Extract the lower XLEN bits of the correct vector element. | |||
3427 | SDValue EltLo = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, Vec); | |||
3428 | ||||
3429 | // To extract the upper XLEN bits of the vector element, shift the first | |||
3430 | // element right by 32 bits and re-extract the lower XLEN bits. | |||
3431 | SDValue ThirtyTwoV = DAG.getNode(RISCVISD::VMV_V_X_VL, DL, ContainerVT, | |||
3432 | DAG.getConstant(32, DL, XLenVT), VL); | |||
3433 | SDValue LShr32 = DAG.getNode(RISCVISD::SRL_VL, DL, ContainerVT, Vec, | |||
3434 | ThirtyTwoV, Mask, VL); | |||
3435 | ||||
3436 | SDValue EltHi = DAG.getNode(RISCVISD::VMV_X_S, DL, XLenVT, LShr32); | |||
3437 | ||||
3438 | Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, EltLo, EltHi)); | |||
3439 | break; | |||
3440 | } | |||
3441 | case ISD::INTRINSIC_WO_CHAIN: { | |||
3442 | unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); | |||
3443 | switch (IntNo) { | |||
3444 | default: | |||
3445 | llvm_unreachable(::llvm::llvm_unreachable_internal("Don't know how to custom type legalize this intrinsic!" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3446) | |||
3446 | "Don't know how to custom type legalize this intrinsic!")::llvm::llvm_unreachable_internal("Don't know how to custom type legalize this intrinsic!" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3446); | |||
3447 | case Intrinsic::riscv_vmv_x_s: { | |||
3448 | EVT VT = N->getValueType(0); | |||
3449 | assert((VT == MVT::i8 || VT == MVT::i16 ||(((VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && "Unexpected custom legalisation!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && \"Unexpected custom legalisation!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3451, __PRETTY_FUNCTION__)) | |||
3450 | (Subtarget.is64Bit() && VT == MVT::i32)) &&(((VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && "Unexpected custom legalisation!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && \"Unexpected custom legalisation!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3451, __PRETTY_FUNCTION__)) | |||
3451 | "Unexpected custom legalisation!")(((VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && "Unexpected custom legalisation!" ) ? static_cast<void> (0) : __assert_fail ("(VT == MVT::i8 || VT == MVT::i16 || (Subtarget.is64Bit() && VT == MVT::i32)) && \"Unexpected custom legalisation!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3451, __PRETTY_FUNCTION__)); | |||
3452 | SDValue Extract = DAG.getNode(RISCVISD::VMV_X_S, DL, | |||
3453 | Subtarget.getXLenVT(), N->getOperand(1)); | |||
3454 | Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Extract)); | |||
3455 | break; | |||
3456 | } | |||
3457 | } | |||
3458 | break; | |||
3459 | } | |||
3460 | case ISD::VECREDUCE_ADD: | |||
3461 | case ISD::VECREDUCE_AND: | |||
3462 | case ISD::VECREDUCE_OR: | |||
3463 | case ISD::VECREDUCE_XOR: | |||
3464 | case ISD::VECREDUCE_SMAX: | |||
3465 | case ISD::VECREDUCE_UMAX: | |||
3466 | case ISD::VECREDUCE_SMIN: | |||
3467 | case ISD::VECREDUCE_UMIN: | |||
3468 | // The custom-lowering for these nodes returns a vector whose first element | |||
3469 | // is the result of the reduction. Extract its first element and let the | |||
3470 | // legalization for EXTRACT_VECTOR_ELT do the rest of the job. | |||
3471 | if (SDValue V = lowerVECREDUCE(SDValue(N, 0), DAG)) | |||
3472 | Results.push_back(V); | |||
3473 | break; | |||
3474 | } | |||
3475 | } | |||
3476 | ||||
3477 | // A structure to hold one of the bit-manipulation patterns below. Together, a | |||
3478 | // SHL and non-SHL pattern may form a bit-manipulation pair on a single source: | |||
3479 | // (or (and (shl x, 1), 0xAAAAAAAA), | |||
3480 | // (and (srl x, 1), 0x55555555)) | |||
3481 | struct RISCVBitmanipPat { | |||
3482 | SDValue Op; | |||
3483 | unsigned ShAmt; | |||
3484 | bool IsSHL; | |||
3485 | ||||
3486 | bool formsPairWith(const RISCVBitmanipPat &Other) const { | |||
3487 | return Op == Other.Op && ShAmt == Other.ShAmt && IsSHL != Other.IsSHL; | |||
3488 | } | |||
3489 | }; | |||
3490 | ||||
3491 | // Matches patterns of the form | |||
3492 | // (and (shl x, C2), (C1 << C2)) | |||
3493 | // (and (srl x, C2), C1) | |||
3494 | // (shl (and x, C1), C2) | |||
3495 | // (srl (and x, (C1 << C2)), C2) | |||
3496 | // Where C2 is a power of 2 and C1 has at least that many leading zeroes. | |||
3497 | // The expected masks for each shift amount are specified in BitmanipMasks where | |||
3498 | // BitmanipMasks[log2(C2)] specifies the expected C1 value. | |||
3499 | // The max allowed shift amount is either XLen/2 or XLen/4 determined by whether | |||
3500 | // BitmanipMasks contains 6 or 5 entries assuming that the maximum possible | |||
3501 | // XLen is 64. | |||
3502 | static Optional<RISCVBitmanipPat> | |||
3503 | matchRISCVBitmanipPat(SDValue Op, ArrayRef<uint64_t> BitmanipMasks) { | |||
3504 | assert((BitmanipMasks.size() == 5 || BitmanipMasks.size() == 6) &&(((BitmanipMasks.size() == 5 || BitmanipMasks.size() == 6) && "Unexpected number of masks") ? static_cast<void> (0) : __assert_fail ("(BitmanipMasks.size() == 5 || BitmanipMasks.size() == 6) && \"Unexpected number of masks\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3505, __PRETTY_FUNCTION__)) | |||
3505 | "Unexpected number of masks")(((BitmanipMasks.size() == 5 || BitmanipMasks.size() == 6) && "Unexpected number of masks") ? static_cast<void> (0) : __assert_fail ("(BitmanipMasks.size() == 5 || BitmanipMasks.size() == 6) && \"Unexpected number of masks\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3505, __PRETTY_FUNCTION__)); | |||
3506 | Optional<uint64_t> Mask; | |||
3507 | // Optionally consume a mask around the shift operation. | |||
3508 | if (Op.getOpcode() == ISD::AND && isa<ConstantSDNode>(Op.getOperand(1))) { | |||
3509 | Mask = Op.getConstantOperandVal(1); | |||
3510 | Op = Op.getOperand(0); | |||
3511 | } | |||
3512 | if (Op.getOpcode() != ISD::SHL && Op.getOpcode() != ISD::SRL) | |||
3513 | return None; | |||
3514 | bool IsSHL = Op.getOpcode() == ISD::SHL; | |||
3515 | ||||
3516 | if (!isa<ConstantSDNode>(Op.getOperand(1))) | |||
3517 | return None; | |||
3518 | uint64_t ShAmt = Op.getConstantOperandVal(1); | |||
3519 | ||||
3520 | unsigned Width = Op.getValueType() == MVT::i64 ? 64 : 32; | |||
3521 | if (ShAmt >= Width && !isPowerOf2_64(ShAmt)) | |||
3522 | return None; | |||
3523 | // If we don't have enough masks for 64 bit, then we must be trying to | |||
3524 | // match SHFL so we're only allowed to shift 1/4 of the width. | |||
3525 | if (BitmanipMasks.size() == 5 && ShAmt >= (Width / 2)) | |||
3526 | return None; | |||
3527 | ||||
3528 | SDValue Src = Op.getOperand(0); | |||
3529 | ||||
3530 | // The expected mask is shifted left when the AND is found around SHL | |||
3531 | // patterns. | |||
3532 | // ((x >> 1) & 0x55555555) | |||
3533 | // ((x << 1) & 0xAAAAAAAA) | |||
3534 | bool SHLExpMask = IsSHL; | |||
3535 | ||||
3536 | if (!Mask) { | |||
3537 | // Sometimes LLVM keeps the mask as an operand of the shift, typically when | |||
3538 | // the mask is all ones: consume that now. | |||
3539 | if (Src.getOpcode() == ISD::AND && isa<ConstantSDNode>(Src.getOperand(1))) { | |||
3540 | Mask = Src.getConstantOperandVal(1); | |||
3541 | Src = Src.getOperand(0); | |||
3542 | // The expected mask is now in fact shifted left for SRL, so reverse the | |||
3543 | // decision. | |||
3544 | // ((x & 0xAAAAAAAA) >> 1) | |||
3545 | // ((x & 0x55555555) << 1) | |||
3546 | SHLExpMask = !SHLExpMask; | |||
3547 | } else { | |||
3548 | // Use a default shifted mask of all-ones if there's no AND, truncated | |||
3549 | // down to the expected width. This simplifies the logic later on. | |||
3550 | Mask = maskTrailingOnes<uint64_t>(Width); | |||
3551 | *Mask &= (IsSHL ? *Mask << ShAmt : *Mask >> ShAmt); | |||
3552 | } | |||
3553 | } | |||
3554 | ||||
3555 | unsigned MaskIdx = Log2_32(ShAmt); | |||
3556 | uint64_t ExpMask = BitmanipMasks[MaskIdx] & maskTrailingOnes<uint64_t>(Width); | |||
3557 | ||||
3558 | if (SHLExpMask) | |||
3559 | ExpMask <<= ShAmt; | |||
3560 | ||||
3561 | if (Mask != ExpMask) | |||
3562 | return None; | |||
3563 | ||||
3564 | return RISCVBitmanipPat{Src, (unsigned)ShAmt, IsSHL}; | |||
3565 | } | |||
3566 | ||||
3567 | // Matches any of the following bit-manipulation patterns: | |||
3568 | // (and (shl x, 1), (0x55555555 << 1)) | |||
3569 | // (and (srl x, 1), 0x55555555) | |||
3570 | // (shl (and x, 0x55555555), 1) | |||
3571 | // (srl (and x, (0x55555555 << 1)), 1) | |||
3572 | // where the shift amount and mask may vary thus: | |||
3573 | // [1] = 0x55555555 / 0xAAAAAAAA | |||
3574 | // [2] = 0x33333333 / 0xCCCCCCCC | |||
3575 | // [4] = 0x0F0F0F0F / 0xF0F0F0F0 | |||
3576 | // [8] = 0x00FF00FF / 0xFF00FF00 | |||
3577 | // [16] = 0x0000FFFF / 0xFFFFFFFF | |||
3578 | // [32] = 0x00000000FFFFFFFF / 0xFFFFFFFF00000000 (for RV64) | |||
3579 | static Optional<RISCVBitmanipPat> matchGREVIPat(SDValue Op) { | |||
3580 | // These are the unshifted masks which we use to match bit-manipulation | |||
3581 | // patterns. They may be shifted left in certain circumstances. | |||
3582 | static const uint64_t BitmanipMasks[] = { | |||
3583 | 0x5555555555555555ULL, 0x3333333333333333ULL, 0x0F0F0F0F0F0F0F0FULL, | |||
3584 | 0x00FF00FF00FF00FFULL, 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL}; | |||
3585 | ||||
3586 | return matchRISCVBitmanipPat(Op, BitmanipMasks); | |||
3587 | } | |||
3588 | ||||
3589 | // Match the following pattern as a GREVI(W) operation | |||
3590 | // (or (BITMANIP_SHL x), (BITMANIP_SRL x)) | |||
3591 | static SDValue combineORToGREV(SDValue Op, SelectionDAG &DAG, | |||
3592 | const RISCVSubtarget &Subtarget) { | |||
3593 | assert(Subtarget.hasStdExtZbp() && "Expected Zbp extenson")((Subtarget.hasStdExtZbp() && "Expected Zbp extenson" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasStdExtZbp() && \"Expected Zbp extenson\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3593, __PRETTY_FUNCTION__)); | |||
3594 | EVT VT = Op.getValueType(); | |||
3595 | ||||
3596 | if (VT == Subtarget.getXLenVT() || (Subtarget.is64Bit() && VT == MVT::i32)) { | |||
3597 | auto LHS = matchGREVIPat(Op.getOperand(0)); | |||
3598 | auto RHS = matchGREVIPat(Op.getOperand(1)); | |||
3599 | if (LHS && RHS && LHS->formsPairWith(*RHS)) { | |||
3600 | SDLoc DL(Op); | |||
3601 | return DAG.getNode( | |||
3602 | RISCVISD::GREVI, DL, VT, LHS->Op, | |||
3603 | DAG.getTargetConstant(LHS->ShAmt, DL, Subtarget.getXLenVT())); | |||
3604 | } | |||
3605 | } | |||
3606 | return SDValue(); | |||
3607 | } | |||
3608 | ||||
3609 | // Matches any the following pattern as a GORCI(W) operation | |||
3610 | // 1. (or (GREVI x, shamt), x) if shamt is a power of 2 | |||
3611 | // 2. (or x, (GREVI x, shamt)) if shamt is a power of 2 | |||
3612 | // 3. (or (or (BITMANIP_SHL x), x), (BITMANIP_SRL x)) | |||
3613 | // Note that with the variant of 3., | |||
3614 | // (or (or (BITMANIP_SHL x), (BITMANIP_SRL x)), x) | |||
3615 | // the inner pattern will first be matched as GREVI and then the outer | |||
3616 | // pattern will be matched to GORC via the first rule above. | |||
3617 | // 4. (or (rotl/rotr x, bitwidth/2), x) | |||
3618 | static SDValue combineORToGORC(SDValue Op, SelectionDAG &DAG, | |||
3619 | const RISCVSubtarget &Subtarget) { | |||
3620 | assert(Subtarget.hasStdExtZbp() && "Expected Zbp extenson")((Subtarget.hasStdExtZbp() && "Expected Zbp extenson" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasStdExtZbp() && \"Expected Zbp extenson\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3620, __PRETTY_FUNCTION__)); | |||
3621 | EVT VT = Op.getValueType(); | |||
3622 | ||||
3623 | if (VT == Subtarget.getXLenVT() || (Subtarget.is64Bit() && VT == MVT::i32)) { | |||
3624 | SDLoc DL(Op); | |||
3625 | SDValue Op0 = Op.getOperand(0); | |||
3626 | SDValue Op1 = Op.getOperand(1); | |||
3627 | ||||
3628 | auto MatchOROfReverse = [&](SDValue Reverse, SDValue X) { | |||
3629 | if (Reverse.getOpcode() == RISCVISD::GREVI && Reverse.getOperand(0) == X && | |||
3630 | isPowerOf2_32(Reverse.getConstantOperandVal(1))) | |||
3631 | return DAG.getNode(RISCVISD::GORCI, DL, VT, X, Reverse.getOperand(1)); | |||
3632 | // We can also form GORCI from ROTL/ROTR by half the bitwidth. | |||
3633 | if ((Reverse.getOpcode() == ISD::ROTL || | |||
3634 | Reverse.getOpcode() == ISD::ROTR) && | |||
3635 | Reverse.getOperand(0) == X && | |||
3636 | isa<ConstantSDNode>(Reverse.getOperand(1))) { | |||
3637 | uint64_t RotAmt = Reverse.getConstantOperandVal(1); | |||
3638 | if (RotAmt == (VT.getSizeInBits() / 2)) | |||
3639 | return DAG.getNode( | |||
3640 | RISCVISD::GORCI, DL, VT, X, | |||
3641 | DAG.getTargetConstant(RotAmt, DL, Subtarget.getXLenVT())); | |||
3642 | } | |||
3643 | return SDValue(); | |||
3644 | }; | |||
3645 | ||||
3646 | // Check for either commutable permutation of (or (GREVI x, shamt), x) | |||
3647 | if (SDValue V = MatchOROfReverse(Op0, Op1)) | |||
3648 | return V; | |||
3649 | if (SDValue V = MatchOROfReverse(Op1, Op0)) | |||
3650 | return V; | |||
3651 | ||||
3652 | // OR is commutable so canonicalize its OR operand to the left | |||
3653 | if (Op0.getOpcode() != ISD::OR && Op1.getOpcode() == ISD::OR) | |||
3654 | std::swap(Op0, Op1); | |||
3655 | if (Op0.getOpcode() != ISD::OR) | |||
3656 | return SDValue(); | |||
3657 | SDValue OrOp0 = Op0.getOperand(0); | |||
3658 | SDValue OrOp1 = Op0.getOperand(1); | |||
3659 | auto LHS = matchGREVIPat(OrOp0); | |||
3660 | // OR is commutable so swap the operands and try again: x might have been | |||
3661 | // on the left | |||
3662 | if (!LHS) { | |||
3663 | std::swap(OrOp0, OrOp1); | |||
3664 | LHS = matchGREVIPat(OrOp0); | |||
3665 | } | |||
3666 | auto RHS = matchGREVIPat(Op1); | |||
3667 | if (LHS && RHS && LHS->formsPairWith(*RHS) && LHS->Op == OrOp1) { | |||
3668 | return DAG.getNode( | |||
3669 | RISCVISD::GORCI, DL, VT, LHS->Op, | |||
3670 | DAG.getTargetConstant(LHS->ShAmt, DL, Subtarget.getXLenVT())); | |||
3671 | } | |||
3672 | } | |||
3673 | return SDValue(); | |||
3674 | } | |||
3675 | ||||
3676 | // Matches any of the following bit-manipulation patterns: | |||
3677 | // (and (shl x, 1), (0x22222222 << 1)) | |||
3678 | // (and (srl x, 1), 0x22222222) | |||
3679 | // (shl (and x, 0x22222222), 1) | |||
3680 | // (srl (and x, (0x22222222 << 1)), 1) | |||
3681 | // where the shift amount and mask may vary thus: | |||
3682 | // [1] = 0x22222222 / 0x44444444 | |||
3683 | // [2] = 0x0C0C0C0C / 0x3C3C3C3C | |||
3684 | // [4] = 0x00F000F0 / 0x0F000F00 | |||
3685 | // [8] = 0x0000FF00 / 0x00FF0000 | |||
3686 | // [16] = 0x00000000FFFF0000 / 0x0000FFFF00000000 (for RV64) | |||
3687 | static Optional<RISCVBitmanipPat> matchSHFLPat(SDValue Op) { | |||
3688 | // These are the unshifted masks which we use to match bit-manipulation | |||
3689 | // patterns. They may be shifted left in certain circumstances. | |||
3690 | static const uint64_t BitmanipMasks[] = { | |||
3691 | 0x2222222222222222ULL, 0x0C0C0C0C0C0C0C0CULL, 0x00F000F000F000F0ULL, | |||
3692 | 0x0000FF000000FF00ULL, 0x00000000FFFF0000ULL}; | |||
3693 | ||||
3694 | return matchRISCVBitmanipPat(Op, BitmanipMasks); | |||
3695 | } | |||
3696 | ||||
3697 | // Match (or (or (SHFL_SHL x), (SHFL_SHR x)), (SHFL_AND x) | |||
3698 | static SDValue combineORToSHFL(SDValue Op, SelectionDAG &DAG, | |||
3699 | const RISCVSubtarget &Subtarget) { | |||
3700 | assert(Subtarget.hasStdExtZbp() && "Expected Zbp extenson")((Subtarget.hasStdExtZbp() && "Expected Zbp extenson" ) ? static_cast<void> (0) : __assert_fail ("Subtarget.hasStdExtZbp() && \"Expected Zbp extenson\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3700, __PRETTY_FUNCTION__)); | |||
3701 | EVT VT = Op.getValueType(); | |||
3702 | ||||
3703 | if (VT != MVT::i32 && VT != Subtarget.getXLenVT()) | |||
3704 | return SDValue(); | |||
3705 | ||||
3706 | SDValue Op0 = Op.getOperand(0); | |||
3707 | SDValue Op1 = Op.getOperand(1); | |||
3708 | ||||
3709 | // Or is commutable so canonicalize the second OR to the LHS. | |||
3710 | if (Op0.getOpcode() != ISD::OR) | |||
3711 | std::swap(Op0, Op1); | |||
3712 | if (Op0.getOpcode() != ISD::OR) | |||
3713 | return SDValue(); | |||
3714 | ||||
3715 | // We found an inner OR, so our operands are the operands of the inner OR | |||
3716 | // and the other operand of the outer OR. | |||
3717 | SDValue A = Op0.getOperand(0); | |||
3718 | SDValue B = Op0.getOperand(1); | |||
3719 | SDValue C = Op1; | |||
3720 | ||||
3721 | auto Match1 = matchSHFLPat(A); | |||
3722 | auto Match2 = matchSHFLPat(B); | |||
3723 | ||||
3724 | // If neither matched, we failed. | |||
3725 | if (!Match1 && !Match2) | |||
3726 | return SDValue(); | |||
3727 | ||||
3728 | // We had at least one match. if one failed, try the remaining C operand. | |||
3729 | if (!Match1) { | |||
3730 | std::swap(A, C); | |||
3731 | Match1 = matchSHFLPat(A); | |||
3732 | if (!Match1) | |||
3733 | return SDValue(); | |||
3734 | } else if (!Match2) { | |||
3735 | std::swap(B, C); | |||
3736 | Match2 = matchSHFLPat(B); | |||
3737 | if (!Match2) | |||
3738 | return SDValue(); | |||
3739 | } | |||
3740 | assert(Match1 && Match2)((Match1 && Match2) ? static_cast<void> (0) : __assert_fail ("Match1 && Match2", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3740, __PRETTY_FUNCTION__)); | |||
3741 | ||||
3742 | // Make sure our matches pair up. | |||
3743 | if (!Match1->formsPairWith(*Match2)) | |||
3744 | return SDValue(); | |||
3745 | ||||
3746 | // All the remains is to make sure C is an AND with the same input, that masks | |||
3747 | // out the bits that are being shuffled. | |||
3748 | if (C.getOpcode() != ISD::AND || !isa<ConstantSDNode>(C.getOperand(1)) || | |||
3749 | C.getOperand(0) != Match1->Op) | |||
3750 | return SDValue(); | |||
3751 | ||||
3752 | uint64_t Mask = C.getConstantOperandVal(1); | |||
3753 | ||||
3754 | static const uint64_t BitmanipMasks[] = { | |||
3755 | 0x9999999999999999ULL, 0xC3C3C3C3C3C3C3C3ULL, 0xF00FF00FF00FF00FULL, | |||
3756 | 0xFF0000FFFF0000FFULL, 0xFFFF00000000FFFFULL, | |||
3757 | }; | |||
3758 | ||||
3759 | unsigned Width = Op.getValueType() == MVT::i64 ? 64 : 32; | |||
3760 | unsigned MaskIdx = Log2_32(Match1->ShAmt); | |||
3761 | uint64_t ExpMask = BitmanipMasks[MaskIdx] & maskTrailingOnes<uint64_t>(Width); | |||
3762 | ||||
3763 | if (Mask != ExpMask) | |||
3764 | return SDValue(); | |||
3765 | ||||
3766 | SDLoc DL(Op); | |||
3767 | return DAG.getNode( | |||
3768 | RISCVISD::SHFLI, DL, VT, Match1->Op, | |||
3769 | DAG.getTargetConstant(Match1->ShAmt, DL, Subtarget.getXLenVT())); | |||
3770 | } | |||
3771 | ||||
3772 | // Combine (GREVI (GREVI x, C2), C1) -> (GREVI x, C1^C2) when C1^C2 is | |||
3773 | // non-zero, and to x when it is. Any repeated GREVI stage undoes itself. | |||
3774 | // Combine (GORCI (GORCI x, C2), C1) -> (GORCI x, C1|C2). Repeated stage does | |||
3775 | // not undo itself, but they are redundant. | |||
3776 | static SDValue combineGREVI_GORCI(SDNode *N, SelectionDAG &DAG) { | |||
3777 | unsigned ShAmt1 = N->getConstantOperandVal(1); | |||
3778 | SDValue Src = N->getOperand(0); | |||
3779 | ||||
3780 | if (Src.getOpcode() != N->getOpcode()) | |||
3781 | return SDValue(); | |||
3782 | ||||
3783 | unsigned ShAmt2 = Src.getConstantOperandVal(1); | |||
3784 | Src = Src.getOperand(0); | |||
3785 | ||||
3786 | unsigned CombinedShAmt; | |||
3787 | if (N->getOpcode() == RISCVISD::GORCI || N->getOpcode() == RISCVISD::GORCIW) | |||
3788 | CombinedShAmt = ShAmt1 | ShAmt2; | |||
3789 | else | |||
3790 | CombinedShAmt = ShAmt1 ^ ShAmt2; | |||
3791 | ||||
3792 | if (CombinedShAmt == 0) | |||
3793 | return Src; | |||
3794 | ||||
3795 | SDLoc DL(N); | |||
3796 | return DAG.getNode(N->getOpcode(), DL, N->getValueType(0), Src, | |||
3797 | DAG.getTargetConstant(CombinedShAmt, DL, | |||
3798 | N->getOperand(1).getValueType())); | |||
3799 | } | |||
3800 | ||||
3801 | SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N, | |||
3802 | DAGCombinerInfo &DCI) const { | |||
3803 | SelectionDAG &DAG = DCI.DAG; | |||
3804 | ||||
3805 | switch (N->getOpcode()) { | |||
3806 | default: | |||
3807 | break; | |||
3808 | case RISCVISD::SplitF64: { | |||
3809 | SDValue Op0 = N->getOperand(0); | |||
3810 | // If the input to SplitF64 is just BuildPairF64 then the operation is | |||
3811 | // redundant. Instead, use BuildPairF64's operands directly. | |||
3812 | if (Op0->getOpcode() == RISCVISD::BuildPairF64) | |||
3813 | return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1)); | |||
3814 | ||||
3815 | SDLoc DL(N); | |||
3816 | ||||
3817 | // It's cheaper to materialise two 32-bit integers than to load a double | |||
3818 | // from the constant pool and transfer it to integer registers through the | |||
3819 | // stack. | |||
3820 | if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) { | |||
3821 | APInt V = C->getValueAPF().bitcastToAPInt(); | |||
3822 | SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32); | |||
3823 | SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32); | |||
3824 | return DCI.CombineTo(N, Lo, Hi); | |||
3825 | } | |||
3826 | ||||
3827 | // This is a target-specific version of a DAGCombine performed in | |||
3828 | // DAGCombiner::visitBITCAST. It performs the equivalent of: | |||
3829 | // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) | |||
3830 | // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) | |||
3831 | if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) || | |||
3832 | !Op0.getNode()->hasOneUse()) | |||
3833 | break; | |||
3834 | SDValue NewSplitF64 = | |||
3835 | DAG.getNode(RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32), | |||
3836 | Op0.getOperand(0)); | |||
3837 | SDValue Lo = NewSplitF64.getValue(0); | |||
3838 | SDValue Hi = NewSplitF64.getValue(1); | |||
3839 | APInt SignBit = APInt::getSignMask(32); | |||
3840 | if (Op0.getOpcode() == ISD::FNEG) { | |||
3841 | SDValue NewHi = DAG.getNode(ISD::XOR, DL, MVT::i32, Hi, | |||
3842 | DAG.getConstant(SignBit, DL, MVT::i32)); | |||
3843 | return DCI.CombineTo(N, Lo, NewHi); | |||
3844 | } | |||
3845 | assert(Op0.getOpcode() == ISD::FABS)((Op0.getOpcode() == ISD::FABS) ? static_cast<void> (0) : __assert_fail ("Op0.getOpcode() == ISD::FABS", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3845, __PRETTY_FUNCTION__)); | |||
3846 | SDValue NewHi = DAG.getNode(ISD::AND, DL, MVT::i32, Hi, | |||
3847 | DAG.getConstant(~SignBit, DL, MVT::i32)); | |||
3848 | return DCI.CombineTo(N, Lo, NewHi); | |||
3849 | } | |||
3850 | case RISCVISD::SLLW: | |||
3851 | case RISCVISD::SRAW: | |||
3852 | case RISCVISD::SRLW: | |||
3853 | case RISCVISD::ROLW: | |||
3854 | case RISCVISD::RORW: { | |||
3855 | // Only the lower 32 bits of LHS and lower 5 bits of RHS are read. | |||
3856 | SDValue LHS = N->getOperand(0); | |||
3857 | SDValue RHS = N->getOperand(1); | |||
3858 | APInt LHSMask = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 32); | |||
3859 | APInt RHSMask = APInt::getLowBitsSet(RHS.getValueSizeInBits(), 5); | |||
3860 | if (SimplifyDemandedBits(N->getOperand(0), LHSMask, DCI) || | |||
3861 | SimplifyDemandedBits(N->getOperand(1), RHSMask, DCI)) { | |||
3862 | if (N->getOpcode() != ISD::DELETED_NODE) | |||
3863 | DCI.AddToWorklist(N); | |||
3864 | return SDValue(N, 0); | |||
3865 | } | |||
3866 | break; | |||
3867 | } | |||
3868 | case RISCVISD::FSL: | |||
3869 | case RISCVISD::FSR: { | |||
3870 | // Only the lower log2(Bitwidth)+1 bits of the the shift amount are read. | |||
3871 | SDValue ShAmt = N->getOperand(2); | |||
3872 | unsigned BitWidth = ShAmt.getValueSizeInBits(); | |||
3873 | assert(isPowerOf2_32(BitWidth) && "Unexpected bit width")((isPowerOf2_32(BitWidth) && "Unexpected bit width") ? static_cast<void> (0) : __assert_fail ("isPowerOf2_32(BitWidth) && \"Unexpected bit width\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3873, __PRETTY_FUNCTION__)); | |||
3874 | APInt ShAmtMask(BitWidth, (BitWidth * 2) - 1); | |||
3875 | if (SimplifyDemandedBits(ShAmt, ShAmtMask, DCI)) { | |||
3876 | if (N->getOpcode() != ISD::DELETED_NODE) | |||
3877 | DCI.AddToWorklist(N); | |||
3878 | return SDValue(N, 0); | |||
3879 | } | |||
3880 | break; | |||
3881 | } | |||
3882 | case RISCVISD::FSLW: | |||
3883 | case RISCVISD::FSRW: { | |||
3884 | // Only the lower 32 bits of Values and lower 6 bits of shift amount are | |||
3885 | // read. | |||
3886 | SDValue Op0 = N->getOperand(0); | |||
3887 | SDValue Op1 = N->getOperand(1); | |||
3888 | SDValue ShAmt = N->getOperand(2); | |||
3889 | APInt OpMask = APInt::getLowBitsSet(Op0.getValueSizeInBits(), 32); | |||
3890 | APInt ShAmtMask = APInt::getLowBitsSet(ShAmt.getValueSizeInBits(), 6); | |||
3891 | if (SimplifyDemandedBits(Op0, OpMask, DCI) || | |||
3892 | SimplifyDemandedBits(Op1, OpMask, DCI) || | |||
3893 | SimplifyDemandedBits(ShAmt, ShAmtMask, DCI)) { | |||
3894 | if (N->getOpcode() != ISD::DELETED_NODE) | |||
3895 | DCI.AddToWorklist(N); | |||
3896 | return SDValue(N, 0); | |||
3897 | } | |||
3898 | break; | |||
3899 | } | |||
3900 | case RISCVISD::GREVIW: | |||
3901 | case RISCVISD::GORCIW: { | |||
3902 | // Only the lower 32 bits of the first operand are read | |||
3903 | SDValue Op0 = N->getOperand(0); | |||
3904 | APInt Mask = APInt::getLowBitsSet(Op0.getValueSizeInBits(), 32); | |||
3905 | if (SimplifyDemandedBits(Op0, Mask, DCI)) { | |||
3906 | if (N->getOpcode() != ISD::DELETED_NODE) | |||
3907 | DCI.AddToWorklist(N); | |||
3908 | return SDValue(N, 0); | |||
3909 | } | |||
3910 | ||||
3911 | return combineGREVI_GORCI(N, DCI.DAG); | |||
3912 | } | |||
3913 | case RISCVISD::FMV_X_ANYEXTW_RV64: { | |||
3914 | SDLoc DL(N); | |||
3915 | SDValue Op0 = N->getOperand(0); | |||
3916 | // If the input to FMV_X_ANYEXTW_RV64 is just FMV_W_X_RV64 then the | |||
3917 | // conversion is unnecessary and can be replaced with an ANY_EXTEND | |||
3918 | // of the FMV_W_X_RV64 operand. | |||
3919 | if (Op0->getOpcode() == RISCVISD::FMV_W_X_RV64) { | |||
3920 | assert(Op0.getOperand(0).getValueType() == MVT::i64 &&((Op0.getOperand(0).getValueType() == MVT::i64 && "Unexpected value type!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getOperand(0).getValueType() == MVT::i64 && \"Unexpected value type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3921, __PRETTY_FUNCTION__)) | |||
3921 | "Unexpected value type!")((Op0.getOperand(0).getValueType() == MVT::i64 && "Unexpected value type!" ) ? static_cast<void> (0) : __assert_fail ("Op0.getOperand(0).getValueType() == MVT::i64 && \"Unexpected value type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3921, __PRETTY_FUNCTION__)); | |||
3922 | return Op0.getOperand(0); | |||
3923 | } | |||
3924 | ||||
3925 | // This is a target-specific version of a DAGCombine performed in | |||
3926 | // DAGCombiner::visitBITCAST. It performs the equivalent of: | |||
3927 | // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) | |||
3928 | // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) | |||
3929 | if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) || | |||
3930 | !Op0.getNode()->hasOneUse()) | |||
3931 | break; | |||
3932 | SDValue NewFMV = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, | |||
3933 | Op0.getOperand(0)); | |||
3934 | APInt SignBit = APInt::getSignMask(32).sext(64); | |||
3935 | if (Op0.getOpcode() == ISD::FNEG) | |||
3936 | return DAG.getNode(ISD::XOR, DL, MVT::i64, NewFMV, | |||
3937 | DAG.getConstant(SignBit, DL, MVT::i64)); | |||
3938 | ||||
3939 | assert(Op0.getOpcode() == ISD::FABS)((Op0.getOpcode() == ISD::FABS) ? static_cast<void> (0) : __assert_fail ("Op0.getOpcode() == ISD::FABS", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 3939, __PRETTY_FUNCTION__)); | |||
3940 | return DAG.getNode(ISD::AND, DL, MVT::i64, NewFMV, | |||
3941 | DAG.getConstant(~SignBit, DL, MVT::i64)); | |||
3942 | } | |||
3943 | case RISCVISD::GREVI: | |||
3944 | case RISCVISD::GORCI: | |||
3945 | return combineGREVI_GORCI(N, DCI.DAG); | |||
3946 | case ISD::OR: | |||
3947 | if (auto GREV = combineORToGREV(SDValue(N, 0), DCI.DAG, Subtarget)) | |||
3948 | return GREV; | |||
3949 | if (auto GORC = combineORToGORC(SDValue(N, 0), DCI.DAG, Subtarget)) | |||
3950 | return GORC; | |||
3951 | if (auto SHFL = combineORToSHFL(SDValue(N, 0), DCI.DAG, Subtarget)) | |||
3952 | return SHFL; | |||
3953 | break; | |||
3954 | case RISCVISD::SELECT_CC: { | |||
3955 | // Transform | |||
3956 | SDValue LHS = N->getOperand(0); | |||
3957 | SDValue RHS = N->getOperand(1); | |||
3958 | auto CCVal = static_cast<ISD::CondCode>(N->getConstantOperandVal(2)); | |||
3959 | if (!ISD::isIntEqualitySetCC(CCVal)) | |||
3960 | break; | |||
3961 | ||||
3962 | // Fold (select_cc (setlt X, Y), 0, ne, trueV, falseV) -> | |||
3963 | // (select_cc X, Y, lt, trueV, falseV) | |||
3964 | // Sometimes the setcc is introduced after select_cc has been formed. | |||
3965 | if (LHS.getOpcode() == ISD::SETCC && isNullConstant(RHS) && | |||
3966 | LHS.getOperand(0).getValueType() == Subtarget.getXLenVT()) { | |||
3967 | // If we're looking for eq 0 instead of ne 0, we need to invert the | |||
3968 | // condition. | |||
3969 | bool Invert = CCVal == ISD::SETEQ; | |||
3970 | CCVal = cast<CondCodeSDNode>(LHS.getOperand(2))->get(); | |||
3971 | if (Invert) | |||
3972 | CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType()); | |||
3973 | ||||
3974 | RHS = LHS.getOperand(1); | |||
3975 | LHS = LHS.getOperand(0); | |||
3976 | normaliseSetCC(LHS, RHS, CCVal); | |||
3977 | ||||
3978 | SDLoc DL(N); | |||
3979 | SDValue TargetCC = DAG.getConstant(CCVal, DL, Subtarget.getXLenVT()); | |||
3980 | return DAG.getNode( | |||
3981 | RISCVISD::SELECT_CC, DL, N->getValueType(0), | |||
3982 | {LHS, RHS, TargetCC, N->getOperand(3), N->getOperand(4)}); | |||
3983 | } | |||
3984 | ||||
3985 | // Fold (select_cc (xor X, Y), 0, eq/ne, trueV, falseV) -> | |||
3986 | // (select_cc X, Y, eq/ne, trueV, falseV) | |||
3987 | if (LHS.getOpcode() == ISD::XOR && isNullConstant(RHS)) | |||
3988 | return DAG.getNode(RISCVISD::SELECT_CC, SDLoc(N), N->getValueType(0), | |||
3989 | {LHS.getOperand(0), LHS.getOperand(1), | |||
3990 | N->getOperand(2), N->getOperand(3), | |||
3991 | N->getOperand(4)}); | |||
3992 | // (select_cc X, 1, setne, trueV, falseV) -> | |||
3993 | // (select_cc X, 0, seteq, trueV, falseV) if we can prove X is 0/1. | |||
3994 | // This can occur when legalizing some floating point comparisons. | |||
3995 | APInt Mask = APInt::getBitsSetFrom(LHS.getValueSizeInBits(), 1); | |||
3996 | if (isOneConstant(RHS) && DAG.MaskedValueIsZero(LHS, Mask)) { | |||
3997 | SDLoc DL(N); | |||
3998 | CCVal = ISD::getSetCCInverse(CCVal, LHS.getValueType()); | |||
3999 | SDValue TargetCC = DAG.getConstant(CCVal, DL, Subtarget.getXLenVT()); | |||
4000 | RHS = DAG.getConstant(0, DL, LHS.getValueType()); | |||
4001 | return DAG.getNode( | |||
4002 | RISCVISD::SELECT_CC, DL, N->getValueType(0), | |||
4003 | {LHS, RHS, TargetCC, N->getOperand(3), N->getOperand(4)}); | |||
4004 | } | |||
4005 | ||||
4006 | break; | |||
4007 | } | |||
4008 | case ISD::SETCC: { | |||
4009 | // (setcc X, 1, setne) -> (setcc X, 0, seteq) if we can prove X is 0/1. | |||
4010 | // Comparing with 0 may allow us to fold into bnez/beqz. | |||
4011 | SDValue LHS = N->getOperand(0); | |||
4012 | SDValue RHS = N->getOperand(1); | |||
4013 | if (LHS.getValueType().isScalableVector()) | |||
4014 | break; | |||
4015 | auto CC = cast<CondCodeSDNode>(N->getOperand(2))->get(); | |||
4016 | APInt Mask = APInt::getBitsSetFrom(LHS.getValueSizeInBits(), 1); | |||
4017 | if (isOneConstant(RHS) && ISD::isIntEqualitySetCC(CC) && | |||
4018 | DAG.MaskedValueIsZero(LHS, Mask)) { | |||
4019 | SDLoc DL(N); | |||
4020 | SDValue Zero = DAG.getConstant(0, DL, LHS.getValueType()); | |||
4021 | CC = ISD::getSetCCInverse(CC, LHS.getValueType()); | |||
4022 | return DAG.getSetCC(DL, N->getValueType(0), LHS, Zero, CC); | |||
4023 | } | |||
4024 | break; | |||
4025 | } | |||
4026 | case ISD::FCOPYSIGN: { | |||
4027 | EVT VT = N->getValueType(0); | |||
4028 | if (!VT.isVector()) | |||
4029 | break; | |||
4030 | // There is a form of VFSGNJ which injects the negated sign of its second | |||
4031 | // operand. Try and bubble any FNEG up after the extend/round to produce | |||
4032 | // this optimized pattern. Avoid modifying cases where FP_ROUND and | |||
4033 | // TRUNC=1. | |||
4034 | SDValue In2 = N->getOperand(1); | |||
4035 | // Avoid cases where the extend/round has multiple uses, as duplicating | |||
4036 | // those is typically more expensive than removing a fneg. | |||
4037 | if (!In2.hasOneUse()) | |||
4038 | break; | |||
4039 | if (In2.getOpcode() != ISD::FP_EXTEND && | |||
4040 | (In2.getOpcode() != ISD::FP_ROUND || In2.getConstantOperandVal(1) != 0)) | |||
4041 | break; | |||
4042 | In2 = In2.getOperand(0); | |||
4043 | if (In2.getOpcode() != ISD::FNEG) | |||
4044 | break; | |||
4045 | SDLoc DL(N); | |||
4046 | SDValue NewFPExtRound = DAG.getFPExtendOrRound(In2.getOperand(0), DL, VT); | |||
4047 | return DAG.getNode(ISD::FCOPYSIGN, DL, VT, N->getOperand(0), | |||
4048 | DAG.getNode(ISD::FNEG, DL, VT, NewFPExtRound)); | |||
4049 | } | |||
4050 | } | |||
4051 | ||||
4052 | return SDValue(); | |||
4053 | } | |||
4054 | ||||
4055 | bool RISCVTargetLowering::isDesirableToCommuteWithShift( | |||
4056 | const SDNode *N, CombineLevel Level) const { | |||
4057 | // The following folds are only desirable if `(OP _, c1 << c2)` can be | |||
4058 | // materialised in fewer instructions than `(OP _, c1)`: | |||
4059 | // | |||
4060 | // (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2) | |||
4061 | // (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2) | |||
4062 | SDValue N0 = N->getOperand(0); | |||
4063 | EVT Ty = N0.getValueType(); | |||
4064 | if (Ty.isScalarInteger() && | |||
4065 | (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR)) { | |||
4066 | auto *C1 = dyn_cast<ConstantSDNode>(N0->getOperand(1)); | |||
4067 | auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)); | |||
4068 | if (C1 && C2) { | |||
4069 | const APInt &C1Int = C1->getAPIntValue(); | |||
4070 | APInt ShiftedC1Int = C1Int << C2->getAPIntValue(); | |||
4071 | ||||
4072 | // We can materialise `c1 << c2` into an add immediate, so it's "free", | |||
4073 | // and the combine should happen, to potentially allow further combines | |||
4074 | // later. | |||
4075 | if (ShiftedC1Int.getMinSignedBits() <= 64 && | |||
4076 | isLegalAddImmediate(ShiftedC1Int.getSExtValue())) | |||
4077 | return true; | |||
4078 | ||||
4079 | // We can materialise `c1` in an add immediate, so it's "free", and the | |||
4080 | // combine should be prevented. | |||
4081 | if (C1Int.getMinSignedBits() <= 64 && | |||
4082 | isLegalAddImmediate(C1Int.getSExtValue())) | |||
4083 | return false; | |||
4084 | ||||
4085 | // Neither constant will fit into an immediate, so find materialisation | |||
4086 | // costs. | |||
4087 | int C1Cost = RISCVMatInt::getIntMatCost(C1Int, Ty.getSizeInBits(), | |||
4088 | Subtarget.is64Bit()); | |||
4089 | int ShiftedC1Cost = RISCVMatInt::getIntMatCost( | |||
4090 | ShiftedC1Int, Ty.getSizeInBits(), Subtarget.is64Bit()); | |||
4091 | ||||
4092 | // Materialising `c1` is cheaper than materialising `c1 << c2`, so the | |||
4093 | // combine should be prevented. | |||
4094 | if (C1Cost < ShiftedC1Cost) | |||
4095 | return false; | |||
4096 | } | |||
4097 | } | |||
4098 | return true; | |||
4099 | } | |||
4100 | ||||
4101 | bool RISCVTargetLowering::targetShrinkDemandedConstant( | |||
4102 | SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, | |||
4103 | TargetLoweringOpt &TLO) const { | |||
4104 | // Delay this optimization as late as possible. | |||
4105 | if (!TLO.LegalOps) | |||
4106 | return false; | |||
4107 | ||||
4108 | EVT VT = Op.getValueType(); | |||
4109 | if (VT.isVector()) | |||
4110 | return false; | |||
4111 | ||||
4112 | // Only handle AND for now. | |||
4113 | if (Op.getOpcode() != ISD::AND) | |||
4114 | return false; | |||
4115 | ||||
4116 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | |||
4117 | if (!C) | |||
4118 | return false; | |||
4119 | ||||
4120 | const APInt &Mask = C->getAPIntValue(); | |||
4121 | ||||
4122 | // Clear all non-demanded bits initially. | |||
4123 | APInt ShrunkMask = Mask & DemandedBits; | |||
4124 | ||||
4125 | // If the shrunk mask fits in sign extended 12 bits, let the target | |||
4126 | // independent code apply it. | |||
4127 | if (ShrunkMask.isSignedIntN(12)) | |||
4128 | return false; | |||
4129 | ||||
4130 | // Try to make a smaller immediate by setting undemanded bits. | |||
4131 | ||||
4132 | // We need to be able to make a negative number through a combination of mask | |||
4133 | // and undemanded bits. | |||
4134 | APInt ExpandedMask = Mask | ~DemandedBits; | |||
4135 | if (!ExpandedMask.isNegative()) | |||
4136 | return false; | |||
4137 | ||||
4138 | // What is the fewest number of bits we need to represent the negative number. | |||
4139 | unsigned MinSignedBits = ExpandedMask.getMinSignedBits(); | |||
4140 | ||||
4141 | // Try to make a 12 bit negative immediate. If that fails try to make a 32 | |||
4142 | // bit negative immediate unless the shrunk immediate already fits in 32 bits. | |||
4143 | APInt NewMask = ShrunkMask; | |||
4144 | if (MinSignedBits <= 12) | |||
4145 | NewMask.setBitsFrom(11); | |||
4146 | else if (MinSignedBits <= 32 && !ShrunkMask.isSignedIntN(32)) | |||
4147 | NewMask.setBitsFrom(31); | |||
4148 | else | |||
4149 | return false; | |||
4150 | ||||
4151 | // Sanity check that our new mask is a subset of the demanded mask. | |||
4152 | assert(NewMask.isSubsetOf(ExpandedMask))((NewMask.isSubsetOf(ExpandedMask)) ? static_cast<void> (0) : __assert_fail ("NewMask.isSubsetOf(ExpandedMask)", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4152, __PRETTY_FUNCTION__)); | |||
4153 | ||||
4154 | // If we aren't changing the mask, just return true to keep it and prevent | |||
4155 | // the caller from optimizing. | |||
4156 | if (NewMask == Mask) | |||
4157 | return true; | |||
4158 | ||||
4159 | // Replace the constant with the new mask. | |||
4160 | SDLoc DL(Op); | |||
4161 | SDValue NewC = TLO.DAG.getConstant(NewMask, DL, VT); | |||
4162 | SDValue NewOp = TLO.DAG.getNode(ISD::AND, DL, VT, Op.getOperand(0), NewC); | |||
4163 | return TLO.CombineTo(Op, NewOp); | |||
4164 | } | |||
4165 | ||||
4166 | void RISCVTargetLowering::computeKnownBitsForTargetNode(const SDValue Op, | |||
4167 | KnownBits &Known, | |||
4168 | const APInt &DemandedElts, | |||
4169 | const SelectionDAG &DAG, | |||
4170 | unsigned Depth) const { | |||
4171 | unsigned BitWidth = Known.getBitWidth(); | |||
4172 | unsigned Opc = Op.getOpcode(); | |||
4173 | assert((Opc >= ISD::BUILTIN_OP_END ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)) | |||
4174 | Opc == ISD::INTRINSIC_WO_CHAIN ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)) | |||
4175 | Opc == ISD::INTRINSIC_W_CHAIN ||(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)) | |||
4176 | Opc == ISD::INTRINSIC_VOID) &&(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)) | |||
4177 | "Should use MaskedValueIsZero if you don't know whether Op"(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)) | |||
4178 | " is a target node!")(((Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID ) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Opc >= ISD::BUILTIN_OP_END || Opc == ISD::INTRINSIC_WO_CHAIN || Opc == ISD::INTRINSIC_W_CHAIN || Opc == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4178, __PRETTY_FUNCTION__)); | |||
4179 | ||||
4180 | Known.resetAll(); | |||
4181 | switch (Opc) { | |||
4182 | default: break; | |||
4183 | case RISCVISD::REMUW: { | |||
4184 | KnownBits Known2; | |||
4185 | Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); | |||
4186 | Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); | |||
4187 | // We only care about the lower 32 bits. | |||
4188 | Known = KnownBits::urem(Known.trunc(32), Known2.trunc(32)); | |||
4189 | // Restore the original width by sign extending. | |||
4190 | Known = Known.sext(BitWidth); | |||
4191 | break; | |||
4192 | } | |||
4193 | case RISCVISD::DIVUW: { | |||
4194 | KnownBits Known2; | |||
4195 | Known = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); | |||
4196 | Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); | |||
4197 | // We only care about the lower 32 bits. | |||
4198 | Known = KnownBits::udiv(Known.trunc(32), Known2.trunc(32)); | |||
4199 | // Restore the original width by sign extending. | |||
4200 | Known = Known.sext(BitWidth); | |||
4201 | break; | |||
4202 | } | |||
4203 | case RISCVISD::READ_VLENB: | |||
4204 | // We assume VLENB is at least 8 bytes. | |||
4205 | // FIXME: The 1.0 draft spec defines minimum VLEN as 128 bits. | |||
4206 | Known.Zero.setLowBits(3); | |||
4207 | break; | |||
4208 | } | |||
4209 | } | |||
4210 | ||||
4211 | unsigned RISCVTargetLowering::ComputeNumSignBitsForTargetNode( | |||
4212 | SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG, | |||
4213 | unsigned Depth) const { | |||
4214 | switch (Op.getOpcode()) { | |||
4215 | default: | |||
4216 | break; | |||
4217 | case RISCVISD::SLLW: | |||
4218 | case RISCVISD::SRAW: | |||
4219 | case RISCVISD::SRLW: | |||
4220 | case RISCVISD::DIVW: | |||
4221 | case RISCVISD::DIVUW: | |||
4222 | case RISCVISD::REMUW: | |||
4223 | case RISCVISD::ROLW: | |||
4224 | case RISCVISD::RORW: | |||
4225 | case RISCVISD::GREVIW: | |||
4226 | case RISCVISD::GORCIW: | |||
4227 | case RISCVISD::FSLW: | |||
4228 | case RISCVISD::FSRW: | |||
4229 | // TODO: As the result is sign-extended, this is conservatively correct. A | |||
4230 | // more precise answer could be calculated for SRAW depending on known | |||
4231 | // bits in the shift amount. | |||
4232 | return 33; | |||
4233 | case RISCVISD::SHFLI: { | |||
4234 | // There is no SHFLIW, but a i64 SHFLI with bit 4 of the control word | |||
4235 | // cleared doesn't affect bit 31. The upper 32 bits will be shuffled, but | |||
4236 | // will stay within the upper 32 bits. If there were more than 32 sign bits | |||
4237 | // before there will be at least 33 sign bits after. | |||
4238 | if (Op.getValueType() == MVT::i64 && | |||
4239 | (Op.getConstantOperandVal(1) & 0x10) == 0) { | |||
4240 | unsigned Tmp = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1); | |||
4241 | if (Tmp > 32) | |||
4242 | return 33; | |||
4243 | } | |||
4244 | break; | |||
4245 | } | |||
4246 | case RISCVISD::VMV_X_S: | |||
4247 | // The number of sign bits of the scalar result is computed by obtaining the | |||
4248 | // element type of the input vector operand, subtracting its width from the | |||
4249 | // XLEN, and then adding one (sign bit within the element type). If the | |||
4250 | // element type is wider than XLen, the least-significant XLEN bits are | |||
4251 | // taken. | |||
4252 | if (Op.getOperand(0).getScalarValueSizeInBits() > Subtarget.getXLen()) | |||
4253 | return 1; | |||
4254 | return Subtarget.getXLen() - Op.getOperand(0).getScalarValueSizeInBits() + 1; | |||
4255 | } | |||
4256 | ||||
4257 | return 1; | |||
4258 | } | |||
4259 | ||||
4260 | static MachineBasicBlock *emitReadCycleWidePseudo(MachineInstr &MI, | |||
4261 | MachineBasicBlock *BB) { | |||
4262 | assert(MI.getOpcode() == RISCV::ReadCycleWide && "Unexpected instruction")((MI.getOpcode() == RISCV::ReadCycleWide && "Unexpected instruction" ) ? static_cast<void> (0) : __assert_fail ("MI.getOpcode() == RISCV::ReadCycleWide && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4262, __PRETTY_FUNCTION__)); | |||
4263 | ||||
4264 | // To read the 64-bit cycle CSR on a 32-bit target, we read the two halves. | |||
4265 | // Should the count have wrapped while it was being read, we need to try | |||
4266 | // again. | |||
4267 | // ... | |||
4268 | // read: | |||
4269 | // rdcycleh x3 # load high word of cycle | |||
4270 | // rdcycle x2 # load low word of cycle | |||
4271 | // rdcycleh x4 # load high word of cycle | |||
4272 | // bne x3, x4, read # check if high word reads match, otherwise try again | |||
4273 | // ... | |||
4274 | ||||
4275 | MachineFunction &MF = *BB->getParent(); | |||
4276 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); | |||
4277 | MachineFunction::iterator It = ++BB->getIterator(); | |||
4278 | ||||
4279 | MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVM_BB); | |||
4280 | MF.insert(It, LoopMBB); | |||
4281 | ||||
4282 | MachineBasicBlock *DoneMBB = MF.CreateMachineBasicBlock(LLVM_BB); | |||
4283 | MF.insert(It, DoneMBB); | |||
4284 | ||||
4285 | // Transfer the remainder of BB and its successor edges to DoneMBB. | |||
4286 | DoneMBB->splice(DoneMBB->begin(), BB, | |||
4287 | std::next(MachineBasicBlock::iterator(MI)), BB->end()); | |||
4288 | DoneMBB->transferSuccessorsAndUpdatePHIs(BB); | |||
4289 | ||||
4290 | BB->addSuccessor(LoopMBB); | |||
4291 | ||||
4292 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
4293 | Register ReadAgainReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass); | |||
4294 | Register LoReg = MI.getOperand(0).getReg(); | |||
4295 | Register HiReg = MI.getOperand(1).getReg(); | |||
4296 | DebugLoc DL = MI.getDebugLoc(); | |||
4297 | ||||
4298 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); | |||
4299 | BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), HiReg) | |||
4300 | .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding) | |||
4301 | .addReg(RISCV::X0); | |||
4302 | BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), LoReg) | |||
4303 | .addImm(RISCVSysReg::lookupSysRegByName("CYCLE")->Encoding) | |||
4304 | .addReg(RISCV::X0); | |||
4305 | BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), ReadAgainReg) | |||
4306 | .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding) | |||
4307 | .addReg(RISCV::X0); | |||
4308 | ||||
4309 | BuildMI(LoopMBB, DL, TII->get(RISCV::BNE)) | |||
4310 | .addReg(HiReg) | |||
4311 | .addReg(ReadAgainReg) | |||
4312 | .addMBB(LoopMBB); | |||
4313 | ||||
4314 | LoopMBB->addSuccessor(LoopMBB); | |||
4315 | LoopMBB->addSuccessor(DoneMBB); | |||
4316 | ||||
4317 | MI.eraseFromParent(); | |||
4318 | ||||
4319 | return DoneMBB; | |||
4320 | } | |||
4321 | ||||
4322 | static MachineBasicBlock *emitSplitF64Pseudo(MachineInstr &MI, | |||
4323 | MachineBasicBlock *BB) { | |||
4324 | assert(MI.getOpcode() == RISCV::SplitF64Pseudo && "Unexpected instruction")((MI.getOpcode() == RISCV::SplitF64Pseudo && "Unexpected instruction" ) ? static_cast<void> (0) : __assert_fail ("MI.getOpcode() == RISCV::SplitF64Pseudo && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4324, __PRETTY_FUNCTION__)); | |||
4325 | ||||
4326 | MachineFunction &MF = *BB->getParent(); | |||
4327 | DebugLoc DL = MI.getDebugLoc(); | |||
4328 | const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo(); | |||
4329 | const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo(); | |||
4330 | Register LoReg = MI.getOperand(0).getReg(); | |||
4331 | Register HiReg = MI.getOperand(1).getReg(); | |||
4332 | Register SrcReg = MI.getOperand(2).getReg(); | |||
4333 | const TargetRegisterClass *SrcRC = &RISCV::FPR64RegClass; | |||
4334 | int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF); | |||
4335 | ||||
4336 | TII.storeRegToStackSlot(*BB, MI, SrcReg, MI.getOperand(2).isKill(), FI, SrcRC, | |||
4337 | RI); | |||
4338 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI); | |||
4339 | MachineMemOperand *MMOLo = | |||
4340 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOLoad, 4, Align(8)); | |||
4341 | MachineMemOperand *MMOHi = MF.getMachineMemOperand( | |||
4342 | MPI.getWithOffset(4), MachineMemOperand::MOLoad, 4, Align(8)); | |||
4343 | BuildMI(*BB, MI, DL, TII.get(RISCV::LW), LoReg) | |||
4344 | .addFrameIndex(FI) | |||
4345 | .addImm(0) | |||
4346 | .addMemOperand(MMOLo); | |||
4347 | BuildMI(*BB, MI, DL, TII.get(RISCV::LW), HiReg) | |||
4348 | .addFrameIndex(FI) | |||
4349 | .addImm(4) | |||
4350 | .addMemOperand(MMOHi); | |||
4351 | MI.eraseFromParent(); // The pseudo instruction is gone now. | |||
4352 | return BB; | |||
4353 | } | |||
4354 | ||||
4355 | static MachineBasicBlock *emitBuildPairF64Pseudo(MachineInstr &MI, | |||
4356 | MachineBasicBlock *BB) { | |||
4357 | assert(MI.getOpcode() == RISCV::BuildPairF64Pseudo &&((MI.getOpcode() == RISCV::BuildPairF64Pseudo && "Unexpected instruction" ) ? static_cast<void> (0) : __assert_fail ("MI.getOpcode() == RISCV::BuildPairF64Pseudo && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4358, __PRETTY_FUNCTION__)) | |||
4358 | "Unexpected instruction")((MI.getOpcode() == RISCV::BuildPairF64Pseudo && "Unexpected instruction" ) ? static_cast<void> (0) : __assert_fail ("MI.getOpcode() == RISCV::BuildPairF64Pseudo && \"Unexpected instruction\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4358, __PRETTY_FUNCTION__)); | |||
4359 | ||||
4360 | MachineFunction &MF = *BB->getParent(); | |||
4361 | DebugLoc DL = MI.getDebugLoc(); | |||
4362 | const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo(); | |||
4363 | const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo(); | |||
4364 | Register DstReg = MI.getOperand(0).getReg(); | |||
4365 | Register LoReg = MI.getOperand(1).getReg(); | |||
4366 | Register HiReg = MI.getOperand(2).getReg(); | |||
4367 | const TargetRegisterClass *DstRC = &RISCV::FPR64RegClass; | |||
4368 | int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex(MF); | |||
4369 | ||||
4370 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI); | |||
4371 | MachineMemOperand *MMOLo = | |||
4372 | MF.getMachineMemOperand(MPI, MachineMemOperand::MOStore, 4, Align(8)); | |||
4373 | MachineMemOperand *MMOHi = MF.getMachineMemOperand( | |||
4374 | MPI.getWithOffset(4), MachineMemOperand::MOStore, 4, Align(8)); | |||
4375 | BuildMI(*BB, MI, DL, TII.get(RISCV::SW)) | |||
4376 | .addReg(LoReg, getKillRegState(MI.getOperand(1).isKill())) | |||
4377 | .addFrameIndex(FI) | |||
4378 | .addImm(0) | |||
4379 | .addMemOperand(MMOLo); | |||
4380 | BuildMI(*BB, MI, DL, TII.get(RISCV::SW)) | |||
4381 | .addReg(HiReg, getKillRegState(MI.getOperand(2).isKill())) | |||
4382 | .addFrameIndex(FI) | |||
4383 | .addImm(4) | |||
4384 | .addMemOperand(MMOHi); | |||
4385 | TII.loadRegFromStackSlot(*BB, MI, DstReg, FI, DstRC, RI); | |||
4386 | MI.eraseFromParent(); // The pseudo instruction is gone now. | |||
4387 | return BB; | |||
4388 | } | |||
4389 | ||||
4390 | static bool isSelectPseudo(MachineInstr &MI) { | |||
4391 | switch (MI.getOpcode()) { | |||
4392 | default: | |||
4393 | return false; | |||
4394 | case RISCV::Select_GPR_Using_CC_GPR: | |||
4395 | case RISCV::Select_FPR16_Using_CC_GPR: | |||
4396 | case RISCV::Select_FPR32_Using_CC_GPR: | |||
4397 | case RISCV::Select_FPR64_Using_CC_GPR: | |||
4398 | return true; | |||
4399 | } | |||
4400 | } | |||
4401 | ||||
4402 | static MachineBasicBlock *emitSelectPseudo(MachineInstr &MI, | |||
4403 | MachineBasicBlock *BB) { | |||
4404 | // To "insert" Select_* instructions, we actually have to insert the triangle | |||
4405 | // control-flow pattern. The incoming instructions know the destination vreg | |||
4406 | // to set, the condition code register to branch on, the true/false values to | |||
4407 | // select between, and the condcode to use to select the appropriate branch. | |||
4408 | // | |||
4409 | // We produce the following control flow: | |||
4410 | // HeadMBB | |||
4411 | // | \ | |||
4412 | // | IfFalseMBB | |||
4413 | // | / | |||
4414 | // TailMBB | |||
4415 | // | |||
4416 | // When we find a sequence of selects we attempt to optimize their emission | |||
4417 | // by sharing the control flow. Currently we only handle cases where we have | |||
4418 | // multiple selects with the exact same condition (same LHS, RHS and CC). | |||
4419 | // The selects may be interleaved with other instructions if the other | |||
4420 | // instructions meet some requirements we deem safe: | |||
4421 | // - They are debug instructions. Otherwise, | |||
4422 | // - They do not have side-effects, do not access memory and their inputs do | |||
4423 | // not depend on the results of the select pseudo-instructions. | |||
4424 | // The TrueV/FalseV operands of the selects cannot depend on the result of | |||
4425 | // previous selects in the sequence. | |||
4426 | // These conditions could be further relaxed. See the X86 target for a | |||
4427 | // related approach and more information. | |||
4428 | Register LHS = MI.getOperand(1).getReg(); | |||
4429 | Register RHS = MI.getOperand(2).getReg(); | |||
4430 | auto CC = static_cast<ISD::CondCode>(MI.getOperand(3).getImm()); | |||
4431 | ||||
4432 | SmallVector<MachineInstr *, 4> SelectDebugValues; | |||
4433 | SmallSet<Register, 4> SelectDests; | |||
4434 | SelectDests.insert(MI.getOperand(0).getReg()); | |||
4435 | ||||
4436 | MachineInstr *LastSelectPseudo = &MI; | |||
4437 | ||||
4438 | for (auto E = BB->end(), SequenceMBBI = MachineBasicBlock::iterator(MI); | |||
4439 | SequenceMBBI != E; ++SequenceMBBI) { | |||
4440 | if (SequenceMBBI->isDebugInstr()) | |||
4441 | continue; | |||
4442 | else if (isSelectPseudo(*SequenceMBBI)) { | |||
4443 | if (SequenceMBBI->getOperand(1).getReg() != LHS || | |||
4444 | SequenceMBBI->getOperand(2).getReg() != RHS || | |||
4445 | SequenceMBBI->getOperand(3).getImm() != CC || | |||
4446 | SelectDests.count(SequenceMBBI->getOperand(4).getReg()) || | |||
4447 | SelectDests.count(SequenceMBBI->getOperand(5).getReg())) | |||
4448 | break; | |||
4449 | LastSelectPseudo = &*SequenceMBBI; | |||
4450 | SequenceMBBI->collectDebugValues(SelectDebugValues); | |||
4451 | SelectDests.insert(SequenceMBBI->getOperand(0).getReg()); | |||
4452 | } else { | |||
4453 | if (SequenceMBBI->hasUnmodeledSideEffects() || | |||
4454 | SequenceMBBI->mayLoadOrStore()) | |||
4455 | break; | |||
4456 | if (llvm::any_of(SequenceMBBI->operands(), [&](MachineOperand &MO) { | |||
4457 | return MO.isReg() && MO.isUse() && SelectDests.count(MO.getReg()); | |||
4458 | })) | |||
4459 | break; | |||
4460 | } | |||
4461 | } | |||
4462 | ||||
4463 | const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo(); | |||
4464 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); | |||
4465 | DebugLoc DL = MI.getDebugLoc(); | |||
4466 | MachineFunction::iterator I = ++BB->getIterator(); | |||
4467 | ||||
4468 | MachineBasicBlock *HeadMBB = BB; | |||
4469 | MachineFunction *F = BB->getParent(); | |||
4470 | MachineBasicBlock *TailMBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
4471 | MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(LLVM_BB); | |||
4472 | ||||
4473 | F->insert(I, IfFalseMBB); | |||
4474 | F->insert(I, TailMBB); | |||
4475 | ||||
4476 | // Transfer debug instructions associated with the selects to TailMBB. | |||
4477 | for (MachineInstr *DebugInstr : SelectDebugValues) { | |||
4478 | TailMBB->push_back(DebugInstr->removeFromParent()); | |||
4479 | } | |||
4480 | ||||
4481 | // Move all instructions after the sequence to TailMBB. | |||
4482 | TailMBB->splice(TailMBB->end(), HeadMBB, | |||
4483 | std::next(LastSelectPseudo->getIterator()), HeadMBB->end()); | |||
4484 | // Update machine-CFG edges by transferring all successors of the current | |||
4485 | // block to the new block which will contain the Phi nodes for the selects. | |||
4486 | TailMBB->transferSuccessorsAndUpdatePHIs(HeadMBB); | |||
4487 | // Set the successors for HeadMBB. | |||
4488 | HeadMBB->addSuccessor(IfFalseMBB); | |||
4489 | HeadMBB->addSuccessor(TailMBB); | |||
4490 | ||||
4491 | // Insert appropriate branch. | |||
4492 | unsigned Opcode = getBranchOpcodeForIntCondCode(CC); | |||
4493 | ||||
4494 | BuildMI(HeadMBB, DL, TII.get(Opcode)) | |||
4495 | .addReg(LHS) | |||
4496 | .addReg(RHS) | |||
4497 | .addMBB(TailMBB); | |||
4498 | ||||
4499 | // IfFalseMBB just falls through to TailMBB. | |||
4500 | IfFalseMBB->addSuccessor(TailMBB); | |||
4501 | ||||
4502 | // Create PHIs for all of the select pseudo-instructions. | |||
4503 | auto SelectMBBI = MI.getIterator(); | |||
4504 | auto SelectEnd = std::next(LastSelectPseudo->getIterator()); | |||
4505 | auto InsertionPoint = TailMBB->begin(); | |||
4506 | while (SelectMBBI != SelectEnd) { | |||
4507 | auto Next = std::next(SelectMBBI); | |||
4508 | if (isSelectPseudo(*SelectMBBI)) { | |||
4509 | // %Result = phi [ %TrueValue, HeadMBB ], [ %FalseValue, IfFalseMBB ] | |||
4510 | BuildMI(*TailMBB, InsertionPoint, SelectMBBI->getDebugLoc(), | |||
4511 | TII.get(RISCV::PHI), SelectMBBI->getOperand(0).getReg()) | |||
4512 | .addReg(SelectMBBI->getOperand(4).getReg()) | |||
4513 | .addMBB(HeadMBB) | |||
4514 | .addReg(SelectMBBI->getOperand(5).getReg()) | |||
4515 | .addMBB(IfFalseMBB); | |||
4516 | SelectMBBI->eraseFromParent(); | |||
4517 | } | |||
4518 | SelectMBBI = Next; | |||
4519 | } | |||
4520 | ||||
4521 | F->getProperties().reset(MachineFunctionProperties::Property::NoPHIs); | |||
4522 | return TailMBB; | |||
4523 | } | |||
4524 | ||||
4525 | static MachineBasicBlock *addVSetVL(MachineInstr &MI, MachineBasicBlock *BB, | |||
4526 | int VLIndex, unsigned SEWIndex, | |||
4527 | RISCVVLMUL VLMul, bool ForceTailAgnostic) { | |||
4528 | MachineFunction &MF = *BB->getParent(); | |||
4529 | DebugLoc DL = MI.getDebugLoc(); | |||
4530 | const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo(); | |||
4531 | ||||
4532 | unsigned SEW = MI.getOperand(SEWIndex).getImm(); | |||
4533 | assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW")((RISCVVType::isValidSEW(SEW) && "Unexpected SEW") ? static_cast <void> (0) : __assert_fail ("RISCVVType::isValidSEW(SEW) && \"Unexpected SEW\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4533, __PRETTY_FUNCTION__)); | |||
4534 | RISCVVSEW ElementWidth = static_cast<RISCVVSEW>(Log2_32(SEW / 8)); | |||
4535 | ||||
4536 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
4537 | ||||
4538 | auto BuildVSETVLI = [&]() { | |||
4539 | if (VLIndex >= 0) { | |||
4540 | Register DestReg = MRI.createVirtualRegister(&RISCV::GPRRegClass); | |||
4541 | Register VLReg = MI.getOperand(VLIndex).getReg(); | |||
4542 | ||||
4543 | // VL might be a compile time constant, but isel would have to put it | |||
4544 | // in a register. See if VL comes from an ADDI X0, imm. | |||
4545 | if (VLReg.isVirtual()) { | |||
4546 | MachineInstr *Def = MRI.getVRegDef(VLReg); | |||
4547 | if (Def && Def->getOpcode() == RISCV::ADDI && | |||
4548 | Def->getOperand(1).getReg() == RISCV::X0 && | |||
4549 | Def->getOperand(2).isImm()) { | |||
4550 | uint64_t Imm = Def->getOperand(2).getImm(); | |||
4551 | // VSETIVLI allows a 5-bit zero extended immediate. | |||
4552 | if (isUInt<5>(Imm)) | |||
4553 | return BuildMI(*BB, MI, DL, TII.get(RISCV::PseudoVSETIVLI)) | |||
4554 | .addReg(DestReg, RegState::Define | RegState::Dead) | |||
4555 | .addImm(Imm); | |||
4556 | } | |||
4557 | } | |||
4558 | ||||
4559 | return BuildMI(*BB, MI, DL, TII.get(RISCV::PseudoVSETVLI)) | |||
4560 | .addReg(DestReg, RegState::Define | RegState::Dead) | |||
4561 | .addReg(VLReg); | |||
4562 | } | |||
4563 | ||||
4564 | // With no VL operator in the pseudo, do not modify VL (rd = X0, rs1 = X0). | |||
4565 | return BuildMI(*BB, MI, DL, TII.get(RISCV::PseudoVSETVLI)) | |||
4566 | .addReg(RISCV::X0, RegState::Define | RegState::Dead) | |||
4567 | .addReg(RISCV::X0, RegState::Kill); | |||
4568 | }; | |||
4569 | ||||
4570 | MachineInstrBuilder MIB = BuildVSETVLI(); | |||
4571 | ||||
4572 | // Default to tail agnostic unless the destination is tied to a source. In | |||
4573 | // that case the user would have some control over the tail values. The tail | |||
4574 | // policy is also ignored on instructions that only update element 0 like | |||
4575 | // vmv.s.x or reductions so use agnostic there to match the common case. | |||
4576 | // FIXME: This is conservatively correct, but we might want to detect that | |||
4577 | // the input is undefined. | |||
4578 | bool TailAgnostic = true; | |||
4579 | unsigned UseOpIdx; | |||
4580 | if (!ForceTailAgnostic && MI.isRegTiedToUseOperand(0, &UseOpIdx)) { | |||
4581 | TailAgnostic = false; | |||
4582 | // If the tied operand is an IMPLICIT_DEF we can keep TailAgnostic. | |||
4583 | const MachineOperand &UseMO = MI.getOperand(UseOpIdx); | |||
4584 | MachineInstr *UseMI = MRI.getVRegDef(UseMO.getReg()); | |||
4585 | if (UseMI && UseMI->isImplicitDef()) | |||
4586 | TailAgnostic = true; | |||
4587 | } | |||
4588 | ||||
4589 | // For simplicity we reuse the vtype representation here. | |||
4590 | MIB.addImm(RISCVVType::encodeVTYPE(VLMul, ElementWidth, | |||
4591 | /*TailAgnostic*/ TailAgnostic, | |||
4592 | /*MaskAgnostic*/ false)); | |||
4593 | ||||
4594 | // Remove (now) redundant operands from pseudo | |||
4595 | if (VLIndex >= 0) { | |||
4596 | MI.getOperand(VLIndex).setReg(RISCV::NoRegister); | |||
4597 | MI.getOperand(VLIndex).setIsKill(false); | |||
4598 | } | |||
4599 | ||||
4600 | return BB; | |||
4601 | } | |||
4602 | ||||
4603 | MachineBasicBlock * | |||
4604 | RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, | |||
4605 | MachineBasicBlock *BB) const { | |||
4606 | uint64_t TSFlags = MI.getDesc().TSFlags; | |||
4607 | ||||
4608 | if (TSFlags & RISCVII::HasSEWOpMask) { | |||
4609 | unsigned NumOperands = MI.getNumExplicitOperands(); | |||
4610 | int VLIndex = (TSFlags & RISCVII::HasVLOpMask) ? NumOperands - 2 : -1; | |||
4611 | unsigned SEWIndex = NumOperands - 1; | |||
4612 | bool ForceTailAgnostic = TSFlags & RISCVII::ForceTailAgnosticMask; | |||
4613 | ||||
4614 | RISCVVLMUL VLMul = static_cast<RISCVVLMUL>((TSFlags & RISCVII::VLMulMask) >> | |||
4615 | RISCVII::VLMulShift); | |||
4616 | return addVSetVL(MI, BB, VLIndex, SEWIndex, VLMul, ForceTailAgnostic); | |||
4617 | } | |||
4618 | ||||
4619 | switch (MI.getOpcode()) { | |||
4620 | default: | |||
4621 | llvm_unreachable("Unexpected instr type to insert")::llvm::llvm_unreachable_internal("Unexpected instr type to insert" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4621); | |||
4622 | case RISCV::ReadCycleWide: | |||
4623 | assert(!Subtarget.is64Bit() &&((!Subtarget.is64Bit() && "ReadCycleWrite is only to be used on riscv32" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && \"ReadCycleWrite is only to be used on riscv32\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4624, __PRETTY_FUNCTION__)) | |||
4624 | "ReadCycleWrite is only to be used on riscv32")((!Subtarget.is64Bit() && "ReadCycleWrite is only to be used on riscv32" ) ? static_cast<void> (0) : __assert_fail ("!Subtarget.is64Bit() && \"ReadCycleWrite is only to be used on riscv32\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4624, __PRETTY_FUNCTION__)); | |||
4625 | return emitReadCycleWidePseudo(MI, BB); | |||
4626 | case RISCV::Select_GPR_Using_CC_GPR: | |||
4627 | case RISCV::Select_FPR16_Using_CC_GPR: | |||
4628 | case RISCV::Select_FPR32_Using_CC_GPR: | |||
4629 | case RISCV::Select_FPR64_Using_CC_GPR: | |||
4630 | return emitSelectPseudo(MI, BB); | |||
4631 | case RISCV::BuildPairF64Pseudo: | |||
4632 | return emitBuildPairF64Pseudo(MI, BB); | |||
4633 | case RISCV::SplitF64Pseudo: | |||
4634 | return emitSplitF64Pseudo(MI, BB); | |||
4635 | } | |||
4636 | } | |||
4637 | ||||
4638 | // Calling Convention Implementation. | |||
4639 | // The expectations for frontend ABI lowering vary from target to target. | |||
4640 | // Ideally, an LLVM frontend would be able to avoid worrying about many ABI | |||
4641 | // details, but this is a longer term goal. For now, we simply try to keep the | |||
4642 | // role of the frontend as simple and well-defined as possible. The rules can | |||
4643 | // be summarised as: | |||
4644 | // * Never split up large scalar arguments. We handle them here. | |||
4645 | // * If a hardfloat calling convention is being used, and the struct may be | |||
4646 | // passed in a pair of registers (fp+fp, int+fp), and both registers are | |||
4647 | // available, then pass as two separate arguments. If either the GPRs or FPRs | |||
4648 | // are exhausted, then pass according to the rule below. | |||
4649 | // * If a struct could never be passed in registers or directly in a stack | |||
4650 | // slot (as it is larger than 2*XLEN and the floating point rules don't | |||
4651 | // apply), then pass it using a pointer with the byval attribute. | |||
4652 | // * If a struct is less than 2*XLEN, then coerce to either a two-element | |||
4653 | // word-sized array or a 2*XLEN scalar (depending on alignment). | |||
4654 | // * The frontend can determine whether a struct is returned by reference or | |||
4655 | // not based on its size and fields. If it will be returned by reference, the | |||
4656 | // frontend must modify the prototype so a pointer with the sret annotation is | |||
4657 | // passed as the first argument. This is not necessary for large scalar | |||
4658 | // returns. | |||
4659 | // * Struct return values and varargs should be coerced to structs containing | |||
4660 | // register-size fields in the same situations they would be for fixed | |||
4661 | // arguments. | |||
4662 | ||||
4663 | static const MCPhysReg ArgGPRs[] = { | |||
4664 | RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13, | |||
4665 | RISCV::X14, RISCV::X15, RISCV::X16, RISCV::X17 | |||
4666 | }; | |||
4667 | static const MCPhysReg ArgFPR16s[] = { | |||
4668 | RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H, | |||
4669 | RISCV::F14_H, RISCV::F15_H, RISCV::F16_H, RISCV::F17_H | |||
4670 | }; | |||
4671 | static const MCPhysReg ArgFPR32s[] = { | |||
4672 | RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F, | |||
4673 | RISCV::F14_F, RISCV::F15_F, RISCV::F16_F, RISCV::F17_F | |||
4674 | }; | |||
4675 | static const MCPhysReg ArgFPR64s[] = { | |||
4676 | RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D, | |||
4677 | RISCV::F14_D, RISCV::F15_D, RISCV::F16_D, RISCV::F17_D | |||
4678 | }; | |||
4679 | // This is an interim calling convention and it may be changed in the future. | |||
4680 | static const MCPhysReg ArgVRs[] = { | |||
4681 | RISCV::V8, RISCV::V9, RISCV::V10, RISCV::V11, RISCV::V12, RISCV::V13, | |||
4682 | RISCV::V14, RISCV::V15, RISCV::V16, RISCV::V17, RISCV::V18, RISCV::V19, | |||
4683 | RISCV::V20, RISCV::V21, RISCV::V22, RISCV::V23}; | |||
4684 | static const MCPhysReg ArgVRM2s[] = {RISCV::V8M2, RISCV::V10M2, RISCV::V12M2, | |||
4685 | RISCV::V14M2, RISCV::V16M2, RISCV::V18M2, | |||
4686 | RISCV::V20M2, RISCV::V22M2}; | |||
4687 | static const MCPhysReg ArgVRM4s[] = {RISCV::V8M4, RISCV::V12M4, RISCV::V16M4, | |||
4688 | RISCV::V20M4}; | |||
4689 | static const MCPhysReg ArgVRM8s[] = {RISCV::V8M8, RISCV::V16M8}; | |||
4690 | ||||
4691 | // Pass a 2*XLEN argument that has been split into two XLEN values through | |||
4692 | // registers or the stack as necessary. | |||
4693 | static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1, | |||
4694 | ISD::ArgFlagsTy ArgFlags1, unsigned ValNo2, | |||
4695 | MVT ValVT2, MVT LocVT2, | |||
4696 | ISD::ArgFlagsTy ArgFlags2) { | |||
4697 | unsigned XLenInBytes = XLen / 8; | |||
4698 | if (Register Reg = State.AllocateReg(ArgGPRs)) { | |||
4699 | // At least one half can be passed via register. | |||
4700 | State.addLoc(CCValAssign::getReg(VA1.getValNo(), VA1.getValVT(), Reg, | |||
4701 | VA1.getLocVT(), CCValAssign::Full)); | |||
4702 | } else { | |||
4703 | // Both halves must be passed on the stack, with proper alignment. | |||
4704 | Align StackAlign = | |||
4705 | std::max(Align(XLenInBytes), ArgFlags1.getNonZeroOrigAlign()); | |||
4706 | State.addLoc( | |||
4707 | CCValAssign::getMem(VA1.getValNo(), VA1.getValVT(), | |||
4708 | State.AllocateStack(XLenInBytes, StackAlign), | |||
4709 | VA1.getLocVT(), CCValAssign::Full)); | |||
4710 | State.addLoc(CCValAssign::getMem( | |||
4711 | ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)), | |||
4712 | LocVT2, CCValAssign::Full)); | |||
4713 | return false; | |||
4714 | } | |||
4715 | ||||
4716 | if (Register Reg = State.AllocateReg(ArgGPRs)) { | |||
4717 | // The second half can also be passed via register. | |||
4718 | State.addLoc( | |||
4719 | CCValAssign::getReg(ValNo2, ValVT2, Reg, LocVT2, CCValAssign::Full)); | |||
4720 | } else { | |||
4721 | // The second half is passed via the stack, without additional alignment. | |||
4722 | State.addLoc(CCValAssign::getMem( | |||
4723 | ValNo2, ValVT2, State.AllocateStack(XLenInBytes, Align(XLenInBytes)), | |||
4724 | LocVT2, CCValAssign::Full)); | |||
4725 | } | |||
4726 | ||||
4727 | return false; | |||
4728 | } | |||
4729 | ||||
4730 | // Implements the RISC-V calling convention. Returns true upon failure. | |||
4731 | static bool CC_RISCV(const DataLayout &DL, RISCVABI::ABI ABI, unsigned ValNo, | |||
4732 | MVT ValVT, MVT LocVT, CCValAssign::LocInfo LocInfo, | |||
4733 | ISD::ArgFlagsTy ArgFlags, CCState &State, bool IsFixed, | |||
4734 | bool IsRet, Type *OrigTy, const RISCVTargetLowering &TLI, | |||
4735 | Optional<unsigned> FirstMaskArgument) { | |||
4736 | unsigned XLen = DL.getLargestLegalIntTypeSizeInBits(); | |||
4737 | assert(XLen == 32 || XLen == 64)((XLen == 32 || XLen == 64) ? static_cast<void> (0) : __assert_fail ("XLen == 32 || XLen == 64", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4737, __PRETTY_FUNCTION__)); | |||
4738 | MVT XLenVT = XLen == 32 ? MVT::i32 : MVT::i64; | |||
4739 | ||||
4740 | // Any return value split in to more than two values can't be returned | |||
4741 | // directly. | |||
4742 | if (IsRet && ValNo > 1) | |||
4743 | return true; | |||
4744 | ||||
4745 | // UseGPRForF16_F32 if targeting one of the soft-float ABIs, if passing a | |||
4746 | // variadic argument, or if no F16/F32 argument registers are available. | |||
4747 | bool UseGPRForF16_F32 = true; | |||
4748 | // UseGPRForF64 if targeting soft-float ABIs or an FLEN=32 ABI, if passing a | |||
4749 | // variadic argument, or if no F64 argument registers are available. | |||
4750 | bool UseGPRForF64 = true; | |||
4751 | ||||
4752 | switch (ABI) { | |||
4753 | default: | |||
4754 | llvm_unreachable("Unexpected ABI")::llvm::llvm_unreachable_internal("Unexpected ABI", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4754); | |||
4755 | case RISCVABI::ABI_ILP32: | |||
4756 | case RISCVABI::ABI_LP64: | |||
4757 | break; | |||
4758 | case RISCVABI::ABI_ILP32F: | |||
4759 | case RISCVABI::ABI_LP64F: | |||
4760 | UseGPRForF16_F32 = !IsFixed; | |||
4761 | break; | |||
4762 | case RISCVABI::ABI_ILP32D: | |||
4763 | case RISCVABI::ABI_LP64D: | |||
4764 | UseGPRForF16_F32 = !IsFixed; | |||
4765 | UseGPRForF64 = !IsFixed; | |||
4766 | break; | |||
4767 | } | |||
4768 | ||||
4769 | // FPR16, FPR32, and FPR64 alias each other. | |||
4770 | if (State.getFirstUnallocated(ArgFPR32s) == array_lengthof(ArgFPR32s)) { | |||
4771 | UseGPRForF16_F32 = true; | |||
4772 | UseGPRForF64 = true; | |||
4773 | } | |||
4774 | ||||
4775 | // From this point on, rely on UseGPRForF16_F32, UseGPRForF64 and | |||
4776 | // similar local variables rather than directly checking against the target | |||
4777 | // ABI. | |||
4778 | ||||
4779 | if (UseGPRForF16_F32 && (ValVT == MVT::f16 || ValVT == MVT::f32)) { | |||
4780 | LocVT = XLenVT; | |||
4781 | LocInfo = CCValAssign::BCvt; | |||
4782 | } else if (UseGPRForF64 && XLen == 64 && ValVT == MVT::f64) { | |||
4783 | LocVT = MVT::i64; | |||
4784 | LocInfo = CCValAssign::BCvt; | |||
4785 | } | |||
4786 | ||||
4787 | // If this is a variadic argument, the RISC-V calling convention requires | |||
4788 | // that it is assigned an 'even' or 'aligned' register if it has 8-byte | |||
4789 | // alignment (RV32) or 16-byte alignment (RV64). An aligned register should | |||
4790 | // be used regardless of whether the original argument was split during | |||
4791 | // legalisation or not. The argument will not be passed by registers if the | |||
4792 | // original type is larger than 2*XLEN, so the register alignment rule does | |||
4793 | // not apply. | |||
4794 | unsigned TwoXLenInBytes = (2 * XLen) / 8; | |||
4795 | if (!IsFixed && ArgFlags.getNonZeroOrigAlign() == TwoXLenInBytes && | |||
4796 | DL.getTypeAllocSize(OrigTy) == TwoXLenInBytes) { | |||
4797 | unsigned RegIdx = State.getFirstUnallocated(ArgGPRs); | |||
4798 | // Skip 'odd' register if necessary. | |||
4799 | if (RegIdx != array_lengthof(ArgGPRs) && RegIdx % 2 == 1) | |||
4800 | State.AllocateReg(ArgGPRs); | |||
4801 | } | |||
4802 | ||||
4803 | SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs(); | |||
4804 | SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags = | |||
4805 | State.getPendingArgFlags(); | |||
4806 | ||||
4807 | assert(PendingLocs.size() == PendingArgFlags.size() &&((PendingLocs.size() == PendingArgFlags.size() && "PendingLocs and PendingArgFlags out of sync" ) ? static_cast<void> (0) : __assert_fail ("PendingLocs.size() == PendingArgFlags.size() && \"PendingLocs and PendingArgFlags out of sync\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4808, __PRETTY_FUNCTION__)) | |||
4808 | "PendingLocs and PendingArgFlags out of sync")((PendingLocs.size() == PendingArgFlags.size() && "PendingLocs and PendingArgFlags out of sync" ) ? static_cast<void> (0) : __assert_fail ("PendingLocs.size() == PendingArgFlags.size() && \"PendingLocs and PendingArgFlags out of sync\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4808, __PRETTY_FUNCTION__)); | |||
4809 | ||||
4810 | // Handle passing f64 on RV32D with a soft float ABI or when floating point | |||
4811 | // registers are exhausted. | |||
4812 | if (UseGPRForF64 && XLen == 32 && ValVT == MVT::f64) { | |||
4813 | assert(!ArgFlags.isSplit() && PendingLocs.empty() &&((!ArgFlags.isSplit() && PendingLocs.empty() && "Can't lower f64 if it is split") ? static_cast<void> ( 0) : __assert_fail ("!ArgFlags.isSplit() && PendingLocs.empty() && \"Can't lower f64 if it is split\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4814, __PRETTY_FUNCTION__)) | |||
4814 | "Can't lower f64 if it is split")((!ArgFlags.isSplit() && PendingLocs.empty() && "Can't lower f64 if it is split") ? static_cast<void> ( 0) : __assert_fail ("!ArgFlags.isSplit() && PendingLocs.empty() && \"Can't lower f64 if it is split\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4814, __PRETTY_FUNCTION__)); | |||
4815 | // Depending on available argument GPRS, f64 may be passed in a pair of | |||
4816 | // GPRs, split between a GPR and the stack, or passed completely on the | |||
4817 | // stack. LowerCall/LowerFormalArguments/LowerReturn must recognise these | |||
4818 | // cases. | |||
4819 | Register Reg = State.AllocateReg(ArgGPRs); | |||
4820 | LocVT = MVT::i32; | |||
4821 | if (!Reg) { | |||
4822 | unsigned StackOffset = State.AllocateStack(8, Align(8)); | |||
4823 | State.addLoc( | |||
4824 | CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo)); | |||
4825 | return false; | |||
4826 | } | |||
4827 | if (!State.AllocateReg(ArgGPRs)) | |||
4828 | State.AllocateStack(4, Align(4)); | |||
4829 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
4830 | return false; | |||
4831 | } | |||
4832 | ||||
4833 | // Split arguments might be passed indirectly, so keep track of the pending | |||
4834 | // values. | |||
4835 | if (ArgFlags.isSplit() || !PendingLocs.empty()) { | |||
4836 | LocVT = XLenVT; | |||
4837 | LocInfo = CCValAssign::Indirect; | |||
4838 | PendingLocs.push_back( | |||
4839 | CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo)); | |||
4840 | PendingArgFlags.push_back(ArgFlags); | |||
4841 | if (!ArgFlags.isSplitEnd()) { | |||
4842 | return false; | |||
4843 | } | |||
4844 | } | |||
4845 | ||||
4846 | // If the split argument only had two elements, it should be passed directly | |||
4847 | // in registers or on the stack. | |||
4848 | if (ArgFlags.isSplitEnd() && PendingLocs.size() <= 2) { | |||
4849 | assert(PendingLocs.size() == 2 && "Unexpected PendingLocs.size()")((PendingLocs.size() == 2 && "Unexpected PendingLocs.size()" ) ? static_cast<void> (0) : __assert_fail ("PendingLocs.size() == 2 && \"Unexpected PendingLocs.size()\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4849, __PRETTY_FUNCTION__)); | |||
4850 | // Apply the normal calling convention rules to the first half of the | |||
4851 | // split argument. | |||
4852 | CCValAssign VA = PendingLocs[0]; | |||
4853 | ISD::ArgFlagsTy AF = PendingArgFlags[0]; | |||
4854 | PendingLocs.clear(); | |||
4855 | PendingArgFlags.clear(); | |||
4856 | return CC_RISCVAssign2XLen(XLen, State, VA, AF, ValNo, ValVT, LocVT, | |||
4857 | ArgFlags); | |||
4858 | } | |||
4859 | ||||
4860 | // Allocate to a register if possible, or else a stack slot. | |||
4861 | Register Reg; | |||
4862 | if (ValVT == MVT::f16 && !UseGPRForF16_F32) | |||
4863 | Reg = State.AllocateReg(ArgFPR16s); | |||
4864 | else if (ValVT == MVT::f32 && !UseGPRForF16_F32) | |||
4865 | Reg = State.AllocateReg(ArgFPR32s); | |||
4866 | else if (ValVT == MVT::f64 && !UseGPRForF64) | |||
4867 | Reg = State.AllocateReg(ArgFPR64s); | |||
4868 | else if (ValVT.isScalableVector()) { | |||
4869 | const TargetRegisterClass *RC = TLI.getRegClassFor(ValVT); | |||
4870 | if (RC == &RISCV::VRRegClass) { | |||
4871 | // Assign the first mask argument to V0. | |||
4872 | // This is an interim calling convention and it may be changed in the | |||
4873 | // future. | |||
4874 | if (FirstMaskArgument.hasValue() && | |||
4875 | ValNo == FirstMaskArgument.getValue()) { | |||
4876 | Reg = State.AllocateReg(RISCV::V0); | |||
4877 | } else { | |||
4878 | Reg = State.AllocateReg(ArgVRs); | |||
4879 | } | |||
4880 | } else if (RC == &RISCV::VRM2RegClass) { | |||
4881 | Reg = State.AllocateReg(ArgVRM2s); | |||
4882 | } else if (RC == &RISCV::VRM4RegClass) { | |||
4883 | Reg = State.AllocateReg(ArgVRM4s); | |||
4884 | } else if (RC == &RISCV::VRM8RegClass) { | |||
4885 | Reg = State.AllocateReg(ArgVRM8s); | |||
4886 | } else { | |||
4887 | llvm_unreachable("Unhandled class register for ValueType")::llvm::llvm_unreachable_internal("Unhandled class register for ValueType" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4887); | |||
4888 | } | |||
4889 | if (!Reg) { | |||
4890 | LocInfo = CCValAssign::Indirect; | |||
4891 | // Try using a GPR to pass the address | |||
4892 | Reg = State.AllocateReg(ArgGPRs); | |||
4893 | LocVT = XLenVT; | |||
4894 | } | |||
4895 | } else | |||
4896 | Reg = State.AllocateReg(ArgGPRs); | |||
4897 | unsigned StackOffset = | |||
4898 | Reg ? 0 : State.AllocateStack(XLen / 8, Align(XLen / 8)); | |||
4899 | ||||
4900 | // If we reach this point and PendingLocs is non-empty, we must be at the | |||
4901 | // end of a split argument that must be passed indirectly. | |||
4902 | if (!PendingLocs.empty()) { | |||
4903 | assert(ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()")((ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()" ) ? static_cast<void> (0) : __assert_fail ("ArgFlags.isSplitEnd() && \"Expected ArgFlags.isSplitEnd()\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4903, __PRETTY_FUNCTION__)); | |||
4904 | assert(PendingLocs.size() > 2 && "Unexpected PendingLocs.size()")((PendingLocs.size() > 2 && "Unexpected PendingLocs.size()" ) ? static_cast<void> (0) : __assert_fail ("PendingLocs.size() > 2 && \"Unexpected PendingLocs.size()\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4904, __PRETTY_FUNCTION__)); | |||
4905 | ||||
4906 | for (auto &It : PendingLocs) { | |||
4907 | if (Reg) | |||
4908 | It.convertToReg(Reg); | |||
4909 | else | |||
4910 | It.convertToMem(StackOffset); | |||
4911 | State.addLoc(It); | |||
4912 | } | |||
4913 | PendingLocs.clear(); | |||
4914 | PendingArgFlags.clear(); | |||
4915 | return false; | |||
4916 | } | |||
4917 | ||||
4918 | assert((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT ||(((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || ( TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector ())) && "Expected an XLenVT or scalable vector types at this stage" ) ? static_cast<void> (0) : __assert_fail ("(!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || (TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector())) && \"Expected an XLenVT or scalable vector types at this stage\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4920, __PRETTY_FUNCTION__)) | |||
4919 | (TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector())) &&(((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || ( TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector ())) && "Expected an XLenVT or scalable vector types at this stage" ) ? static_cast<void> (0) : __assert_fail ("(!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || (TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector())) && \"Expected an XLenVT or scalable vector types at this stage\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4920, __PRETTY_FUNCTION__)) | |||
4920 | "Expected an XLenVT or scalable vector types at this stage")(((!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || ( TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector ())) && "Expected an XLenVT or scalable vector types at this stage" ) ? static_cast<void> (0) : __assert_fail ("(!UseGPRForF16_F32 || !UseGPRForF64 || LocVT == XLenVT || (TLI.getSubtarget().hasStdExtV() && ValVT.isScalableVector())) && \"Expected an XLenVT or scalable vector types at this stage\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4920, __PRETTY_FUNCTION__)); | |||
4921 | ||||
4922 | if (Reg) { | |||
4923 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
4924 | return false; | |||
4925 | } | |||
4926 | ||||
4927 | // When a floating-point value is passed on the stack, no bit-conversion is | |||
4928 | // needed. | |||
4929 | if (ValVT.isFloatingPoint()) { | |||
4930 | LocVT = ValVT; | |||
4931 | LocInfo = CCValAssign::Full; | |||
4932 | } | |||
4933 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo)); | |||
4934 | return false; | |||
4935 | } | |||
4936 | ||||
4937 | template <typename ArgTy> | |||
4938 | static Optional<unsigned> preAssignMask(const ArgTy &Args) { | |||
4939 | for (const auto &ArgIdx : enumerate(Args)) { | |||
4940 | MVT ArgVT = ArgIdx.value().VT; | |||
4941 | if (ArgVT.isScalableVector() && | |||
4942 | ArgVT.getVectorElementType().SimpleTy == MVT::i1) | |||
4943 | return ArgIdx.index(); | |||
4944 | } | |||
4945 | return None; | |||
4946 | } | |||
4947 | ||||
4948 | void RISCVTargetLowering::analyzeInputArgs( | |||
4949 | MachineFunction &MF, CCState &CCInfo, | |||
4950 | const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet) const { | |||
4951 | unsigned NumArgs = Ins.size(); | |||
4952 | FunctionType *FType = MF.getFunction().getFunctionType(); | |||
4953 | ||||
4954 | Optional<unsigned> FirstMaskArgument; | |||
4955 | if (Subtarget.hasStdExtV()) | |||
4956 | FirstMaskArgument = preAssignMask(Ins); | |||
4957 | ||||
4958 | for (unsigned i = 0; i != NumArgs; ++i) { | |||
4959 | MVT ArgVT = Ins[i].VT; | |||
4960 | ISD::ArgFlagsTy ArgFlags = Ins[i].Flags; | |||
4961 | ||||
4962 | Type *ArgTy = nullptr; | |||
4963 | if (IsRet) | |||
4964 | ArgTy = FType->getReturnType(); | |||
4965 | else if (Ins[i].isOrigArg()) | |||
4966 | ArgTy = FType->getParamType(Ins[i].getOrigArgIndex()); | |||
4967 | ||||
4968 | RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI(); | |||
4969 | if (CC_RISCV(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full, | |||
4970 | ArgFlags, CCInfo, /*IsFixed=*/true, IsRet, ArgTy, *this, | |||
4971 | FirstMaskArgument)) { | |||
4972 | LLVM_DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("riscv-lower")) { dbgs() << "InputArg #" << i << " has unhandled type " << EVT(ArgVT).getEVTString() << '\n'; } } while (false) | |||
4973 | << EVT(ArgVT).getEVTString() << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("riscv-lower")) { dbgs() << "InputArg #" << i << " has unhandled type " << EVT(ArgVT).getEVTString() << '\n'; } } while (false); | |||
4974 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 4974); | |||
4975 | } | |||
4976 | } | |||
4977 | } | |||
4978 | ||||
4979 | void RISCVTargetLowering::analyzeOutputArgs( | |||
4980 | MachineFunction &MF, CCState &CCInfo, | |||
4981 | const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet, | |||
4982 | CallLoweringInfo *CLI) const { | |||
4983 | unsigned NumArgs = Outs.size(); | |||
4984 | ||||
4985 | Optional<unsigned> FirstMaskArgument; | |||
4986 | if (Subtarget.hasStdExtV()) | |||
4987 | FirstMaskArgument = preAssignMask(Outs); | |||
4988 | ||||
4989 | for (unsigned i = 0; i != NumArgs; i++) { | |||
4990 | MVT ArgVT = Outs[i].VT; | |||
4991 | ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; | |||
4992 | Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr; | |||
4993 | ||||
4994 | RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI(); | |||
4995 | if (CC_RISCV(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full, | |||
4996 | ArgFlags, CCInfo, Outs[i].IsFixed, IsRet, OrigTy, *this, | |||
4997 | FirstMaskArgument)) { | |||
4998 | LLVM_DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("riscv-lower")) { dbgs() << "OutputArg #" << i << " has unhandled type " << EVT(ArgVT).getEVTString() << "\n"; } } while (false) | |||
4999 | << EVT(ArgVT).getEVTString() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("riscv-lower")) { dbgs() << "OutputArg #" << i << " has unhandled type " << EVT(ArgVT).getEVTString() << "\n"; } } while (false); | |||
5000 | llvm_unreachable(nullptr)::llvm::llvm_unreachable_internal(nullptr, "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5000); | |||
5001 | } | |||
5002 | } | |||
5003 | } | |||
5004 | ||||
5005 | // Convert Val to a ValVT. Should not be called for CCValAssign::Indirect | |||
5006 | // values. | |||
5007 | static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val, | |||
5008 | const CCValAssign &VA, const SDLoc &DL) { | |||
5009 | switch (VA.getLocInfo()) { | |||
5010 | default: | |||
5011 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5011); | |||
5012 | case CCValAssign::Full: | |||
5013 | break; | |||
5014 | case CCValAssign::BCvt: | |||
5015 | if (VA.getLocVT().isInteger() && VA.getValVT() == MVT::f16) | |||
5016 | Val = DAG.getNode(RISCVISD::FMV_H_X, DL, MVT::f16, Val); | |||
5017 | else if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) | |||
5018 | Val = DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, Val); | |||
5019 | else | |||
5020 | Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val); | |||
5021 | break; | |||
5022 | } | |||
5023 | return Val; | |||
5024 | } | |||
5025 | ||||
5026 | // The caller is responsible for loading the full value if the argument is | |||
5027 | // passed with CCValAssign::Indirect. | |||
5028 | static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain, | |||
5029 | const CCValAssign &VA, const SDLoc &DL, | |||
5030 | const RISCVTargetLowering &TLI) { | |||
5031 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5032 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
5033 | EVT LocVT = VA.getLocVT(); | |||
5034 | SDValue Val; | |||
5035 | const TargetRegisterClass *RC = TLI.getRegClassFor(LocVT.getSimpleVT()); | |||
5036 | Register VReg = RegInfo.createVirtualRegister(RC); | |||
5037 | RegInfo.addLiveIn(VA.getLocReg(), VReg); | |||
5038 | Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT); | |||
5039 | ||||
5040 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
5041 | return Val; | |||
5042 | ||||
5043 | return convertLocVTToValVT(DAG, Val, VA, DL); | |||
5044 | } | |||
5045 | ||||
5046 | static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDValue Val, | |||
5047 | const CCValAssign &VA, const SDLoc &DL) { | |||
5048 | EVT LocVT = VA.getLocVT(); | |||
5049 | ||||
5050 | switch (VA.getLocInfo()) { | |||
5051 | default: | |||
5052 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5052); | |||
5053 | case CCValAssign::Full: | |||
5054 | break; | |||
5055 | case CCValAssign::BCvt: | |||
5056 | if (VA.getLocVT().isInteger() && VA.getValVT() == MVT::f16) | |||
5057 | Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTH, DL, VA.getLocVT(), Val); | |||
5058 | else if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) | |||
5059 | Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Val); | |||
5060 | else | |||
5061 | Val = DAG.getNode(ISD::BITCAST, DL, LocVT, Val); | |||
5062 | break; | |||
5063 | } | |||
5064 | return Val; | |||
5065 | } | |||
5066 | ||||
5067 | // The caller is responsible for loading the full value if the argument is | |||
5068 | // passed with CCValAssign::Indirect. | |||
5069 | static SDValue unpackFromMemLoc(SelectionDAG &DAG, SDValue Chain, | |||
5070 | const CCValAssign &VA, const SDLoc &DL) { | |||
5071 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5072 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
5073 | EVT LocVT = VA.getLocVT(); | |||
5074 | EVT ValVT = VA.getValVT(); | |||
5075 | EVT PtrVT = MVT::getIntegerVT(DAG.getDataLayout().getPointerSizeInBits(0)); | |||
5076 | int FI = MFI.CreateFixedObject(ValVT.getSizeInBits() / 8, | |||
5077 | VA.getLocMemOffset(), /*Immutable=*/true); | |||
5078 | SDValue FIN = DAG.getFrameIndex(FI, PtrVT); | |||
5079 | SDValue Val; | |||
5080 | ||||
5081 | ISD::LoadExtType ExtType; | |||
5082 | switch (VA.getLocInfo()) { | |||
5083 | default: | |||
5084 | llvm_unreachable("Unexpected CCValAssign::LocInfo")::llvm::llvm_unreachable_internal("Unexpected CCValAssign::LocInfo" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5084); | |||
5085 | case CCValAssign::Full: | |||
5086 | case CCValAssign::Indirect: | |||
5087 | case CCValAssign::BCvt: | |||
5088 | ExtType = ISD::NON_EXTLOAD; | |||
5089 | break; | |||
5090 | } | |||
5091 | Val = DAG.getExtLoad( | |||
5092 | ExtType, DL, LocVT, Chain, FIN, | |||
5093 | MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), ValVT); | |||
5094 | return Val; | |||
5095 | } | |||
5096 | ||||
5097 | static SDValue unpackF64OnRV32DSoftABI(SelectionDAG &DAG, SDValue Chain, | |||
5098 | const CCValAssign &VA, const SDLoc &DL) { | |||
5099 | assert(VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 &&((VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT:: f64 && "Unexpected VA") ? static_cast<void> (0) : __assert_fail ("VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 && \"Unexpected VA\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5100, __PRETTY_FUNCTION__)) | |||
5100 | "Unexpected VA")((VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT:: f64 && "Unexpected VA") ? static_cast<void> (0) : __assert_fail ("VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 && \"Unexpected VA\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5100, __PRETTY_FUNCTION__)); | |||
5101 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5102 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
5103 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
5104 | ||||
5105 | if (VA.isMemLoc()) { | |||
5106 | // f64 is passed on the stack. | |||
5107 | int FI = MFI.CreateFixedObject(8, VA.getLocMemOffset(), /*Immutable=*/true); | |||
5108 | SDValue FIN = DAG.getFrameIndex(FI, MVT::i32); | |||
5109 | return DAG.getLoad(MVT::f64, DL, Chain, FIN, | |||
5110 | MachinePointerInfo::getFixedStack(MF, FI)); | |||
5111 | } | |||
5112 | ||||
5113 | assert(VA.isRegLoc() && "Expected register VA assignment")((VA.isRegLoc() && "Expected register VA assignment") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Expected register VA assignment\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5113, __PRETTY_FUNCTION__)); | |||
5114 | ||||
5115 | Register LoVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass); | |||
5116 | RegInfo.addLiveIn(VA.getLocReg(), LoVReg); | |||
5117 | SDValue Lo = DAG.getCopyFromReg(Chain, DL, LoVReg, MVT::i32); | |||
5118 | SDValue Hi; | |||
5119 | if (VA.getLocReg() == RISCV::X17) { | |||
5120 | // Second half of f64 is passed on the stack. | |||
5121 | int FI = MFI.CreateFixedObject(4, 0, /*Immutable=*/true); | |||
5122 | SDValue FIN = DAG.getFrameIndex(FI, MVT::i32); | |||
5123 | Hi = DAG.getLoad(MVT::i32, DL, Chain, FIN, | |||
5124 | MachinePointerInfo::getFixedStack(MF, FI)); | |||
5125 | } else { | |||
5126 | // Second half of f64 is passed in another GPR. | |||
5127 | Register HiVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass); | |||
5128 | RegInfo.addLiveIn(VA.getLocReg() + 1, HiVReg); | |||
5129 | Hi = DAG.getCopyFromReg(Chain, DL, HiVReg, MVT::i32); | |||
5130 | } | |||
5131 | return DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, Lo, Hi); | |||
5132 | } | |||
5133 | ||||
5134 | // FastCC has less than 1% performance improvement for some particular | |||
5135 | // benchmark. But theoretically, it may has benenfit for some cases. | |||
5136 | static bool CC_RISCV_FastCC(unsigned ValNo, MVT ValVT, MVT LocVT, | |||
5137 | CCValAssign::LocInfo LocInfo, | |||
5138 | ISD::ArgFlagsTy ArgFlags, CCState &State) { | |||
5139 | ||||
5140 | if (LocVT == MVT::i32 || LocVT == MVT::i64) { | |||
5141 | // X5 and X6 might be used for save-restore libcall. | |||
5142 | static const MCPhysReg GPRList[] = { | |||
5143 | RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13, RISCV::X14, | |||
5144 | RISCV::X15, RISCV::X16, RISCV::X17, RISCV::X7, RISCV::X28, | |||
5145 | RISCV::X29, RISCV::X30, RISCV::X31}; | |||
5146 | if (unsigned Reg = State.AllocateReg(GPRList)) { | |||
5147 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5148 | return false; | |||
5149 | } | |||
5150 | } | |||
5151 | ||||
5152 | if (LocVT == MVT::f16) { | |||
5153 | static const MCPhysReg FPR16List[] = { | |||
5154 | RISCV::F10_H, RISCV::F11_H, RISCV::F12_H, RISCV::F13_H, RISCV::F14_H, | |||
5155 | RISCV::F15_H, RISCV::F16_H, RISCV::F17_H, RISCV::F0_H, RISCV::F1_H, | |||
5156 | RISCV::F2_H, RISCV::F3_H, RISCV::F4_H, RISCV::F5_H, RISCV::F6_H, | |||
5157 | RISCV::F7_H, RISCV::F28_H, RISCV::F29_H, RISCV::F30_H, RISCV::F31_H}; | |||
5158 | if (unsigned Reg = State.AllocateReg(FPR16List)) { | |||
5159 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5160 | return false; | |||
5161 | } | |||
5162 | } | |||
5163 | ||||
5164 | if (LocVT == MVT::f32) { | |||
5165 | static const MCPhysReg FPR32List[] = { | |||
5166 | RISCV::F10_F, RISCV::F11_F, RISCV::F12_F, RISCV::F13_F, RISCV::F14_F, | |||
5167 | RISCV::F15_F, RISCV::F16_F, RISCV::F17_F, RISCV::F0_F, RISCV::F1_F, | |||
5168 | RISCV::F2_F, RISCV::F3_F, RISCV::F4_F, RISCV::F5_F, RISCV::F6_F, | |||
5169 | RISCV::F7_F, RISCV::F28_F, RISCV::F29_F, RISCV::F30_F, RISCV::F31_F}; | |||
5170 | if (unsigned Reg = State.AllocateReg(FPR32List)) { | |||
5171 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5172 | return false; | |||
5173 | } | |||
5174 | } | |||
5175 | ||||
5176 | if (LocVT == MVT::f64) { | |||
5177 | static const MCPhysReg FPR64List[] = { | |||
5178 | RISCV::F10_D, RISCV::F11_D, RISCV::F12_D, RISCV::F13_D, RISCV::F14_D, | |||
5179 | RISCV::F15_D, RISCV::F16_D, RISCV::F17_D, RISCV::F0_D, RISCV::F1_D, | |||
5180 | RISCV::F2_D, RISCV::F3_D, RISCV::F4_D, RISCV::F5_D, RISCV::F6_D, | |||
5181 | RISCV::F7_D, RISCV::F28_D, RISCV::F29_D, RISCV::F30_D, RISCV::F31_D}; | |||
5182 | if (unsigned Reg = State.AllocateReg(FPR64List)) { | |||
5183 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5184 | return false; | |||
5185 | } | |||
5186 | } | |||
5187 | ||||
5188 | if (LocVT == MVT::i32 || LocVT == MVT::f32) { | |||
5189 | unsigned Offset4 = State.AllocateStack(4, Align(4)); | |||
5190 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset4, LocVT, LocInfo)); | |||
5191 | return false; | |||
5192 | } | |||
5193 | ||||
5194 | if (LocVT == MVT::i64 || LocVT == MVT::f64) { | |||
5195 | unsigned Offset5 = State.AllocateStack(8, Align(8)); | |||
5196 | State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset5, LocVT, LocInfo)); | |||
5197 | return false; | |||
5198 | } | |||
5199 | ||||
5200 | return true; // CC didn't match. | |||
5201 | } | |||
5202 | ||||
5203 | static bool CC_RISCV_GHC(unsigned ValNo, MVT ValVT, MVT LocVT, | |||
5204 | CCValAssign::LocInfo LocInfo, | |||
5205 | ISD::ArgFlagsTy ArgFlags, CCState &State) { | |||
5206 | ||||
5207 | if (LocVT == MVT::i32 || LocVT == MVT::i64) { | |||
5208 | // Pass in STG registers: Base, Sp, Hp, R1, R2, R3, R4, R5, R6, R7, SpLim | |||
5209 | // s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 | |||
5210 | static const MCPhysReg GPRList[] = { | |||
5211 | RISCV::X9, RISCV::X18, RISCV::X19, RISCV::X20, RISCV::X21, RISCV::X22, | |||
5212 | RISCV::X23, RISCV::X24, RISCV::X25, RISCV::X26, RISCV::X27}; | |||
5213 | if (unsigned Reg = State.AllocateReg(GPRList)) { | |||
5214 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5215 | return false; | |||
5216 | } | |||
5217 | } | |||
5218 | ||||
5219 | if (LocVT == MVT::f32) { | |||
5220 | // Pass in STG registers: F1, ..., F6 | |||
5221 | // fs0 ... fs5 | |||
5222 | static const MCPhysReg FPR32List[] = {RISCV::F8_F, RISCV::F9_F, | |||
5223 | RISCV::F18_F, RISCV::F19_F, | |||
5224 | RISCV::F20_F, RISCV::F21_F}; | |||
5225 | if (unsigned Reg = State.AllocateReg(FPR32List)) { | |||
5226 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5227 | return false; | |||
5228 | } | |||
5229 | } | |||
5230 | ||||
5231 | if (LocVT == MVT::f64) { | |||
5232 | // Pass in STG registers: D1, ..., D6 | |||
5233 | // fs6 ... fs11 | |||
5234 | static const MCPhysReg FPR64List[] = {RISCV::F22_D, RISCV::F23_D, | |||
5235 | RISCV::F24_D, RISCV::F25_D, | |||
5236 | RISCV::F26_D, RISCV::F27_D}; | |||
5237 | if (unsigned Reg = State.AllocateReg(FPR64List)) { | |||
5238 | State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo)); | |||
5239 | return false; | |||
5240 | } | |||
5241 | } | |||
5242 | ||||
5243 | report_fatal_error("No registers left in GHC calling convention"); | |||
5244 | return true; | |||
5245 | } | |||
5246 | ||||
5247 | // Transform physical registers into virtual registers. | |||
5248 | SDValue RISCVTargetLowering::LowerFormalArguments( | |||
5249 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, | |||
5250 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, | |||
5251 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { | |||
5252 | ||||
5253 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5254 | ||||
5255 | switch (CallConv) { | |||
5256 | default: | |||
5257 | report_fatal_error("Unsupported calling convention"); | |||
5258 | case CallingConv::C: | |||
5259 | case CallingConv::Fast: | |||
5260 | break; | |||
5261 | case CallingConv::GHC: | |||
5262 | if (!MF.getSubtarget().getFeatureBits()[RISCV::FeatureStdExtF] || | |||
5263 | !MF.getSubtarget().getFeatureBits()[RISCV::FeatureStdExtD]) | |||
5264 | report_fatal_error( | |||
5265 | "GHC calling convention requires the F and D instruction set extensions"); | |||
5266 | } | |||
5267 | ||||
5268 | const Function &Func = MF.getFunction(); | |||
5269 | if (Func.hasFnAttribute("interrupt")) { | |||
5270 | if (!Func.arg_empty()) | |||
5271 | report_fatal_error( | |||
5272 | "Functions with the interrupt attribute cannot have arguments!"); | |||
5273 | ||||
5274 | StringRef Kind = | |||
5275 | MF.getFunction().getFnAttribute("interrupt").getValueAsString(); | |||
5276 | ||||
5277 | if (!(Kind == "user" || Kind == "supervisor" || Kind == "machine")) | |||
5278 | report_fatal_error( | |||
5279 | "Function interrupt attribute argument not supported!"); | |||
5280 | } | |||
5281 | ||||
5282 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
5283 | MVT XLenVT = Subtarget.getXLenVT(); | |||
5284 | unsigned XLenInBytes = Subtarget.getXLen() / 8; | |||
5285 | // Used with vargs to acumulate store chains. | |||
5286 | std::vector<SDValue> OutChains; | |||
5287 | ||||
5288 | // Assign locations to all of the incoming arguments. | |||
5289 | SmallVector<CCValAssign, 16> ArgLocs; | |||
5290 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); | |||
5291 | ||||
5292 | if (CallConv == CallingConv::Fast) | |||
5293 | CCInfo.AnalyzeFormalArguments(Ins, CC_RISCV_FastCC); | |||
5294 | else if (CallConv == CallingConv::GHC) | |||
5295 | CCInfo.AnalyzeFormalArguments(Ins, CC_RISCV_GHC); | |||
5296 | else | |||
5297 | analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false); | |||
5298 | ||||
5299 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { | |||
5300 | CCValAssign &VA = ArgLocs[i]; | |||
5301 | SDValue ArgValue; | |||
5302 | // Passing f64 on RV32D with a soft float ABI must be handled as a special | |||
5303 | // case. | |||
5304 | if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) | |||
5305 | ArgValue = unpackF64OnRV32DSoftABI(DAG, Chain, VA, DL); | |||
5306 | else if (VA.isRegLoc()) | |||
5307 | ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL, *this); | |||
5308 | else | |||
5309 | ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL); | |||
5310 | ||||
5311 | if (VA.getLocInfo() == CCValAssign::Indirect) { | |||
5312 | // If the original argument was split and passed by reference (e.g. i128 | |||
5313 | // on RV32), we need to load all parts of it here (using the same | |||
5314 | // address). | |||
5315 | InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue, | |||
5316 | MachinePointerInfo())); | |||
5317 | unsigned ArgIndex = Ins[i].OrigArgIndex; | |||
5318 | assert(Ins[i].PartOffset == 0)((Ins[i].PartOffset == 0) ? static_cast<void> (0) : __assert_fail ("Ins[i].PartOffset == 0", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5318, __PRETTY_FUNCTION__)); | |||
5319 | while (i + 1 != e && Ins[i + 1].OrigArgIndex == ArgIndex) { | |||
5320 | CCValAssign &PartVA = ArgLocs[i + 1]; | |||
5321 | unsigned PartOffset = Ins[i + 1].PartOffset; | |||
5322 | SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue, | |||
5323 | DAG.getIntPtrConstant(PartOffset, DL)); | |||
5324 | InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address, | |||
5325 | MachinePointerInfo())); | |||
5326 | ++i; | |||
5327 | } | |||
5328 | continue; | |||
5329 | } | |||
5330 | InVals.push_back(ArgValue); | |||
5331 | } | |||
5332 | ||||
5333 | if (IsVarArg) { | |||
5334 | ArrayRef<MCPhysReg> ArgRegs = makeArrayRef(ArgGPRs); | |||
5335 | unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs); | |||
5336 | const TargetRegisterClass *RC = &RISCV::GPRRegClass; | |||
5337 | MachineFrameInfo &MFI = MF.getFrameInfo(); | |||
5338 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); | |||
5339 | RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>(); | |||
5340 | ||||
5341 | // Offset of the first variable argument from stack pointer, and size of | |||
5342 | // the vararg save area. For now, the varargs save area is either zero or | |||
5343 | // large enough to hold a0-a7. | |||
5344 | int VaArgOffset, VarArgsSaveSize; | |||
5345 | ||||
5346 | // If all registers are allocated, then all varargs must be passed on the | |||
5347 | // stack and we don't need to save any argregs. | |||
5348 | if (ArgRegs.size() == Idx) { | |||
5349 | VaArgOffset = CCInfo.getNextStackOffset(); | |||
5350 | VarArgsSaveSize = 0; | |||
5351 | } else { | |||
5352 | VarArgsSaveSize = XLenInBytes * (ArgRegs.size() - Idx); | |||
5353 | VaArgOffset = -VarArgsSaveSize; | |||
5354 | } | |||
5355 | ||||
5356 | // Record the frame index of the first variable argument | |||
5357 | // which is a value necessary to VASTART. | |||
5358 | int FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true); | |||
5359 | RVFI->setVarArgsFrameIndex(FI); | |||
5360 | ||||
5361 | // If saving an odd number of registers then create an extra stack slot to | |||
5362 | // ensure that the frame pointer is 2*XLEN-aligned, which in turn ensures | |||
5363 | // offsets to even-numbered registered remain 2*XLEN-aligned. | |||
5364 | if (Idx % 2) { | |||
5365 | MFI.CreateFixedObject(XLenInBytes, VaArgOffset - (int)XLenInBytes, true); | |||
5366 | VarArgsSaveSize += XLenInBytes; | |||
5367 | } | |||
5368 | ||||
5369 | // Copy the integer registers that may have been used for passing varargs | |||
5370 | // to the vararg save area. | |||
5371 | for (unsigned I = Idx; I < ArgRegs.size(); | |||
5372 | ++I, VaArgOffset += XLenInBytes) { | |||
5373 | const Register Reg = RegInfo.createVirtualRegister(RC); | |||
5374 | RegInfo.addLiveIn(ArgRegs[I], Reg); | |||
5375 | SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, XLenVT); | |||
5376 | FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true); | |||
5377 | SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); | |||
5378 | SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff, | |||
5379 | MachinePointerInfo::getFixedStack(MF, FI)); | |||
5380 | cast<StoreSDNode>(Store.getNode()) | |||
5381 | ->getMemOperand() | |||
5382 | ->setValue((Value *)nullptr); | |||
5383 | OutChains.push_back(Store); | |||
5384 | } | |||
5385 | RVFI->setVarArgsSaveSize(VarArgsSaveSize); | |||
5386 | } | |||
5387 | ||||
5388 | // All stores are grouped in one node to allow the matching between | |||
5389 | // the size of Ins and InVals. This only happens for vararg functions. | |||
5390 | if (!OutChains.empty()) { | |||
5391 | OutChains.push_back(Chain); | |||
5392 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains); | |||
5393 | } | |||
5394 | ||||
5395 | return Chain; | |||
5396 | } | |||
5397 | ||||
5398 | /// isEligibleForTailCallOptimization - Check whether the call is eligible | |||
5399 | /// for tail call optimization. | |||
5400 | /// Note: This is modelled after ARM's IsEligibleForTailCallOptimization. | |||
5401 | bool RISCVTargetLowering::isEligibleForTailCallOptimization( | |||
5402 | CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF, | |||
5403 | const SmallVector<CCValAssign, 16> &ArgLocs) const { | |||
5404 | ||||
5405 | auto &Callee = CLI.Callee; | |||
5406 | auto CalleeCC = CLI.CallConv; | |||
5407 | auto &Outs = CLI.Outs; | |||
5408 | auto &Caller = MF.getFunction(); | |||
5409 | auto CallerCC = Caller.getCallingConv(); | |||
5410 | ||||
5411 | // Exception-handling functions need a special set of instructions to | |||
5412 | // indicate a return to the hardware. Tail-calling another function would | |||
5413 | // probably break this. | |||
5414 | // TODO: The "interrupt" attribute isn't currently defined by RISC-V. This | |||
5415 | // should be expanded as new function attributes are introduced. | |||
5416 | if (Caller.hasFnAttribute("interrupt")) | |||
5417 | return false; | |||
5418 | ||||
5419 | // Do not tail call opt if the stack is used to pass parameters. | |||
5420 | if (CCInfo.getNextStackOffset() != 0) | |||
5421 | return false; | |||
5422 | ||||
5423 | // Do not tail call opt if any parameters need to be passed indirectly. | |||
5424 | // Since long doubles (fp128) and i128 are larger than 2*XLEN, they are | |||
5425 | // passed indirectly. So the address of the value will be passed in a | |||
5426 | // register, or if not available, then the address is put on the stack. In | |||
5427 | // order to pass indirectly, space on the stack often needs to be allocated | |||
5428 | // in order to store the value. In this case the CCInfo.getNextStackOffset() | |||
5429 | // != 0 check is not enough and we need to check if any CCValAssign ArgsLocs | |||
5430 | // are passed CCValAssign::Indirect. | |||
5431 | for (auto &VA : ArgLocs) | |||
5432 | if (VA.getLocInfo() == CCValAssign::Indirect) | |||
5433 | return false; | |||
5434 | ||||
5435 | // Do not tail call opt if either caller or callee uses struct return | |||
5436 | // semantics. | |||
5437 | auto IsCallerStructRet = Caller.hasStructRetAttr(); | |||
5438 | auto IsCalleeStructRet = Outs.empty() ? false : Outs[0].Flags.isSRet(); | |||
5439 | if (IsCallerStructRet || IsCalleeStructRet) | |||
5440 | return false; | |||
5441 | ||||
5442 | // Externally-defined functions with weak linkage should not be | |||
5443 | // tail-called. The behaviour of branch instructions in this situation (as | |||
5444 | // used for tail calls) is implementation-defined, so we cannot rely on the | |||
5445 | // linker replacing the tail call with a return. | |||
5446 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { | |||
5447 | const GlobalValue *GV = G->getGlobal(); | |||
5448 | if (GV->hasExternalWeakLinkage()) | |||
5449 | return false; | |||
5450 | } | |||
5451 | ||||
5452 | // The callee has to preserve all registers the caller needs to preserve. | |||
5453 | const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
5454 | const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC); | |||
5455 | if (CalleeCC != CallerCC) { | |||
5456 | const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC); | |||
5457 | if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved)) | |||
5458 | return false; | |||
5459 | } | |||
5460 | ||||
5461 | // Byval parameters hand the function a pointer directly into the stack area | |||
5462 | // we want to reuse during a tail call. Working around this *is* possible | |||
5463 | // but less efficient and uglier in LowerCall. | |||
5464 | for (auto &Arg : Outs) | |||
5465 | if (Arg.Flags.isByVal()) | |||
5466 | return false; | |||
5467 | ||||
5468 | return true; | |||
5469 | } | |||
5470 | ||||
5471 | // Lower a call to a callseq_start + CALL + callseq_end chain, and add input | |||
5472 | // and output parameter nodes. | |||
5473 | SDValue RISCVTargetLowering::LowerCall(CallLoweringInfo &CLI, | |||
5474 | SmallVectorImpl<SDValue> &InVals) const { | |||
5475 | SelectionDAG &DAG = CLI.DAG; | |||
5476 | SDLoc &DL = CLI.DL; | |||
5477 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; | |||
5478 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; | |||
5479 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; | |||
5480 | SDValue Chain = CLI.Chain; | |||
5481 | SDValue Callee = CLI.Callee; | |||
5482 | bool &IsTailCall = CLI.IsTailCall; | |||
5483 | CallingConv::ID CallConv = CLI.CallConv; | |||
5484 | bool IsVarArg = CLI.IsVarArg; | |||
5485 | EVT PtrVT = getPointerTy(DAG.getDataLayout()); | |||
5486 | MVT XLenVT = Subtarget.getXLenVT(); | |||
5487 | ||||
5488 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5489 | ||||
5490 | // Analyze the operands of the call, assigning locations to each operand. | |||
5491 | SmallVector<CCValAssign, 16> ArgLocs; | |||
5492 | CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); | |||
5493 | ||||
5494 | if (CallConv == CallingConv::Fast) | |||
5495 | ArgCCInfo.AnalyzeCallOperands(Outs, CC_RISCV_FastCC); | |||
5496 | else if (CallConv == CallingConv::GHC) | |||
5497 | ArgCCInfo.AnalyzeCallOperands(Outs, CC_RISCV_GHC); | |||
5498 | else | |||
5499 | analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI); | |||
5500 | ||||
5501 | // Check if it's really possible to do a tail call. | |||
5502 | if (IsTailCall) | |||
5503 | IsTailCall = isEligibleForTailCallOptimization(ArgCCInfo, CLI, MF, ArgLocs); | |||
5504 | ||||
5505 | if (IsTailCall) | |||
5506 | ++NumTailCalls; | |||
5507 | else if (CLI.CB && CLI.CB->isMustTailCall()) | |||
5508 | report_fatal_error("failed to perform tail call elimination on a call " | |||
5509 | "site marked musttail"); | |||
5510 | ||||
5511 | // Get a count of how many bytes are to be pushed on the stack. | |||
5512 | unsigned NumBytes = ArgCCInfo.getNextStackOffset(); | |||
5513 | ||||
5514 | // Create local copies for byval args | |||
5515 | SmallVector<SDValue, 8> ByValArgs; | |||
5516 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { | |||
5517 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
5518 | if (!Flags.isByVal()) | |||
5519 | continue; | |||
5520 | ||||
5521 | SDValue Arg = OutVals[i]; | |||
5522 | unsigned Size = Flags.getByValSize(); | |||
5523 | Align Alignment = Flags.getNonZeroByValAlign(); | |||
5524 | ||||
5525 | int FI = | |||
5526 | MF.getFrameInfo().CreateStackObject(Size, Alignment, /*isSS=*/false); | |||
5527 | SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout())); | |||
5528 | SDValue SizeNode = DAG.getConstant(Size, DL, XLenVT); | |||
5529 | ||||
5530 | Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Alignment, | |||
5531 | /*IsVolatile=*/false, | |||
5532 | /*AlwaysInline=*/false, IsTailCall, | |||
5533 | MachinePointerInfo(), MachinePointerInfo()); | |||
5534 | ByValArgs.push_back(FIPtr); | |||
5535 | } | |||
5536 | ||||
5537 | if (!IsTailCall) | |||
5538 | Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL); | |||
5539 | ||||
5540 | // Copy argument values to their designated locations. | |||
5541 | SmallVector<std::pair<Register, SDValue>, 8> RegsToPass; | |||
5542 | SmallVector<SDValue, 8> MemOpChains; | |||
5543 | SDValue StackPtr; | |||
5544 | for (unsigned i = 0, j = 0, e = ArgLocs.size(); i != e; ++i) { | |||
5545 | CCValAssign &VA = ArgLocs[i]; | |||
5546 | SDValue ArgValue = OutVals[i]; | |||
5547 | ISD::ArgFlagsTy Flags = Outs[i].Flags; | |||
5548 | ||||
5549 | // Handle passing f64 on RV32D with a soft float ABI as a special case. | |||
5550 | bool IsF64OnRV32DSoftABI = | |||
5551 | VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64; | |||
5552 | if (IsF64OnRV32DSoftABI && VA.isRegLoc()) { | |||
5553 | SDValue SplitF64 = DAG.getNode( | |||
5554 | RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32), ArgValue); | |||
5555 | SDValue Lo = SplitF64.getValue(0); | |||
5556 | SDValue Hi = SplitF64.getValue(1); | |||
5557 | ||||
5558 | Register RegLo = VA.getLocReg(); | |||
5559 | RegsToPass.push_back(std::make_pair(RegLo, Lo)); | |||
5560 | ||||
5561 | if (RegLo == RISCV::X17) { | |||
5562 | // Second half of f64 is passed on the stack. | |||
5563 | // Work out the address of the stack slot. | |||
5564 | if (!StackPtr.getNode()) | |||
5565 | StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT); | |||
5566 | // Emit the store. | |||
5567 | MemOpChains.push_back( | |||
5568 | DAG.getStore(Chain, DL, Hi, StackPtr, MachinePointerInfo())); | |||
5569 | } else { | |||
5570 | // Second half of f64 is passed in another GPR. | |||
5571 | assert(RegLo < RISCV::X31 && "Invalid register pair")((RegLo < RISCV::X31 && "Invalid register pair") ? static_cast<void> (0) : __assert_fail ("RegLo < RISCV::X31 && \"Invalid register pair\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5571, __PRETTY_FUNCTION__)); | |||
5572 | Register RegHigh = RegLo + 1; | |||
5573 | RegsToPass.push_back(std::make_pair(RegHigh, Hi)); | |||
5574 | } | |||
5575 | continue; | |||
5576 | } | |||
5577 | ||||
5578 | // IsF64OnRV32DSoftABI && VA.isMemLoc() is handled below in the same way | |||
5579 | // as any other MemLoc. | |||
5580 | ||||
5581 | // Promote the value if needed. | |||
5582 | // For now, only handle fully promoted and indirect arguments. | |||
5583 | if (VA.getLocInfo() == CCValAssign::Indirect) { | |||
5584 | // Store the argument in a stack slot and pass its address. | |||
5585 | SDValue SpillSlot = DAG.CreateStackTemporary(Outs[i].ArgVT); | |||
5586 | int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex(); | |||
5587 | MemOpChains.push_back( | |||
5588 | DAG.getStore(Chain, DL, ArgValue, SpillSlot, | |||
5589 | MachinePointerInfo::getFixedStack(MF, FI))); | |||
5590 | // If the original argument was split (e.g. i128), we need | |||
5591 | // to store all parts of it here (and pass just one address). | |||
5592 | unsigned ArgIndex = Outs[i].OrigArgIndex; | |||
5593 | assert(Outs[i].PartOffset == 0)((Outs[i].PartOffset == 0) ? static_cast<void> (0) : __assert_fail ("Outs[i].PartOffset == 0", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5593, __PRETTY_FUNCTION__)); | |||
5594 | while (i + 1 != e && Outs[i + 1].OrigArgIndex == ArgIndex) { | |||
5595 | SDValue PartValue = OutVals[i + 1]; | |||
5596 | unsigned PartOffset = Outs[i + 1].PartOffset; | |||
5597 | SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot, | |||
5598 | DAG.getIntPtrConstant(PartOffset, DL)); | |||
5599 | MemOpChains.push_back( | |||
5600 | DAG.getStore(Chain, DL, PartValue, Address, | |||
5601 | MachinePointerInfo::getFixedStack(MF, FI))); | |||
5602 | ++i; | |||
5603 | } | |||
5604 | ArgValue = SpillSlot; | |||
5605 | } else { | |||
5606 | ArgValue = convertValVTToLocVT(DAG, ArgValue, VA, DL); | |||
5607 | } | |||
5608 | ||||
5609 | // Use local copy if it is a byval arg. | |||
5610 | if (Flags.isByVal()) | |||
5611 | ArgValue = ByValArgs[j++]; | |||
5612 | ||||
5613 | if (VA.isRegLoc()) { | |||
5614 | // Queue up the argument copies and emit them at the end. | |||
5615 | RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue)); | |||
5616 | } else { | |||
5617 | assert(VA.isMemLoc() && "Argument not register or memory")((VA.isMemLoc() && "Argument not register or memory") ? static_cast<void> (0) : __assert_fail ("VA.isMemLoc() && \"Argument not register or memory\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5617, __PRETTY_FUNCTION__)); | |||
5618 | assert(!IsTailCall && "Tail call not allowed if stack is used "((!IsTailCall && "Tail call not allowed if stack is used " "for passing parameters") ? static_cast<void> (0) : __assert_fail ("!IsTailCall && \"Tail call not allowed if stack is used \" \"for passing parameters\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5619, __PRETTY_FUNCTION__)) | |||
5619 | "for passing parameters")((!IsTailCall && "Tail call not allowed if stack is used " "for passing parameters") ? static_cast<void> (0) : __assert_fail ("!IsTailCall && \"Tail call not allowed if stack is used \" \"for passing parameters\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5619, __PRETTY_FUNCTION__)); | |||
5620 | ||||
5621 | // Work out the address of the stack slot. | |||
5622 | if (!StackPtr.getNode()) | |||
5623 | StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT); | |||
5624 | SDValue Address = | |||
5625 | DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, | |||
5626 | DAG.getIntPtrConstant(VA.getLocMemOffset(), DL)); | |||
5627 | ||||
5628 | // Emit the store. | |||
5629 | MemOpChains.push_back( | |||
5630 | DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo())); | |||
5631 | } | |||
5632 | } | |||
5633 | ||||
5634 | // Join the stores, which are independent of one another. | |||
5635 | if (!MemOpChains.empty()) | |||
5636 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains); | |||
5637 | ||||
5638 | SDValue Glue; | |||
5639 | ||||
5640 | // Build a sequence of copy-to-reg nodes, chained and glued together. | |||
5641 | for (auto &Reg : RegsToPass) { | |||
5642 | Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, Glue); | |||
5643 | Glue = Chain.getValue(1); | |||
5644 | } | |||
5645 | ||||
5646 | // Validate that none of the argument registers have been marked as | |||
5647 | // reserved, if so report an error. Do the same for the return address if this | |||
5648 | // is not a tailcall. | |||
5649 | validateCCReservedRegs(RegsToPass, MF); | |||
5650 | if (!IsTailCall && | |||
5651 | MF.getSubtarget<RISCVSubtarget>().isRegisterReservedByUser(RISCV::X1)) | |||
5652 | MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{ | |||
5653 | MF.getFunction(), | |||
5654 | "Return address register required, but has been reserved."}); | |||
5655 | ||||
5656 | // If the callee is a GlobalAddress/ExternalSymbol node, turn it into a | |||
5657 | // TargetGlobalAddress/TargetExternalSymbol node so that legalize won't | |||
5658 | // split it and then direct call can be matched by PseudoCALL. | |||
5659 | if (GlobalAddressSDNode *S = dyn_cast<GlobalAddressSDNode>(Callee)) { | |||
5660 | const GlobalValue *GV = S->getGlobal(); | |||
5661 | ||||
5662 | unsigned OpFlags = RISCVII::MO_CALL; | |||
5663 | if (!getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV)) | |||
5664 | OpFlags = RISCVII::MO_PLT; | |||
5665 | ||||
5666 | Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags); | |||
5667 | } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { | |||
5668 | unsigned OpFlags = RISCVII::MO_CALL; | |||
5669 | ||||
5670 | if (!getTargetMachine().shouldAssumeDSOLocal(*MF.getFunction().getParent(), | |||
5671 | nullptr)) | |||
5672 | OpFlags = RISCVII::MO_PLT; | |||
5673 | ||||
5674 | Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, OpFlags); | |||
5675 | } | |||
5676 | ||||
5677 | // The first call operand is the chain and the second is the target address. | |||
5678 | SmallVector<SDValue, 8> Ops; | |||
5679 | Ops.push_back(Chain); | |||
5680 | Ops.push_back(Callee); | |||
5681 | ||||
5682 | // Add argument registers to the end of the list so that they are | |||
5683 | // known live into the call. | |||
5684 | for (auto &Reg : RegsToPass) | |||
5685 | Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType())); | |||
5686 | ||||
5687 | if (!IsTailCall) { | |||
5688 | // Add a register mask operand representing the call-preserved registers. | |||
5689 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); | |||
5690 | const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv); | |||
5691 | assert(Mask && "Missing call preserved mask for calling convention")((Mask && "Missing call preserved mask for calling convention" ) ? static_cast<void> (0) : __assert_fail ("Mask && \"Missing call preserved mask for calling convention\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5691, __PRETTY_FUNCTION__)); | |||
5692 | Ops.push_back(DAG.getRegisterMask(Mask)); | |||
5693 | } | |||
5694 | ||||
5695 | // Glue the call to the argument copies, if any. | |||
5696 | if (Glue.getNode()) | |||
5697 | Ops.push_back(Glue); | |||
5698 | ||||
5699 | // Emit the call. | |||
5700 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); | |||
5701 | ||||
5702 | if (IsTailCall) { | |||
5703 | MF.getFrameInfo().setHasTailCall(); | |||
5704 | return DAG.getNode(RISCVISD::TAIL, DL, NodeTys, Ops); | |||
5705 | } | |||
5706 | ||||
5707 | Chain = DAG.getNode(RISCVISD::CALL, DL, NodeTys, Ops); | |||
5708 | DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge); | |||
5709 | Glue = Chain.getValue(1); | |||
5710 | ||||
5711 | // Mark the end of the call, which is glued to the call itself. | |||
5712 | Chain = DAG.getCALLSEQ_END(Chain, | |||
5713 | DAG.getConstant(NumBytes, DL, PtrVT, true), | |||
5714 | DAG.getConstant(0, DL, PtrVT, true), | |||
5715 | Glue, DL); | |||
5716 | Glue = Chain.getValue(1); | |||
5717 | ||||
5718 | // Assign locations to each value returned by this call. | |||
5719 | SmallVector<CCValAssign, 16> RVLocs; | |||
5720 | CCState RetCCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); | |||
5721 | analyzeInputArgs(MF, RetCCInfo, Ins, /*IsRet=*/true); | |||
5722 | ||||
5723 | // Copy all of the result registers out of their specified physreg. | |||
5724 | for (auto &VA : RVLocs) { | |||
5725 | // Copy the value out | |||
5726 | SDValue RetValue = | |||
5727 | DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), Glue); | |||
5728 | // Glue the RetValue to the end of the call sequence | |||
5729 | Chain = RetValue.getValue(1); | |||
5730 | Glue = RetValue.getValue(2); | |||
5731 | ||||
5732 | if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) { | |||
5733 | assert(VA.getLocReg() == ArgGPRs[0] && "Unexpected reg assignment")((VA.getLocReg() == ArgGPRs[0] && "Unexpected reg assignment" ) ? static_cast<void> (0) : __assert_fail ("VA.getLocReg() == ArgGPRs[0] && \"Unexpected reg assignment\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5733, __PRETTY_FUNCTION__)); | |||
5734 | SDValue RetValue2 = | |||
5735 | DAG.getCopyFromReg(Chain, DL, ArgGPRs[1], MVT::i32, Glue); | |||
5736 | Chain = RetValue2.getValue(1); | |||
5737 | Glue = RetValue2.getValue(2); | |||
5738 | RetValue = DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, RetValue, | |||
5739 | RetValue2); | |||
5740 | } | |||
5741 | ||||
5742 | RetValue = convertLocVTToValVT(DAG, RetValue, VA, DL); | |||
5743 | ||||
5744 | InVals.push_back(RetValue); | |||
5745 | } | |||
5746 | ||||
5747 | return Chain; | |||
5748 | } | |||
5749 | ||||
5750 | bool RISCVTargetLowering::CanLowerReturn( | |||
5751 | CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg, | |||
5752 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const { | |||
5753 | SmallVector<CCValAssign, 16> RVLocs; | |||
5754 | CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); | |||
5755 | ||||
5756 | Optional<unsigned> FirstMaskArgument; | |||
5757 | if (Subtarget.hasStdExtV()) | |||
5758 | FirstMaskArgument = preAssignMask(Outs); | |||
5759 | ||||
5760 | for (unsigned i = 0, e = Outs.size(); i != e; ++i) { | |||
5761 | MVT VT = Outs[i].VT; | |||
5762 | ISD::ArgFlagsTy ArgFlags = Outs[i].Flags; | |||
5763 | RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI(); | |||
5764 | if (CC_RISCV(MF.getDataLayout(), ABI, i, VT, VT, CCValAssign::Full, | |||
5765 | ArgFlags, CCInfo, /*IsFixed=*/true, /*IsRet=*/true, nullptr, | |||
5766 | *this, FirstMaskArgument)) | |||
5767 | return false; | |||
5768 | } | |||
5769 | return true; | |||
5770 | } | |||
5771 | ||||
5772 | SDValue | |||
5773 | RISCVTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, | |||
5774 | bool IsVarArg, | |||
5775 | const SmallVectorImpl<ISD::OutputArg> &Outs, | |||
5776 | const SmallVectorImpl<SDValue> &OutVals, | |||
5777 | const SDLoc &DL, SelectionDAG &DAG) const { | |||
5778 | const MachineFunction &MF = DAG.getMachineFunction(); | |||
5779 | const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>(); | |||
5780 | ||||
5781 | // Stores the assignment of the return value to a location. | |||
5782 | SmallVector<CCValAssign, 16> RVLocs; | |||
5783 | ||||
5784 | // Info about the registers and stack slot. | |||
5785 | CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, | |||
5786 | *DAG.getContext()); | |||
5787 | ||||
5788 | analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true, | |||
5789 | nullptr); | |||
5790 | ||||
5791 | if (CallConv == CallingConv::GHC && !RVLocs.empty()) | |||
5792 | report_fatal_error("GHC functions return void only"); | |||
5793 | ||||
5794 | SDValue Glue; | |||
5795 | SmallVector<SDValue, 4> RetOps(1, Chain); | |||
5796 | ||||
5797 | // Copy the result values into the output registers. | |||
5798 | for (unsigned i = 0, e = RVLocs.size(); i < e; ++i) { | |||
5799 | SDValue Val = OutVals[i]; | |||
5800 | CCValAssign &VA = RVLocs[i]; | |||
5801 | assert(VA.isRegLoc() && "Can only return in registers!")((VA.isRegLoc() && "Can only return in registers!") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Can only return in registers!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5801, __PRETTY_FUNCTION__)); | |||
5802 | ||||
5803 | if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) { | |||
5804 | // Handle returning f64 on RV32D with a soft float ABI. | |||
5805 | assert(VA.isRegLoc() && "Expected return via registers")((VA.isRegLoc() && "Expected return via registers") ? static_cast<void> (0) : __assert_fail ("VA.isRegLoc() && \"Expected return via registers\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5805, __PRETTY_FUNCTION__)); | |||
5806 | SDValue SplitF64 = DAG.getNode(RISCVISD::SplitF64, DL, | |||
5807 | DAG.getVTList(MVT::i32, MVT::i32), Val); | |||
5808 | SDValue Lo = SplitF64.getValue(0); | |||
5809 | SDValue Hi = SplitF64.getValue(1); | |||
5810 | Register RegLo = VA.getLocReg(); | |||
5811 | assert(RegLo < RISCV::X31 && "Invalid register pair")((RegLo < RISCV::X31 && "Invalid register pair") ? static_cast<void> (0) : __assert_fail ("RegLo < RISCV::X31 && \"Invalid register pair\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 5811, __PRETTY_FUNCTION__)); | |||
5812 | Register RegHi = RegLo + 1; | |||
5813 | ||||
5814 | if (STI.isRegisterReservedByUser(RegLo) || | |||
5815 | STI.isRegisterReservedByUser(RegHi)) | |||
5816 | MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{ | |||
5817 | MF.getFunction(), | |||
5818 | "Return value register required, but has been reserved."}); | |||
5819 | ||||
5820 | Chain = DAG.getCopyToReg(Chain, DL, RegLo, Lo, Glue); | |||
5821 | Glue = Chain.getValue(1); | |||
5822 | RetOps.push_back(DAG.getRegister(RegLo, MVT::i32)); | |||
5823 | Chain = DAG.getCopyToReg(Chain, DL, RegHi, Hi, Glue); | |||
5824 | Glue = Chain.getValue(1); | |||
5825 | RetOps.push_back(DAG.getRegister(RegHi, MVT::i32)); | |||
5826 | } else { | |||
5827 | // Handle a 'normal' return. | |||
5828 | Val = convertValVTToLocVT(DAG, Val, VA, DL); | |||
5829 | Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Glue); | |||
5830 | ||||
5831 | if (STI.isRegisterReservedByUser(VA.getLocReg())) | |||
5832 | MF.getFunction().getContext().diagnose(DiagnosticInfoUnsupported{ | |||
5833 | MF.getFunction(), | |||
5834 | "Return value register required, but has been reserved."}); | |||
5835 | ||||
5836 | // Guarantee that all emitted copies are stuck together. | |||
5837 | Glue = Chain.getValue(1); | |||
5838 | RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT())); | |||
5839 | } | |||
5840 | } | |||
5841 | ||||
5842 | RetOps[0] = Chain; // Update chain. | |||
5843 | ||||
5844 | // Add the glue node if we have it. | |||
5845 | if (Glue.getNode()) { | |||
5846 | RetOps.push_back(Glue); | |||
5847 | } | |||
5848 | ||||
5849 | // Interrupt service routines use different return instructions. | |||
5850 | const Function &Func = DAG.getMachineFunction().getFunction(); | |||
5851 | if (Func.hasFnAttribute("interrupt")) { | |||
5852 | if (!Func.getReturnType()->isVoidTy()) | |||
5853 | report_fatal_error( | |||
5854 | "Functions with the interrupt attribute must have void return type!"); | |||
5855 | ||||
5856 | MachineFunction &MF = DAG.getMachineFunction(); | |||
5857 | StringRef Kind = | |||
5858 | MF.getFunction().getFnAttribute("interrupt").getValueAsString(); | |||
5859 | ||||
5860 | unsigned RetOpc; | |||
5861 | if (Kind == "user") | |||
5862 | RetOpc = RISCVISD::URET_FLAG; | |||
5863 | else if (Kind == "supervisor") | |||
5864 | RetOpc = RISCVISD::SRET_FLAG; | |||
5865 | else | |||
5866 | RetOpc = RISCVISD::MRET_FLAG; | |||
5867 | ||||
5868 | return DAG.getNode(RetOpc, DL, MVT::Other, RetOps); | |||
5869 | } | |||
5870 | ||||
5871 | return DAG.getNode(RISCVISD::RET_FLAG, DL, MVT::Other, RetOps); | |||
5872 | } | |||
5873 | ||||
5874 | void RISCVTargetLowering::validateCCReservedRegs( | |||
5875 | const SmallVectorImpl<std::pair<llvm::Register, llvm::SDValue>> &Regs, | |||
5876 | MachineFunction &MF) const { | |||
5877 | const Function &F = MF.getFunction(); | |||
5878 | const RISCVSubtarget &STI = MF.getSubtarget<RISCVSubtarget>(); | |||
5879 | ||||
5880 | if (llvm::any_of(Regs, [&STI](auto Reg) { | |||
5881 | return STI.isRegisterReservedByUser(Reg.first); | |||
5882 | })) | |||
5883 | F.getContext().diagnose(DiagnosticInfoUnsupported{ | |||
5884 | F, "Argument register required, but has been reserved."}); | |||
5885 | } | |||
5886 | ||||
5887 | bool RISCVTargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const { | |||
5888 | return CI->isTailCall(); | |||
5889 | } | |||
5890 | ||||
5891 | const char *RISCVTargetLowering::getTargetNodeName(unsigned Opcode) const { | |||
5892 | #define NODE_NAME_CASE(NODE) \ | |||
5893 | case RISCVISD::NODE: \ | |||
5894 | return "RISCVISD::" #NODE; | |||
5895 | // clang-format off | |||
5896 | switch ((RISCVISD::NodeType)Opcode) { | |||
5897 | case RISCVISD::FIRST_NUMBER: | |||
5898 | break; | |||
5899 | NODE_NAME_CASE(RET_FLAG) | |||
5900 | NODE_NAME_CASE(URET_FLAG) | |||
5901 | NODE_NAME_CASE(SRET_FLAG) | |||
5902 | NODE_NAME_CASE(MRET_FLAG) | |||
5903 | NODE_NAME_CASE(CALL) | |||
5904 | NODE_NAME_CASE(SELECT_CC) | |||
5905 | NODE_NAME_CASE(BuildPairF64) | |||
5906 | NODE_NAME_CASE(SplitF64) | |||
5907 | NODE_NAME_CASE(TAIL) | |||
5908 | NODE_NAME_CASE(SLLW) | |||
5909 | NODE_NAME_CASE(SRAW) | |||
5910 | NODE_NAME_CASE(SRLW) | |||
5911 | NODE_NAME_CASE(DIVW) | |||
5912 | NODE_NAME_CASE(DIVUW) | |||
5913 | NODE_NAME_CASE(REMUW) | |||
5914 | NODE_NAME_CASE(ROLW) | |||
5915 | NODE_NAME_CASE(RORW) | |||
5916 | NODE_NAME_CASE(FSLW) | |||
5917 | NODE_NAME_CASE(FSRW) | |||
5918 | NODE_NAME_CASE(FSL) | |||
5919 | NODE_NAME_CASE(FSR) | |||
5920 | NODE_NAME_CASE(FMV_H_X) | |||
5921 | NODE_NAME_CASE(FMV_X_ANYEXTH) | |||
5922 | NODE_NAME_CASE(FMV_W_X_RV64) | |||
5923 | NODE_NAME_CASE(FMV_X_ANYEXTW_RV64) | |||
5924 | NODE_NAME_CASE(READ_CYCLE_WIDE) | |||
5925 | NODE_NAME_CASE(GREVI) | |||
5926 | NODE_NAME_CASE(GREVIW) | |||
5927 | NODE_NAME_CASE(GORCI) | |||
5928 | NODE_NAME_CASE(GORCIW) | |||
5929 | NODE_NAME_CASE(SHFLI) | |||
5930 | NODE_NAME_CASE(VMV_V_X_VL) | |||
5931 | NODE_NAME_CASE(VFMV_V_F_VL) | |||
5932 | NODE_NAME_CASE(VMV_X_S) | |||
5933 | NODE_NAME_CASE(VMV_S_XF_VL) | |||
5934 | NODE_NAME_CASE(SPLAT_VECTOR_I64) | |||
5935 | NODE_NAME_CASE(READ_VLENB) | |||
5936 | NODE_NAME_CASE(TRUNCATE_VECTOR_VL) | |||
5937 | NODE_NAME_CASE(VLEFF) | |||
5938 | NODE_NAME_CASE(VLEFF_MASK) | |||
5939 | NODE_NAME_CASE(VSLIDEUP_VL) | |||
5940 | NODE_NAME_CASE(VSLIDEDOWN_VL) | |||
5941 | NODE_NAME_CASE(VID_VL) | |||
5942 | NODE_NAME_CASE(VFNCVT_ROD_VL) | |||
5943 | NODE_NAME_CASE(VECREDUCE_ADD_VL) | |||
5944 | NODE_NAME_CASE(VECREDUCE_UMAX_VL) | |||
5945 | NODE_NAME_CASE(VECREDUCE_SMAX_VL) | |||
5946 | NODE_NAME_CASE(VECREDUCE_UMIN_VL) | |||
5947 | NODE_NAME_CASE(VECREDUCE_SMIN_VL) | |||
5948 | NODE_NAME_CASE(VECREDUCE_AND_VL) | |||
5949 | NODE_NAME_CASE(VECREDUCE_OR_VL) | |||
5950 | NODE_NAME_CASE(VECREDUCE_XOR_VL) | |||
5951 | NODE_NAME_CASE(VECREDUCE_FADD_VL) | |||
5952 | NODE_NAME_CASE(VECREDUCE_SEQ_FADD_VL) | |||
5953 | NODE_NAME_CASE(ADD_VL) | |||
5954 | NODE_NAME_CASE(AND_VL) | |||
5955 | NODE_NAME_CASE(MUL_VL) | |||
5956 | NODE_NAME_CASE(OR_VL) | |||
5957 | NODE_NAME_CASE(SDIV_VL) | |||
5958 | NODE_NAME_CASE(SHL_VL) | |||
5959 | NODE_NAME_CASE(SREM_VL) | |||
5960 | NODE_NAME_CASE(SRA_VL) | |||
5961 | NODE_NAME_CASE(SRL_VL) | |||
5962 | NODE_NAME_CASE(SUB_VL) | |||
5963 | NODE_NAME_CASE(UDIV_VL) | |||
5964 | NODE_NAME_CASE(UREM_VL) | |||
5965 | NODE_NAME_CASE(XOR_VL) | |||
5966 | NODE_NAME_CASE(FADD_VL) | |||
5967 | NODE_NAME_CASE(FSUB_VL) | |||
5968 | NODE_NAME_CASE(FMUL_VL) | |||
5969 | NODE_NAME_CASE(FDIV_VL) | |||
5970 | NODE_NAME_CASE(FNEG_VL) | |||
5971 | NODE_NAME_CASE(FABS_VL) | |||
5972 | NODE_NAME_CASE(FSQRT_VL) | |||
5973 | NODE_NAME_CASE(FMA_VL) | |||
5974 | NODE_NAME_CASE(SMIN_VL) | |||
5975 | NODE_NAME_CASE(SMAX_VL) | |||
5976 | NODE_NAME_CASE(UMIN_VL) | |||
5977 | NODE_NAME_CASE(UMAX_VL) | |||
5978 | NODE_NAME_CASE(MULHS_VL) | |||
5979 | NODE_NAME_CASE(MULHU_VL) | |||
5980 | NODE_NAME_CASE(FP_TO_SINT_VL) | |||
5981 | NODE_NAME_CASE(FP_TO_UINT_VL) | |||
5982 | NODE_NAME_CASE(SINT_TO_FP_VL) | |||
5983 | NODE_NAME_CASE(UINT_TO_FP_VL) | |||
5984 | NODE_NAME_CASE(FP_EXTEND_VL) | |||
5985 | NODE_NAME_CASE(FP_ROUND_VL) | |||
5986 | NODE_NAME_CASE(SETCC_VL) | |||
5987 | NODE_NAME_CASE(VSELECT_VL) | |||
5988 | NODE_NAME_CASE(VMAND_VL) | |||
5989 | NODE_NAME_CASE(VMOR_VL) | |||
5990 | NODE_NAME_CASE(VMXOR_VL) | |||
5991 | NODE_NAME_CASE(VMCLR_VL) | |||
5992 | NODE_NAME_CASE(VMSET_VL) | |||
5993 | NODE_NAME_CASE(VRGATHER_VX_VL) | |||
5994 | NODE_NAME_CASE(VRGATHER_VV_VL) | |||
5995 | NODE_NAME_CASE(VRGATHEREI16_VV_VL) | |||
5996 | NODE_NAME_CASE(VSEXT_VL) | |||
5997 | NODE_NAME_CASE(VZEXT_VL) | |||
5998 | NODE_NAME_CASE(VLE_VL) | |||
5999 | NODE_NAME_CASE(VSE_VL) | |||
6000 | } | |||
6001 | // clang-format on | |||
6002 | return nullptr; | |||
6003 | #undef NODE_NAME_CASE | |||
6004 | } | |||
6005 | ||||
6006 | /// getConstraintType - Given a constraint letter, return the type of | |||
6007 | /// constraint it is for this target. | |||
6008 | RISCVTargetLowering::ConstraintType | |||
6009 | RISCVTargetLowering::getConstraintType(StringRef Constraint) const { | |||
6010 | if (Constraint.size() == 1) { | |||
6011 | switch (Constraint[0]) { | |||
6012 | default: | |||
6013 | break; | |||
6014 | case 'f': | |||
6015 | return C_RegisterClass; | |||
6016 | case 'I': | |||
6017 | case 'J': | |||
6018 | case 'K': | |||
6019 | return C_Immediate; | |||
6020 | case 'A': | |||
6021 | return C_Memory; | |||
6022 | } | |||
6023 | } | |||
6024 | return TargetLowering::getConstraintType(Constraint); | |||
6025 | } | |||
6026 | ||||
6027 | std::pair<unsigned, const TargetRegisterClass *> | |||
6028 | RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, | |||
6029 | StringRef Constraint, | |||
6030 | MVT VT) const { | |||
6031 | // First, see if this is a constraint that directly corresponds to a | |||
6032 | // RISCV register class. | |||
6033 | if (Constraint.size() == 1) { | |||
6034 | switch (Constraint[0]) { | |||
6035 | case 'r': | |||
6036 | return std::make_pair(0U, &RISCV::GPRRegClass); | |||
6037 | case 'f': | |||
6038 | if (Subtarget.hasStdExtZfh() && VT == MVT::f16) | |||
6039 | return std::make_pair(0U, &RISCV::FPR16RegClass); | |||
6040 | if (Subtarget.hasStdExtF() && VT == MVT::f32) | |||
6041 | return std::make_pair(0U, &RISCV::FPR32RegClass); | |||
6042 | if (Subtarget.hasStdExtD() && VT == MVT::f64) | |||
6043 | return std::make_pair(0U, &RISCV::FPR64RegClass); | |||
6044 | break; | |||
6045 | default: | |||
6046 | break; | |||
6047 | } | |||
6048 | } | |||
6049 | ||||
6050 | // Clang will correctly decode the usage of register name aliases into their | |||
6051 | // official names. However, other frontends like `rustc` do not. This allows | |||
6052 | // users of these frontends to use the ABI names for registers in LLVM-style | |||
6053 | // register constraints. | |||
6054 | unsigned XRegFromAlias = StringSwitch<unsigned>(Constraint.lower()) | |||
6055 | .Case("{zero}", RISCV::X0) | |||
6056 | .Case("{ra}", RISCV::X1) | |||
6057 | .Case("{sp}", RISCV::X2) | |||
6058 | .Case("{gp}", RISCV::X3) | |||
6059 | .Case("{tp}", RISCV::X4) | |||
6060 | .Case("{t0}", RISCV::X5) | |||
6061 | .Case("{t1}", RISCV::X6) | |||
6062 | .Case("{t2}", RISCV::X7) | |||
6063 | .Cases("{s0}", "{fp}", RISCV::X8) | |||
6064 | .Case("{s1}", RISCV::X9) | |||
6065 | .Case("{a0}", RISCV::X10) | |||
6066 | .Case("{a1}", RISCV::X11) | |||
6067 | .Case("{a2}", RISCV::X12) | |||
6068 | .Case("{a3}", RISCV::X13) | |||
6069 | .Case("{a4}", RISCV::X14) | |||
6070 | .Case("{a5}", RISCV::X15) | |||
6071 | .Case("{a6}", RISCV::X16) | |||
6072 | .Case("{a7}", RISCV::X17) | |||
6073 | .Case("{s2}", RISCV::X18) | |||
6074 | .Case("{s3}", RISCV::X19) | |||
6075 | .Case("{s4}", RISCV::X20) | |||
6076 | .Case("{s5}", RISCV::X21) | |||
6077 | .Case("{s6}", RISCV::X22) | |||
6078 | .Case("{s7}", RISCV::X23) | |||
6079 | .Case("{s8}", RISCV::X24) | |||
6080 | .Case("{s9}", RISCV::X25) | |||
6081 | .Case("{s10}", RISCV::X26) | |||
6082 | .Case("{s11}", RISCV::X27) | |||
6083 | .Case("{t3}", RISCV::X28) | |||
6084 | .Case("{t4}", RISCV::X29) | |||
6085 | .Case("{t5}", RISCV::X30) | |||
6086 | .Case("{t6}", RISCV::X31) | |||
6087 | .Default(RISCV::NoRegister); | |||
6088 | if (XRegFromAlias != RISCV::NoRegister) | |||
6089 | return std::make_pair(XRegFromAlias, &RISCV::GPRRegClass); | |||
6090 | ||||
6091 | // Since TargetLowering::getRegForInlineAsmConstraint uses the name of the | |||
6092 | // TableGen record rather than the AsmName to choose registers for InlineAsm | |||
6093 | // constraints, plus we want to match those names to the widest floating point | |||
6094 | // register type available, manually select floating point registers here. | |||
6095 | // | |||
6096 | // The second case is the ABI name of the register, so that frontends can also | |||
6097 | // use the ABI names in register constraint lists. | |||
6098 | if (Subtarget.hasStdExtF()) { | |||
6099 | unsigned FReg = StringSwitch<unsigned>(Constraint.lower()) | |||
6100 | .Cases("{f0}", "{ft0}", RISCV::F0_F) | |||
6101 | .Cases("{f1}", "{ft1}", RISCV::F1_F) | |||
6102 | .Cases("{f2}", "{ft2}", RISCV::F2_F) | |||
6103 | .Cases("{f3}", "{ft3}", RISCV::F3_F) | |||
6104 | .Cases("{f4}", "{ft4}", RISCV::F4_F) | |||
6105 | .Cases("{f5}", "{ft5}", RISCV::F5_F) | |||
6106 | .Cases("{f6}", "{ft6}", RISCV::F6_F) | |||
6107 | .Cases("{f7}", "{ft7}", RISCV::F7_F) | |||
6108 | .Cases("{f8}", "{fs0}", RISCV::F8_F) | |||
6109 | .Cases("{f9}", "{fs1}", RISCV::F9_F) | |||
6110 | .Cases("{f10}", "{fa0}", RISCV::F10_F) | |||
6111 | .Cases("{f11}", "{fa1}", RISCV::F11_F) | |||
6112 | .Cases("{f12}", "{fa2}", RISCV::F12_F) | |||
6113 | .Cases("{f13}", "{fa3}", RISCV::F13_F) | |||
6114 | .Cases("{f14}", "{fa4}", RISCV::F14_F) | |||
6115 | .Cases("{f15}", "{fa5}", RISCV::F15_F) | |||
6116 | .Cases("{f16}", "{fa6}", RISCV::F16_F) | |||
6117 | .Cases("{f17}", "{fa7}", RISCV::F17_F) | |||
6118 | .Cases("{f18}", "{fs2}", RISCV::F18_F) | |||
6119 | .Cases("{f19}", "{fs3}", RISCV::F19_F) | |||
6120 | .Cases("{f20}", "{fs4}", RISCV::F20_F) | |||
6121 | .Cases("{f21}", "{fs5}", RISCV::F21_F) | |||
6122 | .Cases("{f22}", "{fs6}", RISCV::F22_F) | |||
6123 | .Cases("{f23}", "{fs7}", RISCV::F23_F) | |||
6124 | .Cases("{f24}", "{fs8}", RISCV::F24_F) | |||
6125 | .Cases("{f25}", "{fs9}", RISCV::F25_F) | |||
6126 | .Cases("{f26}", "{fs10}", RISCV::F26_F) | |||
6127 | .Cases("{f27}", "{fs11}", RISCV::F27_F) | |||
6128 | .Cases("{f28}", "{ft8}", RISCV::F28_F) | |||
6129 | .Cases("{f29}", "{ft9}", RISCV::F29_F) | |||
6130 | .Cases("{f30}", "{ft10}", RISCV::F30_F) | |||
6131 | .Cases("{f31}", "{ft11}", RISCV::F31_F) | |||
6132 | .Default(RISCV::NoRegister); | |||
6133 | if (FReg != RISCV::NoRegister) { | |||
6134 | assert(RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && "Unknown fp-reg")((RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && "Unknown fp-reg") ? static_cast<void> (0) : __assert_fail ("RISCV::F0_F <= FReg && FReg <= RISCV::F31_F && \"Unknown fp-reg\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 6134, __PRETTY_FUNCTION__)); | |||
6135 | if (Subtarget.hasStdExtD()) { | |||
6136 | unsigned RegNo = FReg - RISCV::F0_F; | |||
6137 | unsigned DReg = RISCV::F0_D + RegNo; | |||
6138 | return std::make_pair(DReg, &RISCV::FPR64RegClass); | |||
6139 | } | |||
6140 | return std::make_pair(FReg, &RISCV::FPR32RegClass); | |||
6141 | } | |||
6142 | } | |||
6143 | ||||
6144 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); | |||
6145 | } | |||
6146 | ||||
6147 | unsigned | |||
6148 | RISCVTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const { | |||
6149 | // Currently only support length 1 constraints. | |||
6150 | if (ConstraintCode.size() == 1) { | |||
6151 | switch (ConstraintCode[0]) { | |||
6152 | case 'A': | |||
6153 | return InlineAsm::Constraint_A; | |||
6154 | default: | |||
6155 | break; | |||
6156 | } | |||
6157 | } | |||
6158 | ||||
6159 | return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); | |||
6160 | } | |||
6161 | ||||
6162 | void RISCVTargetLowering::LowerAsmOperandForConstraint( | |||
6163 | SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops, | |||
6164 | SelectionDAG &DAG) const { | |||
6165 | // Currently only support length 1 constraints. | |||
6166 | if (Constraint.length() == 1) { | |||
6167 | switch (Constraint[0]) { | |||
6168 | case 'I': | |||
6169 | // Validate & create a 12-bit signed immediate operand. | |||
6170 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { | |||
6171 | uint64_t CVal = C->getSExtValue(); | |||
6172 | if (isInt<12>(CVal)) | |||
6173 | Ops.push_back( | |||
6174 | DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT())); | |||
6175 | } | |||
6176 | return; | |||
6177 | case 'J': | |||
6178 | // Validate & create an integer zero operand. | |||
6179 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) | |||
6180 | if (C->getZExtValue() == 0) | |||
6181 | Ops.push_back( | |||
6182 | DAG.getTargetConstant(0, SDLoc(Op), Subtarget.getXLenVT())); | |||
6183 | return; | |||
6184 | case 'K': | |||
6185 | // Validate & create a 5-bit unsigned immediate operand. | |||
6186 | if (auto *C = dyn_cast<ConstantSDNode>(Op)) { | |||
6187 | uint64_t CVal = C->getZExtValue(); | |||
6188 | if (isUInt<5>(CVal)) | |||
6189 | Ops.push_back( | |||
6190 | DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT())); | |||
6191 | } | |||
6192 | return; | |||
6193 | default: | |||
6194 | break; | |||
6195 | } | |||
6196 | } | |||
6197 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); | |||
6198 | } | |||
6199 | ||||
6200 | Instruction *RISCVTargetLowering::emitLeadingFence(IRBuilder<> &Builder, | |||
6201 | Instruction *Inst, | |||
6202 | AtomicOrdering Ord) const { | |||
6203 | if (isa<LoadInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent) | |||
6204 | return Builder.CreateFence(Ord); | |||
6205 | if (isa<StoreInst>(Inst) && isReleaseOrStronger(Ord)) | |||
6206 | return Builder.CreateFence(AtomicOrdering::Release); | |||
6207 | return nullptr; | |||
6208 | } | |||
6209 | ||||
6210 | Instruction *RISCVTargetLowering::emitTrailingFence(IRBuilder<> &Builder, | |||
6211 | Instruction *Inst, | |||
6212 | AtomicOrdering Ord) const { | |||
6213 | if (isa<LoadInst>(Inst) && isAcquireOrStronger(Ord)) | |||
6214 | return Builder.CreateFence(AtomicOrdering::Acquire); | |||
6215 | return nullptr; | |||
6216 | } | |||
6217 | ||||
6218 | TargetLowering::AtomicExpansionKind | |||
6219 | RISCVTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { | |||
6220 | // atomicrmw {fadd,fsub} must be expanded to use compare-exchange, as floating | |||
6221 | // point operations can't be used in an lr/sc sequence without breaking the | |||
6222 | // forward-progress guarantee. | |||
6223 | if (AI->isFloatingPointOperation()) | |||
6224 | return AtomicExpansionKind::CmpXChg; | |||
6225 | ||||
6226 | unsigned Size = AI->getType()->getPrimitiveSizeInBits(); | |||
6227 | if (Size == 8 || Size == 16) | |||
6228 | return AtomicExpansionKind::MaskedIntrinsic; | |||
6229 | return AtomicExpansionKind::None; | |||
6230 | } | |||
6231 | ||||
6232 | static Intrinsic::ID | |||
6233 | getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen, AtomicRMWInst::BinOp BinOp) { | |||
6234 | if (XLen == 32) { | |||
6235 | switch (BinOp) { | |||
6236 | default: | |||
6237 | llvm_unreachable("Unexpected AtomicRMW BinOp")::llvm::llvm_unreachable_internal("Unexpected AtomicRMW BinOp" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 6237); | |||
6238 | case AtomicRMWInst::Xchg: | |||
6239 | return Intrinsic::riscv_masked_atomicrmw_xchg_i32; | |||
6240 | case AtomicRMWInst::Add: | |||
6241 | return Intrinsic::riscv_masked_atomicrmw_add_i32; | |||
6242 | case AtomicRMWInst::Sub: | |||
6243 | return Intrinsic::riscv_masked_atomicrmw_sub_i32; | |||
6244 | case AtomicRMWInst::Nand: | |||
6245 | return Intrinsic::riscv_masked_atomicrmw_nand_i32; | |||
6246 | case AtomicRMWInst::Max: | |||
6247 | return Intrinsic::riscv_masked_atomicrmw_max_i32; | |||
6248 | case AtomicRMWInst::Min: | |||
6249 | return Intrinsic::riscv_masked_atomicrmw_min_i32; | |||
6250 | case AtomicRMWInst::UMax: | |||
6251 | return Intrinsic::riscv_masked_atomicrmw_umax_i32; | |||
6252 | case AtomicRMWInst::UMin: | |||
6253 | return Intrinsic::riscv_masked_atomicrmw_umin_i32; | |||
6254 | } | |||
6255 | } | |||
6256 | ||||
6257 | if (XLen == 64) { | |||
6258 | switch (BinOp) { | |||
6259 | default: | |||
6260 | llvm_unreachable("Unexpected AtomicRMW BinOp")::llvm::llvm_unreachable_internal("Unexpected AtomicRMW BinOp" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 6260); | |||
6261 | case AtomicRMWInst::Xchg: | |||
6262 | return Intrinsic::riscv_masked_atomicrmw_xchg_i64; | |||
6263 | case AtomicRMWInst::Add: | |||
6264 | return Intrinsic::riscv_masked_atomicrmw_add_i64; | |||
6265 | case AtomicRMWInst::Sub: | |||
6266 | return Intrinsic::riscv_masked_atomicrmw_sub_i64; | |||
6267 | case AtomicRMWInst::Nand: | |||
6268 | return Intrinsic::riscv_masked_atomicrmw_nand_i64; | |||
6269 | case AtomicRMWInst::Max: | |||
6270 | return Intrinsic::riscv_masked_atomicrmw_max_i64; | |||
6271 | case AtomicRMWInst::Min: | |||
6272 | return Intrinsic::riscv_masked_atomicrmw_min_i64; | |||
6273 | case AtomicRMWInst::UMax: | |||
6274 | return Intrinsic::riscv_masked_atomicrmw_umax_i64; | |||
6275 | case AtomicRMWInst::UMin: | |||
6276 | return Intrinsic::riscv_masked_atomicrmw_umin_i64; | |||
6277 | } | |||
6278 | } | |||
6279 | ||||
6280 | llvm_unreachable("Unexpected XLen\n")::llvm::llvm_unreachable_internal("Unexpected XLen\n", "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/lib/Target/RISCV/RISCVISelLowering.cpp" , 6280); | |||
6281 | } | |||
6282 | ||||
6283 | Value *RISCVTargetLowering::emitMaskedAtomicRMWIntrinsic( | |||
6284 | IRBuilder<> &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr, | |||
6285 | Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const { | |||
6286 | unsigned XLen = Subtarget.getXLen(); | |||
6287 | Value *Ordering = | |||
6288 | Builder.getIntN(XLen, static_cast<uint64_t>(AI->getOrdering())); | |||
6289 | Type *Tys[] = {AlignedAddr->getType()}; | |||
6290 | Function *LrwOpScwLoop = Intrinsic::getDeclaration( | |||
6291 | AI->getModule(), | |||
6292 | getIntrinsicForMaskedAtomicRMWBinOp(XLen, AI->getOperation()), Tys); | |||
6293 | ||||
6294 | if (XLen == 64) { | |||
6295 | Incr = Builder.CreateSExt(Incr, Builder.getInt64Ty()); | |||
6296 | Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty()); | |||
6297 | ShiftAmt = Builder.CreateSExt(ShiftAmt, Builder.getInt64Ty()); | |||
6298 | } | |||
6299 | ||||
6300 | Value *Result; | |||
6301 | ||||
6302 | // Must pass the shift amount needed to sign extend the loaded value prior | |||
6303 | // to performing a signed comparison for min/max. ShiftAmt is the number of | |||
6304 | // bits to shift the value into position. Pass XLen-ShiftAmt-ValWidth, which | |||
6305 | // is the number of bits to left+right shift the value in order to | |||
6306 | // sign-extend. | |||
6307 | if (AI->getOperation() == AtomicRMWInst::Min || | |||
6308 | AI->getOperation() == AtomicRMWInst::Max) { | |||
6309 | const DataLayout &DL = AI->getModule()->getDataLayout(); | |||
6310 | unsigned ValWidth = | |||
6311 | DL.getTypeStoreSizeInBits(AI->getValOperand()->getType()); | |||
6312 | Value *SextShamt = | |||
6313 | Builder.CreateSub(Builder.getIntN(XLen, XLen - ValWidth), ShiftAmt); | |||
6314 | Result = Builder.CreateCall(LrwOpScwLoop, | |||
6315 | {AlignedAddr, Incr, Mask, SextShamt, Ordering}); | |||
6316 | } else { | |||
6317 | Result = | |||
6318 | Builder.CreateCall(LrwOpScwLoop, {AlignedAddr, Incr, Mask, Ordering}); | |||
6319 | } | |||
6320 | ||||
6321 | if (XLen == 64) | |||
6322 | Result = Builder.CreateTrunc(Result, Builder.getInt32Ty()); | |||
6323 | return Result; | |||
6324 | } | |||
6325 | ||||
6326 | TargetLowering::AtomicExpansionKind | |||
6327 | RISCVTargetLowering::shouldExpandAtomicCmpXchgInIR( | |||
6328 | AtomicCmpXchgInst *CI) const { | |||
6329 | unsigned Size = CI->getCompareOperand()->getType()->getPrimitiveSizeInBits(); | |||
6330 | if (Size == 8 || Size == 16) | |||
6331 | return AtomicExpansionKind::MaskedIntrinsic; | |||
6332 | return AtomicExpansionKind::None; | |||
6333 | } | |||
6334 | ||||
6335 | Value *RISCVTargetLowering::emitMaskedAtomicCmpXchgIntrinsic( | |||
6336 | IRBuilder<> &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr, | |||
6337 | Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const { | |||
6338 | unsigned XLen = Subtarget.getXLen(); | |||
6339 | Value *Ordering = Builder.getIntN(XLen, static_cast<uint64_t>(Ord)); | |||
6340 | Intrinsic::ID CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i32; | |||
6341 | if (XLen == 64) { | |||
6342 | CmpVal = Builder.CreateSExt(CmpVal, Builder.getInt64Ty()); | |||
6343 | NewVal = Builder.CreateSExt(NewVal, Builder.getInt64Ty()); | |||
6344 | Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty()); | |||
6345 | CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i64; | |||
6346 | } | |||
6347 | Type *Tys[] = {AlignedAddr->getType()}; | |||
6348 | Function *MaskedCmpXchg = | |||
6349 | Intrinsic::getDeclaration(CI->getModule(), CmpXchgIntrID, Tys); | |||
6350 | Value *Result = Builder.CreateCall( | |||
6351 | MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering}); | |||
6352 | if (XLen == 64) | |||
6353 | Result = Builder.CreateTrunc(Result, Builder.getInt32Ty()); | |||
6354 | return Result; | |||
6355 | } | |||
6356 | ||||
6357 | bool RISCVTargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF, | |||
6358 | EVT VT) const { | |||
6359 | VT = VT.getScalarType(); | |||
6360 | ||||
6361 | if (!VT.isSimple()) | |||
6362 | return false; | |||
6363 | ||||
6364 | switch (VT.getSimpleVT().SimpleTy) { | |||
6365 | case MVT::f16: | |||
6366 | return Subtarget.hasStdExtZfh(); | |||
6367 | case MVT::f32: | |||
6368 | return Subtarget.hasStdExtF(); | |||
6369 | case MVT::f64: | |||
6370 | return Subtarget.hasStdExtD(); | |||
6371 | default: | |||
6372 | break; | |||
6373 | } | |||
6374 | ||||
6375 | return false; | |||
6376 | } | |||
6377 | ||||
6378 | Register RISCVTargetLowering::getExceptionPointerRegister( | |||
6379 | const Constant *PersonalityFn) const { | |||
6380 | return RISCV::X10; | |||
6381 | } | |||
6382 | ||||
6383 | Register RISCVTargetLowering::getExceptionSelectorRegister( | |||
6384 | const Constant *PersonalityFn) const { | |||
6385 | return RISCV::X11; | |||
6386 | } | |||
6387 | ||||
6388 | bool RISCVTargetLowering::shouldExtendTypeInLibCall(EVT Type) const { | |||
6389 | // Return false to suppress the unnecessary extensions if the LibCall | |||
6390 | // arguments or return value is f32 type for LP64 ABI. | |||
6391 | RISCVABI::ABI ABI = Subtarget.getTargetABI(); | |||
6392 | if (ABI == RISCVABI::ABI_LP64 && (Type == MVT::f32)) | |||
6393 | return false; | |||
6394 | ||||
6395 | return true; | |||
6396 | } | |||
6397 | ||||
6398 | bool RISCVTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const { | |||
6399 | if (Subtarget.is64Bit() && Type == MVT::i32) | |||
6400 | return true; | |||
6401 | ||||
6402 | return IsSigned; | |||
6403 | } | |||
6404 | ||||
6405 | bool RISCVTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT, | |||
6406 | SDValue C) const { | |||
6407 | // Check integral scalar types. | |||
6408 | if (VT.isScalarInteger()) { | |||
6409 | // Omit the optimization if the sub target has the M extension and the data | |||
6410 | // size exceeds XLen. | |||
6411 | if (Subtarget.hasStdExtM() && VT.getSizeInBits() > Subtarget.getXLen()) | |||
6412 | return false; | |||
6413 | if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) { | |||
6414 | // Break the MUL to a SLLI and an ADD/SUB. | |||
6415 | const APInt &Imm = ConstNode->getAPIntValue(); | |||
6416 | if ((Imm + 1).isPowerOf2() || (Imm - 1).isPowerOf2() || | |||
6417 | (1 - Imm).isPowerOf2() || (-1 - Imm).isPowerOf2()) | |||
6418 | return true; | |||
6419 | // Omit the following optimization if the sub target has the M extension | |||
6420 | // and the data size >= XLen. | |||
6421 | if (Subtarget.hasStdExtM() && VT.getSizeInBits() >= Subtarget.getXLen()) | |||
6422 | return false; | |||
6423 | // Break the MUL to two SLLI instructions and an ADD/SUB, if Imm needs | |||
6424 | // a pair of LUI/ADDI. | |||
6425 | if (!Imm.isSignedIntN(12) && Imm.countTrailingZeros() < 12) { | |||
6426 | APInt ImmS = Imm.ashr(Imm.countTrailingZeros()); | |||
6427 | if ((ImmS + 1).isPowerOf2() || (ImmS - 1).isPowerOf2() || | |||
6428 | (1 - ImmS).isPowerOf2()) | |||
6429 | return true; | |||
6430 | } | |||
6431 | } | |||
6432 | } | |||
6433 | ||||
6434 | return false; | |||
6435 | } | |||
6436 | ||||
6437 | bool RISCVTargetLowering::useRVVForFixedLengthVectorVT(MVT VT) const { | |||
6438 | if (!Subtarget.useRVVForFixedLengthVectors()) | |||
6439 | return false; | |||
6440 | ||||
6441 | if (!VT.isFixedLengthVector()) | |||
6442 | return false; | |||
6443 | ||||
6444 | // Don't use RVV for vectors we cannot scalarize if required. | |||
6445 | switch (VT.getVectorElementType().SimpleTy) { | |||
6446 | // i1 is supported but has different rules. | |||
6447 | default: | |||
6448 | return false; | |||
6449 | case MVT::i1: | |||
6450 | // Masks can only use a single register. | |||
6451 | if (VT.getVectorNumElements() > Subtarget.getMinRVVVectorSizeInBits()) | |||
6452 | return false; | |||
6453 | break; | |||
6454 | case MVT::i8: | |||
6455 | case MVT::i16: | |||
6456 | case MVT::i32: | |||
6457 | case MVT::i64: | |||
6458 | break; | |||
6459 | case MVT::f16: | |||
6460 | if (!Subtarget.hasStdExtZfh()) | |||
6461 | return false; | |||
6462 | break; | |||
6463 | case MVT::f32: | |||
6464 | if (!Subtarget.hasStdExtF()) | |||
6465 | return false; | |||
6466 | break; | |||
6467 | case MVT::f64: | |||
6468 | if (!Subtarget.hasStdExtD()) | |||
6469 | return false; | |||
6470 | break; | |||
6471 | } | |||
6472 | ||||
6473 | unsigned LMul = Subtarget.getLMULForFixedLengthVector(VT); | |||
6474 | // Don't use RVV for types that don't fit. | |||
6475 | if (LMul > Subtarget.getMaxLMULForFixedLengthVectors()) | |||
6476 | return false; | |||
6477 | ||||
6478 | // TODO: Perhaps an artificial restriction, but worth having whilst getting | |||
6479 | // the base fixed length RVV support in place. | |||
6480 | if (!VT.isPow2VectorType()) | |||
6481 | return false; | |||
6482 | ||||
6483 | return true; | |||
6484 | } | |||
6485 | ||||
6486 | bool RISCVTargetLowering::allowsMisalignedMemoryAccesses( | |||
6487 | EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags, | |||
6488 | bool *Fast) const { | |||
6489 | if (!VT.isScalableVector()) | |||
6490 | return false; | |||
6491 | ||||
6492 | EVT ElemVT = VT.getVectorElementType(); | |||
6493 | if (Alignment >= ElemVT.getStoreSize()) { | |||
6494 | if (Fast) | |||
6495 | *Fast = true; | |||
6496 | return true; | |||
6497 | } | |||
6498 | ||||
6499 | return false; | |||
6500 | } | |||
6501 | ||||
6502 | #define GET_REGISTER_MATCHER | |||
6503 | #include "RISCVGenAsmMatcher.inc" | |||
6504 | ||||
6505 | Register | |||
6506 | RISCVTargetLowering::getRegisterByName(const char *RegName, LLT VT, | |||
6507 | const MachineFunction &MF) const { | |||
6508 | Register Reg = MatchRegisterAltName(RegName); | |||
6509 | if (Reg == RISCV::NoRegister) | |||
6510 | Reg = MatchRegisterName(RegName); | |||
6511 | if (Reg == RISCV::NoRegister) | |||
6512 | report_fatal_error( | |||
6513 | Twine("Invalid register name \"" + StringRef(RegName) + "\".")); | |||
6514 | BitVector ReservedRegs = Subtarget.getRegisterInfo()->getReservedRegs(MF); | |||
6515 | if (!ReservedRegs.test(Reg) && !Subtarget.isRegisterReservedByUser(Reg)) | |||
6516 | report_fatal_error(Twine("Trying to obtain non-reserved register \"" + | |||
6517 | StringRef(RegName) + "\".")); | |||
6518 | return Reg; | |||
6519 | } | |||
6520 | ||||
6521 | namespace llvm { | |||
6522 | namespace RISCVVIntrinsicsTable { | |||
6523 | ||||
6524 | #define GET_RISCVVIntrinsicsTable_IMPL | |||
6525 | #include "RISCVGenSearchableTables.inc" | |||
6526 | ||||
6527 | } // namespace RISCVVIntrinsicsTable | |||
6528 | ||||
6529 | } // namespace llvm |
1 | //===- llvm/Support/Casting.h - Allow flexible, checked, casts --*- C++ -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the isa<X>(), cast<X>(), dyn_cast<X>(), cast_or_null<X>(), |
10 | // and dyn_cast_or_null<X>() templates. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #ifndef LLVM_SUPPORT_CASTING_H |
15 | #define LLVM_SUPPORT_CASTING_H |
16 | |
17 | #include "llvm/Support/Compiler.h" |
18 | #include "llvm/Support/type_traits.h" |
19 | #include <cassert> |
20 | #include <memory> |
21 | #include <type_traits> |
22 | |
23 | namespace llvm { |
24 | |
25 | //===----------------------------------------------------------------------===// |
26 | // isa<x> Support Templates |
27 | //===----------------------------------------------------------------------===// |
28 | |
29 | // Define a template that can be specialized by smart pointers to reflect the |
30 | // fact that they are automatically dereferenced, and are not involved with the |
31 | // template selection process... the default implementation is a noop. |
32 | // |
33 | template<typename From> struct simplify_type { |
34 | using SimpleType = From; // The real type this represents... |
35 | |
36 | // An accessor to get the real value... |
37 | static SimpleType &getSimplifiedValue(From &Val) { return Val; } |
38 | }; |
39 | |
40 | template<typename From> struct simplify_type<const From> { |
41 | using NonConstSimpleType = typename simplify_type<From>::SimpleType; |
42 | using SimpleType = |
43 | typename add_const_past_pointer<NonConstSimpleType>::type; |
44 | using RetType = |
45 | typename add_lvalue_reference_if_not_pointer<SimpleType>::type; |
46 | |
47 | static RetType getSimplifiedValue(const From& Val) { |
48 | return simplify_type<From>::getSimplifiedValue(const_cast<From&>(Val)); |
49 | } |
50 | }; |
51 | |
52 | // The core of the implementation of isa<X> is here; To and From should be |
53 | // the names of classes. This template can be specialized to customize the |
54 | // implementation of isa<> without rewriting it from scratch. |
55 | template <typename To, typename From, typename Enabler = void> |
56 | struct isa_impl { |
57 | static inline bool doit(const From &Val) { |
58 | return To::classof(&Val); |
59 | } |
60 | }; |
61 | |
62 | /// Always allow upcasts, and perform no dynamic check for them. |
63 | template <typename To, typename From> |
64 | struct isa_impl<To, From, std::enable_if_t<std::is_base_of<To, From>::value>> { |
65 | static inline bool doit(const From &) { return true; } |
66 | }; |
67 | |
68 | template <typename To, typename From> struct isa_impl_cl { |
69 | static inline bool doit(const From &Val) { |
70 | return isa_impl<To, From>::doit(Val); |
71 | } |
72 | }; |
73 | |
74 | template <typename To, typename From> struct isa_impl_cl<To, const From> { |
75 | static inline bool doit(const From &Val) { |
76 | return isa_impl<To, From>::doit(Val); |
77 | } |
78 | }; |
79 | |
80 | template <typename To, typename From> |
81 | struct isa_impl_cl<To, const std::unique_ptr<From>> { |
82 | static inline bool doit(const std::unique_ptr<From> &Val) { |
83 | assert(Val && "isa<> used on a null pointer")((Val && "isa<> used on a null pointer") ? static_cast <void> (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 83, __PRETTY_FUNCTION__)); |
84 | return isa_impl_cl<To, From>::doit(*Val); |
85 | } |
86 | }; |
87 | |
88 | template <typename To, typename From> struct isa_impl_cl<To, From*> { |
89 | static inline bool doit(const From *Val) { |
90 | assert(Val && "isa<> used on a null pointer")((Val && "isa<> used on a null pointer") ? static_cast <void> (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 90, __PRETTY_FUNCTION__)); |
91 | return isa_impl<To, From>::doit(*Val); |
92 | } |
93 | }; |
94 | |
95 | template <typename To, typename From> struct isa_impl_cl<To, From*const> { |
96 | static inline bool doit(const From *Val) { |
97 | assert(Val && "isa<> used on a null pointer")((Val && "isa<> used on a null pointer") ? static_cast <void> (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 97, __PRETTY_FUNCTION__)); |
98 | return isa_impl<To, From>::doit(*Val); |
99 | } |
100 | }; |
101 | |
102 | template <typename To, typename From> struct isa_impl_cl<To, const From*> { |
103 | static inline bool doit(const From *Val) { |
104 | assert(Val && "isa<> used on a null pointer")((Val && "isa<> used on a null pointer") ? static_cast <void> (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 104, __PRETTY_FUNCTION__)); |
105 | return isa_impl<To, From>::doit(*Val); |
106 | } |
107 | }; |
108 | |
109 | template <typename To, typename From> struct isa_impl_cl<To, const From*const> { |
110 | static inline bool doit(const From *Val) { |
111 | assert(Val && "isa<> used on a null pointer")((Val && "isa<> used on a null pointer") ? static_cast <void> (0) : __assert_fail ("Val && \"isa<> used on a null pointer\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 111, __PRETTY_FUNCTION__)); |
112 | return isa_impl<To, From>::doit(*Val); |
113 | } |
114 | }; |
115 | |
116 | template<typename To, typename From, typename SimpleFrom> |
117 | struct isa_impl_wrap { |
118 | // When From != SimplifiedType, we can simplify the type some more by using |
119 | // the simplify_type template. |
120 | static bool doit(const From &Val) { |
121 | return isa_impl_wrap<To, SimpleFrom, |
122 | typename simplify_type<SimpleFrom>::SimpleType>::doit( |
123 | simplify_type<const From>::getSimplifiedValue(Val)); |
124 | } |
125 | }; |
126 | |
127 | template<typename To, typename FromTy> |
128 | struct isa_impl_wrap<To, FromTy, FromTy> { |
129 | // When From == SimpleType, we are as simple as we are going to get. |
130 | static bool doit(const FromTy &Val) { |
131 | return isa_impl_cl<To,FromTy>::doit(Val); |
132 | } |
133 | }; |
134 | |
135 | // isa<X> - Return true if the parameter to the template is an instance of one |
136 | // of the template type arguments. Used like this: |
137 | // |
138 | // if (isa<Type>(myVal)) { ... } |
139 | // if (isa<Type0, Type1, Type2>(myVal)) { ... } |
140 | // |
141 | template <class X, class Y> LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isa(const Y &Val) { |
142 | return isa_impl_wrap<X, const Y, |
143 | typename simplify_type<const Y>::SimpleType>::doit(Val); |
144 | } |
145 | |
146 | template <typename First, typename Second, typename... Rest, typename Y> |
147 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isa(const Y &Val) { |
148 | return isa<First>(Val) || isa<Second, Rest...>(Val); |
149 | } |
150 | |
151 | // isa_and_nonnull<X> - Functionally identical to isa, except that a null value |
152 | // is accepted. |
153 | // |
154 | template <typename... X, class Y> |
155 | LLVM_NODISCARD[[clang::warn_unused_result]] inline bool isa_and_nonnull(const Y &Val) { |
156 | if (!Val) |
157 | return false; |
158 | return isa<X...>(Val); |
159 | } |
160 | |
161 | //===----------------------------------------------------------------------===// |
162 | // cast<x> Support Templates |
163 | //===----------------------------------------------------------------------===// |
164 | |
165 | template<class To, class From> struct cast_retty; |
166 | |
167 | // Calculate what type the 'cast' function should return, based on a requested |
168 | // type of To and a source type of From. |
169 | template<class To, class From> struct cast_retty_impl { |
170 | using ret_type = To &; // Normal case, return Ty& |
171 | }; |
172 | template<class To, class From> struct cast_retty_impl<To, const From> { |
173 | using ret_type = const To &; // Normal case, return Ty& |
174 | }; |
175 | |
176 | template<class To, class From> struct cast_retty_impl<To, From*> { |
177 | using ret_type = To *; // Pointer arg case, return Ty* |
178 | }; |
179 | |
180 | template<class To, class From> struct cast_retty_impl<To, const From*> { |
181 | using ret_type = const To *; // Constant pointer arg case, return const Ty* |
182 | }; |
183 | |
184 | template<class To, class From> struct cast_retty_impl<To, const From*const> { |
185 | using ret_type = const To *; // Constant pointer arg case, return const Ty* |
186 | }; |
187 | |
188 | template <class To, class From> |
189 | struct cast_retty_impl<To, std::unique_ptr<From>> { |
190 | private: |
191 | using PointerType = typename cast_retty_impl<To, From *>::ret_type; |
192 | using ResultType = std::remove_pointer_t<PointerType>; |
193 | |
194 | public: |
195 | using ret_type = std::unique_ptr<ResultType>; |
196 | }; |
197 | |
198 | template<class To, class From, class SimpleFrom> |
199 | struct cast_retty_wrap { |
200 | // When the simplified type and the from type are not the same, use the type |
201 | // simplifier to reduce the type, then reuse cast_retty_impl to get the |
202 | // resultant type. |
203 | using ret_type = typename cast_retty<To, SimpleFrom>::ret_type; |
204 | }; |
205 | |
206 | template<class To, class FromTy> |
207 | struct cast_retty_wrap<To, FromTy, FromTy> { |
208 | // When the simplified type is equal to the from type, use it directly. |
209 | using ret_type = typename cast_retty_impl<To,FromTy>::ret_type; |
210 | }; |
211 | |
212 | template<class To, class From> |
213 | struct cast_retty { |
214 | using ret_type = typename cast_retty_wrap< |
215 | To, From, typename simplify_type<From>::SimpleType>::ret_type; |
216 | }; |
217 | |
218 | // Ensure the non-simple values are converted using the simplify_type template |
219 | // that may be specialized by smart pointers... |
220 | // |
221 | template<class To, class From, class SimpleFrom> struct cast_convert_val { |
222 | // This is not a simple type, use the template to simplify it... |
223 | static typename cast_retty<To, From>::ret_type doit(From &Val) { |
224 | return cast_convert_val<To, SimpleFrom, |
225 | typename simplify_type<SimpleFrom>::SimpleType>::doit( |
226 | simplify_type<From>::getSimplifiedValue(Val)); |
227 | } |
228 | }; |
229 | |
230 | template<class To, class FromTy> struct cast_convert_val<To,FromTy,FromTy> { |
231 | // This _is_ a simple type, just cast it. |
232 | static typename cast_retty<To, FromTy>::ret_type doit(const FromTy &Val) { |
233 | typename cast_retty<To, FromTy>::ret_type Res2 |
234 | = (typename cast_retty<To, FromTy>::ret_type)const_cast<FromTy&>(Val); |
235 | return Res2; |
236 | } |
237 | }; |
238 | |
239 | template <class X> struct is_simple_type { |
240 | static const bool value = |
241 | std::is_same<X, typename simplify_type<X>::SimpleType>::value; |
242 | }; |
243 | |
244 | // cast<X> - Return the argument parameter cast to the specified type. This |
245 | // casting operator asserts that the type is correct, so it does not return null |
246 | // on failure. It does not allow a null argument (use cast_or_null for that). |
247 | // It is typically used like this: |
248 | // |
249 | // cast<Instruction>(myVal)->getParent() |
250 | // |
251 | template <class X, class Y> |
252 | inline std::enable_if_t<!is_simple_type<Y>::value, |
253 | typename cast_retty<X, const Y>::ret_type> |
254 | cast(const Y &Val) { |
255 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 255, __PRETTY_FUNCTION__)); |
256 | return cast_convert_val< |
257 | X, const Y, typename simplify_type<const Y>::SimpleType>::doit(Val); |
258 | } |
259 | |
260 | template <class X, class Y> |
261 | inline typename cast_retty<X, Y>::ret_type cast(Y &Val) { |
262 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 262, __PRETTY_FUNCTION__)); |
263 | return cast_convert_val<X, Y, |
264 | typename simplify_type<Y>::SimpleType>::doit(Val); |
265 | } |
266 | |
267 | template <class X, class Y> |
268 | inline typename cast_retty<X, Y *>::ret_type cast(Y *Val) { |
269 | assert(isa<X>(Val) && "cast<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 269, __PRETTY_FUNCTION__)); |
270 | return cast_convert_val<X, Y*, |
271 | typename simplify_type<Y*>::SimpleType>::doit(Val); |
272 | } |
273 | |
274 | template <class X, class Y> |
275 | inline typename cast_retty<X, std::unique_ptr<Y>>::ret_type |
276 | cast(std::unique_ptr<Y> &&Val) { |
277 | assert(isa<X>(Val.get()) && "cast<Ty>() argument of incompatible type!")((isa<X>(Val.get()) && "cast<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val.get()) && \"cast<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 277, __PRETTY_FUNCTION__)); |
278 | using ret_type = typename cast_retty<X, std::unique_ptr<Y>>::ret_type; |
279 | return ret_type( |
280 | cast_convert_val<X, Y *, typename simplify_type<Y *>::SimpleType>::doit( |
281 | Val.release())); |
282 | } |
283 | |
284 | // cast_or_null<X> - Functionally identical to cast, except that a null value is |
285 | // accepted. |
286 | // |
287 | template <class X, class Y> |
288 | LLVM_NODISCARD[[clang::warn_unused_result]] inline std::enable_if_t< |
289 | !is_simple_type<Y>::value, typename cast_retty<X, const Y>::ret_type> |
290 | cast_or_null(const Y &Val) { |
291 | if (!Val) |
292 | return nullptr; |
293 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 293, __PRETTY_FUNCTION__)); |
294 | return cast<X>(Val); |
295 | } |
296 | |
297 | template <class X, class Y> |
298 | LLVM_NODISCARD[[clang::warn_unused_result]] inline std::enable_if_t<!is_simple_type<Y>::value, |
299 | typename cast_retty<X, Y>::ret_type> |
300 | cast_or_null(Y &Val) { |
301 | if (!Val) |
302 | return nullptr; |
303 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 303, __PRETTY_FUNCTION__)); |
304 | return cast<X>(Val); |
305 | } |
306 | |
307 | template <class X, class Y> |
308 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type |
309 | cast_or_null(Y *Val) { |
310 | if (!Val) return nullptr; |
311 | assert(isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!")((isa<X>(Val) && "cast_or_null<Ty>() argument of incompatible type!" ) ? static_cast<void> (0) : __assert_fail ("isa<X>(Val) && \"cast_or_null<Ty>() argument of incompatible type!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/Support/Casting.h" , 311, __PRETTY_FUNCTION__)); |
312 | return cast<X>(Val); |
313 | } |
314 | |
315 | template <class X, class Y> |
316 | inline typename cast_retty<X, std::unique_ptr<Y>>::ret_type |
317 | cast_or_null(std::unique_ptr<Y> &&Val) { |
318 | if (!Val) |
319 | return nullptr; |
320 | return cast<X>(std::move(Val)); |
321 | } |
322 | |
323 | // dyn_cast<X> - Return the argument parameter cast to the specified type. This |
324 | // casting operator returns null if the argument is of the wrong type, so it can |
325 | // be used to test for a type as well as cast if successful. This should be |
326 | // used in the context of an if statement like this: |
327 | // |
328 | // if (const Instruction *I = dyn_cast<Instruction>(myVal)) { ... } |
329 | // |
330 | |
331 | template <class X, class Y> |
332 | LLVM_NODISCARD[[clang::warn_unused_result]] inline std::enable_if_t< |
333 | !is_simple_type<Y>::value, typename cast_retty<X, const Y>::ret_type> |
334 | dyn_cast(const Y &Val) { |
335 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
336 | } |
337 | |
338 | template <class X, class Y> |
339 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y>::ret_type dyn_cast(Y &Val) { |
340 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
341 | } |
342 | |
343 | template <class X, class Y> |
344 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type dyn_cast(Y *Val) { |
345 | return isa<X>(Val) ? cast<X>(Val) : nullptr; |
346 | } |
347 | |
348 | // dyn_cast_or_null<X> - Functionally identical to dyn_cast, except that a null |
349 | // value is accepted. |
350 | // |
351 | template <class X, class Y> |
352 | LLVM_NODISCARD[[clang::warn_unused_result]] inline std::enable_if_t< |
353 | !is_simple_type<Y>::value, typename cast_retty<X, const Y>::ret_type> |
354 | dyn_cast_or_null(const Y &Val) { |
355 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
356 | } |
357 | |
358 | template <class X, class Y> |
359 | LLVM_NODISCARD[[clang::warn_unused_result]] inline std::enable_if_t<!is_simple_type<Y>::value, |
360 | typename cast_retty<X, Y>::ret_type> |
361 | dyn_cast_or_null(Y &Val) { |
362 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
363 | } |
364 | |
365 | template <class X, class Y> |
366 | LLVM_NODISCARD[[clang::warn_unused_result]] inline typename cast_retty<X, Y *>::ret_type |
367 | dyn_cast_or_null(Y *Val) { |
368 | return (Val && isa<X>(Val)) ? cast<X>(Val) : nullptr; |
369 | } |
370 | |
371 | // unique_dyn_cast<X> - Given a unique_ptr<Y>, try to return a unique_ptr<X>, |
372 | // taking ownership of the input pointer iff isa<X>(Val) is true. If the |
373 | // cast is successful, From refers to nullptr on exit and the casted value |
374 | // is returned. If the cast is unsuccessful, the function returns nullptr |
375 | // and From is unchanged. |
376 | template <class X, class Y> |
377 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast(std::unique_ptr<Y> &Val) |
378 | -> decltype(cast<X>(Val)) { |
379 | if (!isa<X>(Val)) |
380 | return nullptr; |
381 | return cast<X>(std::move(Val)); |
382 | } |
383 | |
384 | template <class X, class Y> |
385 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast(std::unique_ptr<Y> &&Val) { |
386 | return unique_dyn_cast<X, Y>(Val); |
387 | } |
388 | |
389 | // dyn_cast_or_null<X> - Functionally identical to unique_dyn_cast, except that |
390 | // a null value is accepted. |
391 | template <class X, class Y> |
392 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast_or_null(std::unique_ptr<Y> &Val) |
393 | -> decltype(cast<X>(Val)) { |
394 | if (!Val) |
395 | return nullptr; |
396 | return unique_dyn_cast<X, Y>(Val); |
397 | } |
398 | |
399 | template <class X, class Y> |
400 | LLVM_NODISCARD[[clang::warn_unused_result]] inline auto unique_dyn_cast_or_null(std::unique_ptr<Y> &&Val) { |
401 | return unique_dyn_cast_or_null<X, Y>(Val); |
402 | } |
403 | |
404 | } // end namespace llvm |
405 | |
406 | #endif // LLVM_SUPPORT_CASTING_H |
1 | //===- llvm/CodeGen/SelectionDAGNodes.h - SelectionDAG Nodes ----*- C++ -*-===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This file declares the SDNode class and derived classes, which are used to | |||
10 | // represent the nodes and operations present in a SelectionDAG. These nodes | |||
11 | // and operations are machine code level operations, with some similarities to | |||
12 | // the GCC RTL representation. | |||
13 | // | |||
14 | // Clients should include the SelectionDAG.h file instead of this file directly. | |||
15 | // | |||
16 | //===----------------------------------------------------------------------===// | |||
17 | ||||
18 | #ifndef LLVM_CODEGEN_SELECTIONDAGNODES_H | |||
19 | #define LLVM_CODEGEN_SELECTIONDAGNODES_H | |||
20 | ||||
21 | #include "llvm/ADT/APFloat.h" | |||
22 | #include "llvm/ADT/ArrayRef.h" | |||
23 | #include "llvm/ADT/BitVector.h" | |||
24 | #include "llvm/ADT/FoldingSet.h" | |||
25 | #include "llvm/ADT/GraphTraits.h" | |||
26 | #include "llvm/ADT/SmallPtrSet.h" | |||
27 | #include "llvm/ADT/SmallVector.h" | |||
28 | #include "llvm/ADT/ilist_node.h" | |||
29 | #include "llvm/ADT/iterator.h" | |||
30 | #include "llvm/ADT/iterator_range.h" | |||
31 | #include "llvm/CodeGen/ISDOpcodes.h" | |||
32 | #include "llvm/CodeGen/MachineMemOperand.h" | |||
33 | #include "llvm/CodeGen/Register.h" | |||
34 | #include "llvm/CodeGen/ValueTypes.h" | |||
35 | #include "llvm/IR/Constants.h" | |||
36 | #include "llvm/IR/DebugLoc.h" | |||
37 | #include "llvm/IR/Instruction.h" | |||
38 | #include "llvm/IR/Instructions.h" | |||
39 | #include "llvm/IR/Metadata.h" | |||
40 | #include "llvm/IR/Operator.h" | |||
41 | #include "llvm/Support/AlignOf.h" | |||
42 | #include "llvm/Support/AtomicOrdering.h" | |||
43 | #include "llvm/Support/Casting.h" | |||
44 | #include "llvm/Support/ErrorHandling.h" | |||
45 | #include "llvm/Support/MachineValueType.h" | |||
46 | #include "llvm/Support/TypeSize.h" | |||
47 | #include <algorithm> | |||
48 | #include <cassert> | |||
49 | #include <climits> | |||
50 | #include <cstddef> | |||
51 | #include <cstdint> | |||
52 | #include <cstring> | |||
53 | #include <iterator> | |||
54 | #include <string> | |||
55 | #include <tuple> | |||
56 | ||||
57 | namespace llvm { | |||
58 | ||||
59 | class APInt; | |||
60 | class Constant; | |||
61 | template <typename T> struct DenseMapInfo; | |||
62 | class GlobalValue; | |||
63 | class MachineBasicBlock; | |||
64 | class MachineConstantPoolValue; | |||
65 | class MCSymbol; | |||
66 | class raw_ostream; | |||
67 | class SDNode; | |||
68 | class SelectionDAG; | |||
69 | class Type; | |||
70 | class Value; | |||
71 | ||||
72 | void checkForCycles(const SDNode *N, const SelectionDAG *DAG = nullptr, | |||
73 | bool force = false); | |||
74 | ||||
75 | /// This represents a list of ValueType's that has been intern'd by | |||
76 | /// a SelectionDAG. Instances of this simple value class are returned by | |||
77 | /// SelectionDAG::getVTList(...). | |||
78 | /// | |||
79 | struct SDVTList { | |||
80 | const EVT *VTs; | |||
81 | unsigned int NumVTs; | |||
82 | }; | |||
83 | ||||
84 | namespace ISD { | |||
85 | ||||
86 | /// Node predicates | |||
87 | ||||
88 | /// If N is a BUILD_VECTOR or SPLAT_VECTOR node whose elements are all the | |||
89 | /// same constant or undefined, return true and return the constant value in | |||
90 | /// \p SplatValue. | |||
91 | bool isConstantSplatVector(const SDNode *N, APInt &SplatValue); | |||
92 | ||||
93 | /// Return true if the specified node is a BUILD_VECTOR or SPLAT_VECTOR where | |||
94 | /// all of the elements are ~0 or undef. If \p BuildVectorOnly is set to | |||
95 | /// true, it only checks BUILD_VECTOR. | |||
96 | bool isConstantSplatVectorAllOnes(const SDNode *N, | |||
97 | bool BuildVectorOnly = false); | |||
98 | ||||
99 | /// Return true if the specified node is a BUILD_VECTOR or SPLAT_VECTOR where | |||
100 | /// all of the elements are 0 or undef. If \p BuildVectorOnly is set to true, it | |||
101 | /// only checks BUILD_VECTOR. | |||
102 | bool isConstantSplatVectorAllZeros(const SDNode *N, | |||
103 | bool BuildVectorOnly = false); | |||
104 | ||||
105 | /// Return true if the specified node is a BUILD_VECTOR where all of the | |||
106 | /// elements are ~0 or undef. | |||
107 | bool isBuildVectorAllOnes(const SDNode *N); | |||
108 | ||||
109 | /// Return true if the specified node is a BUILD_VECTOR where all of the | |||
110 | /// elements are 0 or undef. | |||
111 | bool isBuildVectorAllZeros(const SDNode *N); | |||
112 | ||||
113 | /// Return true if the specified node is a BUILD_VECTOR node of all | |||
114 | /// ConstantSDNode or undef. | |||
115 | bool isBuildVectorOfConstantSDNodes(const SDNode *N); | |||
116 | ||||
117 | /// Return true if the specified node is a BUILD_VECTOR node of all | |||
118 | /// ConstantFPSDNode or undef. | |||
119 | bool isBuildVectorOfConstantFPSDNodes(const SDNode *N); | |||
120 | ||||
121 | /// Return true if the node has at least one operand and all operands of the | |||
122 | /// specified node are ISD::UNDEF. | |||
123 | bool allOperandsUndef(const SDNode *N); | |||
124 | ||||
125 | } // end namespace ISD | |||
126 | ||||
127 | //===----------------------------------------------------------------------===// | |||
128 | /// Unlike LLVM values, Selection DAG nodes may return multiple | |||
129 | /// values as the result of a computation. Many nodes return multiple values, | |||
130 | /// from loads (which define a token and a return value) to ADDC (which returns | |||
131 | /// a result and a carry value), to calls (which may return an arbitrary number | |||
132 | /// of values). | |||
133 | /// | |||
134 | /// As such, each use of a SelectionDAG computation must indicate the node that | |||
135 | /// computes it as well as which return value to use from that node. This pair | |||
136 | /// of information is represented with the SDValue value type. | |||
137 | /// | |||
138 | class SDValue { | |||
139 | friend struct DenseMapInfo<SDValue>; | |||
140 | ||||
141 | SDNode *Node = nullptr; // The node defining the value we are using. | |||
142 | unsigned ResNo = 0; // Which return value of the node we are using. | |||
143 | ||||
144 | public: | |||
145 | SDValue() = default; | |||
146 | SDValue(SDNode *node, unsigned resno); | |||
147 | ||||
148 | /// get the index which selects a specific result in the SDNode | |||
149 | unsigned getResNo() const { return ResNo; } | |||
150 | ||||
151 | /// get the SDNode which holds the desired result | |||
152 | SDNode *getNode() const { return Node; } | |||
153 | ||||
154 | /// set the SDNode | |||
155 | void setNode(SDNode *N) { Node = N; } | |||
156 | ||||
157 | inline SDNode *operator->() const { return Node; } | |||
158 | ||||
159 | bool operator==(const SDValue &O) const { | |||
160 | return Node == O.Node && ResNo == O.ResNo; | |||
161 | } | |||
162 | bool operator!=(const SDValue &O) const { | |||
163 | return !operator==(O); | |||
164 | } | |||
165 | bool operator<(const SDValue &O) const { | |||
166 | return std::tie(Node, ResNo) < std::tie(O.Node, O.ResNo); | |||
167 | } | |||
168 | explicit operator bool() const { | |||
169 | return Node != nullptr; | |||
170 | } | |||
171 | ||||
172 | SDValue getValue(unsigned R) const { | |||
173 | return SDValue(Node, R); | |||
174 | } | |||
175 | ||||
176 | /// Return true if this node is an operand of N. | |||
177 | bool isOperandOf(const SDNode *N) const; | |||
178 | ||||
179 | /// Return the ValueType of the referenced return value. | |||
180 | inline EVT getValueType() const; | |||
181 | ||||
182 | /// Return the simple ValueType of the referenced return value. | |||
183 | MVT getSimpleValueType() const { | |||
184 | return getValueType().getSimpleVT(); | |||
185 | } | |||
186 | ||||
187 | /// Returns the size of the value in bits. | |||
188 | /// | |||
189 | /// If the value type is a scalable vector type, the scalable property will | |||
190 | /// be set and the runtime size will be a positive integer multiple of the | |||
191 | /// base size. | |||
192 | TypeSize getValueSizeInBits() const { | |||
193 | return getValueType().getSizeInBits(); | |||
194 | } | |||
195 | ||||
196 | uint64_t getScalarValueSizeInBits() const { | |||
197 | return getValueType().getScalarType().getFixedSizeInBits(); | |||
198 | } | |||
199 | ||||
200 | // Forwarding methods - These forward to the corresponding methods in SDNode. | |||
201 | inline unsigned getOpcode() const; | |||
202 | inline unsigned getNumOperands() const; | |||
203 | inline const SDValue &getOperand(unsigned i) const; | |||
204 | inline uint64_t getConstantOperandVal(unsigned i) const; | |||
205 | inline const APInt &getConstantOperandAPInt(unsigned i) const; | |||
206 | inline bool isTargetMemoryOpcode() const; | |||
207 | inline bool isTargetOpcode() const; | |||
208 | inline bool isMachineOpcode() const; | |||
209 | inline bool isUndef() const; | |||
210 | inline unsigned getMachineOpcode() const; | |||
211 | inline const DebugLoc &getDebugLoc() const; | |||
212 | inline void dump() const; | |||
213 | inline void dump(const SelectionDAG *G) const; | |||
214 | inline void dumpr() const; | |||
215 | inline void dumpr(const SelectionDAG *G) const; | |||
216 | ||||
217 | /// Return true if this operand (which must be a chain) reaches the | |||
218 | /// specified operand without crossing any side-effecting instructions. | |||
219 | /// In practice, this looks through token factors and non-volatile loads. | |||
220 | /// In order to remain efficient, this only | |||
221 | /// looks a couple of nodes in, it does not do an exhaustive search. | |||
222 | bool reachesChainWithoutSideEffects(SDValue Dest, | |||
223 | unsigned Depth = 2) const; | |||
224 | ||||
225 | /// Return true if there are no nodes using value ResNo of Node. | |||
226 | inline bool use_empty() const; | |||
227 | ||||
228 | /// Return true if there is exactly one node using value ResNo of Node. | |||
229 | inline bool hasOneUse() const; | |||
230 | }; | |||
231 | ||||
232 | template<> struct DenseMapInfo<SDValue> { | |||
233 | static inline SDValue getEmptyKey() { | |||
234 | SDValue V; | |||
235 | V.ResNo = -1U; | |||
236 | return V; | |||
237 | } | |||
238 | ||||
239 | static inline SDValue getTombstoneKey() { | |||
240 | SDValue V; | |||
241 | V.ResNo = -2U; | |||
242 | return V; | |||
243 | } | |||
244 | ||||
245 | static unsigned getHashValue(const SDValue &Val) { | |||
246 | return ((unsigned)((uintptr_t)Val.getNode() >> 4) ^ | |||
247 | (unsigned)((uintptr_t)Val.getNode() >> 9)) + Val.getResNo(); | |||
248 | } | |||
249 | ||||
250 | static bool isEqual(const SDValue &LHS, const SDValue &RHS) { | |||
251 | return LHS == RHS; | |||
252 | } | |||
253 | }; | |||
254 | ||||
255 | /// Allow casting operators to work directly on | |||
256 | /// SDValues as if they were SDNode*'s. | |||
257 | template<> struct simplify_type<SDValue> { | |||
258 | using SimpleType = SDNode *; | |||
259 | ||||
260 | static SimpleType getSimplifiedValue(SDValue &Val) { | |||
261 | return Val.getNode(); | |||
262 | } | |||
263 | }; | |||
264 | template<> struct simplify_type<const SDValue> { | |||
265 | using SimpleType = /*const*/ SDNode *; | |||
266 | ||||
267 | static SimpleType getSimplifiedValue(const SDValue &Val) { | |||
268 | return Val.getNode(); | |||
269 | } | |||
270 | }; | |||
271 | ||||
272 | /// Represents a use of a SDNode. This class holds an SDValue, | |||
273 | /// which records the SDNode being used and the result number, a | |||
274 | /// pointer to the SDNode using the value, and Next and Prev pointers, | |||
275 | /// which link together all the uses of an SDNode. | |||
276 | /// | |||
277 | class SDUse { | |||
278 | /// Val - The value being used. | |||
279 | SDValue Val; | |||
280 | /// User - The user of this value. | |||
281 | SDNode *User = nullptr; | |||
282 | /// Prev, Next - Pointers to the uses list of the SDNode referred by | |||
283 | /// this operand. | |||
284 | SDUse **Prev = nullptr; | |||
285 | SDUse *Next = nullptr; | |||
286 | ||||
287 | public: | |||
288 | SDUse() = default; | |||
289 | SDUse(const SDUse &U) = delete; | |||
290 | SDUse &operator=(const SDUse &) = delete; | |||
291 | ||||
292 | /// Normally SDUse will just implicitly convert to an SDValue that it holds. | |||
293 | operator const SDValue&() const { return Val; } | |||
294 | ||||
295 | /// If implicit conversion to SDValue doesn't work, the get() method returns | |||
296 | /// the SDValue. | |||
297 | const SDValue &get() const { return Val; } | |||
298 | ||||
299 | /// This returns the SDNode that contains this Use. | |||
300 | SDNode *getUser() { return User; } | |||
301 | ||||
302 | /// Get the next SDUse in the use list. | |||
303 | SDUse *getNext() const { return Next; } | |||
304 | ||||
305 | /// Convenience function for get().getNode(). | |||
306 | SDNode *getNode() const { return Val.getNode(); } | |||
307 | /// Convenience function for get().getResNo(). | |||
308 | unsigned getResNo() const { return Val.getResNo(); } | |||
309 | /// Convenience function for get().getValueType(). | |||
310 | EVT getValueType() const { return Val.getValueType(); } | |||
311 | ||||
312 | /// Convenience function for get().operator== | |||
313 | bool operator==(const SDValue &V) const { | |||
314 | return Val == V; | |||
315 | } | |||
316 | ||||
317 | /// Convenience function for get().operator!= | |||
318 | bool operator!=(const SDValue &V) const { | |||
319 | return Val != V; | |||
320 | } | |||
321 | ||||
322 | /// Convenience function for get().operator< | |||
323 | bool operator<(const SDValue &V) const { | |||
324 | return Val < V; | |||
325 | } | |||
326 | ||||
327 | private: | |||
328 | friend class SelectionDAG; | |||
329 | friend class SDNode; | |||
330 | // TODO: unfriend HandleSDNode once we fix its operand handling. | |||
331 | friend class HandleSDNode; | |||
332 | ||||
333 | void setUser(SDNode *p) { User = p; } | |||
334 | ||||
335 | /// Remove this use from its existing use list, assign it the | |||
336 | /// given value, and add it to the new value's node's use list. | |||
337 | inline void set(const SDValue &V); | |||
338 | /// Like set, but only supports initializing a newly-allocated | |||
339 | /// SDUse with a non-null value. | |||
340 | inline void setInitial(const SDValue &V); | |||
341 | /// Like set, but only sets the Node portion of the value, | |||
342 | /// leaving the ResNo portion unmodified. | |||
343 | inline void setNode(SDNode *N); | |||
344 | ||||
345 | void addToList(SDUse **List) { | |||
346 | Next = *List; | |||
347 | if (Next) Next->Prev = &Next; | |||
348 | Prev = List; | |||
349 | *List = this; | |||
350 | } | |||
351 | ||||
352 | void removeFromList() { | |||
353 | *Prev = Next; | |||
354 | if (Next) Next->Prev = Prev; | |||
355 | } | |||
356 | }; | |||
357 | ||||
358 | /// simplify_type specializations - Allow casting operators to work directly on | |||
359 | /// SDValues as if they were SDNode*'s. | |||
360 | template<> struct simplify_type<SDUse> { | |||
361 | using SimpleType = SDNode *; | |||
362 | ||||
363 | static SimpleType getSimplifiedValue(SDUse &Val) { | |||
364 | return Val.getNode(); | |||
365 | } | |||
366 | }; | |||
367 | ||||
368 | /// These are IR-level optimization flags that may be propagated to SDNodes. | |||
369 | /// TODO: This data structure should be shared by the IR optimizer and the | |||
370 | /// the backend. | |||
371 | struct SDNodeFlags { | |||
372 | private: | |||
373 | bool NoUnsignedWrap : 1; | |||
374 | bool NoSignedWrap : 1; | |||
375 | bool Exact : 1; | |||
376 | bool NoNaNs : 1; | |||
377 | bool NoInfs : 1; | |||
378 | bool NoSignedZeros : 1; | |||
379 | bool AllowReciprocal : 1; | |||
380 | bool AllowContract : 1; | |||
381 | bool ApproximateFuncs : 1; | |||
382 | bool AllowReassociation : 1; | |||
383 | ||||
384 | // We assume instructions do not raise floating-point exceptions by default, | |||
385 | // and only those marked explicitly may do so. We could choose to represent | |||
386 | // this via a positive "FPExcept" flags like on the MI level, but having a | |||
387 | // negative "NoFPExcept" flag here (that defaults to true) makes the flag | |||
388 | // intersection logic more straightforward. | |||
389 | bool NoFPExcept : 1; | |||
390 | ||||
391 | public: | |||
392 | /// Default constructor turns off all optimization flags. | |||
393 | SDNodeFlags() | |||
394 | : NoUnsignedWrap(false), NoSignedWrap(false), Exact(false), NoNaNs(false), | |||
395 | NoInfs(false), NoSignedZeros(false), AllowReciprocal(false), | |||
396 | AllowContract(false), ApproximateFuncs(false), | |||
397 | AllowReassociation(false), NoFPExcept(false) {} | |||
398 | ||||
399 | /// Propagate the fast-math-flags from an IR FPMathOperator. | |||
400 | void copyFMF(const FPMathOperator &FPMO) { | |||
401 | setNoNaNs(FPMO.hasNoNaNs()); | |||
402 | setNoInfs(FPMO.hasNoInfs()); | |||
403 | setNoSignedZeros(FPMO.hasNoSignedZeros()); | |||
404 | setAllowReciprocal(FPMO.hasAllowReciprocal()); | |||
405 | setAllowContract(FPMO.hasAllowContract()); | |||
406 | setApproximateFuncs(FPMO.hasApproxFunc()); | |||
407 | setAllowReassociation(FPMO.hasAllowReassoc()); | |||
408 | } | |||
409 | ||||
410 | // These are mutators for each flag. | |||
411 | void setNoUnsignedWrap(bool b) { NoUnsignedWrap = b; } | |||
412 | void setNoSignedWrap(bool b) { NoSignedWrap = b; } | |||
413 | void setExact(bool b) { Exact = b; } | |||
414 | void setNoNaNs(bool b) { NoNaNs = b; } | |||
415 | void setNoInfs(bool b) { NoInfs = b; } | |||
416 | void setNoSignedZeros(bool b) { NoSignedZeros = b; } | |||
417 | void setAllowReciprocal(bool b) { AllowReciprocal = b; } | |||
418 | void setAllowContract(bool b) { AllowContract = b; } | |||
419 | void setApproximateFuncs(bool b) { ApproximateFuncs = b; } | |||
420 | void setAllowReassociation(bool b) { AllowReassociation = b; } | |||
421 | void setNoFPExcept(bool b) { NoFPExcept = b; } | |||
422 | ||||
423 | // These are accessors for each flag. | |||
424 | bool hasNoUnsignedWrap() const { return NoUnsignedWrap; } | |||
425 | bool hasNoSignedWrap() const { return NoSignedWrap; } | |||
426 | bool hasExact() const { return Exact; } | |||
427 | bool hasNoNaNs() const { return NoNaNs; } | |||
428 | bool hasNoInfs() const { return NoInfs; } | |||
429 | bool hasNoSignedZeros() const { return NoSignedZeros; } | |||
430 | bool hasAllowReciprocal() const { return AllowReciprocal; } | |||
431 | bool hasAllowContract() const { return AllowContract; } | |||
432 | bool hasApproximateFuncs() const { return ApproximateFuncs; } | |||
433 | bool hasAllowReassociation() const { return AllowReassociation; } | |||
434 | bool hasNoFPExcept() const { return NoFPExcept; } | |||
435 | ||||
436 | /// Clear any flags in this flag set that aren't also set in Flags. All | |||
437 | /// flags will be cleared if Flags are undefined. | |||
438 | void intersectWith(const SDNodeFlags Flags) { | |||
439 | NoUnsignedWrap &= Flags.NoUnsignedWrap; | |||
440 | NoSignedWrap &= Flags.NoSignedWrap; | |||
441 | Exact &= Flags.Exact; | |||
442 | NoNaNs &= Flags.NoNaNs; | |||
443 | NoInfs &= Flags.NoInfs; | |||
444 | NoSignedZeros &= Flags.NoSignedZeros; | |||
445 | AllowReciprocal &= Flags.AllowReciprocal; | |||
446 | AllowContract &= Flags.AllowContract; | |||
447 | ApproximateFuncs &= Flags.ApproximateFuncs; | |||
448 | AllowReassociation &= Flags.AllowReassociation; | |||
449 | NoFPExcept &= Flags.NoFPExcept; | |||
450 | } | |||
451 | }; | |||
452 | ||||
453 | /// Represents one node in the SelectionDAG. | |||
454 | /// | |||
455 | class SDNode : public FoldingSetNode, public ilist_node<SDNode> { | |||
456 | private: | |||
457 | /// The operation that this node performs. | |||
458 | int16_t NodeType; | |||
459 | ||||
460 | protected: | |||
461 | // We define a set of mini-helper classes to help us interpret the bits in our | |||
462 | // SubclassData. These are designed to fit within a uint16_t so they pack | |||
463 | // with NodeType. | |||
464 | ||||
465 | #if defined(_AIX) && (!defined(__GNUC__4) || defined(__ibmxl__)) | |||
466 | // Except for GCC; by default, AIX compilers store bit-fields in 4-byte words | |||
467 | // and give the `pack` pragma push semantics. | |||
468 | #define BEGIN_TWO_BYTE_PACK() _Pragma("pack(2)")pack(2) | |||
469 | #define END_TWO_BYTE_PACK() _Pragma("pack(pop)")pack(pop) | |||
470 | #else | |||
471 | #define BEGIN_TWO_BYTE_PACK() | |||
472 | #define END_TWO_BYTE_PACK() | |||
473 | #endif | |||
474 | ||||
475 | BEGIN_TWO_BYTE_PACK() | |||
476 | class SDNodeBitfields { | |||
477 | friend class SDNode; | |||
478 | friend class MemIntrinsicSDNode; | |||
479 | friend class MemSDNode; | |||
480 | friend class SelectionDAG; | |||
481 | ||||
482 | uint16_t HasDebugValue : 1; | |||
483 | uint16_t IsMemIntrinsic : 1; | |||
484 | uint16_t IsDivergent : 1; | |||
485 | }; | |||
486 | enum { NumSDNodeBits = 3 }; | |||
487 | ||||
488 | class ConstantSDNodeBitfields { | |||
489 | friend class ConstantSDNode; | |||
490 | ||||
491 | uint16_t : NumSDNodeBits; | |||
492 | ||||
493 | uint16_t IsOpaque : 1; | |||
494 | }; | |||
495 | ||||
496 | class MemSDNodeBitfields { | |||
497 | friend class MemSDNode; | |||
498 | friend class MemIntrinsicSDNode; | |||
499 | friend class AtomicSDNode; | |||
500 | ||||
501 | uint16_t : NumSDNodeBits; | |||
502 | ||||
503 | uint16_t IsVolatile : 1; | |||
504 | uint16_t IsNonTemporal : 1; | |||
505 | uint16_t IsDereferenceable : 1; | |||
506 | uint16_t IsInvariant : 1; | |||
507 | }; | |||
508 | enum { NumMemSDNodeBits = NumSDNodeBits + 4 }; | |||
509 | ||||
510 | class LSBaseSDNodeBitfields { | |||
511 | friend class LSBaseSDNode; | |||
512 | friend class MaskedLoadStoreSDNode; | |||
513 | friend class MaskedGatherScatterSDNode; | |||
514 | ||||
515 | uint16_t : NumMemSDNodeBits; | |||
516 | ||||
517 | // This storage is shared between disparate class hierarchies to hold an | |||
518 | // enumeration specific to the class hierarchy in use. | |||
519 | // LSBaseSDNode => enum ISD::MemIndexedMode | |||
520 | // MaskedLoadStoreBaseSDNode => enum ISD::MemIndexedMode | |||
521 | // MaskedGatherScatterSDNode => enum ISD::MemIndexType | |||
522 | uint16_t AddressingMode : 3; | |||
523 | }; | |||
524 | enum { NumLSBaseSDNodeBits = NumMemSDNodeBits + 3 }; | |||
525 | ||||
526 | class LoadSDNodeBitfields { | |||
527 | friend class LoadSDNode; | |||
528 | friend class MaskedLoadSDNode; | |||
529 | friend class MaskedGatherSDNode; | |||
530 | ||||
531 | uint16_t : NumLSBaseSDNodeBits; | |||
532 | ||||
533 | uint16_t ExtTy : 2; // enum ISD::LoadExtType | |||
534 | uint16_t IsExpanding : 1; | |||
535 | }; | |||
536 | ||||
537 | class StoreSDNodeBitfields { | |||
538 | friend class StoreSDNode; | |||
539 | friend class MaskedStoreSDNode; | |||
540 | friend class MaskedScatterSDNode; | |||
541 | ||||
542 | uint16_t : NumLSBaseSDNodeBits; | |||
543 | ||||
544 | uint16_t IsTruncating : 1; | |||
545 | uint16_t IsCompressing : 1; | |||
546 | }; | |||
547 | ||||
548 | union { | |||
549 | char RawSDNodeBits[sizeof(uint16_t)]; | |||
550 | SDNodeBitfields SDNodeBits; | |||
551 | ConstantSDNodeBitfields ConstantSDNodeBits; | |||
552 | MemSDNodeBitfields MemSDNodeBits; | |||
553 | LSBaseSDNodeBitfields LSBaseSDNodeBits; | |||
554 | LoadSDNodeBitfields LoadSDNodeBits; | |||
555 | StoreSDNodeBitfields StoreSDNodeBits; | |||
556 | }; | |||
557 | END_TWO_BYTE_PACK() | |||
558 | #undef BEGIN_TWO_BYTE_PACK | |||
559 | #undef END_TWO_BYTE_PACK | |||
560 | ||||
561 | // RawSDNodeBits must cover the entirety of the union. This means that all of | |||
562 | // the union's members must have size <= RawSDNodeBits. We write the RHS as | |||
563 | // "2" instead of sizeof(RawSDNodeBits) because MSVC can't handle the latter. | |||
564 | static_assert(sizeof(SDNodeBitfields) <= 2, "field too wide"); | |||
565 | static_assert(sizeof(ConstantSDNodeBitfields) <= 2, "field too wide"); | |||
566 | static_assert(sizeof(MemSDNodeBitfields) <= 2, "field too wide"); | |||
567 | static_assert(sizeof(LSBaseSDNodeBitfields) <= 2, "field too wide"); | |||
568 | static_assert(sizeof(LoadSDNodeBitfields) <= 2, "field too wide"); | |||
569 | static_assert(sizeof(StoreSDNodeBitfields) <= 2, "field too wide"); | |||
570 | ||||
571 | private: | |||
572 | friend class SelectionDAG; | |||
573 | // TODO: unfriend HandleSDNode once we fix its operand handling. | |||
574 | friend class HandleSDNode; | |||
575 | ||||
576 | /// Unique id per SDNode in the DAG. | |||
577 | int NodeId = -1; | |||
578 | ||||
579 | /// The values that are used by this operation. | |||
580 | SDUse *OperandList = nullptr; | |||
581 | ||||
582 | /// The types of the values this node defines. SDNode's may | |||
583 | /// define multiple values simultaneously. | |||
584 | const EVT *ValueList; | |||
585 | ||||
586 | /// List of uses for this SDNode. | |||
587 | SDUse *UseList = nullptr; | |||
588 | ||||
589 | /// The number of entries in the Operand/Value list. | |||
590 | unsigned short NumOperands = 0; | |||
591 | unsigned short NumValues; | |||
592 | ||||
593 | // The ordering of the SDNodes. It roughly corresponds to the ordering of the | |||
594 | // original LLVM instructions. | |||
595 | // This is used for turning off scheduling, because we'll forgo | |||
596 | // the normal scheduling algorithms and output the instructions according to | |||
597 | // this ordering. | |||
598 | unsigned IROrder; | |||
599 | ||||
600 | /// Source line information. | |||
601 | DebugLoc debugLoc; | |||
602 | ||||
603 | /// Return a pointer to the specified value type. | |||
604 | static const EVT *getValueTypeList(EVT VT); | |||
605 | ||||
606 | SDNodeFlags Flags; | |||
607 | ||||
608 | public: | |||
609 | /// Unique and persistent id per SDNode in the DAG. | |||
610 | /// Used for debug printing. | |||
611 | uint16_t PersistentId; | |||
612 | ||||
613 | //===--------------------------------------------------------------------===// | |||
614 | // Accessors | |||
615 | // | |||
616 | ||||
617 | /// Return the SelectionDAG opcode value for this node. For | |||
618 | /// pre-isel nodes (those for which isMachineOpcode returns false), these | |||
619 | /// are the opcode values in the ISD and <target>ISD namespaces. For | |||
620 | /// post-isel opcodes, see getMachineOpcode. | |||
621 | unsigned getOpcode() const { return (unsigned short)NodeType; } | |||
622 | ||||
623 | /// Test if this node has a target-specific opcode (in the | |||
624 | /// \<target\>ISD namespace). | |||
625 | bool isTargetOpcode() const { return NodeType >= ISD::BUILTIN_OP_END; } | |||
626 | ||||
627 | /// Test if this node has a target-specific opcode that may raise | |||
628 | /// FP exceptions (in the \<target\>ISD namespace and greater than | |||
629 | /// FIRST_TARGET_STRICTFP_OPCODE). Note that all target memory | |||
630 | /// opcode are currently automatically considered to possibly raise | |||
631 | /// FP exceptions as well. | |||
632 | bool isTargetStrictFPOpcode() const { | |||
633 | return NodeType >= ISD::FIRST_TARGET_STRICTFP_OPCODE; | |||
634 | } | |||
635 | ||||
636 | /// Test if this node has a target-specific | |||
637 | /// memory-referencing opcode (in the \<target\>ISD namespace and | |||
638 | /// greater than FIRST_TARGET_MEMORY_OPCODE). | |||
639 | bool isTargetMemoryOpcode() const { | |||
640 | return NodeType >= ISD::FIRST_TARGET_MEMORY_OPCODE; | |||
641 | } | |||
642 | ||||
643 | /// Return true if the type of the node type undefined. | |||
644 | bool isUndef() const { return NodeType == ISD::UNDEF; } | |||
645 | ||||
646 | /// Test if this node is a memory intrinsic (with valid pointer information). | |||
647 | /// INTRINSIC_W_CHAIN and INTRINSIC_VOID nodes are sometimes created for | |||
648 | /// non-memory intrinsics (with chains) that are not really instances of | |||
649 | /// MemSDNode. For such nodes, we need some extra state to determine the | |||
650 | /// proper classof relationship. | |||
651 | bool isMemIntrinsic() const { | |||
652 | return (NodeType == ISD::INTRINSIC_W_CHAIN || | |||
653 | NodeType == ISD::INTRINSIC_VOID) && | |||
654 | SDNodeBits.IsMemIntrinsic; | |||
655 | } | |||
656 | ||||
657 | /// Test if this node is a strict floating point pseudo-op. | |||
658 | bool isStrictFPOpcode() { | |||
659 | switch (NodeType) { | |||
660 | default: | |||
661 | return false; | |||
662 | case ISD::STRICT_FP16_TO_FP: | |||
663 | case ISD::STRICT_FP_TO_FP16: | |||
664 | #define DAG_INSTRUCTION(NAME, NARG, ROUND_MODE, INTRINSIC, DAGN) \ | |||
665 | case ISD::STRICT_##DAGN: | |||
666 | #include "llvm/IR/ConstrainedOps.def" | |||
667 | return true; | |||
668 | } | |||
669 | } | |||
670 | ||||
671 | /// Test if this node has a post-isel opcode, directly | |||
672 | /// corresponding to a MachineInstr opcode. | |||
673 | bool isMachineOpcode() const { return NodeType < 0; } | |||
674 | ||||
675 | /// This may only be called if isMachineOpcode returns | |||
676 | /// true. It returns the MachineInstr opcode value that the node's opcode | |||
677 | /// corresponds to. | |||
678 | unsigned getMachineOpcode() const { | |||
679 | assert(isMachineOpcode() && "Not a MachineInstr opcode!")((isMachineOpcode() && "Not a MachineInstr opcode!") ? static_cast<void> (0) : __assert_fail ("isMachineOpcode() && \"Not a MachineInstr opcode!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 679, __PRETTY_FUNCTION__)); | |||
680 | return ~NodeType; | |||
681 | } | |||
682 | ||||
683 | bool getHasDebugValue() const { return SDNodeBits.HasDebugValue; } | |||
684 | void setHasDebugValue(bool b) { SDNodeBits.HasDebugValue = b; } | |||
685 | ||||
686 | bool isDivergent() const { return SDNodeBits.IsDivergent; } | |||
687 | ||||
688 | /// Return true if there are no uses of this node. | |||
689 | bool use_empty() const { return UseList == nullptr; } | |||
690 | ||||
691 | /// Return true if there is exactly one use of this node. | |||
692 | bool hasOneUse() const { return hasSingleElement(uses()); } | |||
693 | ||||
694 | /// Return the number of uses of this node. This method takes | |||
695 | /// time proportional to the number of uses. | |||
696 | size_t use_size() const { return std::distance(use_begin(), use_end()); } | |||
697 | ||||
698 | /// Return the unique node id. | |||
699 | int getNodeId() const { return NodeId; } | |||
700 | ||||
701 | /// Set unique node id. | |||
702 | void setNodeId(int Id) { NodeId = Id; } | |||
703 | ||||
704 | /// Return the node ordering. | |||
705 | unsigned getIROrder() const { return IROrder; } | |||
706 | ||||
707 | /// Set the node ordering. | |||
708 | void setIROrder(unsigned Order) { IROrder = Order; } | |||
709 | ||||
710 | /// Return the source location info. | |||
711 | const DebugLoc &getDebugLoc() const { return debugLoc; } | |||
712 | ||||
713 | /// Set source location info. Try to avoid this, putting | |||
714 | /// it in the constructor is preferable. | |||
715 | void setDebugLoc(DebugLoc dl) { debugLoc = std::move(dl); } | |||
716 | ||||
717 | /// This class provides iterator support for SDUse | |||
718 | /// operands that use a specific SDNode. | |||
719 | class use_iterator | |||
720 | : public std::iterator<std::forward_iterator_tag, SDUse, ptrdiff_t> { | |||
721 | friend class SDNode; | |||
722 | ||||
723 | SDUse *Op = nullptr; | |||
724 | ||||
725 | explicit use_iterator(SDUse *op) : Op(op) {} | |||
726 | ||||
727 | public: | |||
728 | using reference = std::iterator<std::forward_iterator_tag, | |||
729 | SDUse, ptrdiff_t>::reference; | |||
730 | using pointer = std::iterator<std::forward_iterator_tag, | |||
731 | SDUse, ptrdiff_t>::pointer; | |||
732 | ||||
733 | use_iterator() = default; | |||
734 | use_iterator(const use_iterator &I) : Op(I.Op) {} | |||
735 | ||||
736 | bool operator==(const use_iterator &x) const { | |||
737 | return Op == x.Op; | |||
738 | } | |||
739 | bool operator!=(const use_iterator &x) const { | |||
740 | return !operator==(x); | |||
741 | } | |||
742 | ||||
743 | /// Return true if this iterator is at the end of uses list. | |||
744 | bool atEnd() const { return Op == nullptr; } | |||
745 | ||||
746 | // Iterator traversal: forward iteration only. | |||
747 | use_iterator &operator++() { // Preincrement | |||
748 | assert(Op && "Cannot increment end iterator!")((Op && "Cannot increment end iterator!") ? static_cast <void> (0) : __assert_fail ("Op && \"Cannot increment end iterator!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 748, __PRETTY_FUNCTION__)); | |||
749 | Op = Op->getNext(); | |||
750 | return *this; | |||
751 | } | |||
752 | ||||
753 | use_iterator operator++(int) { // Postincrement | |||
754 | use_iterator tmp = *this; ++*this; return tmp; | |||
755 | } | |||
756 | ||||
757 | /// Retrieve a pointer to the current user node. | |||
758 | SDNode *operator*() const { | |||
759 | assert(Op && "Cannot dereference end iterator!")((Op && "Cannot dereference end iterator!") ? static_cast <void> (0) : __assert_fail ("Op && \"Cannot dereference end iterator!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 759, __PRETTY_FUNCTION__)); | |||
760 | return Op->getUser(); | |||
761 | } | |||
762 | ||||
763 | SDNode *operator->() const { return operator*(); } | |||
764 | ||||
765 | SDUse &getUse() const { return *Op; } | |||
766 | ||||
767 | /// Retrieve the operand # of this use in its user. | |||
768 | unsigned getOperandNo() const { | |||
769 | assert(Op && "Cannot dereference end iterator!")((Op && "Cannot dereference end iterator!") ? static_cast <void> (0) : __assert_fail ("Op && \"Cannot dereference end iterator!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 769, __PRETTY_FUNCTION__)); | |||
770 | return (unsigned)(Op - Op->getUser()->OperandList); | |||
771 | } | |||
772 | }; | |||
773 | ||||
774 | /// Provide iteration support to walk over all uses of an SDNode. | |||
775 | use_iterator use_begin() const { | |||
776 | return use_iterator(UseList); | |||
777 | } | |||
778 | ||||
779 | static use_iterator use_end() { return use_iterator(nullptr); } | |||
780 | ||||
781 | inline iterator_range<use_iterator> uses() { | |||
782 | return make_range(use_begin(), use_end()); | |||
783 | } | |||
784 | inline iterator_range<use_iterator> uses() const { | |||
785 | return make_range(use_begin(), use_end()); | |||
786 | } | |||
787 | ||||
788 | /// Return true if there are exactly NUSES uses of the indicated value. | |||
789 | /// This method ignores uses of other values defined by this operation. | |||
790 | bool hasNUsesOfValue(unsigned NUses, unsigned Value) const; | |||
791 | ||||
792 | /// Return true if there are any use of the indicated value. | |||
793 | /// This method ignores uses of other values defined by this operation. | |||
794 | bool hasAnyUseOfValue(unsigned Value) const; | |||
795 | ||||
796 | /// Return true if this node is the only use of N. | |||
797 | bool isOnlyUserOf(const SDNode *N) const; | |||
798 | ||||
799 | /// Return true if this node is an operand of N. | |||
800 | bool isOperandOf(const SDNode *N) const; | |||
801 | ||||
802 | /// Return true if this node is a predecessor of N. | |||
803 | /// NOTE: Implemented on top of hasPredecessor and every bit as | |||
804 | /// expensive. Use carefully. | |||
805 | bool isPredecessorOf(const SDNode *N) const { | |||
806 | return N->hasPredecessor(this); | |||
807 | } | |||
808 | ||||
809 | /// Return true if N is a predecessor of this node. | |||
810 | /// N is either an operand of this node, or can be reached by recursively | |||
811 | /// traversing up the operands. | |||
812 | /// NOTE: This is an expensive method. Use it carefully. | |||
813 | bool hasPredecessor(const SDNode *N) const; | |||
814 | ||||
815 | /// Returns true if N is a predecessor of any node in Worklist. This | |||
816 | /// helper keeps Visited and Worklist sets externally to allow unions | |||
817 | /// searches to be performed in parallel, caching of results across | |||
818 | /// queries and incremental addition to Worklist. Stops early if N is | |||
819 | /// found but will resume. Remember to clear Visited and Worklists | |||
820 | /// if DAG changes. MaxSteps gives a maximum number of nodes to visit before | |||
821 | /// giving up. The TopologicalPrune flag signals that positive NodeIds are | |||
822 | /// topologically ordered (Operands have strictly smaller node id) and search | |||
823 | /// can be pruned leveraging this. | |||
824 | static bool hasPredecessorHelper(const SDNode *N, | |||
825 | SmallPtrSetImpl<const SDNode *> &Visited, | |||
826 | SmallVectorImpl<const SDNode *> &Worklist, | |||
827 | unsigned int MaxSteps = 0, | |||
828 | bool TopologicalPrune = false) { | |||
829 | SmallVector<const SDNode *, 8> DeferredNodes; | |||
830 | if (Visited.count(N)) | |||
831 | return true; | |||
832 | ||||
833 | // Node Id's are assigned in three places: As a topological | |||
834 | // ordering (> 0), during legalization (results in values set to | |||
835 | // 0), new nodes (set to -1). If N has a topolgical id then we | |||
836 | // know that all nodes with ids smaller than it cannot be | |||
837 | // successors and we need not check them. Filter out all node | |||
838 | // that can't be matches. We add them to the worklist before exit | |||
839 | // in case of multiple calls. Note that during selection the topological id | |||
840 | // may be violated if a node's predecessor is selected before it. We mark | |||
841 | // this at selection negating the id of unselected successors and | |||
842 | // restricting topological pruning to positive ids. | |||
843 | ||||
844 | int NId = N->getNodeId(); | |||
845 | // If we Invalidated the Id, reconstruct original NId. | |||
846 | if (NId < -1) | |||
847 | NId = -(NId + 1); | |||
848 | ||||
849 | bool Found = false; | |||
850 | while (!Worklist.empty()) { | |||
851 | const SDNode *M = Worklist.pop_back_val(); | |||
852 | int MId = M->getNodeId(); | |||
853 | if (TopologicalPrune && M->getOpcode() != ISD::TokenFactor && (NId > 0) && | |||
854 | (MId > 0) && (MId < NId)) { | |||
855 | DeferredNodes.push_back(M); | |||
856 | continue; | |||
857 | } | |||
858 | for (const SDValue &OpV : M->op_values()) { | |||
859 | SDNode *Op = OpV.getNode(); | |||
860 | if (Visited.insert(Op).second) | |||
861 | Worklist.push_back(Op); | |||
862 | if (Op == N) | |||
863 | Found = true; | |||
864 | } | |||
865 | if (Found) | |||
866 | break; | |||
867 | if (MaxSteps != 0 && Visited.size() >= MaxSteps) | |||
868 | break; | |||
869 | } | |||
870 | // Push deferred nodes back on worklist. | |||
871 | Worklist.append(DeferredNodes.begin(), DeferredNodes.end()); | |||
872 | // If we bailed early, conservatively return found. | |||
873 | if (MaxSteps != 0 && Visited.size() >= MaxSteps) | |||
874 | return true; | |||
875 | return Found; | |||
876 | } | |||
877 | ||||
878 | /// Return true if all the users of N are contained in Nodes. | |||
879 | /// NOTE: Requires at least one match, but doesn't require them all. | |||
880 | static bool areOnlyUsersOf(ArrayRef<const SDNode *> Nodes, const SDNode *N); | |||
881 | ||||
882 | /// Return the number of values used by this operation. | |||
883 | unsigned getNumOperands() const { return NumOperands; } | |||
884 | ||||
885 | /// Return the maximum number of operands that a SDNode can hold. | |||
886 | static constexpr size_t getMaxNumOperands() { | |||
887 | return std::numeric_limits<decltype(SDNode::NumOperands)>::max(); | |||
888 | } | |||
889 | ||||
890 | /// Helper method returns the integer value of a ConstantSDNode operand. | |||
891 | inline uint64_t getConstantOperandVal(unsigned Num) const; | |||
892 | ||||
893 | /// Helper method returns the APInt of a ConstantSDNode operand. | |||
894 | inline const APInt &getConstantOperandAPInt(unsigned Num) const; | |||
895 | ||||
896 | const SDValue &getOperand(unsigned Num) const { | |||
897 | assert(Num < NumOperands && "Invalid child # of SDNode!")((Num < NumOperands && "Invalid child # of SDNode!" ) ? static_cast<void> (0) : __assert_fail ("Num < NumOperands && \"Invalid child # of SDNode!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 897, __PRETTY_FUNCTION__)); | |||
898 | return OperandList[Num]; | |||
899 | } | |||
900 | ||||
901 | using op_iterator = SDUse *; | |||
902 | ||||
903 | op_iterator op_begin() const { return OperandList; } | |||
904 | op_iterator op_end() const { return OperandList+NumOperands; } | |||
905 | ArrayRef<SDUse> ops() const { return makeArrayRef(op_begin(), op_end()); } | |||
906 | ||||
907 | /// Iterator for directly iterating over the operand SDValue's. | |||
908 | struct value_op_iterator | |||
909 | : iterator_adaptor_base<value_op_iterator, op_iterator, | |||
910 | std::random_access_iterator_tag, SDValue, | |||
911 | ptrdiff_t, value_op_iterator *, | |||
912 | value_op_iterator *> { | |||
913 | explicit value_op_iterator(SDUse *U = nullptr) | |||
914 | : iterator_adaptor_base(U) {} | |||
915 | ||||
916 | const SDValue &operator*() const { return I->get(); } | |||
917 | }; | |||
918 | ||||
919 | iterator_range<value_op_iterator> op_values() const { | |||
920 | return make_range(value_op_iterator(op_begin()), | |||
921 | value_op_iterator(op_end())); | |||
922 | } | |||
923 | ||||
924 | SDVTList getVTList() const { | |||
925 | SDVTList X = { ValueList, NumValues }; | |||
926 | return X; | |||
927 | } | |||
928 | ||||
929 | /// If this node has a glue operand, return the node | |||
930 | /// to which the glue operand points. Otherwise return NULL. | |||
931 | SDNode *getGluedNode() const { | |||
932 | if (getNumOperands() != 0 && | |||
933 | getOperand(getNumOperands()-1).getValueType() == MVT::Glue) | |||
934 | return getOperand(getNumOperands()-1).getNode(); | |||
935 | return nullptr; | |||
936 | } | |||
937 | ||||
938 | /// If this node has a glue value with a user, return | |||
939 | /// the user (there is at most one). Otherwise return NULL. | |||
940 | SDNode *getGluedUser() const { | |||
941 | for (use_iterator UI = use_begin(), UE = use_end(); UI != UE; ++UI) | |||
942 | if (UI.getUse().get().getValueType() == MVT::Glue) | |||
943 | return *UI; | |||
944 | return nullptr; | |||
945 | } | |||
946 | ||||
947 | SDNodeFlags getFlags() const { return Flags; } | |||
948 | void setFlags(SDNodeFlags NewFlags) { Flags = NewFlags; } | |||
949 | ||||
950 | /// Clear any flags in this node that aren't also set in Flags. | |||
951 | /// If Flags is not in a defined state then this has no effect. | |||
952 | void intersectFlagsWith(const SDNodeFlags Flags); | |||
953 | ||||
954 | /// Return the number of values defined/returned by this operator. | |||
955 | unsigned getNumValues() const { return NumValues; } | |||
956 | ||||
957 | /// Return the type of a specified result. | |||
958 | EVT getValueType(unsigned ResNo) const { | |||
959 | assert(ResNo < NumValues && "Illegal result number!")((ResNo < NumValues && "Illegal result number!") ? static_cast<void> (0) : __assert_fail ("ResNo < NumValues && \"Illegal result number!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 959, __PRETTY_FUNCTION__)); | |||
960 | return ValueList[ResNo]; | |||
961 | } | |||
962 | ||||
963 | /// Return the type of a specified result as a simple type. | |||
964 | MVT getSimpleValueType(unsigned ResNo) const { | |||
965 | return getValueType(ResNo).getSimpleVT(); | |||
966 | } | |||
967 | ||||
968 | /// Returns MVT::getSizeInBits(getValueType(ResNo)). | |||
969 | /// | |||
970 | /// If the value type is a scalable vector type, the scalable property will | |||
971 | /// be set and the runtime size will be a positive integer multiple of the | |||
972 | /// base size. | |||
973 | TypeSize getValueSizeInBits(unsigned ResNo) const { | |||
974 | return getValueType(ResNo).getSizeInBits(); | |||
975 | } | |||
976 | ||||
977 | using value_iterator = const EVT *; | |||
978 | ||||
979 | value_iterator value_begin() const { return ValueList; } | |||
980 | value_iterator value_end() const { return ValueList+NumValues; } | |||
981 | iterator_range<value_iterator> values() const { | |||
982 | return llvm::make_range(value_begin(), value_end()); | |||
983 | } | |||
984 | ||||
985 | /// Return the opcode of this operation for printing. | |||
986 | std::string getOperationName(const SelectionDAG *G = nullptr) const; | |||
987 | static const char* getIndexedModeName(ISD::MemIndexedMode AM); | |||
988 | void print_types(raw_ostream &OS, const SelectionDAG *G) const; | |||
989 | void print_details(raw_ostream &OS, const SelectionDAG *G) const; | |||
990 | void print(raw_ostream &OS, const SelectionDAG *G = nullptr) const; | |||
991 | void printr(raw_ostream &OS, const SelectionDAG *G = nullptr) const; | |||
992 | ||||
993 | /// Print a SelectionDAG node and all children down to | |||
994 | /// the leaves. The given SelectionDAG allows target-specific nodes | |||
995 | /// to be printed in human-readable form. Unlike printr, this will | |||
996 | /// print the whole DAG, including children that appear multiple | |||
997 | /// times. | |||
998 | /// | |||
999 | void printrFull(raw_ostream &O, const SelectionDAG *G = nullptr) const; | |||
1000 | ||||
1001 | /// Print a SelectionDAG node and children up to | |||
1002 | /// depth "depth." The given SelectionDAG allows target-specific | |||
1003 | /// nodes to be printed in human-readable form. Unlike printr, this | |||
1004 | /// will print children that appear multiple times wherever they are | |||
1005 | /// used. | |||
1006 | /// | |||
1007 | void printrWithDepth(raw_ostream &O, const SelectionDAG *G = nullptr, | |||
1008 | unsigned depth = 100) const; | |||
1009 | ||||
1010 | /// Dump this node, for debugging. | |||
1011 | void dump() const; | |||
1012 | ||||
1013 | /// Dump (recursively) this node and its use-def subgraph. | |||
1014 | void dumpr() const; | |||
1015 | ||||
1016 | /// Dump this node, for debugging. | |||
1017 | /// The given SelectionDAG allows target-specific nodes to be printed | |||
1018 | /// in human-readable form. | |||
1019 | void dump(const SelectionDAG *G) const; | |||
1020 | ||||
1021 | /// Dump (recursively) this node and its use-def subgraph. | |||
1022 | /// The given SelectionDAG allows target-specific nodes to be printed | |||
1023 | /// in human-readable form. | |||
1024 | void dumpr(const SelectionDAG *G) const; | |||
1025 | ||||
1026 | /// printrFull to dbgs(). The given SelectionDAG allows | |||
1027 | /// target-specific nodes to be printed in human-readable form. | |||
1028 | /// Unlike dumpr, this will print the whole DAG, including children | |||
1029 | /// that appear multiple times. | |||
1030 | void dumprFull(const SelectionDAG *G = nullptr) const; | |||
1031 | ||||
1032 | /// printrWithDepth to dbgs(). The given | |||
1033 | /// SelectionDAG allows target-specific nodes to be printed in | |||
1034 | /// human-readable form. Unlike dumpr, this will print children | |||
1035 | /// that appear multiple times wherever they are used. | |||
1036 | /// | |||
1037 | void dumprWithDepth(const SelectionDAG *G = nullptr, | |||
1038 | unsigned depth = 100) const; | |||
1039 | ||||
1040 | /// Gather unique data for the node. | |||
1041 | void Profile(FoldingSetNodeID &ID) const; | |||
1042 | ||||
1043 | /// This method should only be used by the SDUse class. | |||
1044 | void addUse(SDUse &U) { U.addToList(&UseList); } | |||
1045 | ||||
1046 | protected: | |||
1047 | static SDVTList getSDVTList(EVT VT) { | |||
1048 | SDVTList Ret = { getValueTypeList(VT), 1 }; | |||
1049 | return Ret; | |||
1050 | } | |||
1051 | ||||
1052 | /// Create an SDNode. | |||
1053 | /// | |||
1054 | /// SDNodes are created without any operands, and never own the operand | |||
1055 | /// storage. To add operands, see SelectionDAG::createOperands. | |||
1056 | SDNode(unsigned Opc, unsigned Order, DebugLoc dl, SDVTList VTs) | |||
1057 | : NodeType(Opc), ValueList(VTs.VTs), NumValues(VTs.NumVTs), | |||
1058 | IROrder(Order), debugLoc(std::move(dl)) { | |||
1059 | memset(&RawSDNodeBits, 0, sizeof(RawSDNodeBits)); | |||
1060 | assert(debugLoc.hasTrivialDestructor() && "Expected trivial destructor")((debugLoc.hasTrivialDestructor() && "Expected trivial destructor" ) ? static_cast<void> (0) : __assert_fail ("debugLoc.hasTrivialDestructor() && \"Expected trivial destructor\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1060, __PRETTY_FUNCTION__)); | |||
1061 | assert(NumValues == VTs.NumVTs &&((NumValues == VTs.NumVTs && "NumValues wasn't wide enough for its operands!" ) ? static_cast<void> (0) : __assert_fail ("NumValues == VTs.NumVTs && \"NumValues wasn't wide enough for its operands!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1062, __PRETTY_FUNCTION__)) | |||
1062 | "NumValues wasn't wide enough for its operands!")((NumValues == VTs.NumVTs && "NumValues wasn't wide enough for its operands!" ) ? static_cast<void> (0) : __assert_fail ("NumValues == VTs.NumVTs && \"NumValues wasn't wide enough for its operands!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1062, __PRETTY_FUNCTION__)); | |||
1063 | } | |||
1064 | ||||
1065 | /// Release the operands and set this node to have zero operands. | |||
1066 | void DropOperands(); | |||
1067 | }; | |||
1068 | ||||
1069 | /// Wrapper class for IR location info (IR ordering and DebugLoc) to be passed | |||
1070 | /// into SDNode creation functions. | |||
1071 | /// When an SDNode is created from the DAGBuilder, the DebugLoc is extracted | |||
1072 | /// from the original Instruction, and IROrder is the ordinal position of | |||
1073 | /// the instruction. | |||
1074 | /// When an SDNode is created after the DAG is being built, both DebugLoc and | |||
1075 | /// the IROrder are propagated from the original SDNode. | |||
1076 | /// So SDLoc class provides two constructors besides the default one, one to | |||
1077 | /// be used by the DAGBuilder, the other to be used by others. | |||
1078 | class SDLoc { | |||
1079 | private: | |||
1080 | DebugLoc DL; | |||
1081 | int IROrder = 0; | |||
1082 | ||||
1083 | public: | |||
1084 | SDLoc() = default; | |||
1085 | SDLoc(const SDNode *N) : DL(N->getDebugLoc()), IROrder(N->getIROrder()) {} | |||
1086 | SDLoc(const SDValue V) : SDLoc(V.getNode()) {} | |||
1087 | SDLoc(const Instruction *I, int Order) : IROrder(Order) { | |||
1088 | assert(Order >= 0 && "bad IROrder")((Order >= 0 && "bad IROrder") ? static_cast<void > (0) : __assert_fail ("Order >= 0 && \"bad IROrder\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1088, __PRETTY_FUNCTION__)); | |||
1089 | if (I) | |||
1090 | DL = I->getDebugLoc(); | |||
1091 | } | |||
1092 | ||||
1093 | unsigned getIROrder() const { return IROrder; } | |||
1094 | const DebugLoc &getDebugLoc() const { return DL; } | |||
1095 | }; | |||
1096 | ||||
1097 | // Define inline functions from the SDValue class. | |||
1098 | ||||
1099 | inline SDValue::SDValue(SDNode *node, unsigned resno) | |||
1100 | : Node(node), ResNo(resno) { | |||
1101 | // Explicitly check for !ResNo to avoid use-after-free, because there are | |||
1102 | // callers that use SDValue(N, 0) with a deleted N to indicate successful | |||
1103 | // combines. | |||
1104 | assert((!Node || !ResNo || ResNo < Node->getNumValues()) &&(((!Node || !ResNo || ResNo < Node->getNumValues()) && "Invalid result number for the given node!") ? static_cast< void> (0) : __assert_fail ("(!Node || !ResNo || ResNo < Node->getNumValues()) && \"Invalid result number for the given node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1105, __PRETTY_FUNCTION__)) | |||
1105 | "Invalid result number for the given node!")(((!Node || !ResNo || ResNo < Node->getNumValues()) && "Invalid result number for the given node!") ? static_cast< void> (0) : __assert_fail ("(!Node || !ResNo || ResNo < Node->getNumValues()) && \"Invalid result number for the given node!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1105, __PRETTY_FUNCTION__)); | |||
1106 | assert(ResNo < -2U && "Cannot use result numbers reserved for DenseMaps.")((ResNo < -2U && "Cannot use result numbers reserved for DenseMaps." ) ? static_cast<void> (0) : __assert_fail ("ResNo < -2U && \"Cannot use result numbers reserved for DenseMaps.\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1106, __PRETTY_FUNCTION__)); | |||
1107 | } | |||
1108 | ||||
1109 | inline unsigned SDValue::getOpcode() const { | |||
1110 | return Node->getOpcode(); | |||
| ||||
1111 | } | |||
1112 | ||||
1113 | inline EVT SDValue::getValueType() const { | |||
1114 | return Node->getValueType(ResNo); | |||
1115 | } | |||
1116 | ||||
1117 | inline unsigned SDValue::getNumOperands() const { | |||
1118 | return Node->getNumOperands(); | |||
1119 | } | |||
1120 | ||||
1121 | inline const SDValue &SDValue::getOperand(unsigned i) const { | |||
1122 | return Node->getOperand(i); | |||
1123 | } | |||
1124 | ||||
1125 | inline uint64_t SDValue::getConstantOperandVal(unsigned i) const { | |||
1126 | return Node->getConstantOperandVal(i); | |||
1127 | } | |||
1128 | ||||
1129 | inline const APInt &SDValue::getConstantOperandAPInt(unsigned i) const { | |||
1130 | return Node->getConstantOperandAPInt(i); | |||
1131 | } | |||
1132 | ||||
1133 | inline bool SDValue::isTargetOpcode() const { | |||
1134 | return Node->isTargetOpcode(); | |||
1135 | } | |||
1136 | ||||
1137 | inline bool SDValue::isTargetMemoryOpcode() const { | |||
1138 | return Node->isTargetMemoryOpcode(); | |||
1139 | } | |||
1140 | ||||
1141 | inline bool SDValue::isMachineOpcode() const { | |||
1142 | return Node->isMachineOpcode(); | |||
1143 | } | |||
1144 | ||||
1145 | inline unsigned SDValue::getMachineOpcode() const { | |||
1146 | return Node->getMachineOpcode(); | |||
1147 | } | |||
1148 | ||||
1149 | inline bool SDValue::isUndef() const { | |||
1150 | return Node->isUndef(); | |||
1151 | } | |||
1152 | ||||
1153 | inline bool SDValue::use_empty() const { | |||
1154 | return !Node->hasAnyUseOfValue(ResNo); | |||
1155 | } | |||
1156 | ||||
1157 | inline bool SDValue::hasOneUse() const { | |||
1158 | return Node->hasNUsesOfValue(1, ResNo); | |||
1159 | } | |||
1160 | ||||
1161 | inline const DebugLoc &SDValue::getDebugLoc() const { | |||
1162 | return Node->getDebugLoc(); | |||
1163 | } | |||
1164 | ||||
1165 | inline void SDValue::dump() const { | |||
1166 | return Node->dump(); | |||
1167 | } | |||
1168 | ||||
1169 | inline void SDValue::dump(const SelectionDAG *G) const { | |||
1170 | return Node->dump(G); | |||
1171 | } | |||
1172 | ||||
1173 | inline void SDValue::dumpr() const { | |||
1174 | return Node->dumpr(); | |||
1175 | } | |||
1176 | ||||
1177 | inline void SDValue::dumpr(const SelectionDAG *G) const { | |||
1178 | return Node->dumpr(G); | |||
1179 | } | |||
1180 | ||||
1181 | // Define inline functions from the SDUse class. | |||
1182 | ||||
1183 | inline void SDUse::set(const SDValue &V) { | |||
1184 | if (Val.getNode()) removeFromList(); | |||
1185 | Val = V; | |||
1186 | if (V.getNode()) V.getNode()->addUse(*this); | |||
1187 | } | |||
1188 | ||||
1189 | inline void SDUse::setInitial(const SDValue &V) { | |||
1190 | Val = V; | |||
1191 | V.getNode()->addUse(*this); | |||
1192 | } | |||
1193 | ||||
1194 | inline void SDUse::setNode(SDNode *N) { | |||
1195 | if (Val.getNode()) removeFromList(); | |||
1196 | Val.setNode(N); | |||
1197 | if (N) N->addUse(*this); | |||
1198 | } | |||
1199 | ||||
1200 | /// This class is used to form a handle around another node that | |||
1201 | /// is persistent and is updated across invocations of replaceAllUsesWith on its | |||
1202 | /// operand. This node should be directly created by end-users and not added to | |||
1203 | /// the AllNodes list. | |||
1204 | class HandleSDNode : public SDNode { | |||
1205 | SDUse Op; | |||
1206 | ||||
1207 | public: | |||
1208 | explicit HandleSDNode(SDValue X) | |||
1209 | : SDNode(ISD::HANDLENODE, 0, DebugLoc(), getSDVTList(MVT::Other)) { | |||
1210 | // HandleSDNodes are never inserted into the DAG, so they won't be | |||
1211 | // auto-numbered. Use ID 65535 as a sentinel. | |||
1212 | PersistentId = 0xffff; | |||
1213 | ||||
1214 | // Manually set up the operand list. This node type is special in that it's | |||
1215 | // always stack allocated and SelectionDAG does not manage its operands. | |||
1216 | // TODO: This should either (a) not be in the SDNode hierarchy, or (b) not | |||
1217 | // be so special. | |||
1218 | Op.setUser(this); | |||
1219 | Op.setInitial(X); | |||
1220 | NumOperands = 1; | |||
1221 | OperandList = &Op; | |||
1222 | } | |||
1223 | ~HandleSDNode(); | |||
1224 | ||||
1225 | const SDValue &getValue() const { return Op; } | |||
1226 | }; | |||
1227 | ||||
1228 | class AddrSpaceCastSDNode : public SDNode { | |||
1229 | private: | |||
1230 | unsigned SrcAddrSpace; | |||
1231 | unsigned DestAddrSpace; | |||
1232 | ||||
1233 | public: | |||
1234 | AddrSpaceCastSDNode(unsigned Order, const DebugLoc &dl, EVT VT, | |||
1235 | unsigned SrcAS, unsigned DestAS); | |||
1236 | ||||
1237 | unsigned getSrcAddressSpace() const { return SrcAddrSpace; } | |||
1238 | unsigned getDestAddressSpace() const { return DestAddrSpace; } | |||
1239 | ||||
1240 | static bool classof(const SDNode *N) { | |||
1241 | return N->getOpcode() == ISD::ADDRSPACECAST; | |||
1242 | } | |||
1243 | }; | |||
1244 | ||||
1245 | /// This is an abstract virtual class for memory operations. | |||
1246 | class MemSDNode : public SDNode { | |||
1247 | private: | |||
1248 | // VT of in-memory value. | |||
1249 | EVT MemoryVT; | |||
1250 | ||||
1251 | protected: | |||
1252 | /// Memory reference information. | |||
1253 | MachineMemOperand *MMO; | |||
1254 | ||||
1255 | public: | |||
1256 | MemSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
1257 | EVT memvt, MachineMemOperand *MMO); | |||
1258 | ||||
1259 | bool readMem() const { return MMO->isLoad(); } | |||
1260 | bool writeMem() const { return MMO->isStore(); } | |||
1261 | ||||
1262 | /// Returns alignment and volatility of the memory access | |||
1263 | Align getOriginalAlign() const { return MMO->getBaseAlign(); } | |||
1264 | Align getAlign() const { return MMO->getAlign(); } | |||
1265 | LLVM_ATTRIBUTE_DEPRECATED(unsigned getOriginalAlignment() const,[[deprecated("Use getOriginalAlign() instead")]] unsigned getOriginalAlignment () const | |||
1266 | "Use getOriginalAlign() instead")[[deprecated("Use getOriginalAlign() instead")]] unsigned getOriginalAlignment () const { | |||
1267 | return MMO->getBaseAlign().value(); | |||
1268 | } | |||
1269 | // FIXME: Remove once transition to getAlign is over. | |||
1270 | unsigned getAlignment() const { return MMO->getAlign().value(); } | |||
1271 | ||||
1272 | /// Return the SubclassData value, without HasDebugValue. This contains an | |||
1273 | /// encoding of the volatile flag, as well as bits used by subclasses. This | |||
1274 | /// function should only be used to compute a FoldingSetNodeID value. | |||
1275 | /// The HasDebugValue bit is masked out because CSE map needs to match | |||
1276 | /// nodes with debug info with nodes without debug info. Same is about | |||
1277 | /// isDivergent bit. | |||
1278 | unsigned getRawSubclassData() const { | |||
1279 | uint16_t Data; | |||
1280 | union { | |||
1281 | char RawSDNodeBits[sizeof(uint16_t)]; | |||
1282 | SDNodeBitfields SDNodeBits; | |||
1283 | }; | |||
1284 | memcpy(&RawSDNodeBits, &this->RawSDNodeBits, sizeof(this->RawSDNodeBits)); | |||
1285 | SDNodeBits.HasDebugValue = 0; | |||
1286 | SDNodeBits.IsDivergent = false; | |||
1287 | memcpy(&Data, &RawSDNodeBits, sizeof(RawSDNodeBits)); | |||
1288 | return Data; | |||
1289 | } | |||
1290 | ||||
1291 | bool isVolatile() const { return MemSDNodeBits.IsVolatile; } | |||
1292 | bool isNonTemporal() const { return MemSDNodeBits.IsNonTemporal; } | |||
1293 | bool isDereferenceable() const { return MemSDNodeBits.IsDereferenceable; } | |||
1294 | bool isInvariant() const { return MemSDNodeBits.IsInvariant; } | |||
1295 | ||||
1296 | // Returns the offset from the location of the access. | |||
1297 | int64_t getSrcValueOffset() const { return MMO->getOffset(); } | |||
1298 | ||||
1299 | /// Returns the AA info that describes the dereference. | |||
1300 | AAMDNodes getAAInfo() const { return MMO->getAAInfo(); } | |||
1301 | ||||
1302 | /// Returns the Ranges that describes the dereference. | |||
1303 | const MDNode *getRanges() const { return MMO->getRanges(); } | |||
1304 | ||||
1305 | /// Returns the synchronization scope ID for this memory operation. | |||
1306 | SyncScope::ID getSyncScopeID() const { return MMO->getSyncScopeID(); } | |||
1307 | ||||
1308 | /// Return the atomic ordering requirements for this memory operation. For | |||
1309 | /// cmpxchg atomic operations, return the atomic ordering requirements when | |||
1310 | /// store occurs. | |||
1311 | AtomicOrdering getOrdering() const { return MMO->getOrdering(); } | |||
1312 | ||||
1313 | /// Return true if the memory operation ordering is Unordered or higher. | |||
1314 | bool isAtomic() const { return MMO->isAtomic(); } | |||
1315 | ||||
1316 | /// Returns true if the memory operation doesn't imply any ordering | |||
1317 | /// constraints on surrounding memory operations beyond the normal memory | |||
1318 | /// aliasing rules. | |||
1319 | bool isUnordered() const { return MMO->isUnordered(); } | |||
1320 | ||||
1321 | /// Returns true if the memory operation is neither atomic or volatile. | |||
1322 | bool isSimple() const { return !isAtomic() && !isVolatile(); } | |||
1323 | ||||
1324 | /// Return the type of the in-memory value. | |||
1325 | EVT getMemoryVT() const { return MemoryVT; } | |||
1326 | ||||
1327 | /// Return a MachineMemOperand object describing the memory | |||
1328 | /// reference performed by operation. | |||
1329 | MachineMemOperand *getMemOperand() const { return MMO; } | |||
1330 | ||||
1331 | const MachinePointerInfo &getPointerInfo() const { | |||
1332 | return MMO->getPointerInfo(); | |||
1333 | } | |||
1334 | ||||
1335 | /// Return the address space for the associated pointer | |||
1336 | unsigned getAddressSpace() const { | |||
1337 | return getPointerInfo().getAddrSpace(); | |||
1338 | } | |||
1339 | ||||
1340 | /// Update this MemSDNode's MachineMemOperand information | |||
1341 | /// to reflect the alignment of NewMMO, if it has a greater alignment. | |||
1342 | /// This must only be used when the new alignment applies to all users of | |||
1343 | /// this MachineMemOperand. | |||
1344 | void refineAlignment(const MachineMemOperand *NewMMO) { | |||
1345 | MMO->refineAlignment(NewMMO); | |||
1346 | } | |||
1347 | ||||
1348 | const SDValue &getChain() const { return getOperand(0); } | |||
1349 | ||||
1350 | const SDValue &getBasePtr() const { | |||
1351 | switch (getOpcode()) { | |||
1352 | case ISD::STORE: | |||
1353 | case ISD::MSTORE: | |||
1354 | return getOperand(2); | |||
1355 | case ISD::MGATHER: | |||
1356 | case ISD::MSCATTER: | |||
1357 | return getOperand(3); | |||
1358 | default: | |||
1359 | return getOperand(1); | |||
1360 | } | |||
1361 | } | |||
1362 | ||||
1363 | // Methods to support isa and dyn_cast | |||
1364 | static bool classof(const SDNode *N) { | |||
1365 | // For some targets, we lower some target intrinsics to a MemIntrinsicNode | |||
1366 | // with either an intrinsic or a target opcode. | |||
1367 | return N->getOpcode() == ISD::LOAD || | |||
1368 | N->getOpcode() == ISD::STORE || | |||
1369 | N->getOpcode() == ISD::PREFETCH || | |||
1370 | N->getOpcode() == ISD::ATOMIC_CMP_SWAP || | |||
1371 | N->getOpcode() == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS || | |||
1372 | N->getOpcode() == ISD::ATOMIC_SWAP || | |||
1373 | N->getOpcode() == ISD::ATOMIC_LOAD_ADD || | |||
1374 | N->getOpcode() == ISD::ATOMIC_LOAD_SUB || | |||
1375 | N->getOpcode() == ISD::ATOMIC_LOAD_AND || | |||
1376 | N->getOpcode() == ISD::ATOMIC_LOAD_CLR || | |||
1377 | N->getOpcode() == ISD::ATOMIC_LOAD_OR || | |||
1378 | N->getOpcode() == ISD::ATOMIC_LOAD_XOR || | |||
1379 | N->getOpcode() == ISD::ATOMIC_LOAD_NAND || | |||
1380 | N->getOpcode() == ISD::ATOMIC_LOAD_MIN || | |||
1381 | N->getOpcode() == ISD::ATOMIC_LOAD_MAX || | |||
1382 | N->getOpcode() == ISD::ATOMIC_LOAD_UMIN || | |||
1383 | N->getOpcode() == ISD::ATOMIC_LOAD_UMAX || | |||
1384 | N->getOpcode() == ISD::ATOMIC_LOAD_FADD || | |||
1385 | N->getOpcode() == ISD::ATOMIC_LOAD_FSUB || | |||
1386 | N->getOpcode() == ISD::ATOMIC_LOAD || | |||
1387 | N->getOpcode() == ISD::ATOMIC_STORE || | |||
1388 | N->getOpcode() == ISD::MLOAD || | |||
1389 | N->getOpcode() == ISD::MSTORE || | |||
1390 | N->getOpcode() == ISD::MGATHER || | |||
1391 | N->getOpcode() == ISD::MSCATTER || | |||
1392 | N->isMemIntrinsic() || | |||
1393 | N->isTargetMemoryOpcode(); | |||
1394 | } | |||
1395 | }; | |||
1396 | ||||
1397 | /// This is an SDNode representing atomic operations. | |||
1398 | class AtomicSDNode : public MemSDNode { | |||
1399 | public: | |||
1400 | AtomicSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl, SDVTList VTL, | |||
1401 | EVT MemVT, MachineMemOperand *MMO) | |||
1402 | : MemSDNode(Opc, Order, dl, VTL, MemVT, MMO) { | |||
1403 | assert(((Opc != ISD::ATOMIC_LOAD && Opc != ISD::ATOMIC_STORE) ||((((Opc != ISD::ATOMIC_LOAD && Opc != ISD::ATOMIC_STORE ) || MMO->isAtomic()) && "then why are we using an AtomicSDNode?" ) ? static_cast<void> (0) : __assert_fail ("((Opc != ISD::ATOMIC_LOAD && Opc != ISD::ATOMIC_STORE) || MMO->isAtomic()) && \"then why are we using an AtomicSDNode?\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1404, __PRETTY_FUNCTION__)) | |||
1404 | MMO->isAtomic()) && "then why are we using an AtomicSDNode?")((((Opc != ISD::ATOMIC_LOAD && Opc != ISD::ATOMIC_STORE ) || MMO->isAtomic()) && "then why are we using an AtomicSDNode?" ) ? static_cast<void> (0) : __assert_fail ("((Opc != ISD::ATOMIC_LOAD && Opc != ISD::ATOMIC_STORE) || MMO->isAtomic()) && \"then why are we using an AtomicSDNode?\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1404, __PRETTY_FUNCTION__)); | |||
1405 | } | |||
1406 | ||||
1407 | const SDValue &getBasePtr() const { return getOperand(1); } | |||
1408 | const SDValue &getVal() const { return getOperand(2); } | |||
1409 | ||||
1410 | /// Returns true if this SDNode represents cmpxchg atomic operation, false | |||
1411 | /// otherwise. | |||
1412 | bool isCompareAndSwap() const { | |||
1413 | unsigned Op = getOpcode(); | |||
1414 | return Op == ISD::ATOMIC_CMP_SWAP || | |||
1415 | Op == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS; | |||
1416 | } | |||
1417 | ||||
1418 | /// For cmpxchg atomic operations, return the atomic ordering requirements | |||
1419 | /// when store does not occur. | |||
1420 | AtomicOrdering getFailureOrdering() const { | |||
1421 | assert(isCompareAndSwap() && "Must be cmpxchg operation")((isCompareAndSwap() && "Must be cmpxchg operation") ? static_cast<void> (0) : __assert_fail ("isCompareAndSwap() && \"Must be cmpxchg operation\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1421, __PRETTY_FUNCTION__)); | |||
1422 | return MMO->getFailureOrdering(); | |||
1423 | } | |||
1424 | ||||
1425 | // Methods to support isa and dyn_cast | |||
1426 | static bool classof(const SDNode *N) { | |||
1427 | return N->getOpcode() == ISD::ATOMIC_CMP_SWAP || | |||
1428 | N->getOpcode() == ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS || | |||
1429 | N->getOpcode() == ISD::ATOMIC_SWAP || | |||
1430 | N->getOpcode() == ISD::ATOMIC_LOAD_ADD || | |||
1431 | N->getOpcode() == ISD::ATOMIC_LOAD_SUB || | |||
1432 | N->getOpcode() == ISD::ATOMIC_LOAD_AND || | |||
1433 | N->getOpcode() == ISD::ATOMIC_LOAD_CLR || | |||
1434 | N->getOpcode() == ISD::ATOMIC_LOAD_OR || | |||
1435 | N->getOpcode() == ISD::ATOMIC_LOAD_XOR || | |||
1436 | N->getOpcode() == ISD::ATOMIC_LOAD_NAND || | |||
1437 | N->getOpcode() == ISD::ATOMIC_LOAD_MIN || | |||
1438 | N->getOpcode() == ISD::ATOMIC_LOAD_MAX || | |||
1439 | N->getOpcode() == ISD::ATOMIC_LOAD_UMIN || | |||
1440 | N->getOpcode() == ISD::ATOMIC_LOAD_UMAX || | |||
1441 | N->getOpcode() == ISD::ATOMIC_LOAD_FADD || | |||
1442 | N->getOpcode() == ISD::ATOMIC_LOAD_FSUB || | |||
1443 | N->getOpcode() == ISD::ATOMIC_LOAD || | |||
1444 | N->getOpcode() == ISD::ATOMIC_STORE; | |||
1445 | } | |||
1446 | }; | |||
1447 | ||||
1448 | /// This SDNode is used for target intrinsics that touch | |||
1449 | /// memory and need an associated MachineMemOperand. Its opcode may be | |||
1450 | /// INTRINSIC_VOID, INTRINSIC_W_CHAIN, PREFETCH, or a target-specific opcode | |||
1451 | /// with a value not less than FIRST_TARGET_MEMORY_OPCODE. | |||
1452 | class MemIntrinsicSDNode : public MemSDNode { | |||
1453 | public: | |||
1454 | MemIntrinsicSDNode(unsigned Opc, unsigned Order, const DebugLoc &dl, | |||
1455 | SDVTList VTs, EVT MemoryVT, MachineMemOperand *MMO) | |||
1456 | : MemSDNode(Opc, Order, dl, VTs, MemoryVT, MMO) { | |||
1457 | SDNodeBits.IsMemIntrinsic = true; | |||
1458 | } | |||
1459 | ||||
1460 | // Methods to support isa and dyn_cast | |||
1461 | static bool classof(const SDNode *N) { | |||
1462 | // We lower some target intrinsics to their target opcode | |||
1463 | // early a node with a target opcode can be of this class | |||
1464 | return N->isMemIntrinsic() || | |||
1465 | N->getOpcode() == ISD::PREFETCH || | |||
1466 | N->isTargetMemoryOpcode(); | |||
1467 | } | |||
1468 | }; | |||
1469 | ||||
1470 | /// This SDNode is used to implement the code generator | |||
1471 | /// support for the llvm IR shufflevector instruction. It combines elements | |||
1472 | /// from two input vectors into a new input vector, with the selection and | |||
1473 | /// ordering of elements determined by an array of integers, referred to as | |||
1474 | /// the shuffle mask. For input vectors of width N, mask indices of 0..N-1 | |||
1475 | /// refer to elements from the LHS input, and indices from N to 2N-1 the RHS. | |||
1476 | /// An index of -1 is treated as undef, such that the code generator may put | |||
1477 | /// any value in the corresponding element of the result. | |||
1478 | class ShuffleVectorSDNode : public SDNode { | |||
1479 | // The memory for Mask is owned by the SelectionDAG's OperandAllocator, and | |||
1480 | // is freed when the SelectionDAG object is destroyed. | |||
1481 | const int *Mask; | |||
1482 | ||||
1483 | protected: | |||
1484 | friend class SelectionDAG; | |||
1485 | ||||
1486 | ShuffleVectorSDNode(EVT VT, unsigned Order, const DebugLoc &dl, const int *M) | |||
1487 | : SDNode(ISD::VECTOR_SHUFFLE, Order, dl, getSDVTList(VT)), Mask(M) {} | |||
1488 | ||||
1489 | public: | |||
1490 | ArrayRef<int> getMask() const { | |||
1491 | EVT VT = getValueType(0); | |||
1492 | return makeArrayRef(Mask, VT.getVectorNumElements()); | |||
1493 | } | |||
1494 | ||||
1495 | int getMaskElt(unsigned Idx) const { | |||
1496 | assert(Idx < getValueType(0).getVectorNumElements() && "Idx out of range!")((Idx < getValueType(0).getVectorNumElements() && "Idx out of range!" ) ? static_cast<void> (0) : __assert_fail ("Idx < getValueType(0).getVectorNumElements() && \"Idx out of range!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1496, __PRETTY_FUNCTION__)); | |||
1497 | return Mask[Idx]; | |||
1498 | } | |||
1499 | ||||
1500 | bool isSplat() const { return isSplatMask(Mask, getValueType(0)); } | |||
1501 | ||||
1502 | int getSplatIndex() const { | |||
1503 | assert(isSplat() && "Cannot get splat index for non-splat!")((isSplat() && "Cannot get splat index for non-splat!" ) ? static_cast<void> (0) : __assert_fail ("isSplat() && \"Cannot get splat index for non-splat!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1503, __PRETTY_FUNCTION__)); | |||
1504 | EVT VT = getValueType(0); | |||
1505 | for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) | |||
1506 | if (Mask[i] >= 0) | |||
1507 | return Mask[i]; | |||
1508 | ||||
1509 | // We can choose any index value here and be correct because all elements | |||
1510 | // are undefined. Return 0 for better potential for callers to simplify. | |||
1511 | return 0; | |||
1512 | } | |||
1513 | ||||
1514 | static bool isSplatMask(const int *Mask, EVT VT); | |||
1515 | ||||
1516 | /// Change values in a shuffle permute mask assuming | |||
1517 | /// the two vector operands have swapped position. | |||
1518 | static void commuteMask(MutableArrayRef<int> Mask) { | |||
1519 | unsigned NumElems = Mask.size(); | |||
1520 | for (unsigned i = 0; i != NumElems; ++i) { | |||
1521 | int idx = Mask[i]; | |||
1522 | if (idx < 0) | |||
1523 | continue; | |||
1524 | else if (idx < (int)NumElems) | |||
1525 | Mask[i] = idx + NumElems; | |||
1526 | else | |||
1527 | Mask[i] = idx - NumElems; | |||
1528 | } | |||
1529 | } | |||
1530 | ||||
1531 | static bool classof(const SDNode *N) { | |||
1532 | return N->getOpcode() == ISD::VECTOR_SHUFFLE; | |||
1533 | } | |||
1534 | }; | |||
1535 | ||||
1536 | class ConstantSDNode : public SDNode { | |||
1537 | friend class SelectionDAG; | |||
1538 | ||||
1539 | const ConstantInt *Value; | |||
1540 | ||||
1541 | ConstantSDNode(bool isTarget, bool isOpaque, const ConstantInt *val, EVT VT) | |||
1542 | : SDNode(isTarget ? ISD::TargetConstant : ISD::Constant, 0, DebugLoc(), | |||
1543 | getSDVTList(VT)), | |||
1544 | Value(val) { | |||
1545 | ConstantSDNodeBits.IsOpaque = isOpaque; | |||
1546 | } | |||
1547 | ||||
1548 | public: | |||
1549 | const ConstantInt *getConstantIntValue() const { return Value; } | |||
1550 | const APInt &getAPIntValue() const { return Value->getValue(); } | |||
1551 | uint64_t getZExtValue() const { return Value->getZExtValue(); } | |||
1552 | int64_t getSExtValue() const { return Value->getSExtValue(); } | |||
1553 | uint64_t getLimitedValue(uint64_t Limit = UINT64_MAX(18446744073709551615UL)) { | |||
1554 | return Value->getLimitedValue(Limit); | |||
1555 | } | |||
1556 | MaybeAlign getMaybeAlignValue() const { return Value->getMaybeAlignValue(); } | |||
1557 | Align getAlignValue() const { return Value->getAlignValue(); } | |||
1558 | ||||
1559 | bool isOne() const { return Value->isOne(); } | |||
1560 | bool isNullValue() const { return Value->isZero(); } | |||
1561 | bool isAllOnesValue() const { return Value->isMinusOne(); } | |||
1562 | ||||
1563 | bool isOpaque() const { return ConstantSDNodeBits.IsOpaque; } | |||
1564 | ||||
1565 | static bool classof(const SDNode *N) { | |||
1566 | return N->getOpcode() == ISD::Constant || | |||
1567 | N->getOpcode() == ISD::TargetConstant; | |||
1568 | } | |||
1569 | }; | |||
1570 | ||||
1571 | uint64_t SDNode::getConstantOperandVal(unsigned Num) const { | |||
1572 | return cast<ConstantSDNode>(getOperand(Num))->getZExtValue(); | |||
1573 | } | |||
1574 | ||||
1575 | const APInt &SDNode::getConstantOperandAPInt(unsigned Num) const { | |||
1576 | return cast<ConstantSDNode>(getOperand(Num))->getAPIntValue(); | |||
1577 | } | |||
1578 | ||||
1579 | class ConstantFPSDNode : public SDNode { | |||
1580 | friend class SelectionDAG; | |||
1581 | ||||
1582 | const ConstantFP *Value; | |||
1583 | ||||
1584 | ConstantFPSDNode(bool isTarget, const ConstantFP *val, EVT VT) | |||
1585 | : SDNode(isTarget ? ISD::TargetConstantFP : ISD::ConstantFP, 0, | |||
1586 | DebugLoc(), getSDVTList(VT)), | |||
1587 | Value(val) {} | |||
1588 | ||||
1589 | public: | |||
1590 | const APFloat& getValueAPF() const { return Value->getValueAPF(); } | |||
1591 | const ConstantFP *getConstantFPValue() const { return Value; } | |||
1592 | ||||
1593 | /// Return true if the value is positive or negative zero. | |||
1594 | bool isZero() const { return Value->isZero(); } | |||
1595 | ||||
1596 | /// Return true if the value is a NaN. | |||
1597 | bool isNaN() const { return Value->isNaN(); } | |||
1598 | ||||
1599 | /// Return true if the value is an infinity | |||
1600 | bool isInfinity() const { return Value->isInfinity(); } | |||
1601 | ||||
1602 | /// Return true if the value is negative. | |||
1603 | bool isNegative() const { return Value->isNegative(); } | |||
1604 | ||||
1605 | /// We don't rely on operator== working on double values, as | |||
1606 | /// it returns true for things that are clearly not equal, like -0.0 and 0.0. | |||
1607 | /// As such, this method can be used to do an exact bit-for-bit comparison of | |||
1608 | /// two floating point values. | |||
1609 | ||||
1610 | /// We leave the version with the double argument here because it's just so | |||
1611 | /// convenient to write "2.0" and the like. Without this function we'd | |||
1612 | /// have to duplicate its logic everywhere it's called. | |||
1613 | bool isExactlyValue(double V) const { | |||
1614 | return Value->getValueAPF().isExactlyValue(V); | |||
1615 | } | |||
1616 | bool isExactlyValue(const APFloat& V) const; | |||
1617 | ||||
1618 | static bool isValueValidForType(EVT VT, const APFloat& Val); | |||
1619 | ||||
1620 | static bool classof(const SDNode *N) { | |||
1621 | return N->getOpcode() == ISD::ConstantFP || | |||
1622 | N->getOpcode() == ISD::TargetConstantFP; | |||
1623 | } | |||
1624 | }; | |||
1625 | ||||
1626 | /// Returns true if \p V is a constant integer zero. | |||
1627 | bool isNullConstant(SDValue V); | |||
1628 | ||||
1629 | /// Returns true if \p V is an FP constant with a value of positive zero. | |||
1630 | bool isNullFPConstant(SDValue V); | |||
1631 | ||||
1632 | /// Returns true if \p V is an integer constant with all bits set. | |||
1633 | bool isAllOnesConstant(SDValue V); | |||
1634 | ||||
1635 | /// Returns true if \p V is a constant integer one. | |||
1636 | bool isOneConstant(SDValue V); | |||
1637 | ||||
1638 | /// Return the non-bitcasted source operand of \p V if it exists. | |||
1639 | /// If \p V is not a bitcasted value, it is returned as-is. | |||
1640 | SDValue peekThroughBitcasts(SDValue V); | |||
1641 | ||||
1642 | /// Return the non-bitcasted and one-use source operand of \p V if it exists. | |||
1643 | /// If \p V is not a bitcasted one-use value, it is returned as-is. | |||
1644 | SDValue peekThroughOneUseBitcasts(SDValue V); | |||
1645 | ||||
1646 | /// Return the non-extracted vector source operand of \p V if it exists. | |||
1647 | /// If \p V is not an extracted subvector, it is returned as-is. | |||
1648 | SDValue peekThroughExtractSubvectors(SDValue V); | |||
1649 | ||||
1650 | /// Returns true if \p V is a bitwise not operation. Assumes that an all ones | |||
1651 | /// constant is canonicalized to be operand 1. | |||
1652 | bool isBitwiseNot(SDValue V, bool AllowUndefs = false); | |||
1653 | ||||
1654 | /// Returns the SDNode if it is a constant splat BuildVector or constant int. | |||
1655 | ConstantSDNode *isConstOrConstSplat(SDValue N, bool AllowUndefs = false, | |||
1656 | bool AllowTruncation = false); | |||
1657 | ||||
1658 | /// Returns the SDNode if it is a demanded constant splat BuildVector or | |||
1659 | /// constant int. | |||
1660 | ConstantSDNode *isConstOrConstSplat(SDValue N, const APInt &DemandedElts, | |||
1661 | bool AllowUndefs = false, | |||
1662 | bool AllowTruncation = false); | |||
1663 | ||||
1664 | /// Returns the SDNode if it is a constant splat BuildVector or constant float. | |||
1665 | ConstantFPSDNode *isConstOrConstSplatFP(SDValue N, bool AllowUndefs = false); | |||
1666 | ||||
1667 | /// Returns the SDNode if it is a demanded constant splat BuildVector or | |||
1668 | /// constant float. | |||
1669 | ConstantFPSDNode *isConstOrConstSplatFP(SDValue N, const APInt &DemandedElts, | |||
1670 | bool AllowUndefs = false); | |||
1671 | ||||
1672 | /// Return true if the value is a constant 0 integer or a splatted vector of | |||
1673 | /// a constant 0 integer (with no undefs by default). | |||
1674 | /// Build vector implicit truncation is not an issue for null values. | |||
1675 | bool isNullOrNullSplat(SDValue V, bool AllowUndefs = false); | |||
1676 | ||||
1677 | /// Return true if the value is a constant 1 integer or a splatted vector of a | |||
1678 | /// constant 1 integer (with no undefs). | |||
1679 | /// Does not permit build vector implicit truncation. | |||
1680 | bool isOneOrOneSplat(SDValue V, bool AllowUndefs = false); | |||
1681 | ||||
1682 | /// Return true if the value is a constant -1 integer or a splatted vector of a | |||
1683 | /// constant -1 integer (with no undefs). | |||
1684 | /// Does not permit build vector implicit truncation. | |||
1685 | bool isAllOnesOrAllOnesSplat(SDValue V, bool AllowUndefs = false); | |||
1686 | ||||
1687 | class GlobalAddressSDNode : public SDNode { | |||
1688 | friend class SelectionDAG; | |||
1689 | ||||
1690 | const GlobalValue *TheGlobal; | |||
1691 | int64_t Offset; | |||
1692 | unsigned TargetFlags; | |||
1693 | ||||
1694 | GlobalAddressSDNode(unsigned Opc, unsigned Order, const DebugLoc &DL, | |||
1695 | const GlobalValue *GA, EVT VT, int64_t o, | |||
1696 | unsigned TF); | |||
1697 | ||||
1698 | public: | |||
1699 | const GlobalValue *getGlobal() const { return TheGlobal; } | |||
1700 | int64_t getOffset() const { return Offset; } | |||
1701 | unsigned getTargetFlags() const { return TargetFlags; } | |||
1702 | // Return the address space this GlobalAddress belongs to. | |||
1703 | unsigned getAddressSpace() const; | |||
1704 | ||||
1705 | static bool classof(const SDNode *N) { | |||
1706 | return N->getOpcode() == ISD::GlobalAddress || | |||
1707 | N->getOpcode() == ISD::TargetGlobalAddress || | |||
1708 | N->getOpcode() == ISD::GlobalTLSAddress || | |||
1709 | N->getOpcode() == ISD::TargetGlobalTLSAddress; | |||
1710 | } | |||
1711 | }; | |||
1712 | ||||
1713 | class FrameIndexSDNode : public SDNode { | |||
1714 | friend class SelectionDAG; | |||
1715 | ||||
1716 | int FI; | |||
1717 | ||||
1718 | FrameIndexSDNode(int fi, EVT VT, bool isTarg) | |||
1719 | : SDNode(isTarg ? ISD::TargetFrameIndex : ISD::FrameIndex, | |||
1720 | 0, DebugLoc(), getSDVTList(VT)), FI(fi) { | |||
1721 | } | |||
1722 | ||||
1723 | public: | |||
1724 | int getIndex() const { return FI; } | |||
1725 | ||||
1726 | static bool classof(const SDNode *N) { | |||
1727 | return N->getOpcode() == ISD::FrameIndex || | |||
1728 | N->getOpcode() == ISD::TargetFrameIndex; | |||
1729 | } | |||
1730 | }; | |||
1731 | ||||
1732 | /// This SDNode is used for LIFETIME_START/LIFETIME_END values, which indicate | |||
1733 | /// the offet and size that are started/ended in the underlying FrameIndex. | |||
1734 | class LifetimeSDNode : public SDNode { | |||
1735 | friend class SelectionDAG; | |||
1736 | int64_t Size; | |||
1737 | int64_t Offset; // -1 if offset is unknown. | |||
1738 | ||||
1739 | LifetimeSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl, | |||
1740 | SDVTList VTs, int64_t Size, int64_t Offset) | |||
1741 | : SDNode(Opcode, Order, dl, VTs), Size(Size), Offset(Offset) {} | |||
1742 | public: | |||
1743 | int64_t getFrameIndex() const { | |||
1744 | return cast<FrameIndexSDNode>(getOperand(1))->getIndex(); | |||
1745 | } | |||
1746 | ||||
1747 | bool hasOffset() const { return Offset >= 0; } | |||
1748 | int64_t getOffset() const { | |||
1749 | assert(hasOffset() && "offset is unknown")((hasOffset() && "offset is unknown") ? static_cast< void> (0) : __assert_fail ("hasOffset() && \"offset is unknown\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1749, __PRETTY_FUNCTION__)); | |||
1750 | return Offset; | |||
1751 | } | |||
1752 | int64_t getSize() const { | |||
1753 | assert(hasOffset() && "offset is unknown")((hasOffset() && "offset is unknown") ? static_cast< void> (0) : __assert_fail ("hasOffset() && \"offset is unknown\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1753, __PRETTY_FUNCTION__)); | |||
1754 | return Size; | |||
1755 | } | |||
1756 | ||||
1757 | // Methods to support isa and dyn_cast | |||
1758 | static bool classof(const SDNode *N) { | |||
1759 | return N->getOpcode() == ISD::LIFETIME_START || | |||
1760 | N->getOpcode() == ISD::LIFETIME_END; | |||
1761 | } | |||
1762 | }; | |||
1763 | ||||
1764 | /// This SDNode is used for PSEUDO_PROBE values, which are the function guid and | |||
1765 | /// the index of the basic block being probed. A pseudo probe serves as a place | |||
1766 | /// holder and will be removed at the end of compilation. It does not have any | |||
1767 | /// operand because we do not want the instruction selection to deal with any. | |||
1768 | class PseudoProbeSDNode : public SDNode { | |||
1769 | friend class SelectionDAG; | |||
1770 | uint64_t Guid; | |||
1771 | uint64_t Index; | |||
1772 | uint32_t Attributes; | |||
1773 | ||||
1774 | PseudoProbeSDNode(unsigned Opcode, unsigned Order, const DebugLoc &Dl, | |||
1775 | SDVTList VTs, uint64_t Guid, uint64_t Index, uint32_t Attr) | |||
1776 | : SDNode(Opcode, Order, Dl, VTs), Guid(Guid), Index(Index), | |||
1777 | Attributes(Attr) {} | |||
1778 | ||||
1779 | public: | |||
1780 | uint64_t getGuid() const { return Guid; } | |||
1781 | uint64_t getIndex() const { return Index; } | |||
1782 | uint32_t getAttributes() const { return Attributes; } | |||
1783 | ||||
1784 | // Methods to support isa and dyn_cast | |||
1785 | static bool classof(const SDNode *N) { | |||
1786 | return N->getOpcode() == ISD::PSEUDO_PROBE; | |||
1787 | } | |||
1788 | }; | |||
1789 | ||||
1790 | class JumpTableSDNode : public SDNode { | |||
1791 | friend class SelectionDAG; | |||
1792 | ||||
1793 | int JTI; | |||
1794 | unsigned TargetFlags; | |||
1795 | ||||
1796 | JumpTableSDNode(int jti, EVT VT, bool isTarg, unsigned TF) | |||
1797 | : SDNode(isTarg ? ISD::TargetJumpTable : ISD::JumpTable, | |||
1798 | 0, DebugLoc(), getSDVTList(VT)), JTI(jti), TargetFlags(TF) { | |||
1799 | } | |||
1800 | ||||
1801 | public: | |||
1802 | int getIndex() const { return JTI; } | |||
1803 | unsigned getTargetFlags() const { return TargetFlags; } | |||
1804 | ||||
1805 | static bool classof(const SDNode *N) { | |||
1806 | return N->getOpcode() == ISD::JumpTable || | |||
1807 | N->getOpcode() == ISD::TargetJumpTable; | |||
1808 | } | |||
1809 | }; | |||
1810 | ||||
1811 | class ConstantPoolSDNode : public SDNode { | |||
1812 | friend class SelectionDAG; | |||
1813 | ||||
1814 | union { | |||
1815 | const Constant *ConstVal; | |||
1816 | MachineConstantPoolValue *MachineCPVal; | |||
1817 | } Val; | |||
1818 | int Offset; // It's a MachineConstantPoolValue if top bit is set. | |||
1819 | Align Alignment; // Minimum alignment requirement of CP. | |||
1820 | unsigned TargetFlags; | |||
1821 | ||||
1822 | ConstantPoolSDNode(bool isTarget, const Constant *c, EVT VT, int o, | |||
1823 | Align Alignment, unsigned TF) | |||
1824 | : SDNode(isTarget ? ISD::TargetConstantPool : ISD::ConstantPool, 0, | |||
1825 | DebugLoc(), getSDVTList(VT)), | |||
1826 | Offset(o), Alignment(Alignment), TargetFlags(TF) { | |||
1827 | assert(Offset >= 0 && "Offset is too large")((Offset >= 0 && "Offset is too large") ? static_cast <void> (0) : __assert_fail ("Offset >= 0 && \"Offset is too large\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1827, __PRETTY_FUNCTION__)); | |||
1828 | Val.ConstVal = c; | |||
1829 | } | |||
1830 | ||||
1831 | ConstantPoolSDNode(bool isTarget, MachineConstantPoolValue *v, EVT VT, int o, | |||
1832 | Align Alignment, unsigned TF) | |||
1833 | : SDNode(isTarget ? ISD::TargetConstantPool : ISD::ConstantPool, 0, | |||
1834 | DebugLoc(), getSDVTList(VT)), | |||
1835 | Offset(o), Alignment(Alignment), TargetFlags(TF) { | |||
1836 | assert(Offset >= 0 && "Offset is too large")((Offset >= 0 && "Offset is too large") ? static_cast <void> (0) : __assert_fail ("Offset >= 0 && \"Offset is too large\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1836, __PRETTY_FUNCTION__)); | |||
1837 | Val.MachineCPVal = v; | |||
1838 | Offset |= 1 << (sizeof(unsigned)*CHAR_BIT8-1); | |||
1839 | } | |||
1840 | ||||
1841 | public: | |||
1842 | bool isMachineConstantPoolEntry() const { | |||
1843 | return Offset < 0; | |||
1844 | } | |||
1845 | ||||
1846 | const Constant *getConstVal() const { | |||
1847 | assert(!isMachineConstantPoolEntry() && "Wrong constantpool type")((!isMachineConstantPoolEntry() && "Wrong constantpool type" ) ? static_cast<void> (0) : __assert_fail ("!isMachineConstantPoolEntry() && \"Wrong constantpool type\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1847, __PRETTY_FUNCTION__)); | |||
1848 | return Val.ConstVal; | |||
1849 | } | |||
1850 | ||||
1851 | MachineConstantPoolValue *getMachineCPVal() const { | |||
1852 | assert(isMachineConstantPoolEntry() && "Wrong constantpool type")((isMachineConstantPoolEntry() && "Wrong constantpool type" ) ? static_cast<void> (0) : __assert_fail ("isMachineConstantPoolEntry() && \"Wrong constantpool type\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 1852, __PRETTY_FUNCTION__)); | |||
1853 | return Val.MachineCPVal; | |||
1854 | } | |||
1855 | ||||
1856 | int getOffset() const { | |||
1857 | return Offset & ~(1 << (sizeof(unsigned)*CHAR_BIT8-1)); | |||
1858 | } | |||
1859 | ||||
1860 | // Return the alignment of this constant pool object, which is either 0 (for | |||
1861 | // default alignment) or the desired value. | |||
1862 | Align getAlign() const { return Alignment; } | |||
1863 | unsigned getTargetFlags() const { return TargetFlags; } | |||
1864 | ||||
1865 | Type *getType() const; | |||
1866 | ||||
1867 | static bool classof(const SDNode *N) { | |||
1868 | return N->getOpcode() == ISD::ConstantPool || | |||
1869 | N->getOpcode() == ISD::TargetConstantPool; | |||
1870 | } | |||
1871 | }; | |||
1872 | ||||
1873 | /// Completely target-dependent object reference. | |||
1874 | class TargetIndexSDNode : public SDNode { | |||
1875 | friend class SelectionDAG; | |||
1876 | ||||
1877 | unsigned TargetFlags; | |||
1878 | int Index; | |||
1879 | int64_t Offset; | |||
1880 | ||||
1881 | public: | |||
1882 | TargetIndexSDNode(int Idx, EVT VT, int64_t Ofs, unsigned TF) | |||
1883 | : SDNode(ISD::TargetIndex, 0, DebugLoc(), getSDVTList(VT)), | |||
1884 | TargetFlags(TF), Index(Idx), Offset(Ofs) {} | |||
1885 | ||||
1886 | unsigned getTargetFlags() const { return TargetFlags; } | |||
1887 | int getIndex() const { return Index; } | |||
1888 | int64_t getOffset() const { return Offset; } | |||
1889 | ||||
1890 | static bool classof(const SDNode *N) { | |||
1891 | return N->getOpcode() == ISD::TargetIndex; | |||
1892 | } | |||
1893 | }; | |||
1894 | ||||
1895 | class BasicBlockSDNode : public SDNode { | |||
1896 | friend class SelectionDAG; | |||
1897 | ||||
1898 | MachineBasicBlock *MBB; | |||
1899 | ||||
1900 | /// Debug info is meaningful and potentially useful here, but we create | |||
1901 | /// blocks out of order when they're jumped to, which makes it a bit | |||
1902 | /// harder. Let's see if we need it first. | |||
1903 | explicit BasicBlockSDNode(MachineBasicBlock *mbb) | |||
1904 | : SDNode(ISD::BasicBlock, 0, DebugLoc(), getSDVTList(MVT::Other)), MBB(mbb) | |||
1905 | {} | |||
1906 | ||||
1907 | public: | |||
1908 | MachineBasicBlock *getBasicBlock() const { return MBB; } | |||
1909 | ||||
1910 | static bool classof(const SDNode *N) { | |||
1911 | return N->getOpcode() == ISD::BasicBlock; | |||
1912 | } | |||
1913 | }; | |||
1914 | ||||
1915 | /// A "pseudo-class" with methods for operating on BUILD_VECTORs. | |||
1916 | class BuildVectorSDNode : public SDNode { | |||
1917 | public: | |||
1918 | // These are constructed as SDNodes and then cast to BuildVectorSDNodes. | |||
1919 | explicit BuildVectorSDNode() = delete; | |||
1920 | ||||
1921 | /// Check if this is a constant splat, and if so, find the | |||
1922 | /// smallest element size that splats the vector. If MinSplatBits is | |||
1923 | /// nonzero, the element size must be at least that large. Note that the | |||
1924 | /// splat element may be the entire vector (i.e., a one element vector). | |||
1925 | /// Returns the splat element value in SplatValue. Any undefined bits in | |||
1926 | /// that value are zero, and the corresponding bits in the SplatUndef mask | |||
1927 | /// are set. The SplatBitSize value is set to the splat element size in | |||
1928 | /// bits. HasAnyUndefs is set to true if any bits in the vector are | |||
1929 | /// undefined. isBigEndian describes the endianness of the target. | |||
1930 | bool isConstantSplat(APInt &SplatValue, APInt &SplatUndef, | |||
1931 | unsigned &SplatBitSize, bool &HasAnyUndefs, | |||
1932 | unsigned MinSplatBits = 0, | |||
1933 | bool isBigEndian = false) const; | |||
1934 | ||||
1935 | /// Returns the demanded splatted value or a null value if this is not a | |||
1936 | /// splat. | |||
1937 | /// | |||
1938 | /// The DemandedElts mask indicates the elements that must be in the splat. | |||
1939 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
1940 | /// the vector width and set the bits where elements are undef. | |||
1941 | SDValue getSplatValue(const APInt &DemandedElts, | |||
1942 | BitVector *UndefElements = nullptr) const; | |||
1943 | ||||
1944 | /// Returns the splatted value or a null value if this is not a splat. | |||
1945 | /// | |||
1946 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
1947 | /// the vector width and set the bits where elements are undef. | |||
1948 | SDValue getSplatValue(BitVector *UndefElements = nullptr) const; | |||
1949 | ||||
1950 | /// Find the shortest repeating sequence of values in the build vector. | |||
1951 | /// | |||
1952 | /// e.g. { u, X, u, X, u, u, X, u } -> { X } | |||
1953 | /// { X, Y, u, Y, u, u, X, u } -> { X, Y } | |||
1954 | /// | |||
1955 | /// Currently this must be a power-of-2 build vector. | |||
1956 | /// The DemandedElts mask indicates the elements that must be present, | |||
1957 | /// undemanded elements in Sequence may be null (SDValue()). If passed a | |||
1958 | /// non-null UndefElements bitvector, it will resize it to match the original | |||
1959 | /// vector width and set the bits where elements are undef. If result is | |||
1960 | /// false, Sequence will be empty. | |||
1961 | bool getRepeatedSequence(const APInt &DemandedElts, | |||
1962 | SmallVectorImpl<SDValue> &Sequence, | |||
1963 | BitVector *UndefElements = nullptr) const; | |||
1964 | ||||
1965 | /// Find the shortest repeating sequence of values in the build vector. | |||
1966 | /// | |||
1967 | /// e.g. { u, X, u, X, u, u, X, u } -> { X } | |||
1968 | /// { X, Y, u, Y, u, u, X, u } -> { X, Y } | |||
1969 | /// | |||
1970 | /// Currently this must be a power-of-2 build vector. | |||
1971 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
1972 | /// the original vector width and set the bits where elements are undef. | |||
1973 | /// If result is false, Sequence will be empty. | |||
1974 | bool getRepeatedSequence(SmallVectorImpl<SDValue> &Sequence, | |||
1975 | BitVector *UndefElements = nullptr) const; | |||
1976 | ||||
1977 | /// Returns the demanded splatted constant or null if this is not a constant | |||
1978 | /// splat. | |||
1979 | /// | |||
1980 | /// The DemandedElts mask indicates the elements that must be in the splat. | |||
1981 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
1982 | /// the vector width and set the bits where elements are undef. | |||
1983 | ConstantSDNode * | |||
1984 | getConstantSplatNode(const APInt &DemandedElts, | |||
1985 | BitVector *UndefElements = nullptr) const; | |||
1986 | ||||
1987 | /// Returns the splatted constant or null if this is not a constant | |||
1988 | /// splat. | |||
1989 | /// | |||
1990 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
1991 | /// the vector width and set the bits where elements are undef. | |||
1992 | ConstantSDNode * | |||
1993 | getConstantSplatNode(BitVector *UndefElements = nullptr) const; | |||
1994 | ||||
1995 | /// Returns the demanded splatted constant FP or null if this is not a | |||
1996 | /// constant FP splat. | |||
1997 | /// | |||
1998 | /// The DemandedElts mask indicates the elements that must be in the splat. | |||
1999 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
2000 | /// the vector width and set the bits where elements are undef. | |||
2001 | ConstantFPSDNode * | |||
2002 | getConstantFPSplatNode(const APInt &DemandedElts, | |||
2003 | BitVector *UndefElements = nullptr) const; | |||
2004 | ||||
2005 | /// Returns the splatted constant FP or null if this is not a constant | |||
2006 | /// FP splat. | |||
2007 | /// | |||
2008 | /// If passed a non-null UndefElements bitvector, it will resize it to match | |||
2009 | /// the vector width and set the bits where elements are undef. | |||
2010 | ConstantFPSDNode * | |||
2011 | getConstantFPSplatNode(BitVector *UndefElements = nullptr) const; | |||
2012 | ||||
2013 | /// If this is a constant FP splat and the splatted constant FP is an | |||
2014 | /// exact power or 2, return the log base 2 integer value. Otherwise, | |||
2015 | /// return -1. | |||
2016 | /// | |||
2017 | /// The BitWidth specifies the necessary bit precision. | |||
2018 | int32_t getConstantFPSplatPow2ToLog2Int(BitVector *UndefElements, | |||
2019 | uint32_t BitWidth) const; | |||
2020 | ||||
2021 | bool isConstant() const; | |||
2022 | ||||
2023 | static bool classof(const SDNode *N) { | |||
2024 | return N->getOpcode() == ISD::BUILD_VECTOR; | |||
2025 | } | |||
2026 | }; | |||
2027 | ||||
2028 | /// An SDNode that holds an arbitrary LLVM IR Value. This is | |||
2029 | /// used when the SelectionDAG needs to make a simple reference to something | |||
2030 | /// in the LLVM IR representation. | |||
2031 | /// | |||
2032 | class SrcValueSDNode : public SDNode { | |||
2033 | friend class SelectionDAG; | |||
2034 | ||||
2035 | const Value *V; | |||
2036 | ||||
2037 | /// Create a SrcValue for a general value. | |||
2038 | explicit SrcValueSDNode(const Value *v) | |||
2039 | : SDNode(ISD::SRCVALUE, 0, DebugLoc(), getSDVTList(MVT::Other)), V(v) {} | |||
2040 | ||||
2041 | public: | |||
2042 | /// Return the contained Value. | |||
2043 | const Value *getValue() const { return V; } | |||
2044 | ||||
2045 | static bool classof(const SDNode *N) { | |||
2046 | return N->getOpcode() == ISD::SRCVALUE; | |||
2047 | } | |||
2048 | }; | |||
2049 | ||||
2050 | class MDNodeSDNode : public SDNode { | |||
2051 | friend class SelectionDAG; | |||
2052 | ||||
2053 | const MDNode *MD; | |||
2054 | ||||
2055 | explicit MDNodeSDNode(const MDNode *md) | |||
2056 | : SDNode(ISD::MDNODE_SDNODE, 0, DebugLoc(), getSDVTList(MVT::Other)), MD(md) | |||
2057 | {} | |||
2058 | ||||
2059 | public: | |||
2060 | const MDNode *getMD() const { return MD; } | |||
2061 | ||||
2062 | static bool classof(const SDNode *N) { | |||
2063 | return N->getOpcode() == ISD::MDNODE_SDNODE; | |||
2064 | } | |||
2065 | }; | |||
2066 | ||||
2067 | class RegisterSDNode : public SDNode { | |||
2068 | friend class SelectionDAG; | |||
2069 | ||||
2070 | Register Reg; | |||
2071 | ||||
2072 | RegisterSDNode(Register reg, EVT VT) | |||
2073 | : SDNode(ISD::Register, 0, DebugLoc(), getSDVTList(VT)), Reg(reg) {} | |||
2074 | ||||
2075 | public: | |||
2076 | Register getReg() const { return Reg; } | |||
2077 | ||||
2078 | static bool classof(const SDNode *N) { | |||
2079 | return N->getOpcode() == ISD::Register; | |||
2080 | } | |||
2081 | }; | |||
2082 | ||||
2083 | class RegisterMaskSDNode : public SDNode { | |||
2084 | friend class SelectionDAG; | |||
2085 | ||||
2086 | // The memory for RegMask is not owned by the node. | |||
2087 | const uint32_t *RegMask; | |||
2088 | ||||
2089 | RegisterMaskSDNode(const uint32_t *mask) | |||
2090 | : SDNode(ISD::RegisterMask, 0, DebugLoc(), getSDVTList(MVT::Untyped)), | |||
2091 | RegMask(mask) {} | |||
2092 | ||||
2093 | public: | |||
2094 | const uint32_t *getRegMask() const { return RegMask; } | |||
2095 | ||||
2096 | static bool classof(const SDNode *N) { | |||
2097 | return N->getOpcode() == ISD::RegisterMask; | |||
2098 | } | |||
2099 | }; | |||
2100 | ||||
2101 | class BlockAddressSDNode : public SDNode { | |||
2102 | friend class SelectionDAG; | |||
2103 | ||||
2104 | const BlockAddress *BA; | |||
2105 | int64_t Offset; | |||
2106 | unsigned TargetFlags; | |||
2107 | ||||
2108 | BlockAddressSDNode(unsigned NodeTy, EVT VT, const BlockAddress *ba, | |||
2109 | int64_t o, unsigned Flags) | |||
2110 | : SDNode(NodeTy, 0, DebugLoc(), getSDVTList(VT)), | |||
2111 | BA(ba), Offset(o), TargetFlags(Flags) {} | |||
2112 | ||||
2113 | public: | |||
2114 | const BlockAddress *getBlockAddress() const { return BA; } | |||
2115 | int64_t getOffset() const { return Offset; } | |||
2116 | unsigned getTargetFlags() const { return TargetFlags; } | |||
2117 | ||||
2118 | static bool classof(const SDNode *N) { | |||
2119 | return N->getOpcode() == ISD::BlockAddress || | |||
2120 | N->getOpcode() == ISD::TargetBlockAddress; | |||
2121 | } | |||
2122 | }; | |||
2123 | ||||
2124 | class LabelSDNode : public SDNode { | |||
2125 | friend class SelectionDAG; | |||
2126 | ||||
2127 | MCSymbol *Label; | |||
2128 | ||||
2129 | LabelSDNode(unsigned Opcode, unsigned Order, const DebugLoc &dl, MCSymbol *L) | |||
2130 | : SDNode(Opcode, Order, dl, getSDVTList(MVT::Other)), Label(L) { | |||
2131 | assert(LabelSDNode::classof(this) && "not a label opcode")((LabelSDNode::classof(this) && "not a label opcode") ? static_cast<void> (0) : __assert_fail ("LabelSDNode::classof(this) && \"not a label opcode\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2131, __PRETTY_FUNCTION__)); | |||
2132 | } | |||
2133 | ||||
2134 | public: | |||
2135 | MCSymbol *getLabel() const { return Label; } | |||
2136 | ||||
2137 | static bool classof(const SDNode *N) { | |||
2138 | return N->getOpcode() == ISD::EH_LABEL || | |||
2139 | N->getOpcode() == ISD::ANNOTATION_LABEL; | |||
2140 | } | |||
2141 | }; | |||
2142 | ||||
2143 | class ExternalSymbolSDNode : public SDNode { | |||
2144 | friend class SelectionDAG; | |||
2145 | ||||
2146 | const char *Symbol; | |||
2147 | unsigned TargetFlags; | |||
2148 | ||||
2149 | ExternalSymbolSDNode(bool isTarget, const char *Sym, unsigned TF, EVT VT) | |||
2150 | : SDNode(isTarget ? ISD::TargetExternalSymbol : ISD::ExternalSymbol, 0, | |||
2151 | DebugLoc(), getSDVTList(VT)), | |||
2152 | Symbol(Sym), TargetFlags(TF) {} | |||
2153 | ||||
2154 | public: | |||
2155 | const char *getSymbol() const { return Symbol; } | |||
2156 | unsigned getTargetFlags() const { return TargetFlags; } | |||
2157 | ||||
2158 | static bool classof(const SDNode *N) { | |||
2159 | return N->getOpcode() == ISD::ExternalSymbol || | |||
2160 | N->getOpcode() == ISD::TargetExternalSymbol; | |||
2161 | } | |||
2162 | }; | |||
2163 | ||||
2164 | class MCSymbolSDNode : public SDNode { | |||
2165 | friend class SelectionDAG; | |||
2166 | ||||
2167 | MCSymbol *Symbol; | |||
2168 | ||||
2169 | MCSymbolSDNode(MCSymbol *Symbol, EVT VT) | |||
2170 | : SDNode(ISD::MCSymbol, 0, DebugLoc(), getSDVTList(VT)), Symbol(Symbol) {} | |||
2171 | ||||
2172 | public: | |||
2173 | MCSymbol *getMCSymbol() const { return Symbol; } | |||
2174 | ||||
2175 | static bool classof(const SDNode *N) { | |||
2176 | return N->getOpcode() == ISD::MCSymbol; | |||
2177 | } | |||
2178 | }; | |||
2179 | ||||
2180 | class CondCodeSDNode : public SDNode { | |||
2181 | friend class SelectionDAG; | |||
2182 | ||||
2183 | ISD::CondCode Condition; | |||
2184 | ||||
2185 | explicit CondCodeSDNode(ISD::CondCode Cond) | |||
2186 | : SDNode(ISD::CONDCODE, 0, DebugLoc(), getSDVTList(MVT::Other)), | |||
2187 | Condition(Cond) {} | |||
2188 | ||||
2189 | public: | |||
2190 | ISD::CondCode get() const { return Condition; } | |||
2191 | ||||
2192 | static bool classof(const SDNode *N) { | |||
2193 | return N->getOpcode() == ISD::CONDCODE; | |||
2194 | } | |||
2195 | }; | |||
2196 | ||||
2197 | /// This class is used to represent EVT's, which are used | |||
2198 | /// to parameterize some operations. | |||
2199 | class VTSDNode : public SDNode { | |||
2200 | friend class SelectionDAG; | |||
2201 | ||||
2202 | EVT ValueType; | |||
2203 | ||||
2204 | explicit VTSDNode(EVT VT) | |||
2205 | : SDNode(ISD::VALUETYPE, 0, DebugLoc(), getSDVTList(MVT::Other)), | |||
2206 | ValueType(VT) {} | |||
2207 | ||||
2208 | public: | |||
2209 | EVT getVT() const { return ValueType; } | |||
2210 | ||||
2211 | static bool classof(const SDNode *N) { | |||
2212 | return N->getOpcode() == ISD::VALUETYPE; | |||
2213 | } | |||
2214 | }; | |||
2215 | ||||
2216 | /// Base class for LoadSDNode and StoreSDNode | |||
2217 | class LSBaseSDNode : public MemSDNode { | |||
2218 | public: | |||
2219 | LSBaseSDNode(ISD::NodeType NodeTy, unsigned Order, const DebugLoc &dl, | |||
2220 | SDVTList VTs, ISD::MemIndexedMode AM, EVT MemVT, | |||
2221 | MachineMemOperand *MMO) | |||
2222 | : MemSDNode(NodeTy, Order, dl, VTs, MemVT, MMO) { | |||
2223 | LSBaseSDNodeBits.AddressingMode = AM; | |||
2224 | assert(getAddressingMode() == AM && "Value truncated")((getAddressingMode() == AM && "Value truncated") ? static_cast <void> (0) : __assert_fail ("getAddressingMode() == AM && \"Value truncated\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2224, __PRETTY_FUNCTION__)); | |||
2225 | } | |||
2226 | ||||
2227 | const SDValue &getOffset() const { | |||
2228 | return getOperand(getOpcode() == ISD::LOAD ? 2 : 3); | |||
2229 | } | |||
2230 | ||||
2231 | /// Return the addressing mode for this load or store: | |||
2232 | /// unindexed, pre-inc, pre-dec, post-inc, or post-dec. | |||
2233 | ISD::MemIndexedMode getAddressingMode() const { | |||
2234 | return static_cast<ISD::MemIndexedMode>(LSBaseSDNodeBits.AddressingMode); | |||
2235 | } | |||
2236 | ||||
2237 | /// Return true if this is a pre/post inc/dec load/store. | |||
2238 | bool isIndexed() const { return getAddressingMode() != ISD::UNINDEXED; } | |||
2239 | ||||
2240 | /// Return true if this is NOT a pre/post inc/dec load/store. | |||
2241 | bool isUnindexed() const { return getAddressingMode() == ISD::UNINDEXED; } | |||
2242 | ||||
2243 | static bool classof(const SDNode *N) { | |||
2244 | return N->getOpcode() == ISD::LOAD || | |||
2245 | N->getOpcode() == ISD::STORE; | |||
2246 | } | |||
2247 | }; | |||
2248 | ||||
2249 | /// This class is used to represent ISD::LOAD nodes. | |||
2250 | class LoadSDNode : public LSBaseSDNode { | |||
2251 | friend class SelectionDAG; | |||
2252 | ||||
2253 | LoadSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2254 | ISD::MemIndexedMode AM, ISD::LoadExtType ETy, EVT MemVT, | |||
2255 | MachineMemOperand *MMO) | |||
2256 | : LSBaseSDNode(ISD::LOAD, Order, dl, VTs, AM, MemVT, MMO) { | |||
2257 | LoadSDNodeBits.ExtTy = ETy; | |||
2258 | assert(readMem() && "Load MachineMemOperand is not a load!")((readMem() && "Load MachineMemOperand is not a load!" ) ? static_cast<void> (0) : __assert_fail ("readMem() && \"Load MachineMemOperand is not a load!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2258, __PRETTY_FUNCTION__)); | |||
2259 | assert(!writeMem() && "Load MachineMemOperand is a store!")((!writeMem() && "Load MachineMemOperand is a store!" ) ? static_cast<void> (0) : __assert_fail ("!writeMem() && \"Load MachineMemOperand is a store!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2259, __PRETTY_FUNCTION__)); | |||
2260 | } | |||
2261 | ||||
2262 | public: | |||
2263 | /// Return whether this is a plain node, | |||
2264 | /// or one of the varieties of value-extending loads. | |||
2265 | ISD::LoadExtType getExtensionType() const { | |||
2266 | return static_cast<ISD::LoadExtType>(LoadSDNodeBits.ExtTy); | |||
2267 | } | |||
2268 | ||||
2269 | const SDValue &getBasePtr() const { return getOperand(1); } | |||
2270 | const SDValue &getOffset() const { return getOperand(2); } | |||
2271 | ||||
2272 | static bool classof(const SDNode *N) { | |||
2273 | return N->getOpcode() == ISD::LOAD; | |||
2274 | } | |||
2275 | }; | |||
2276 | ||||
2277 | /// This class is used to represent ISD::STORE nodes. | |||
2278 | class StoreSDNode : public LSBaseSDNode { | |||
2279 | friend class SelectionDAG; | |||
2280 | ||||
2281 | StoreSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2282 | ISD::MemIndexedMode AM, bool isTrunc, EVT MemVT, | |||
2283 | MachineMemOperand *MMO) | |||
2284 | : LSBaseSDNode(ISD::STORE, Order, dl, VTs, AM, MemVT, MMO) { | |||
2285 | StoreSDNodeBits.IsTruncating = isTrunc; | |||
2286 | assert(!readMem() && "Store MachineMemOperand is a load!")((!readMem() && "Store MachineMemOperand is a load!") ? static_cast<void> (0) : __assert_fail ("!readMem() && \"Store MachineMemOperand is a load!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2286, __PRETTY_FUNCTION__)); | |||
2287 | assert(writeMem() && "Store MachineMemOperand is not a store!")((writeMem() && "Store MachineMemOperand is not a store!" ) ? static_cast<void> (0) : __assert_fail ("writeMem() && \"Store MachineMemOperand is not a store!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2287, __PRETTY_FUNCTION__)); | |||
2288 | } | |||
2289 | ||||
2290 | public: | |||
2291 | /// Return true if the op does a truncation before store. | |||
2292 | /// For integers this is the same as doing a TRUNCATE and storing the result. | |||
2293 | /// For floats, it is the same as doing an FP_ROUND and storing the result. | |||
2294 | bool isTruncatingStore() const { return StoreSDNodeBits.IsTruncating; } | |||
2295 | void setTruncatingStore(bool Truncating) { | |||
2296 | StoreSDNodeBits.IsTruncating = Truncating; | |||
2297 | } | |||
2298 | ||||
2299 | const SDValue &getValue() const { return getOperand(1); } | |||
2300 | const SDValue &getBasePtr() const { return getOperand(2); } | |||
2301 | const SDValue &getOffset() const { return getOperand(3); } | |||
2302 | ||||
2303 | static bool classof(const SDNode *N) { | |||
2304 | return N->getOpcode() == ISD::STORE; | |||
2305 | } | |||
2306 | }; | |||
2307 | ||||
2308 | /// This base class is used to represent MLOAD and MSTORE nodes | |||
2309 | class MaskedLoadStoreSDNode : public MemSDNode { | |||
2310 | public: | |||
2311 | friend class SelectionDAG; | |||
2312 | ||||
2313 | MaskedLoadStoreSDNode(ISD::NodeType NodeTy, unsigned Order, | |||
2314 | const DebugLoc &dl, SDVTList VTs, | |||
2315 | ISD::MemIndexedMode AM, EVT MemVT, | |||
2316 | MachineMemOperand *MMO) | |||
2317 | : MemSDNode(NodeTy, Order, dl, VTs, MemVT, MMO) { | |||
2318 | LSBaseSDNodeBits.AddressingMode = AM; | |||
2319 | assert(getAddressingMode() == AM && "Value truncated")((getAddressingMode() == AM && "Value truncated") ? static_cast <void> (0) : __assert_fail ("getAddressingMode() == AM && \"Value truncated\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2319, __PRETTY_FUNCTION__)); | |||
2320 | } | |||
2321 | ||||
2322 | // MaskedLoadSDNode (Chain, ptr, offset, mask, passthru) | |||
2323 | // MaskedStoreSDNode (Chain, data, ptr, offset, mask) | |||
2324 | // Mask is a vector of i1 elements | |||
2325 | const SDValue &getOffset() const { | |||
2326 | return getOperand(getOpcode() == ISD::MLOAD ? 2 : 3); | |||
2327 | } | |||
2328 | const SDValue &getMask() const { | |||
2329 | return getOperand(getOpcode() == ISD::MLOAD ? 3 : 4); | |||
2330 | } | |||
2331 | ||||
2332 | /// Return the addressing mode for this load or store: | |||
2333 | /// unindexed, pre-inc, pre-dec, post-inc, or post-dec. | |||
2334 | ISD::MemIndexedMode getAddressingMode() const { | |||
2335 | return static_cast<ISD::MemIndexedMode>(LSBaseSDNodeBits.AddressingMode); | |||
2336 | } | |||
2337 | ||||
2338 | /// Return true if this is a pre/post inc/dec load/store. | |||
2339 | bool isIndexed() const { return getAddressingMode() != ISD::UNINDEXED; } | |||
2340 | ||||
2341 | /// Return true if this is NOT a pre/post inc/dec load/store. | |||
2342 | bool isUnindexed() const { return getAddressingMode() == ISD::UNINDEXED; } | |||
2343 | ||||
2344 | static bool classof(const SDNode *N) { | |||
2345 | return N->getOpcode() == ISD::MLOAD || | |||
2346 | N->getOpcode() == ISD::MSTORE; | |||
2347 | } | |||
2348 | }; | |||
2349 | ||||
2350 | /// This class is used to represent an MLOAD node | |||
2351 | class MaskedLoadSDNode : public MaskedLoadStoreSDNode { | |||
2352 | public: | |||
2353 | friend class SelectionDAG; | |||
2354 | ||||
2355 | MaskedLoadSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2356 | ISD::MemIndexedMode AM, ISD::LoadExtType ETy, | |||
2357 | bool IsExpanding, EVT MemVT, MachineMemOperand *MMO) | |||
2358 | : MaskedLoadStoreSDNode(ISD::MLOAD, Order, dl, VTs, AM, MemVT, MMO) { | |||
2359 | LoadSDNodeBits.ExtTy = ETy; | |||
2360 | LoadSDNodeBits.IsExpanding = IsExpanding; | |||
2361 | } | |||
2362 | ||||
2363 | ISD::LoadExtType getExtensionType() const { | |||
2364 | return static_cast<ISD::LoadExtType>(LoadSDNodeBits.ExtTy); | |||
2365 | } | |||
2366 | ||||
2367 | const SDValue &getBasePtr() const { return getOperand(1); } | |||
2368 | const SDValue &getOffset() const { return getOperand(2); } | |||
2369 | const SDValue &getMask() const { return getOperand(3); } | |||
2370 | const SDValue &getPassThru() const { return getOperand(4); } | |||
2371 | ||||
2372 | static bool classof(const SDNode *N) { | |||
2373 | return N->getOpcode() == ISD::MLOAD; | |||
2374 | } | |||
2375 | ||||
2376 | bool isExpandingLoad() const { return LoadSDNodeBits.IsExpanding; } | |||
2377 | }; | |||
2378 | ||||
2379 | /// This class is used to represent an MSTORE node | |||
2380 | class MaskedStoreSDNode : public MaskedLoadStoreSDNode { | |||
2381 | public: | |||
2382 | friend class SelectionDAG; | |||
2383 | ||||
2384 | MaskedStoreSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2385 | ISD::MemIndexedMode AM, bool isTrunc, bool isCompressing, | |||
2386 | EVT MemVT, MachineMemOperand *MMO) | |||
2387 | : MaskedLoadStoreSDNode(ISD::MSTORE, Order, dl, VTs, AM, MemVT, MMO) { | |||
2388 | StoreSDNodeBits.IsTruncating = isTrunc; | |||
2389 | StoreSDNodeBits.IsCompressing = isCompressing; | |||
2390 | } | |||
2391 | ||||
2392 | /// Return true if the op does a truncation before store. | |||
2393 | /// For integers this is the same as doing a TRUNCATE and storing the result. | |||
2394 | /// For floats, it is the same as doing an FP_ROUND and storing the result. | |||
2395 | bool isTruncatingStore() const { return StoreSDNodeBits.IsTruncating; } | |||
2396 | ||||
2397 | /// Returns true if the op does a compression to the vector before storing. | |||
2398 | /// The node contiguously stores the active elements (integers or floats) | |||
2399 | /// in src (those with their respective bit set in writemask k) to unaligned | |||
2400 | /// memory at base_addr. | |||
2401 | bool isCompressingStore() const { return StoreSDNodeBits.IsCompressing; } | |||
2402 | ||||
2403 | const SDValue &getValue() const { return getOperand(1); } | |||
2404 | const SDValue &getBasePtr() const { return getOperand(2); } | |||
2405 | const SDValue &getOffset() const { return getOperand(3); } | |||
2406 | const SDValue &getMask() const { return getOperand(4); } | |||
2407 | ||||
2408 | static bool classof(const SDNode *N) { | |||
2409 | return N->getOpcode() == ISD::MSTORE; | |||
2410 | } | |||
2411 | }; | |||
2412 | ||||
2413 | /// This is a base class used to represent | |||
2414 | /// MGATHER and MSCATTER nodes | |||
2415 | /// | |||
2416 | class MaskedGatherScatterSDNode : public MemSDNode { | |||
2417 | public: | |||
2418 | friend class SelectionDAG; | |||
2419 | ||||
2420 | MaskedGatherScatterSDNode(ISD::NodeType NodeTy, unsigned Order, | |||
2421 | const DebugLoc &dl, SDVTList VTs, EVT MemVT, | |||
2422 | MachineMemOperand *MMO, ISD::MemIndexType IndexType) | |||
2423 | : MemSDNode(NodeTy, Order, dl, VTs, MemVT, MMO) { | |||
2424 | LSBaseSDNodeBits.AddressingMode = IndexType; | |||
2425 | assert(getIndexType() == IndexType && "Value truncated")((getIndexType() == IndexType && "Value truncated") ? static_cast<void> (0) : __assert_fail ("getIndexType() == IndexType && \"Value truncated\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2425, __PRETTY_FUNCTION__)); | |||
2426 | } | |||
2427 | ||||
2428 | /// How is Index applied to BasePtr when computing addresses. | |||
2429 | ISD::MemIndexType getIndexType() const { | |||
2430 | return static_cast<ISD::MemIndexType>(LSBaseSDNodeBits.AddressingMode); | |||
2431 | } | |||
2432 | void setIndexType(ISD::MemIndexType IndexType) { | |||
2433 | LSBaseSDNodeBits.AddressingMode = IndexType; | |||
2434 | } | |||
2435 | bool isIndexScaled() const { | |||
2436 | return (getIndexType() == ISD::SIGNED_SCALED) || | |||
2437 | (getIndexType() == ISD::UNSIGNED_SCALED); | |||
2438 | } | |||
2439 | bool isIndexSigned() const { | |||
2440 | return (getIndexType() == ISD::SIGNED_SCALED) || | |||
2441 | (getIndexType() == ISD::SIGNED_UNSCALED); | |||
2442 | } | |||
2443 | ||||
2444 | // In the both nodes address is Op1, mask is Op2: | |||
2445 | // MaskedGatherSDNode (Chain, passthru, mask, base, index, scale) | |||
2446 | // MaskedScatterSDNode (Chain, value, mask, base, index, scale) | |||
2447 | // Mask is a vector of i1 elements | |||
2448 | const SDValue &getBasePtr() const { return getOperand(3); } | |||
2449 | const SDValue &getIndex() const { return getOperand(4); } | |||
2450 | const SDValue &getMask() const { return getOperand(2); } | |||
2451 | const SDValue &getScale() const { return getOperand(5); } | |||
2452 | ||||
2453 | static bool classof(const SDNode *N) { | |||
2454 | return N->getOpcode() == ISD::MGATHER || | |||
2455 | N->getOpcode() == ISD::MSCATTER; | |||
2456 | } | |||
2457 | }; | |||
2458 | ||||
2459 | /// This class is used to represent an MGATHER node | |||
2460 | /// | |||
2461 | class MaskedGatherSDNode : public MaskedGatherScatterSDNode { | |||
2462 | public: | |||
2463 | friend class SelectionDAG; | |||
2464 | ||||
2465 | MaskedGatherSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2466 | EVT MemVT, MachineMemOperand *MMO, | |||
2467 | ISD::MemIndexType IndexType, ISD::LoadExtType ETy) | |||
2468 | : MaskedGatherScatterSDNode(ISD::MGATHER, Order, dl, VTs, MemVT, MMO, | |||
2469 | IndexType) { | |||
2470 | LoadSDNodeBits.ExtTy = ETy; | |||
2471 | } | |||
2472 | ||||
2473 | const SDValue &getPassThru() const { return getOperand(1); } | |||
2474 | ||||
2475 | ISD::LoadExtType getExtensionType() const { | |||
2476 | return ISD::LoadExtType(LoadSDNodeBits.ExtTy); | |||
2477 | } | |||
2478 | ||||
2479 | static bool classof(const SDNode *N) { | |||
2480 | return N->getOpcode() == ISD::MGATHER; | |||
2481 | } | |||
2482 | }; | |||
2483 | ||||
2484 | /// This class is used to represent an MSCATTER node | |||
2485 | /// | |||
2486 | class MaskedScatterSDNode : public MaskedGatherScatterSDNode { | |||
2487 | public: | |||
2488 | friend class SelectionDAG; | |||
2489 | ||||
2490 | MaskedScatterSDNode(unsigned Order, const DebugLoc &dl, SDVTList VTs, | |||
2491 | EVT MemVT, MachineMemOperand *MMO, | |||
2492 | ISD::MemIndexType IndexType, bool IsTrunc) | |||
2493 | : MaskedGatherScatterSDNode(ISD::MSCATTER, Order, dl, VTs, MemVT, MMO, | |||
2494 | IndexType) { | |||
2495 | StoreSDNodeBits.IsTruncating = IsTrunc; | |||
2496 | } | |||
2497 | ||||
2498 | /// Return true if the op does a truncation before store. | |||
2499 | /// For integers this is the same as doing a TRUNCATE and storing the result. | |||
2500 | /// For floats, it is the same as doing an FP_ROUND and storing the result. | |||
2501 | bool isTruncatingStore() const { return StoreSDNodeBits.IsTruncating; } | |||
2502 | ||||
2503 | const SDValue &getValue() const { return getOperand(1); } | |||
2504 | ||||
2505 | static bool classof(const SDNode *N) { | |||
2506 | return N->getOpcode() == ISD::MSCATTER; | |||
2507 | } | |||
2508 | }; | |||
2509 | ||||
2510 | /// An SDNode that represents everything that will be needed | |||
2511 | /// to construct a MachineInstr. These nodes are created during the | |||
2512 | /// instruction selection proper phase. | |||
2513 | /// | |||
2514 | /// Note that the only supported way to set the `memoperands` is by calling the | |||
2515 | /// `SelectionDAG::setNodeMemRefs` function as the memory management happens | |||
2516 | /// inside the DAG rather than in the node. | |||
2517 | class MachineSDNode : public SDNode { | |||
2518 | private: | |||
2519 | friend class SelectionDAG; | |||
2520 | ||||
2521 | MachineSDNode(unsigned Opc, unsigned Order, const DebugLoc &DL, SDVTList VTs) | |||
2522 | : SDNode(Opc, Order, DL, VTs) {} | |||
2523 | ||||
2524 | // We use a pointer union between a single `MachineMemOperand` pointer and | |||
2525 | // a pointer to an array of `MachineMemOperand` pointers. This is null when | |||
2526 | // the number of these is zero, the single pointer variant used when the | |||
2527 | // number is one, and the array is used for larger numbers. | |||
2528 | // | |||
2529 | // The array is allocated via the `SelectionDAG`'s allocator and so will | |||
2530 | // always live until the DAG is cleaned up and doesn't require ownership here. | |||
2531 | // | |||
2532 | // We can't use something simpler like `TinyPtrVector` here because `SDNode` | |||
2533 | // subclasses aren't managed in a conforming C++ manner. See the comments on | |||
2534 | // `SelectionDAG::MorphNodeTo` which details what all goes on, but the | |||
2535 | // constraint here is that these don't manage memory with their constructor or | |||
2536 | // destructor and can be initialized to a good state even if they start off | |||
2537 | // uninitialized. | |||
2538 | PointerUnion<MachineMemOperand *, MachineMemOperand **> MemRefs = {}; | |||
2539 | ||||
2540 | // Note that this could be folded into the above `MemRefs` member if doing so | |||
2541 | // is advantageous at some point. We don't need to store this in most cases. | |||
2542 | // However, at the moment this doesn't appear to make the allocation any | |||
2543 | // smaller and makes the code somewhat simpler to read. | |||
2544 | int NumMemRefs = 0; | |||
2545 | ||||
2546 | public: | |||
2547 | using mmo_iterator = ArrayRef<MachineMemOperand *>::const_iterator; | |||
2548 | ||||
2549 | ArrayRef<MachineMemOperand *> memoperands() const { | |||
2550 | // Special case the common cases. | |||
2551 | if (NumMemRefs == 0) | |||
2552 | return {}; | |||
2553 | if (NumMemRefs == 1) | |||
2554 | return makeArrayRef(MemRefs.getAddrOfPtr1(), 1); | |||
2555 | ||||
2556 | // Otherwise we have an actual array. | |||
2557 | return makeArrayRef(MemRefs.get<MachineMemOperand **>(), NumMemRefs); | |||
2558 | } | |||
2559 | mmo_iterator memoperands_begin() const { return memoperands().begin(); } | |||
2560 | mmo_iterator memoperands_end() const { return memoperands().end(); } | |||
2561 | bool memoperands_empty() const { return memoperands().empty(); } | |||
2562 | ||||
2563 | /// Clear out the memory reference descriptor list. | |||
2564 | void clearMemRefs() { | |||
2565 | MemRefs = nullptr; | |||
2566 | NumMemRefs = 0; | |||
2567 | } | |||
2568 | ||||
2569 | static bool classof(const SDNode *N) { | |||
2570 | return N->isMachineOpcode(); | |||
2571 | } | |||
2572 | }; | |||
2573 | ||||
2574 | /// An SDNode that records if a register contains a value that is guaranteed to | |||
2575 | /// be aligned accordingly. | |||
2576 | class AssertAlignSDNode : public SDNode { | |||
2577 | Align Alignment; | |||
2578 | ||||
2579 | public: | |||
2580 | AssertAlignSDNode(unsigned Order, const DebugLoc &DL, EVT VT, Align A) | |||
2581 | : SDNode(ISD::AssertAlign, Order, DL, getSDVTList(VT)), Alignment(A) {} | |||
2582 | ||||
2583 | Align getAlign() const { return Alignment; } | |||
2584 | ||||
2585 | static bool classof(const SDNode *N) { | |||
2586 | return N->getOpcode() == ISD::AssertAlign; | |||
2587 | } | |||
2588 | }; | |||
2589 | ||||
2590 | class SDNodeIterator : public std::iterator<std::forward_iterator_tag, | |||
2591 | SDNode, ptrdiff_t> { | |||
2592 | const SDNode *Node; | |||
2593 | unsigned Operand; | |||
2594 | ||||
2595 | SDNodeIterator(const SDNode *N, unsigned Op) : Node(N), Operand(Op) {} | |||
2596 | ||||
2597 | public: | |||
2598 | bool operator==(const SDNodeIterator& x) const { | |||
2599 | return Operand == x.Operand; | |||
2600 | } | |||
2601 | bool operator!=(const SDNodeIterator& x) const { return !operator==(x); } | |||
2602 | ||||
2603 | pointer operator*() const { | |||
2604 | return Node->getOperand(Operand).getNode(); | |||
2605 | } | |||
2606 | pointer operator->() const { return operator*(); } | |||
2607 | ||||
2608 | SDNodeIterator& operator++() { // Preincrement | |||
2609 | ++Operand; | |||
2610 | return *this; | |||
2611 | } | |||
2612 | SDNodeIterator operator++(int) { // Postincrement | |||
2613 | SDNodeIterator tmp = *this; ++*this; return tmp; | |||
2614 | } | |||
2615 | size_t operator-(SDNodeIterator Other) const { | |||
2616 | assert(Node == Other.Node &&((Node == Other.Node && "Cannot compare iterators of two different nodes!" ) ? static_cast<void> (0) : __assert_fail ("Node == Other.Node && \"Cannot compare iterators of two different nodes!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2617, __PRETTY_FUNCTION__)) | |||
2617 | "Cannot compare iterators of two different nodes!")((Node == Other.Node && "Cannot compare iterators of two different nodes!" ) ? static_cast<void> (0) : __assert_fail ("Node == Other.Node && \"Cannot compare iterators of two different nodes!\"" , "/build/llvm-toolchain-snapshot-13~++20210310111112+6f912a2cd491/llvm/include/llvm/CodeGen/SelectionDAGNodes.h" , 2617, __PRETTY_FUNCTION__)); | |||
2618 | return Operand - Other.Operand; | |||
2619 | } | |||
2620 | ||||
2621 | static SDNodeIterator begin(const SDNode *N) { return SDNodeIterator(N, 0); } | |||
2622 | static SDNodeIterator end (const SDNode *N) { | |||
2623 | return SDNodeIterator(N, N->getNumOperands()); | |||
2624 | } | |||
2625 | ||||
2626 | unsigned getOperand() const { return Operand; } | |||
2627 | const SDNode *getNode() const { return Node; } | |||
2628 | }; | |||
2629 | ||||
2630 | template <> struct GraphTraits<SDNode*> { | |||
2631 | using NodeRef = SDNode *; | |||
2632 | using ChildIteratorType = SDNodeIterator; | |||
2633 | ||||
2634 | static NodeRef getEntryNode(SDNode *N) { return N; } | |||
2635 | ||||
2636 | static ChildIteratorType child_begin(NodeRef N) { | |||
2637 | return SDNodeIterator::begin(N); | |||
2638 | } | |||
2639 | ||||
2640 | static ChildIteratorType child_end(NodeRef N) { | |||
2641 | return SDNodeIterator::end(N); | |||
2642 | } | |||
2643 | }; | |||
2644 | ||||
2645 | /// A representation of the largest SDNode, for use in sizeof(). | |||
2646 | /// | |||
2647 | /// This needs to be a union because the largest node differs on 32 bit systems | |||
2648 | /// with 4 and 8 byte pointer alignment, respectively. | |||
2649 | using LargestSDNode = AlignedCharArrayUnion<AtomicSDNode, TargetIndexSDNode, | |||
2650 | BlockAddressSDNode, | |||
2651 | GlobalAddressSDNode, | |||
2652 | PseudoProbeSDNode>; | |||
2653 | ||||
2654 | /// The SDNode class with the greatest alignment requirement. | |||
2655 | using MostAlignedSDNode = GlobalAddressSDNode; | |||
2656 | ||||
2657 | namespace ISD { | |||
2658 | ||||
2659 | /// Returns true if the specified node is a non-extending and unindexed load. | |||
2660 | inline bool isNormalLoad(const SDNode *N) { | |||
2661 | const LoadSDNode *Ld = dyn_cast<LoadSDNode>(N); | |||
2662 | return Ld && Ld->getExtensionType() == ISD::NON_EXTLOAD && | |||
2663 | Ld->getAddressingMode() == ISD::UNINDEXED; | |||
2664 | } | |||
2665 | ||||
2666 | /// Returns true if the specified node is a non-extending load. | |||
2667 | inline bool isNON_EXTLoad(const SDNode *N) { | |||
2668 | return isa<LoadSDNode>(N) && | |||
2669 | cast<LoadSDNode>(N)->getExtensionType() == ISD::NON_EXTLOAD; | |||
2670 | } | |||
2671 | ||||
2672 | /// Returns true if the specified node is a EXTLOAD. | |||
2673 | inline bool isEXTLoad(const SDNode *N) { | |||
2674 | return isa<LoadSDNode>(N) && | |||
2675 | cast<LoadSDNode>(N)->getExtensionType() == ISD::EXTLOAD; | |||
2676 | } | |||
2677 | ||||
2678 | /// Returns true if the specified node is a SEXTLOAD. | |||
2679 | inline bool isSEXTLoad(const SDNode *N) { | |||
2680 | return isa<LoadSDNode>(N) && | |||
2681 | cast<LoadSDNode>(N)->getExtensionType() == ISD::SEXTLOAD; | |||
2682 | } | |||
2683 | ||||
2684 | /// Returns true if the specified node is a ZEXTLOAD. | |||
2685 | inline bool isZEXTLoad(const SDNode *N) { | |||
2686 | return isa<LoadSDNode>(N) && | |||
2687 | cast<LoadSDNode>(N)->getExtensionType() == ISD::ZEXTLOAD; | |||
2688 | } | |||
2689 | ||||
2690 | /// Returns true if the specified node is an unindexed load. | |||
2691 | inline bool isUNINDEXEDLoad(const SDNode *N) { | |||
2692 | return isa<LoadSDNode>(N) && | |||
2693 | cast<LoadSDNode>(N)->getAddressingMode() == ISD::UNINDEXED; | |||
2694 | } | |||
2695 | ||||
2696 | /// Returns true if the specified node is a non-truncating | |||
2697 | /// and unindexed store. | |||
2698 | inline bool isNormalStore(const SDNode *N) { | |||
2699 | const StoreSDNode *St = dyn_cast<StoreSDNode>(N); | |||
2700 | return St && !St->isTruncatingStore() && | |||
2701 | St->getAddressingMode() == ISD::UNINDEXED; | |||
2702 | } | |||
2703 | ||||
2704 | /// Returns true if the specified node is a non-truncating store. | |||
2705 | inline bool isNON_TRUNCStore(const SDNode *N) { | |||
2706 | return isa<StoreSDNode>(N) && !cast<StoreSDNode>(N)->isTruncatingStore(); | |||
2707 | } | |||
2708 | ||||
2709 | /// Returns true if the specified node is a truncating store. | |||
2710 | inline bool isTRUNCStore(const SDNode *N) { | |||
2711 | return isa<StoreSDNode>(N) && cast<StoreSDNode>(N)->isTruncatingStore(); | |||
2712 | } | |||
2713 | ||||
2714 | /// Returns true if the specified node is an unindexed store. | |||
2715 | inline bool isUNINDEXEDStore(const SDNode *N) { | |||
2716 | return isa<StoreSDNode>(N) && | |||
2717 | cast<StoreSDNode>(N)->getAddressingMode() == ISD::UNINDEXED; | |||
2718 | } | |||
2719 | ||||
2720 | /// Attempt to match a unary predicate against a scalar/splat constant or | |||
2721 | /// every element of a constant BUILD_VECTOR. | |||
2722 | /// If AllowUndef is true, then UNDEF elements will pass nullptr to Match. | |||
2723 | bool matchUnaryPredicate(SDValue Op, | |||
2724 | std::function<bool(ConstantSDNode *)> Match, | |||
2725 | bool AllowUndefs = false); | |||
2726 | ||||
2727 | /// Attempt to match a binary predicate against a pair of scalar/splat | |||
2728 | /// constants or every element of a pair of constant BUILD_VECTORs. | |||
2729 | /// If AllowUndef is true, then UNDEF elements will pass nullptr to Match. | |||
2730 | /// If AllowTypeMismatch is true then RetType + ArgTypes don't need to match. | |||
2731 | bool matchBinaryPredicate( | |||
2732 | SDValue LHS, SDValue RHS, | |||
2733 | std::function<bool(ConstantSDNode *, ConstantSDNode *)> Match, | |||
2734 | bool AllowUndefs = false, bool AllowTypeMismatch = false); | |||
2735 | ||||
2736 | /// Returns true if the specified value is the overflow result from one | |||
2737 | /// of the overflow intrinsic nodes. | |||
2738 | inline bool isOverflowIntrOpRes(SDValue Op) { | |||
2739 | unsigned Opc = Op.getOpcode(); | |||
2740 | return (Op.getResNo() == 1 && | |||
2741 | (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO || | |||
2742 | Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)); | |||
2743 | } | |||
2744 | ||||
2745 | } // end namespace ISD | |||
2746 | ||||
2747 | } // end namespace llvm | |||
2748 | ||||
2749 | #endif // LLVM_CODEGEN_SELECTIONDAGNODES_H |