1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUCallLowering.h"
19 #include "AMDGPUFrameLowering.h"
20 #include "AMDGPUIntrinsicInfo.h"
21 #include "AMDGPURegisterInfo.h"
22 #include "AMDGPUSubtarget.h"
23 #include "R600MachineFunctionInfo.h"
24 #include "SIMachineFunctionInfo.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/Support/KnownBits.h"
33 #include "SIInstrInfo.h"
36 static bool allocateKernArg(unsigned ValNo, MVT ValVT, MVT LocVT,
37 CCValAssign::LocInfo LocInfo,
38 ISD::ArgFlagsTy ArgFlags, CCState &State) {
39 MachineFunction &MF = State.getMachineFunction();
40 AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
42 uint64_t Offset = MFI->allocateKernArg(LocVT.getStoreSize(),
43 ArgFlags.getOrigAlign());
44 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, Offset, LocVT, LocInfo));
48 static bool allocateCCRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
49 CCValAssign::LocInfo LocInfo,
50 ISD::ArgFlagsTy ArgFlags, CCState &State,
51 const TargetRegisterClass *RC,
53 ArrayRef<MCPhysReg> RegList = makeArrayRef(RC->begin(), NumRegs);
54 unsigned RegResult = State.AllocateReg(RegList);
55 if (RegResult == AMDGPU::NoRegister)
58 State.addLoc(CCValAssign::getReg(ValNo, ValVT, RegResult, LocVT, LocInfo));
62 static bool allocateSGPRTuple(unsigned ValNo, MVT ValVT, MVT LocVT,
63 CCValAssign::LocInfo LocInfo,
64 ISD::ArgFlagsTy ArgFlags, CCState &State) {
65 switch (LocVT.SimpleTy) {
71 return allocateCCRegs(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State,
72 &AMDGPU::SGPR_64RegClass, 20);
79 #include "AMDGPUGenCallingConv.inc"
81 // Find a larger type to do a load / store of a vector with.
82 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
83 unsigned StoreSize = VT.getStoreSizeInBits();
85 return EVT::getIntegerVT(Ctx, StoreSize);
87 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
88 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
91 AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
92 const AMDGPUSubtarget &STI)
93 : TargetLowering(TM), Subtarget(&STI) {
94 AMDGPUASI = AMDGPU::getAMDGPUAS(TM);
95 // Lower floating point store/load to integer store/load to reduce the number
96 // of patterns in tablegen.
97 setOperationAction(ISD::LOAD, MVT::f32, Promote);
98 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
100 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
101 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
103 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
104 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
106 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
107 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
109 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
110 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
112 setOperationAction(ISD::LOAD, MVT::i64, Promote);
113 AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
115 setOperationAction(ISD::LOAD, MVT::v2i64, Promote);
116 AddPromotedToType(ISD::LOAD, MVT::v2i64, MVT::v4i32);
118 setOperationAction(ISD::LOAD, MVT::f64, Promote);
119 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::v2i32);
121 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
122 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v4i32);
124 // There are no 64-bit extloads. These should be done as a 32-bit extload and
125 // an extension to 64-bit.
126 for (MVT VT : MVT::integer_valuetypes()) {
127 setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
128 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
129 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
132 for (MVT VT : MVT::integer_valuetypes()) {
136 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
137 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Legal);
138 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Legal);
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
141 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
142 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i8, Legal);
143 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i16, Legal);
144 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
146 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
147 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i8, Legal);
148 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i16, Legal);
149 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
152 for (MVT VT : MVT::integer_vector_valuetypes()) {
153 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
154 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
155 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
156 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
157 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
158 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
159 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
160 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
161 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
162 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
163 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
164 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
167 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
168 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
169 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
170 setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Expand);
172 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
173 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
174 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
175 setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f32, Expand);
177 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
178 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
179 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
180 setLoadExtAction(ISD::EXTLOAD, MVT::v8f64, MVT::v8f16, Expand);
182 setOperationAction(ISD::STORE, MVT::f32, Promote);
183 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
185 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
186 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
188 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
189 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
191 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
192 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
194 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
195 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
197 setOperationAction(ISD::STORE, MVT::i64, Promote);
198 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
200 setOperationAction(ISD::STORE, MVT::v2i64, Promote);
201 AddPromotedToType(ISD::STORE, MVT::v2i64, MVT::v4i32);
203 setOperationAction(ISD::STORE, MVT::f64, Promote);
204 AddPromotedToType(ISD::STORE, MVT::f64, MVT::v2i32);
206 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
207 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v4i32);
209 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
210 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
211 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
212 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
214 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
215 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Expand);
216 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Expand);
217 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Expand);
219 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
220 setTruncStoreAction(MVT::v2f32, MVT::v2f16, Expand);
221 setTruncStoreAction(MVT::v4f32, MVT::v4f16, Expand);
222 setTruncStoreAction(MVT::v8f32, MVT::v8f16, Expand);
224 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
225 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
227 setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
228 setTruncStoreAction(MVT::v2f64, MVT::v2f16, Expand);
230 setTruncStoreAction(MVT::v4f64, MVT::v4f32, Expand);
231 setTruncStoreAction(MVT::v4f64, MVT::v4f16, Expand);
233 setTruncStoreAction(MVT::v8f64, MVT::v8f32, Expand);
234 setTruncStoreAction(MVT::v8f64, MVT::v8f16, Expand);
237 setOperationAction(ISD::Constant, MVT::i32, Legal);
238 setOperationAction(ISD::Constant, MVT::i64, Legal);
239 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
240 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
242 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
243 setOperationAction(ISD::BRIND, MVT::Other, Expand);
245 // This is totally unsupported, just custom lower to produce an error.
246 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
248 // Library functions. These default to Expand, but we have instructions
250 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
251 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
252 setOperationAction(ISD::FPOW, MVT::f32, Legal);
253 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
254 setOperationAction(ISD::FABS, MVT::f32, Legal);
255 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
256 setOperationAction(ISD::FRINT, MVT::f32, Legal);
257 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
258 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
259 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
261 setOperationAction(ISD::FROUND, MVT::f32, Custom);
262 setOperationAction(ISD::FROUND, MVT::f64, Custom);
264 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
265 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
267 setOperationAction(ISD::FREM, MVT::f32, Custom);
268 setOperationAction(ISD::FREM, MVT::f64, Custom);
270 // v_mad_f32 does not support denormals according to some sources.
271 if (!Subtarget->hasFP32Denormals())
272 setOperationAction(ISD::FMAD, MVT::f32, Legal);
274 // Expand to fneg + fadd.
275 setOperationAction(ISD::FSUB, MVT::f64, Expand);
277 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
278 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
279 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
280 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
281 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
282 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
283 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
284 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
285 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
286 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
288 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
289 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
290 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
291 setOperationAction(ISD::FRINT, MVT::f64, Custom);
292 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
295 if (!Subtarget->hasBFI()) {
296 // fcopysign can be done in a single instruction with BFI.
297 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
298 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
301 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
302 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
303 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
305 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
306 for (MVT VT : ScalarIntVTs) {
307 // These should use [SU]DIVREM, so set them to expand
308 setOperationAction(ISD::SDIV, VT, Expand);
309 setOperationAction(ISD::UDIV, VT, Expand);
310 setOperationAction(ISD::SREM, VT, Expand);
311 setOperationAction(ISD::UREM, VT, Expand);
313 // GPU does not have divrem function for signed or unsigned.
314 setOperationAction(ISD::SDIVREM, VT, Custom);
315 setOperationAction(ISD::UDIVREM, VT, Custom);
317 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
318 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
319 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
321 setOperationAction(ISD::BSWAP, VT, Expand);
322 setOperationAction(ISD::CTTZ, VT, Expand);
323 setOperationAction(ISD::CTLZ, VT, Expand);
326 if (!Subtarget->hasBCNT(32))
327 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
329 if (!Subtarget->hasBCNT(64))
330 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
332 // The hardware supports 32-bit ROTR, but not ROTL.
333 setOperationAction(ISD::ROTL, MVT::i32, Expand);
334 setOperationAction(ISD::ROTL, MVT::i64, Expand);
335 setOperationAction(ISD::ROTR, MVT::i64, Expand);
337 setOperationAction(ISD::MUL, MVT::i64, Expand);
338 setOperationAction(ISD::MULHU, MVT::i64, Expand);
339 setOperationAction(ISD::MULHS, MVT::i64, Expand);
340 setOperationAction(ISD::UDIV, MVT::i32, Expand);
341 setOperationAction(ISD::UREM, MVT::i32, Expand);
342 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
343 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
344 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
345 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
346 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
348 setOperationAction(ISD::SMIN, MVT::i32, Legal);
349 setOperationAction(ISD::UMIN, MVT::i32, Legal);
350 setOperationAction(ISD::SMAX, MVT::i32, Legal);
351 setOperationAction(ISD::UMAX, MVT::i32, Legal);
353 if (Subtarget->hasFFBH())
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Custom);
356 if (Subtarget->hasFFBL())
357 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Legal);
359 setOperationAction(ISD::CTLZ, MVT::i64, Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
362 // We only really have 32-bit BFE instructions (and 16-bit on VI).
364 // On SI+ there are 64-bit BFEs, but they are scalar only and there isn't any
365 // effort to match them now. We want this to be false for i64 cases when the
366 // extraction isn't restricted to the upper or lower half. Ideally we would
367 // have some pass reduce 64-bit extracts to 32-bit if possible. Extracts that
368 // span the midpoint are probably relatively rare, so don't worry about them
370 if (Subtarget->hasBFE())
371 setHasExtractBitsInsn(true);
373 static const MVT::SimpleValueType VectorIntTypes[] = {
374 MVT::v2i32, MVT::v4i32
377 for (MVT VT : VectorIntTypes) {
378 // Expand the following operations for the current type by default.
379 setOperationAction(ISD::ADD, VT, Expand);
380 setOperationAction(ISD::AND, VT, Expand);
381 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
382 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
383 setOperationAction(ISD::MUL, VT, Expand);
384 setOperationAction(ISD::MULHU, VT, Expand);
385 setOperationAction(ISD::MULHS, VT, Expand);
386 setOperationAction(ISD::OR, VT, Expand);
387 setOperationAction(ISD::SHL, VT, Expand);
388 setOperationAction(ISD::SRA, VT, Expand);
389 setOperationAction(ISD::SRL, VT, Expand);
390 setOperationAction(ISD::ROTL, VT, Expand);
391 setOperationAction(ISD::ROTR, VT, Expand);
392 setOperationAction(ISD::SUB, VT, Expand);
393 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
394 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
395 setOperationAction(ISD::SDIV, VT, Expand);
396 setOperationAction(ISD::UDIV, VT, Expand);
397 setOperationAction(ISD::SREM, VT, Expand);
398 setOperationAction(ISD::UREM, VT, Expand);
399 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
400 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
401 setOperationAction(ISD::SDIVREM, VT, Custom);
402 setOperationAction(ISD::UDIVREM, VT, Expand);
403 setOperationAction(ISD::ADDC, VT, Expand);
404 setOperationAction(ISD::SUBC, VT, Expand);
405 setOperationAction(ISD::ADDE, VT, Expand);
406 setOperationAction(ISD::SUBE, VT, Expand);
407 setOperationAction(ISD::SELECT, VT, Expand);
408 setOperationAction(ISD::VSELECT, VT, Expand);
409 setOperationAction(ISD::SELECT_CC, VT, Expand);
410 setOperationAction(ISD::XOR, VT, Expand);
411 setOperationAction(ISD::BSWAP, VT, Expand);
412 setOperationAction(ISD::CTPOP, VT, Expand);
413 setOperationAction(ISD::CTTZ, VT, Expand);
414 setOperationAction(ISD::CTLZ, VT, Expand);
415 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
418 static const MVT::SimpleValueType FloatVectorTypes[] = {
419 MVT::v2f32, MVT::v4f32
422 for (MVT VT : FloatVectorTypes) {
423 setOperationAction(ISD::FABS, VT, Expand);
424 setOperationAction(ISD::FMINNUM, VT, Expand);
425 setOperationAction(ISD::FMAXNUM, VT, Expand);
426 setOperationAction(ISD::FADD, VT, Expand);
427 setOperationAction(ISD::FCEIL, VT, Expand);
428 setOperationAction(ISD::FCOS, VT, Expand);
429 setOperationAction(ISD::FDIV, VT, Expand);
430 setOperationAction(ISD::FEXP2, VT, Expand);
431 setOperationAction(ISD::FLOG2, VT, Expand);
432 setOperationAction(ISD::FREM, VT, Expand);
433 setOperationAction(ISD::FPOW, VT, Expand);
434 setOperationAction(ISD::FFLOOR, VT, Expand);
435 setOperationAction(ISD::FTRUNC, VT, Expand);
436 setOperationAction(ISD::FMUL, VT, Expand);
437 setOperationAction(ISD::FMA, VT, Expand);
438 setOperationAction(ISD::FRINT, VT, Expand);
439 setOperationAction(ISD::FNEARBYINT, VT, Expand);
440 setOperationAction(ISD::FSQRT, VT, Expand);
441 setOperationAction(ISD::FSIN, VT, Expand);
442 setOperationAction(ISD::FSUB, VT, Expand);
443 setOperationAction(ISD::FNEG, VT, Expand);
444 setOperationAction(ISD::VSELECT, VT, Expand);
445 setOperationAction(ISD::SELECT_CC, VT, Expand);
446 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
447 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
450 // This causes using an unrolled select operation rather than expansion with
451 // bit operations. This is in general better, but the alternative using BFI
452 // instructions may be better if the select sources are SGPRs.
453 setOperationAction(ISD::SELECT, MVT::v2f32, Promote);
454 AddPromotedToType(ISD::SELECT, MVT::v2f32, MVT::v2i32);
456 setOperationAction(ISD::SELECT, MVT::v4f32, Promote);
457 AddPromotedToType(ISD::SELECT, MVT::v4f32, MVT::v4i32);
459 // There are no libcalls of any kind.
460 for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I)
461 setLibcallName(static_cast<RTLIB::Libcall>(I), nullptr);
463 setBooleanContents(ZeroOrNegativeOneBooleanContent);
464 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
466 setSchedulingPreference(Sched::RegPressure);
467 setJumpIsExpensive(true);
469 // FIXME: This is only partially true. If we have to do vector compares, any
470 // SGPR pair can be a condition register. If we have a uniform condition, we
471 // are better off doing SALU operations, where there is only one SCC. For now,
472 // we don't have a way of knowing during instruction selection if a condition
473 // will be uniform and we always use vector compares. Assume we are using
474 // vector compares until that is fixed.
475 setHasMultipleConditionRegisters(true);
477 // SI at least has hardware support for floating point exceptions, but no way
478 // of using or handling them is implemented. They are also optional in OpenCL
480 setHasFloatingPointExceptions(Subtarget->hasFPExceptions());
482 PredictableSelectIsExpensive = false;
484 // We want to find all load dependencies for long chains of stores to enable
485 // merging into very wide vectors. The problem is with vectors with > 4
486 // elements. MergeConsecutiveStores will attempt to merge these because x8/x16
487 // vectors are a legal type, even though we have to split the loads
488 // usually. When we can more precisely specify load legality per address
489 // space, we should be able to make FindBetterChain/MergeConsecutiveStores
490 // smarter so that they can figure out what to do in 2 iterations without all
491 // N > 4 stores on the same chain.
492 GatherAllAliasesMaxDepth = 16;
494 // memcpy/memmove/memset are expanded in the IR, so we shouldn't need to worry
495 // about these during lowering.
496 MaxStoresPerMemcpy = 0xffffffff;
497 MaxStoresPerMemmove = 0xffffffff;
498 MaxStoresPerMemset = 0xffffffff;
500 setTargetDAGCombine(ISD::BITCAST);
501 setTargetDAGCombine(ISD::SHL);
502 setTargetDAGCombine(ISD::SRA);
503 setTargetDAGCombine(ISD::SRL);
504 setTargetDAGCombine(ISD::MUL);
505 setTargetDAGCombine(ISD::MULHU);
506 setTargetDAGCombine(ISD::MULHS);
507 setTargetDAGCombine(ISD::SELECT);
508 setTargetDAGCombine(ISD::SELECT_CC);
509 setTargetDAGCombine(ISD::STORE);
510 setTargetDAGCombine(ISD::FADD);
511 setTargetDAGCombine(ISD::FSUB);
512 setTargetDAGCombine(ISD::FNEG);
513 setTargetDAGCombine(ISD::FABS);
516 //===----------------------------------------------------------------------===//
517 // Target Information
518 //===----------------------------------------------------------------------===//
521 static bool fnegFoldsIntoOp(unsigned Opc) {
533 case ISD::FNEARBYINT:
535 case AMDGPUISD::RCP_LEGACY:
536 case AMDGPUISD::SIN_HW:
537 case AMDGPUISD::FMUL_LEGACY:
538 case AMDGPUISD::FMIN_LEGACY:
539 case AMDGPUISD::FMAX_LEGACY:
546 /// \p returns true if the operation will definitely need to use a 64-bit
547 /// encoding, and thus will use a VOP3 encoding regardless of the source
550 static bool opMustUseVOP3Encoding(const SDNode *N, MVT VT) {
551 return N->getNumOperands() > 2 || VT == MVT::f64;
554 // Most FP instructions support source modifiers, but this could be refined
557 static bool hasSourceMods(const SDNode *N) {
558 if (isa<MemSDNode>(N))
561 switch (N->getOpcode()) {
567 case AMDGPUISD::INTERP_P1:
568 case AMDGPUISD::INTERP_P2:
569 case AMDGPUISD::DIV_SCALE:
571 // TODO: Should really be looking at the users of the bitcast. These are
572 // problematic because bitcasts are used to legalize all stores to integer
581 bool AMDGPUTargetLowering::allUsesHaveSourceMods(const SDNode *N,
582 unsigned CostThreshold) {
583 // Some users (such as 3-operand FMA/MAD) must use a VOP3 encoding, and thus
584 // it is truly free to use a source modifier in all cases. If there are
585 // multiple users but for each one will necessitate using VOP3, there will be
586 // a code size increase. Try to avoid increasing code size unless we know it
587 // will save on the instruction count.
588 unsigned NumMayIncreaseSize = 0;
589 MVT VT = N->getValueType(0).getScalarType().getSimpleVT();
591 // XXX - Should this limit number of uses to check?
592 for (const SDNode *U : N->uses()) {
593 if (!hasSourceMods(U))
596 if (!opMustUseVOP3Encoding(U, VT)) {
597 if (++NumMayIncreaseSize > CostThreshold)
605 MVT AMDGPUTargetLowering::getVectorIdxTy(const DataLayout &) const {
609 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
613 // The backend supports 32 and 64 bit floating point immediates.
614 // FIXME: Why are we reporting vectors of FP immediates as legal?
615 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
616 EVT ScalarVT = VT.getScalarType();
617 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64 ||
618 (ScalarVT == MVT::f16 && Subtarget->has16BitInsts()));
621 // We don't want to shrink f64 / f32 constants.
622 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
623 EVT ScalarVT = VT.getScalarType();
624 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
627 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
631 unsigned NewSize = NewVT.getStoreSizeInBits();
633 // If we are reducing to a 32-bit load, this is always better.
637 EVT OldVT = N->getValueType(0);
638 unsigned OldSize = OldVT.getStoreSizeInBits();
640 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
641 // extloads, so doing one requires using a buffer_load. In cases where we
642 // still couldn't use a scalar load, using the wider load shouldn't really
645 // If the old size already had to be an extload, there's no harm in continuing
646 // to reduce the width.
647 return (OldSize < 32);
650 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
653 assert(LoadTy.getSizeInBits() == CastTy.getSizeInBits());
655 if (LoadTy.getScalarType() == MVT::i32)
658 unsigned LScalarSize = LoadTy.getScalarSizeInBits();
659 unsigned CastScalarSize = CastTy.getScalarSizeInBits();
661 return (LScalarSize < CastScalarSize) ||
662 (CastScalarSize >= 32);
665 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
666 // profitable with the expansion for 64-bit since it's generally good to
668 // FIXME: These should really have the size as a parameter.
669 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
673 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
677 //===---------------------------------------------------------------------===//
679 //===---------------------------------------------------------------------===//
681 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
682 assert(VT.isFloatingPoint());
684 // Packed operations do not have a fabs modifier.
685 return VT == MVT::f32 || VT == MVT::f64 ||
686 (Subtarget->has16BitInsts() && VT == MVT::f16);
689 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
690 assert(VT.isFloatingPoint());
691 return VT == MVT::f32 || VT == MVT::f64 ||
692 (Subtarget->has16BitInsts() && VT == MVT::f16) ||
693 (Subtarget->hasVOP3PInsts() && VT == MVT::v2f16);
696 bool AMDGPUTargetLowering:: storeOfVectorConstantIsCheap(EVT MemVT,
702 bool AMDGPUTargetLowering::aggressivelyPreferBuildVectorSources(EVT VecVT) const {
703 // There are few operations which truly have vector input operands. Any vector
704 // operation is going to involve operations on each component, and a
705 // build_vector will be a copy per element, so it always makes sense to use a
706 // build_vector input in place of the extracted element to avoid a copy into a
709 // We should probably only do this if all users are extracts only, but this
710 // should be the common case.
714 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
715 // Truncate is just accessing a subregister.
717 unsigned SrcSize = Source.getSizeInBits();
718 unsigned DestSize = Dest.getSizeInBits();
720 return DestSize < SrcSize && DestSize % 32 == 0 ;
723 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
724 // Truncate is just accessing a subregister.
726 unsigned SrcSize = Source->getScalarSizeInBits();
727 unsigned DestSize = Dest->getScalarSizeInBits();
729 if (DestSize== 16 && Subtarget->has16BitInsts())
730 return SrcSize >= 32;
732 return DestSize < SrcSize && DestSize % 32 == 0;
735 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
736 unsigned SrcSize = Src->getScalarSizeInBits();
737 unsigned DestSize = Dest->getScalarSizeInBits();
739 if (SrcSize == 16 && Subtarget->has16BitInsts())
740 return DestSize >= 32;
742 return SrcSize == 32 && DestSize == 64;
745 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
746 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
747 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
748 // this will enable reducing 64-bit operations the 32-bit, which is always
752 return Dest == MVT::i32 ||Dest == MVT::i64 ;
754 return Src == MVT::i32 && Dest == MVT::i64;
757 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
758 return isZExtFree(Val.getValueType(), VT2);
761 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
762 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
763 // limited number of native 64-bit operations. Shrinking an operation to fit
764 // in a single 32-bit register should always be helpful. As currently used,
765 // this is much less general than the name suggests, and is only used in
766 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
767 // not profitable, and may actually be harmful.
768 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
771 //===---------------------------------------------------------------------===//
772 // TargetLowering Callbacks
773 //===---------------------------------------------------------------------===//
775 CCAssignFn *AMDGPUCallLowering::CCAssignFnForCall(CallingConv::ID CC,
776 bool IsVarArg) const {
780 /// The SelectionDAGBuilder will automatically promote function arguments
781 /// with illegal types. However, this does not work for the AMDGPU targets
782 /// since the function arguments are stored in memory as these illegal types.
783 /// In order to handle this properly we need to get the original types sizes
784 /// from the LLVM IR Function and fixup the ISD:InputArg values before
785 /// passing them to AnalyzeFormalArguments()
787 /// When the SelectionDAGBuilder computes the Ins, it takes care of splitting
788 /// input values across multiple registers. Each item in the Ins array
789 /// represents a single value that will be stored in regsters. Ins[x].VT is
790 /// the value type of the value that will be stored in the register, so
791 /// whatever SDNode we lower the argument to needs to be this type.
793 /// In order to correctly lower the arguments we need to know the size of each
794 /// argument. Since Ins[x].VT gives us the size of the register that will
795 /// hold the value, we need to look at Ins[x].ArgVT to see the 'real' type
796 /// for the orignal function argument so that we can deduce the correct memory
797 /// type to use for Ins[x]. In most cases the correct memory type will be
798 /// Ins[x].ArgVT. However, this will not always be the case. If, for example,
799 /// we have a kernel argument of type v8i8, this argument will be split into
800 /// 8 parts and each part will be represented by its own item in the Ins array.
801 /// For each part the Ins[x].ArgVT will be the v8i8, which is the full type of
802 /// the argument before it was split. From this, we deduce that the memory type
803 /// for each individual part is i8. We pass the memory type as LocVT to the
804 /// calling convention analysis function and the register type (Ins[x].VT) as
806 void AMDGPUTargetLowering::analyzeFormalArgumentsCompute(CCState &State,
807 const SmallVectorImpl<ISD::InputArg> &Ins) const {
808 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
809 const ISD::InputArg &In = Ins[i];
812 unsigned NumRegs = getNumRegisters(State.getContext(), In.ArgVT);
814 if (!Subtarget->isAmdHsaOS() &&
815 (In.ArgVT == MVT::i16 || In.ArgVT == MVT::i8 || In.ArgVT == MVT::f16)) {
816 // The ABI says the caller will extend these values to 32-bits.
817 MemVT = In.ArgVT.isInteger() ? MVT::i32 : MVT::f32;
818 } else if (NumRegs == 1) {
819 // This argument is not split, so the IR type is the memory type.
820 assert(!In.Flags.isSplit());
821 if (In.ArgVT.isExtended()) {
822 // We have an extended type, like i24, so we should just use the register type
827 } else if (In.ArgVT.isVector() && In.VT.isVector() &&
828 In.ArgVT.getScalarType() == In.VT.getScalarType()) {
829 assert(In.ArgVT.getVectorNumElements() > In.VT.getVectorNumElements());
830 // We have a vector value which has been split into a vector with
831 // the same scalar type, but fewer elements. This should handle
832 // all the floating-point vector types.
834 } else if (In.ArgVT.isVector() &&
835 In.ArgVT.getVectorNumElements() == NumRegs) {
836 // This arg has been split so that each element is stored in a separate
838 MemVT = In.ArgVT.getScalarType();
839 } else if (In.ArgVT.isExtended()) {
840 // We have an extended type, like i65.
843 unsigned MemoryBits = In.ArgVT.getStoreSizeInBits() / NumRegs;
844 assert(In.ArgVT.getStoreSizeInBits() % NumRegs == 0);
845 if (In.VT.isInteger()) {
846 MemVT = EVT::getIntegerVT(State.getContext(), MemoryBits);
847 } else if (In.VT.isVector()) {
848 assert(!In.VT.getScalarType().isFloatingPoint());
849 unsigned NumElements = In.VT.getVectorNumElements();
850 assert(MemoryBits % NumElements == 0);
851 // This vector type has been split into another vector type with
852 // a different elements size.
853 EVT ScalarVT = EVT::getIntegerVT(State.getContext(),
854 MemoryBits / NumElements);
855 MemVT = EVT::getVectorVT(State.getContext(), ScalarVT, NumElements);
857 llvm_unreachable("cannot deduce memory type.");
861 // Convert one element vectors to scalar.
862 if (MemVT.isVector() && MemVT.getVectorNumElements() == 1)
863 MemVT = MemVT.getScalarType();
865 if (MemVT.isExtended()) {
866 // This should really only happen if we have vec3 arguments
867 assert(MemVT.isVector() && MemVT.getVectorNumElements() == 3);
868 MemVT = MemVT.getPow2VectorType(State.getContext());
871 assert(MemVT.isSimple());
872 allocateKernArg(i, In.VT, MemVT.getSimpleVT(), CCValAssign::Full, In.Flags,
877 void AMDGPUTargetLowering::AnalyzeReturn(CCState &State,
878 const SmallVectorImpl<ISD::OutputArg> &Outs) const {
880 State.AnalyzeReturn(Outs, RetCC_SI);
884 AMDGPUTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
886 const SmallVectorImpl<ISD::OutputArg> &Outs,
887 const SmallVectorImpl<SDValue> &OutVals,
888 const SDLoc &DL, SelectionDAG &DAG) const {
889 return DAG.getNode(AMDGPUISD::ENDPGM, DL, MVT::Other, Chain);
892 //===---------------------------------------------------------------------===//
893 // Target specific lowering
894 //===---------------------------------------------------------------------===//
896 /// Selects the correct CCAssignFn for a given CallingConvention value.
897 CCAssignFn *AMDGPUTargetLowering::CCAssignFnForCall(CallingConv::ID CC,
901 case CallingConv::AMDGPU_KERNEL:
902 case CallingConv::SPIR_KERNEL:
903 return CC_AMDGPU_Kernel;
904 case CallingConv::AMDGPU_VS:
905 case CallingConv::AMDGPU_HS:
906 case CallingConv::AMDGPU_GS:
907 case CallingConv::AMDGPU_PS:
908 case CallingConv::AMDGPU_CS:
911 report_fatal_error("Unsupported calling convention.");
915 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
916 SmallVectorImpl<SDValue> &InVals) const {
917 SDValue Callee = CLI.Callee;
918 SelectionDAG &DAG = CLI.DAG;
920 const Function &Fn = *DAG.getMachineFunction().getFunction();
922 StringRef FuncName("<unknown>");
924 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
925 FuncName = G->getSymbol();
926 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
927 FuncName = G->getGlobal()->getName();
929 DiagnosticInfoUnsupported NoCalls(
930 Fn, "unsupported call to function " + FuncName, CLI.DL.getDebugLoc());
931 DAG.getContext()->diagnose(NoCalls);
933 if (!CLI.IsTailCall) {
934 for (unsigned I = 0, E = CLI.Ins.size(); I != E; ++I)
935 InVals.push_back(DAG.getUNDEF(CLI.Ins[I].VT));
938 return DAG.getEntryNode();
941 SDValue AMDGPUTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
942 SelectionDAG &DAG) const {
943 const Function &Fn = *DAG.getMachineFunction().getFunction();
945 DiagnosticInfoUnsupported NoDynamicAlloca(Fn, "unsupported dynamic alloca",
946 SDLoc(Op).getDebugLoc());
947 DAG.getContext()->diagnose(NoDynamicAlloca);
948 auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
949 return DAG.getMergeValues(Ops, SDLoc());
952 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
953 SelectionDAG &DAG) const {
954 switch (Op.getOpcode()) {
956 Op->print(errs(), &DAG);
957 llvm_unreachable("Custom lowering code for this"
958 "instruction is not implemented yet!");
960 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
961 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
962 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
963 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
964 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
965 case ISD::FREM: return LowerFREM(Op, DAG);
966 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
967 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
968 case ISD::FRINT: return LowerFRINT(Op, DAG);
969 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
970 case ISD::FROUND: return LowerFROUND(Op, DAG);
971 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
972 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
973 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
974 case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
975 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
976 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
978 case ISD::CTLZ_ZERO_UNDEF:
979 return LowerCTLZ(Op, DAG);
980 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
985 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
986 SmallVectorImpl<SDValue> &Results,
987 SelectionDAG &DAG) const {
988 switch (N->getOpcode()) {
989 case ISD::SIGN_EXTEND_INREG:
990 // Different parts of legalization seem to interpret which type of
991 // sign_extend_inreg is the one to check for custom lowering. The extended
992 // from type is what really matters, but some places check for custom
993 // lowering of the result type. This results in trying to use
994 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
995 // nothing here and let the illegal result integer be handled normally.
1002 static bool hasDefinedInitializer(const GlobalValue *GV) {
1003 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
1004 if (!GVar || !GVar->hasInitializer())
1007 return !isa<UndefValue>(GVar->getInitializer());
1010 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
1012 SelectionDAG &DAG) const {
1014 const DataLayout &DL = DAG.getDataLayout();
1015 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
1016 const GlobalValue *GV = G->getGlobal();
1018 if (G->getAddressSpace() == AMDGPUASI.LOCAL_ADDRESS) {
1019 // XXX: What does the value of G->getOffset() mean?
1020 assert(G->getOffset() == 0 &&
1021 "Do not know what to do with an non-zero offset");
1023 // TODO: We could emit code to handle the initialization somewhere.
1024 if (!hasDefinedInitializer(GV)) {
1025 unsigned Offset = MFI->allocateLDSGlobal(DL, *GV);
1026 return DAG.getConstant(Offset, SDLoc(Op), Op.getValueType());
1030 const Function &Fn = *DAG.getMachineFunction().getFunction();
1031 DiagnosticInfoUnsupported BadInit(
1032 Fn, "unsupported initializer for address space", SDLoc(Op).getDebugLoc());
1033 DAG.getContext()->diagnose(BadInit);
1037 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
1038 SelectionDAG &DAG) const {
1039 SmallVector<SDValue, 8> Args;
1041 for (const SDUse &U : Op->ops())
1042 DAG.ExtractVectorElements(U.get(), Args);
1044 return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1047 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
1048 SelectionDAG &DAG) const {
1050 SmallVector<SDValue, 8> Args;
1051 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1052 EVT VT = Op.getValueType();
1053 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
1054 VT.getVectorNumElements());
1056 return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Args);
1059 /// \brief Generate Min/Max node
1060 SDValue AMDGPUTargetLowering::combineFMinMaxLegacy(const SDLoc &DL, EVT VT,
1061 SDValue LHS, SDValue RHS,
1062 SDValue True, SDValue False,
1064 DAGCombinerInfo &DCI) const {
1065 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1068 SelectionDAG &DAG = DCI.DAG;
1069 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1078 case ISD::SETFALSE2:
1087 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1088 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1094 // Ordered. Assume ordered for undefined.
1096 // Only do this after legalization to avoid interfering with other combines
1097 // which might occur.
1098 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1099 !DCI.isCalledByLegalizer())
1102 // We need to permute the operands to get the correct NaN behavior. The
1103 // selected operand is the second one based on the failing compare with NaN,
1104 // so permute it based on the compare type the hardware uses.
1106 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1107 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1112 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1113 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1119 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1120 !DCI.isCalledByLegalizer())
1124 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1125 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1127 case ISD::SETCC_INVALID:
1128 llvm_unreachable("Invalid setcc condcode!");
1133 std::pair<SDValue, SDValue>
1134 AMDGPUTargetLowering::split64BitValue(SDValue Op, SelectionDAG &DAG) const {
1137 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1139 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1140 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1142 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1143 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1145 return std::make_pair(Lo, Hi);
1148 SDValue AMDGPUTargetLowering::getLoHalf64(SDValue Op, SelectionDAG &DAG) const {
1151 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1152 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1153 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1156 SDValue AMDGPUTargetLowering::getHiHalf64(SDValue Op, SelectionDAG &DAG) const {
1159 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op);
1160 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1161 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1164 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1165 SelectionDAG &DAG) const {
1166 LoadSDNode *Load = cast<LoadSDNode>(Op);
1167 EVT VT = Op.getValueType();
1170 // If this is a 2 element vector, we really want to scalarize and not create
1171 // weird 1 element vectors.
1172 if (VT.getVectorNumElements() == 2)
1173 return scalarizeVectorLoad(Load, DAG);
1175 SDValue BasePtr = Load->getBasePtr();
1176 EVT PtrVT = BasePtr.getValueType();
1177 EVT MemVT = Load->getMemoryVT();
1180 const MachinePointerInfo &SrcValue = Load->getMemOperand()->getPointerInfo();
1183 EVT LoMemVT, HiMemVT;
1186 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1187 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1188 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1190 unsigned Size = LoMemVT.getStoreSize();
1191 unsigned BaseAlign = Load->getAlignment();
1192 unsigned HiAlign = MinAlign(BaseAlign, Size);
1194 SDValue LoLoad = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1195 Load->getChain(), BasePtr, SrcValue, LoMemVT,
1196 BaseAlign, Load->getMemOperand()->getFlags());
1197 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1198 DAG.getConstant(Size, SL, PtrVT));
1200 DAG.getExtLoad(Load->getExtensionType(), SL, HiVT, Load->getChain(),
1201 HiPtr, SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1202 HiMemVT, HiAlign, Load->getMemOperand()->getFlags());
1205 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1206 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1207 LoLoad.getValue(1), HiLoad.getValue(1))
1210 return DAG.getMergeValues(Ops, SL);
1213 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1214 SelectionDAG &DAG) const {
1215 StoreSDNode *Store = cast<StoreSDNode>(Op);
1216 SDValue Val = Store->getValue();
1217 EVT VT = Val.getValueType();
1219 // If this is a 2 element vector, we really want to scalarize and not create
1220 // weird 1 element vectors.
1221 if (VT.getVectorNumElements() == 2)
1222 return scalarizeVectorStore(Store, DAG);
1224 EVT MemVT = Store->getMemoryVT();
1225 SDValue Chain = Store->getChain();
1226 SDValue BasePtr = Store->getBasePtr();
1230 EVT LoMemVT, HiMemVT;
1233 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1234 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1235 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1237 EVT PtrVT = BasePtr.getValueType();
1238 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1239 DAG.getConstant(LoMemVT.getStoreSize(), SL,
1242 const MachinePointerInfo &SrcValue = Store->getMemOperand()->getPointerInfo();
1243 unsigned BaseAlign = Store->getAlignment();
1244 unsigned Size = LoMemVT.getStoreSize();
1245 unsigned HiAlign = MinAlign(BaseAlign, Size);
1248 DAG.getTruncStore(Chain, SL, Lo, BasePtr, SrcValue, LoMemVT, BaseAlign,
1249 Store->getMemOperand()->getFlags());
1251 DAG.getTruncStore(Chain, SL, Hi, HiPtr, SrcValue.getWithOffset(Size),
1252 HiMemVT, HiAlign, Store->getMemOperand()->getFlags());
1254 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1257 // This is a shortcut for integer division because we have fast i32<->f32
1258 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1259 // float is enough to accurately represent up to a 24-bit signed integer.
1260 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG,
1263 EVT VT = Op.getValueType();
1264 SDValue LHS = Op.getOperand(0);
1265 SDValue RHS = Op.getOperand(1);
1266 MVT IntVT = MVT::i32;
1267 MVT FltVT = MVT::f32;
1269 unsigned LHSSignBits = DAG.ComputeNumSignBits(LHS);
1270 if (LHSSignBits < 9)
1273 unsigned RHSSignBits = DAG.ComputeNumSignBits(RHS);
1274 if (RHSSignBits < 9)
1277 unsigned BitSize = VT.getSizeInBits();
1278 unsigned SignBits = std::min(LHSSignBits, RHSSignBits);
1279 unsigned DivBits = BitSize - SignBits;
1283 ISD::NodeType ToFp = Sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1284 ISD::NodeType ToInt = Sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1286 SDValue jq = DAG.getConstant(1, DL, IntVT);
1289 // char|short jq = ia ^ ib;
1290 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1292 // jq = jq >> (bitsize - 2)
1293 jq = DAG.getNode(ISD::SRA, DL, VT, jq,
1294 DAG.getConstant(BitSize - 2, DL, VT));
1297 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1300 // int ia = (int)LHS;
1303 // int ib, (int)RHS;
1306 // float fa = (float)ia;
1307 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1309 // float fb = (float)ib;
1310 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1312 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1313 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1316 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1318 // float fqneg = -fq;
1319 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1321 // float fr = mad(fqneg, fb, fa);
1322 unsigned OpCode = Subtarget->hasFP32Denormals() ?
1323 (unsigned)AMDGPUISD::FMAD_FTZ :
1324 (unsigned)ISD::FMAD;
1325 SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
1327 // int iq = (int)fq;
1328 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1331 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1334 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1336 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1338 // int cv = fr >= fb;
1339 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1341 // jq = (cv ? jq : 0);
1342 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1345 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1347 // Rem needs compensation, it's easier to recompute it
1348 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1349 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1351 // Truncate to number of bits this divide really is.
1354 = DAG.getValueType(EVT::getIntegerVT(*DAG.getContext(), DivBits));
1355 Div = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Div, InRegSize);
1356 Rem = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, Rem, InRegSize);
1358 SDValue TruncMask = DAG.getConstant((UINT64_C(1) << DivBits) - 1, DL, VT);
1359 Div = DAG.getNode(ISD::AND, DL, VT, Div, TruncMask);
1360 Rem = DAG.getNode(ISD::AND, DL, VT, Rem, TruncMask);
1363 return DAG.getMergeValues({ Div, Rem }, DL);
1366 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1368 SmallVectorImpl<SDValue> &Results) const {
1369 assert(Op.getValueType() == MVT::i64);
1372 EVT VT = Op.getValueType();
1373 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1375 SDValue one = DAG.getConstant(1, DL, HalfVT);
1376 SDValue zero = DAG.getConstant(0, DL, HalfVT);
1379 SDValue LHS = Op.getOperand(0);
1380 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1381 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1383 SDValue RHS = Op.getOperand(1);
1384 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1385 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1387 if (VT == MVT::i64 &&
1388 DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1389 DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1391 SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1394 SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(0), zero});
1395 SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), zero});
1397 Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
1398 Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
1402 // Get Speculative values
1403 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1404 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1406 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1407 SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {REM_Lo, zero});
1408 REM = DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM);
1410 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1411 SDValue DIV_Lo = zero;
1413 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1415 for (unsigned i = 0; i < halfBitWidth; ++i) {
1416 const unsigned bitPos = halfBitWidth - i - 1;
1417 SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1418 // Get value of high bit
1419 SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1420 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1421 HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1424 REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1426 REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1428 SDValue BIT = DAG.getConstant(1ULL << bitPos, DL, HalfVT);
1429 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1431 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1434 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1435 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1438 SDValue DIV = DAG.getBuildVector(MVT::v2i32, DL, {DIV_Lo, DIV_Hi});
1439 DIV = DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV);
1440 Results.push_back(DIV);
1441 Results.push_back(REM);
1444 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1445 SelectionDAG &DAG) const {
1447 EVT VT = Op.getValueType();
1449 if (VT == MVT::i64) {
1450 SmallVector<SDValue, 2> Results;
1451 LowerUDIVREM64(Op, DAG, Results);
1452 return DAG.getMergeValues(Results, DL);
1455 if (VT == MVT::i32) {
1456 if (SDValue Res = LowerDIVREM24(Op, DAG, false))
1460 SDValue Num = Op.getOperand(0);
1461 SDValue Den = Op.getOperand(1);
1463 // RCP = URECIP(Den) = 2^32 / Den + e
1464 // e is rounding error.
1465 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1467 // RCP_LO = mul(RCP, Den) */
1468 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1470 // RCP_HI = mulhu (RCP, Den) */
1471 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1473 // NEG_RCP_LO = -RCP_LO
1474 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1477 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1478 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1481 // Calculate the rounding error from the URECIP instruction
1482 // E = mulhu(ABS_RCP_LO, RCP)
1483 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1485 // RCP_A_E = RCP + E
1486 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1488 // RCP_S_E = RCP - E
1489 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1491 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1492 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1495 // Quotient = mulhu(Tmp0, Num)
1496 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1498 // Num_S_Remainder = Quotient * Den
1499 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1501 // Remainder = Num - Num_S_Remainder
1502 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1504 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1505 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1506 DAG.getConstant(-1, DL, VT),
1507 DAG.getConstant(0, DL, VT),
1509 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1510 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1512 DAG.getConstant(-1, DL, VT),
1513 DAG.getConstant(0, DL, VT),
1515 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1516 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1519 // Calculate Division result:
1521 // Quotient_A_One = Quotient + 1
1522 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1523 DAG.getConstant(1, DL, VT));
1525 // Quotient_S_One = Quotient - 1
1526 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1527 DAG.getConstant(1, DL, VT));
1529 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1530 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1531 Quotient, Quotient_A_One, ISD::SETEQ);
1533 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1534 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1535 Quotient_S_One, Div, ISD::SETEQ);
1537 // Calculate Rem result:
1539 // Remainder_S_Den = Remainder - Den
1540 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1542 // Remainder_A_Den = Remainder + Den
1543 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1545 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1546 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1547 Remainder, Remainder_S_Den, ISD::SETEQ);
1549 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1550 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1551 Remainder_A_Den, Rem, ISD::SETEQ);
1556 return DAG.getMergeValues(Ops, DL);
1559 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1560 SelectionDAG &DAG) const {
1562 EVT VT = Op.getValueType();
1564 SDValue LHS = Op.getOperand(0);
1565 SDValue RHS = Op.getOperand(1);
1567 SDValue Zero = DAG.getConstant(0, DL, VT);
1568 SDValue NegOne = DAG.getConstant(-1, DL, VT);
1570 if (VT == MVT::i32) {
1571 if (SDValue Res = LowerDIVREM24(Op, DAG, true))
1575 if (VT == MVT::i64 &&
1576 DAG.ComputeNumSignBits(LHS) > 32 &&
1577 DAG.ComputeNumSignBits(RHS) > 32) {
1578 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1581 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1582 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1583 SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1586 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1587 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1589 return DAG.getMergeValues(Res, DL);
1592 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1593 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1594 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1595 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1597 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1598 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1600 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1601 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1603 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1604 SDValue Rem = Div.getValue(1);
1606 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1607 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1609 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1610 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1616 return DAG.getMergeValues(Res, DL);
1619 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1620 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1622 EVT VT = Op.getValueType();
1623 SDValue X = Op.getOperand(0);
1624 SDValue Y = Op.getOperand(1);
1626 // TODO: Should this propagate fast-math-flags?
1628 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1629 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1630 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1632 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1635 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1637 SDValue Src = Op.getOperand(0);
1639 // result = trunc(src)
1640 // if (src > 0.0 && src != result)
1643 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1645 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
1646 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1649 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
1651 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1652 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1653 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1655 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1656 // TODO: Should this propagate fast-math-flags?
1657 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1660 static SDValue extractF64Exponent(SDValue Hi, const SDLoc &SL,
1661 SelectionDAG &DAG) {
1662 const unsigned FractBits = 52;
1663 const unsigned ExpBits = 11;
1665 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1667 DAG.getConstant(FractBits - 32, SL, MVT::i32),
1668 DAG.getConstant(ExpBits, SL, MVT::i32));
1669 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1670 DAG.getConstant(1023, SL, MVT::i32));
1675 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1677 SDValue Src = Op.getOperand(0);
1679 assert(Op.getValueType() == MVT::f64);
1681 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1682 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1684 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1686 // Extract the upper half, since this is where we will find the sign and
1688 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1690 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
1692 const unsigned FractBits = 52;
1694 // Extract the sign bit.
1695 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
1696 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1698 // Extend back to to 64-bits.
1699 SDValue SignBit64 = DAG.getBuildVector(MVT::v2i32, SL, {Zero, SignBit});
1700 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1702 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1703 const SDValue FractMask
1704 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
1706 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1707 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1708 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1711 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
1713 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
1715 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1716 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1718 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1719 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1721 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1724 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1726 SDValue Src = Op.getOperand(0);
1728 assert(Op.getValueType() == MVT::f64);
1730 APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
1731 SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
1732 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1734 // TODO: Should this propagate fast-math-flags?
1736 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1737 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1739 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1741 APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
1742 SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
1745 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
1746 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1748 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1751 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1752 // FNEARBYINT and FRINT are the same, except in their handling of FP
1753 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1754 // rint, so just treat them as equivalent.
1755 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1758 // XXX - May require not supporting f32 denormals?
1760 // Don't handle v2f16. The extra instructions to scalarize and repack around the
1761 // compare and vselect end up producing worse code than scalarizing the whole
1763 SDValue AMDGPUTargetLowering::LowerFROUND32_16(SDValue Op, SelectionDAG &DAG) const {
1765 SDValue X = Op.getOperand(0);
1766 EVT VT = Op.getValueType();
1768 SDValue T = DAG.getNode(ISD::FTRUNC, SL, VT, X);
1770 // TODO: Should this propagate fast-math-flags?
1772 SDValue Diff = DAG.getNode(ISD::FSUB, SL, VT, X, T);
1774 SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, VT, Diff);
1776 const SDValue Zero = DAG.getConstantFP(0.0, SL, VT);
1777 const SDValue One = DAG.getConstantFP(1.0, SL, VT);
1778 const SDValue Half = DAG.getConstantFP(0.5, SL, VT);
1780 SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, VT, One, X);
1783 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
1785 SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
1787 SDValue Sel = DAG.getNode(ISD::SELECT, SL, VT, Cmp, SignOne, Zero);
1789 return DAG.getNode(ISD::FADD, SL, VT, T, Sel);
1792 SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
1794 SDValue X = Op.getOperand(0);
1796 SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
1798 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1799 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1800 const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
1801 const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
1803 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i32);
1805 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
1807 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
1809 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
1811 const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
1814 SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
1815 SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
1816 DAG.getConstant(INT64_C(0x0008000000000000), SL,
1820 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
1821 SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
1822 DAG.getConstant(0, SL, MVT::i64), Tmp0,
1825 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
1826 D, DAG.getConstant(0, SL, MVT::i64));
1827 SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
1829 K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
1830 K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
1832 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1833 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1834 SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
1836 SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
1838 DAG.getConstantFP(1.0, SL, MVT::f64),
1839 DAG.getConstantFP(0.0, SL, MVT::f64));
1841 SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
1843 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
1844 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
1849 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
1850 EVT VT = Op.getValueType();
1852 if (VT == MVT::f32 || VT == MVT::f16)
1853 return LowerFROUND32_16(Op, DAG);
1856 return LowerFROUND64(Op, DAG);
1858 llvm_unreachable("unhandled type");
1861 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1863 SDValue Src = Op.getOperand(0);
1865 // result = trunc(src);
1866 // if (src < 0.0 && src != result)
1869 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1871 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
1872 const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
1875 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::f64);
1877 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1878 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1879 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1881 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1882 // TODO: Should this propagate fast-math-flags?
1883 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1886 SDValue AMDGPUTargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
1888 SDValue Src = Op.getOperand(0);
1889 bool ZeroUndef = Op.getOpcode() == ISD::CTLZ_ZERO_UNDEF;
1891 if (ZeroUndef && Src.getValueType() == MVT::i32)
1892 return DAG.getNode(AMDGPUISD::FFBH_U32, SL, MVT::i32, Src);
1894 SDValue Vec = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1896 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1897 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1899 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, Zero);
1900 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Vec, One);
1902 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
1903 *DAG.getContext(), MVT::i32);
1905 SDValue Hi0 = DAG.getSetCC(SL, SetCCVT, Hi, Zero, ISD::SETEQ);
1907 SDValue CtlzLo = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SL, MVT::i32, Lo);
1908 SDValue CtlzHi = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SL, MVT::i32, Hi);
1910 const SDValue Bits32 = DAG.getConstant(32, SL, MVT::i32);
1911 SDValue Add = DAG.getNode(ISD::ADD, SL, MVT::i32, CtlzLo, Bits32);
1913 // ctlz(x) = hi_32(x) == 0 ? ctlz(lo_32(x)) + 32 : ctlz(hi_32(x))
1914 SDValue NewCtlz = DAG.getNode(ISD::SELECT, SL, MVT::i32, Hi0, Add, CtlzHi);
1917 // Test if the full 64-bit input is zero.
1919 // FIXME: DAG combines turn what should be an s_and_b64 into a v_or_b32,
1920 // which we probably don't want.
1921 SDValue Lo0 = DAG.getSetCC(SL, SetCCVT, Lo, Zero, ISD::SETEQ);
1922 SDValue SrcIsZero = DAG.getNode(ISD::AND, SL, SetCCVT, Lo0, Hi0);
1924 // TODO: If i64 setcc is half rate, it can result in 1 fewer instruction
1925 // with the same cycles, otherwise it is slower.
1926 // SDValue SrcIsZero = DAG.getSetCC(SL, SetCCVT, Src,
1927 // DAG.getConstant(0, SL, MVT::i64), ISD::SETEQ);
1929 const SDValue Bits32 = DAG.getConstant(64, SL, MVT::i32);
1931 // The instruction returns -1 for 0 input, but the defined intrinsic
1932 // behavior is to return the number of bits.
1933 NewCtlz = DAG.getNode(ISD::SELECT, SL, MVT::i32,
1934 SrcIsZero, Bits32, NewCtlz);
1937 return DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i64, NewCtlz);
1940 SDValue AMDGPUTargetLowering::LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG,
1941 bool Signed) const {
1945 // uint lz = clz(u);
1946 // uint e = (u != 0) ? 127U + 63U - lz : 0;
1947 // u = (u << lz) & 0x7fffffffffffffffUL;
1948 // ulong t = u & 0xffffffffffUL;
1949 // uint v = (e << 23) | (uint)(u >> 40);
1950 // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U);
1951 // return as_float(v + r);
1956 // long s = l >> 63;
1957 // float r = cul2f((l + s) ^ s);
1958 // return s ? -r : r;
1962 SDValue Src = Op.getOperand(0);
1967 const SDValue SignBit = DAG.getConstant(63, SL, MVT::i64);
1968 S = DAG.getNode(ISD::SRA, SL, MVT::i64, L, SignBit);
1970 SDValue LPlusS = DAG.getNode(ISD::ADD, SL, MVT::i64, L, S);
1971 L = DAG.getNode(ISD::XOR, SL, MVT::i64, LPlusS, S);
1974 EVT SetCCVT = getSetCCResultType(DAG.getDataLayout(),
1975 *DAG.getContext(), MVT::f32);
1978 SDValue ZeroI32 = DAG.getConstant(0, SL, MVT::i32);
1979 SDValue ZeroI64 = DAG.getConstant(0, SL, MVT::i64);
1980 SDValue LZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SL, MVT::i64, L);
1981 LZ = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LZ);
1983 SDValue K = DAG.getConstant(127U + 63U, SL, MVT::i32);
1984 SDValue E = DAG.getSelect(SL, MVT::i32,
1985 DAG.getSetCC(SL, SetCCVT, L, ZeroI64, ISD::SETNE),
1986 DAG.getNode(ISD::SUB, SL, MVT::i32, K, LZ),
1989 SDValue U = DAG.getNode(ISD::AND, SL, MVT::i64,
1990 DAG.getNode(ISD::SHL, SL, MVT::i64, L, LZ),
1991 DAG.getConstant((-1ULL) >> 1, SL, MVT::i64));
1993 SDValue T = DAG.getNode(ISD::AND, SL, MVT::i64, U,
1994 DAG.getConstant(0xffffffffffULL, SL, MVT::i64));
1996 SDValue UShl = DAG.getNode(ISD::SRL, SL, MVT::i64,
1997 U, DAG.getConstant(40, SL, MVT::i64));
1999 SDValue V = DAG.getNode(ISD::OR, SL, MVT::i32,
2000 DAG.getNode(ISD::SHL, SL, MVT::i32, E, DAG.getConstant(23, SL, MVT::i32)),
2001 DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, UShl));
2003 SDValue C = DAG.getConstant(0x8000000000ULL, SL, MVT::i64);
2004 SDValue RCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETUGT);
2005 SDValue TCmp = DAG.getSetCC(SL, SetCCVT, T, C, ISD::SETEQ);
2007 SDValue One = DAG.getConstant(1, SL, MVT::i32);
2009 SDValue VTrunc1 = DAG.getNode(ISD::AND, SL, MVT::i32, V, One);
2011 SDValue R = DAG.getSelect(SL, MVT::i32,
2014 DAG.getSelect(SL, MVT::i32, TCmp, VTrunc1, ZeroI32));
2015 R = DAG.getNode(ISD::ADD, SL, MVT::i32, V, R);
2016 R = DAG.getNode(ISD::BITCAST, SL, MVT::f32, R);
2021 SDValue RNeg = DAG.getNode(ISD::FNEG, SL, MVT::f32, R);
2022 return DAG.getSelect(SL, MVT::f32, DAG.getSExtOrTrunc(S, SL, SetCCVT), RNeg, R);
2025 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2026 bool Signed) const {
2028 SDValue Src = Op.getOperand(0);
2030 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2032 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2033 DAG.getConstant(0, SL, MVT::i32));
2034 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2035 DAG.getConstant(1, SL, MVT::i32));
2037 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2040 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2042 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2043 DAG.getConstant(32, SL, MVT::i32));
2044 // TODO: Should this propagate fast-math-flags?
2045 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2048 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2049 SelectionDAG &DAG) const {
2050 assert(Op.getOperand(0).getValueType() == MVT::i64 &&
2051 "operation should be legal");
2053 // TODO: Factor out code common with LowerSINT_TO_FP.
2055 EVT DestVT = Op.getValueType();
2056 if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2058 SDValue Src = Op.getOperand(0);
2060 SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2061 SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2063 DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2068 if (DestVT == MVT::f32)
2069 return LowerINT_TO_FP32(Op, DAG, false);
2071 assert(DestVT == MVT::f64);
2072 return LowerINT_TO_FP64(Op, DAG, false);
2075 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2076 SelectionDAG &DAG) const {
2077 assert(Op.getOperand(0).getValueType() == MVT::i64 &&
2078 "operation should be legal");
2080 // TODO: Factor out code common with LowerUINT_TO_FP.
2082 EVT DestVT = Op.getValueType();
2083 if (Subtarget->has16BitInsts() && DestVT == MVT::f16) {
2085 SDValue Src = Op.getOperand(0);
2087 SDValue IntToFp32 = DAG.getNode(Op.getOpcode(), DL, MVT::f32, Src);
2088 SDValue FPRoundFlag = DAG.getIntPtrConstant(0, SDLoc(Op));
2090 DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, IntToFp32, FPRoundFlag);
2095 if (DestVT == MVT::f32)
2096 return LowerINT_TO_FP32(Op, DAG, true);
2098 assert(DestVT == MVT::f64);
2099 return LowerINT_TO_FP64(Op, DAG, true);
2102 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2103 bool Signed) const {
2106 SDValue Src = Op.getOperand(0);
2108 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2110 SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
2112 SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
2114 // TODO: Should this propagate fast-math-flags?
2115 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2117 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2120 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2122 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2123 MVT::i32, FloorMul);
2124 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2126 SDValue Result = DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi});
2128 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2131 SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
2133 SDValue N0 = Op.getOperand(0);
2135 // Convert to target node to get known bits
2136 if (N0.getValueType() == MVT::f32)
2137 return DAG.getNode(AMDGPUISD::FP_TO_FP16, DL, Op.getValueType(), N0);
2139 if (getTargetMachine().Options.UnsafeFPMath) {
2140 // There is a generic expand for FP_TO_FP16 with unsafe fast math.
2144 assert(N0.getSimpleValueType() == MVT::f64);
2146 // f64 -> f16 conversion using round-to-nearest-even rounding mode.
2147 const unsigned ExpMask = 0x7ff;
2148 const unsigned ExpBiasf64 = 1023;
2149 const unsigned ExpBiasf16 = 15;
2150 SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
2151 SDValue One = DAG.getConstant(1, DL, MVT::i32);
2152 SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
2153 SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
2154 DAG.getConstant(32, DL, MVT::i64));
2155 UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
2156 U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
2157 SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2158 DAG.getConstant(20, DL, MVT::i64));
2159 E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
2160 DAG.getConstant(ExpMask, DL, MVT::i32));
2161 // Subtract the fp64 exponent bias (1023) to get the real exponent and
2162 // add the f16 bias (15) to get the biased exponent for the f16 format.
2163 E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
2164 DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));
2166 SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2167 DAG.getConstant(8, DL, MVT::i32));
2168 M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
2169 DAG.getConstant(0xffe, DL, MVT::i32));
2171 SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
2172 DAG.getConstant(0x1ff, DL, MVT::i32));
2173 MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);
2175 SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
2176 M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);
2178 // (M != 0 ? 0x0200 : 0) | 0x7c00;
2179 SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
2180 DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
2181 Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));
2183 // N = M | (E << 12);
2184 SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2185 DAG.getNode(ISD::SHL, DL, MVT::i32, E,
2186 DAG.getConstant(12, DL, MVT::i32)));
2188 // B = clamp(1-E, 0, 13);
2189 SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
2191 SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
2192 B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
2193 DAG.getConstant(13, DL, MVT::i32));
2195 SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
2196 DAG.getConstant(0x1000, DL, MVT::i32));
2198 SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
2199 SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
2200 SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
2201 D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);
2203 SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
2204 SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
2205 DAG.getConstant(0x7, DL, MVT::i32));
2206 V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
2207 DAG.getConstant(2, DL, MVT::i32));
2208 SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
2209 One, Zero, ISD::SETEQ);
2210 SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
2211 One, Zero, ISD::SETGT);
2212 V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
2213 V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);
2215 V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
2216 DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
2217 V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
2220 // Extract the sign bit.
2221 SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
2222 DAG.getConstant(16, DL, MVT::i32));
2223 Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
2224 DAG.getConstant(0x8000, DL, MVT::i32));
2226 V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
2227 return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
2230 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2231 SelectionDAG &DAG) const {
2232 SDValue Src = Op.getOperand(0);
2234 // TODO: Factor out code common with LowerFP_TO_UINT.
2236 EVT SrcVT = Src.getValueType();
2237 if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
2240 SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
2242 DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
2247 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2248 return LowerFP64_TO_INT(Op, DAG, true);
2253 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2254 SelectionDAG &DAG) const {
2255 SDValue Src = Op.getOperand(0);
2257 // TODO: Factor out code common with LowerFP_TO_SINT.
2259 EVT SrcVT = Src.getValueType();
2260 if (Subtarget->has16BitInsts() && SrcVT == MVT::f16) {
2263 SDValue FPExtend = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Src);
2265 DAG.getNode(Op.getOpcode(), DL, MVT::i64, FPExtend);
2270 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2271 return LowerFP64_TO_INT(Op, DAG, false);
2276 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2277 SelectionDAG &DAG) const {
2278 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2279 MVT VT = Op.getSimpleValueType();
2280 MVT ScalarVT = VT.getScalarType();
2282 assert(VT.isVector());
2284 SDValue Src = Op.getOperand(0);
2287 // TODO: Don't scalarize on Evergreen?
2288 unsigned NElts = VT.getVectorNumElements();
2289 SmallVector<SDValue, 8> Args;
2290 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2292 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2293 for (unsigned I = 0; I < NElts; ++I)
2294 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2296 return DAG.getBuildVector(VT, DL, Args);
2299 //===----------------------------------------------------------------------===//
2300 // Custom DAG optimizations
2301 //===----------------------------------------------------------------------===//
2303 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2305 EVT VT = Op.getValueType();
2306 DAG.computeKnownBits(Op, Known);
2308 return (VT.getSizeInBits() - Known.countMinLeadingZeros()) <= 24;
2311 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2312 EVT VT = Op.getValueType();
2314 // In order for this to be a signed 24-bit value, bit 23, must
2316 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2317 // as unsigned 24-bit values.
2318 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2321 static bool simplifyI24(SDNode *Node24, unsigned OpIdx,
2322 TargetLowering::DAGCombinerInfo &DCI) {
2324 SelectionDAG &DAG = DCI.DAG;
2325 SDValue Op = Node24->getOperand(OpIdx);
2326 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2327 EVT VT = Op.getValueType();
2329 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2330 APInt KnownZero, KnownOne;
2331 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2332 if (TLI.SimplifyDemandedBits(Node24, OpIdx, Demanded, DCI, TLO))
2338 template <typename IntTy>
2339 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0, uint32_t Offset,
2340 uint32_t Width, const SDLoc &DL) {
2341 if (Width + Offset < 32) {
2342 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2343 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2344 return DAG.getConstant(Result, DL, MVT::i32);
2347 return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2350 static bool hasVolatileUser(SDNode *Val) {
2351 for (SDNode *U : Val->uses()) {
2352 if (MemSDNode *M = dyn_cast<MemSDNode>(U)) {
2353 if (M->isVolatile())
2361 bool AMDGPUTargetLowering::shouldCombineMemoryType(EVT VT) const {
2362 // i32 vectors are the canonical memory type.
2363 if (VT.getScalarType() == MVT::i32 || isTypeLegal(VT))
2366 if (!VT.isByteSized())
2369 unsigned Size = VT.getStoreSize();
2371 if ((Size == 1 || Size == 2 || Size == 4) && !VT.isVector())
2374 if (Size == 3 || (Size > 4 && (Size % 4 != 0)))
2380 // Replace load of an illegal type with a store of a bitcast to a friendlier
2382 SDValue AMDGPUTargetLowering::performLoadCombine(SDNode *N,
2383 DAGCombinerInfo &DCI) const {
2384 if (!DCI.isBeforeLegalize())
2387 LoadSDNode *LN = cast<LoadSDNode>(N);
2388 if (LN->isVolatile() || !ISD::isNormalLoad(LN) || hasVolatileUser(LN))
2392 SelectionDAG &DAG = DCI.DAG;
2393 EVT VT = LN->getMemoryVT();
2395 unsigned Size = VT.getStoreSize();
2396 unsigned Align = LN->getAlignment();
2397 if (Align < Size && isTypeLegal(VT)) {
2399 unsigned AS = LN->getAddressSpace();
2401 // Expand unaligned loads earlier than legalization. Due to visitation order
2402 // problems during legalization, the emitted instructions to pack and unpack
2403 // the bytes again are not eliminated in the case of an unaligned copy.
2404 if (!allowsMisalignedMemoryAccesses(VT, AS, Align, &IsFast)) {
2406 return scalarizeVectorLoad(LN, DAG);
2409 std::tie(Ops[0], Ops[1]) = expandUnalignedLoad(LN, DAG);
2410 return DAG.getMergeValues(Ops, SDLoc(N));
2417 if (!shouldCombineMemoryType(VT))
2420 EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2423 = DAG.getLoad(NewVT, SL, LN->getChain(),
2424 LN->getBasePtr(), LN->getMemOperand());
2426 SDValue BC = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad);
2427 DCI.CombineTo(N, BC, NewLoad.getValue(1));
2428 return SDValue(N, 0);
2431 // Replace store of an illegal type with a store of a bitcast to a friendlier
2433 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2434 DAGCombinerInfo &DCI) const {
2435 if (!DCI.isBeforeLegalize())
2438 StoreSDNode *SN = cast<StoreSDNode>(N);
2439 if (SN->isVolatile() || !ISD::isNormalStore(SN))
2442 EVT VT = SN->getMemoryVT();
2443 unsigned Size = VT.getStoreSize();
2446 SelectionDAG &DAG = DCI.DAG;
2447 unsigned Align = SN->getAlignment();
2448 if (Align < Size && isTypeLegal(VT)) {
2450 unsigned AS = SN->getAddressSpace();
2452 // Expand unaligned stores earlier than legalization. Due to visitation
2453 // order problems during legalization, the emitted instructions to pack and
2454 // unpack the bytes again are not eliminated in the case of an unaligned
2456 if (!allowsMisalignedMemoryAccesses(VT, AS, Align, &IsFast)) {
2458 return scalarizeVectorStore(SN, DAG);
2460 return expandUnalignedStore(SN, DAG);
2467 if (!shouldCombineMemoryType(VT))
2470 EVT NewVT = getEquivalentMemType(*DAG.getContext(), VT);
2471 SDValue Val = SN->getValue();
2473 //DCI.AddToWorklist(Val.getNode());
2475 bool OtherUses = !Val.hasOneUse();
2476 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, NewVT, Val);
2478 SDValue CastBack = DAG.getNode(ISD::BITCAST, SL, VT, CastVal);
2479 DAG.ReplaceAllUsesOfValueWith(Val, CastBack);
2482 return DAG.getStore(SN->getChain(), SL, CastVal,
2483 SN->getBasePtr(), SN->getMemOperand());
2486 SDValue AMDGPUTargetLowering::performClampCombine(SDNode *N,
2487 DAGCombinerInfo &DCI) const {
2488 ConstantFPSDNode *CSrc = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
2492 const APFloat &F = CSrc->getValueAPF();
2493 APFloat Zero = APFloat::getZero(F.getSemantics());
2494 APFloat::cmpResult Cmp0 = F.compare(Zero);
2495 if (Cmp0 == APFloat::cmpLessThan ||
2496 (Cmp0 == APFloat::cmpUnordered && Subtarget->enableDX10Clamp())) {
2497 return DCI.DAG.getConstantFP(Zero, SDLoc(N), N->getValueType(0));
2500 APFloat One(F.getSemantics(), "1.0");
2501 APFloat::cmpResult Cmp1 = F.compare(One);
2502 if (Cmp1 == APFloat::cmpGreaterThan)
2503 return DCI.DAG.getConstantFP(One, SDLoc(N), N->getValueType(0));
2505 return SDValue(CSrc, 0);
2508 /// Split the 64-bit value \p LHS into two 32-bit components, and perform the
2509 /// binary operation \p Opc to it with the corresponding constant operands.
2510 SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
2511 DAGCombinerInfo &DCI, const SDLoc &SL,
2512 unsigned Opc, SDValue LHS,
2513 uint32_t ValLo, uint32_t ValHi) const {
2514 SelectionDAG &DAG = DCI.DAG;
2516 std::tie(Lo, Hi) = split64BitValue(LHS, DAG);
2518 SDValue LoRHS = DAG.getConstant(ValLo, SL, MVT::i32);
2519 SDValue HiRHS = DAG.getConstant(ValHi, SL, MVT::i32);
2521 SDValue LoAnd = DAG.getNode(Opc, SL, MVT::i32, Lo, LoRHS);
2522 SDValue HiAnd = DAG.getNode(Opc, SL, MVT::i32, Hi, HiRHS);
2524 // Re-visit the ands. It's possible we eliminated one of them and it could
2525 // simplify the vector.
2526 DCI.AddToWorklist(Lo.getNode());
2527 DCI.AddToWorklist(Hi.getNode());
2529 SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {LoAnd, HiAnd});
2530 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
2533 SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
2534 DAGCombinerInfo &DCI) const {
2535 if (N->getValueType(0) != MVT::i64)
2538 // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
2540 // On some subtargets, 64-bit shift is a quarter rate instruction. In the
2541 // common case, splitting this into a move and a 32-bit shift is faster and
2542 // the same code size.
2543 const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
2547 unsigned RHSVal = RHS->getZExtValue();
2551 SDValue LHS = N->getOperand(0);
2554 SelectionDAG &DAG = DCI.DAG;
2556 SDValue ShiftAmt = DAG.getConstant(RHSVal - 32, SL, MVT::i32);
2558 SDValue Lo = DAG.getNode(ISD::TRUNCATE, SL, MVT::i32, LHS);
2559 SDValue NewShift = DAG.getNode(ISD::SHL, SL, MVT::i32, Lo, ShiftAmt);
2561 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2563 SDValue Vec = DAG.getBuildVector(MVT::v2i32, SL, {Zero, NewShift});
2564 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Vec);
2567 SDValue AMDGPUTargetLowering::performSraCombine(SDNode *N,
2568 DAGCombinerInfo &DCI) const {
2569 if (N->getValueType(0) != MVT::i64)
2572 const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
2576 SelectionDAG &DAG = DCI.DAG;
2578 unsigned RHSVal = RHS->getZExtValue();
2580 // (sra i64:x, 32) -> build_pair x, (sra hi_32(x), 31)
2582 SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
2583 SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
2584 DAG.getConstant(31, SL, MVT::i32));
2586 SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {Hi, NewShift});
2587 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
2590 // (sra i64:x, 63) -> build_pair (sra hi_32(x), 31), (sra hi_32(x), 31)
2592 SDValue Hi = getHiHalf64(N->getOperand(0), DAG);
2593 SDValue NewShift = DAG.getNode(ISD::SRA, SL, MVT::i32, Hi,
2594 DAG.getConstant(31, SL, MVT::i32));
2595 SDValue BuildVec = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, NewShift});
2596 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildVec);
2602 SDValue AMDGPUTargetLowering::performSrlCombine(SDNode *N,
2603 DAGCombinerInfo &DCI) const {
2604 if (N->getValueType(0) != MVT::i64)
2607 const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N->getOperand(1));
2611 unsigned ShiftAmt = RHS->getZExtValue();
2615 // srl i64:x, C for C >= 32
2617 // build_pair (srl hi_32(x), C - 32), 0
2619 SelectionDAG &DAG = DCI.DAG;
2622 SDValue One = DAG.getConstant(1, SL, MVT::i32);
2623 SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2625 SDValue VecOp = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, N->getOperand(0));
2626 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32,
2629 SDValue NewConst = DAG.getConstant(ShiftAmt - 32, SL, MVT::i32);
2630 SDValue NewShift = DAG.getNode(ISD::SRL, SL, MVT::i32, Hi, NewConst);
2632 SDValue BuildPair = DAG.getBuildVector(MVT::v2i32, SL, {NewShift, Zero});
2634 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, BuildPair);
2637 // We need to specifically handle i64 mul here to avoid unnecessary conversion
2638 // instructions. If we only match on the legalized i64 mul expansion,
2639 // SimplifyDemandedBits will be unable to remove them because there will be
2640 // multiple uses due to the separate mul + mulh[su].
2641 static SDValue getMul24(SelectionDAG &DAG, const SDLoc &SL,
2642 SDValue N0, SDValue N1, unsigned Size, bool Signed) {
2644 unsigned MulOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
2645 return DAG.getNode(MulOpc, SL, MVT::i32, N0, N1);
2648 // Because we want to eliminate extension instructions before the
2649 // operation, we need to create a single user here (i.e. not the separate
2650 // mul_lo + mul_hi) so that SimplifyDemandedBits will deal with it.
2652 unsigned MulOpc = Signed ? AMDGPUISD::MUL_LOHI_I24 : AMDGPUISD::MUL_LOHI_U24;
2654 SDValue Mul = DAG.getNode(MulOpc, SL,
2655 DAG.getVTList(MVT::i32, MVT::i32), N0, N1);
2657 return DAG.getNode(ISD::BUILD_PAIR, SL, MVT::i64,
2658 Mul.getValue(0), Mul.getValue(1));
2661 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2662 DAGCombinerInfo &DCI) const {
2663 EVT VT = N->getValueType(0);
2665 unsigned Size = VT.getSizeInBits();
2666 if (VT.isVector() || Size > 64)
2669 // There are i16 integer mul/mad.
2670 if (Subtarget->has16BitInsts() && VT.getScalarType().bitsLE(MVT::i16))
2673 SelectionDAG &DAG = DCI.DAG;
2676 SDValue N0 = N->getOperand(0);
2677 SDValue N1 = N->getOperand(1);
2680 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2681 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2682 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2683 Mul = getMul24(DAG, DL, N0, N1, Size, false);
2684 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2685 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2686 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2687 Mul = getMul24(DAG, DL, N0, N1, Size, true);
2692 // We need to use sext even for MUL_U24, because MUL_U24 is used
2693 // for signed multiply of 8 and 16-bit types.
2694 return DAG.getSExtOrTrunc(Mul, DL, VT);
2697 SDValue AMDGPUTargetLowering::performMulhsCombine(SDNode *N,
2698 DAGCombinerInfo &DCI) const {
2699 EVT VT = N->getValueType(0);
2701 if (!Subtarget->hasMulI24() || VT.isVector())
2704 SelectionDAG &DAG = DCI.DAG;
2707 SDValue N0 = N->getOperand(0);
2708 SDValue N1 = N->getOperand(1);
2710 if (!isI24(N0, DAG) || !isI24(N1, DAG))
2713 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2714 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2716 SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_I24, DL, MVT::i32, N0, N1);
2717 DCI.AddToWorklist(Mulhi.getNode());
2718 return DAG.getSExtOrTrunc(Mulhi, DL, VT);
2721 SDValue AMDGPUTargetLowering::performMulhuCombine(SDNode *N,
2722 DAGCombinerInfo &DCI) const {
2723 EVT VT = N->getValueType(0);
2725 if (!Subtarget->hasMulU24() || VT.isVector() || VT.getSizeInBits() > 32)
2728 SelectionDAG &DAG = DCI.DAG;
2731 SDValue N0 = N->getOperand(0);
2732 SDValue N1 = N->getOperand(1);
2734 if (!isU24(N0, DAG) || !isU24(N1, DAG))
2737 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2738 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2740 SDValue Mulhi = DAG.getNode(AMDGPUISD::MULHI_U24, DL, MVT::i32, N0, N1);
2741 DCI.AddToWorklist(Mulhi.getNode());
2742 return DAG.getZExtOrTrunc(Mulhi, DL, VT);
2745 SDValue AMDGPUTargetLowering::performMulLoHi24Combine(
2746 SDNode *N, DAGCombinerInfo &DCI) const {
2747 SelectionDAG &DAG = DCI.DAG;
2749 // Simplify demanded bits before splitting into multiple users.
2750 if (simplifyI24(N, 0, DCI) || simplifyI24(N, 1, DCI))
2753 SDValue N0 = N->getOperand(0);
2754 SDValue N1 = N->getOperand(1);
2756 bool Signed = (N->getOpcode() == AMDGPUISD::MUL_LOHI_I24);
2758 unsigned MulLoOpc = Signed ? AMDGPUISD::MUL_I24 : AMDGPUISD::MUL_U24;
2759 unsigned MulHiOpc = Signed ? AMDGPUISD::MULHI_I24 : AMDGPUISD::MULHI_U24;
2763 SDValue MulLo = DAG.getNode(MulLoOpc, SL, MVT::i32, N0, N1);
2764 SDValue MulHi = DAG.getNode(MulHiOpc, SL, MVT::i32, N0, N1);
2765 return DAG.getMergeValues({ MulLo, MulHi }, SL);
2768 static bool isNegativeOne(SDValue Val) {
2769 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val))
2770 return C->isAllOnesValue();
2774 static bool isCtlzOpc(unsigned Opc) {
2775 return Opc == ISD::CTLZ || Opc == ISD::CTLZ_ZERO_UNDEF;
2778 SDValue AMDGPUTargetLowering::getFFBH_U32(SelectionDAG &DAG,
2780 const SDLoc &DL) const {
2781 EVT VT = Op.getValueType();
2782 EVT LegalVT = getTypeToTransformTo(*DAG.getContext(), VT);
2783 if (LegalVT != MVT::i32 && (Subtarget->has16BitInsts() &&
2784 LegalVT != MVT::i16))
2788 Op = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Op);
2790 SDValue FFBH = DAG.getNode(AMDGPUISD::FFBH_U32, DL, MVT::i32, Op);
2792 FFBH = DAG.getNode(ISD::TRUNCATE, DL, VT, FFBH);
2797 // The native instructions return -1 on 0 input. Optimize out a select that
2798 // produces -1 on 0.
2800 // TODO: If zero is not undef, we could also do this if the output is compared
2801 // against the bitwidth.
2803 // TODO: Should probably combine against FFBH_U32 instead of ctlz directly.
2804 SDValue AMDGPUTargetLowering::performCtlzCombine(const SDLoc &SL, SDValue Cond,
2805 SDValue LHS, SDValue RHS,
2806 DAGCombinerInfo &DCI) const {
2807 ConstantSDNode *CmpRhs = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2808 if (!CmpRhs || !CmpRhs->isNullValue())
2811 SelectionDAG &DAG = DCI.DAG;
2812 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
2813 SDValue CmpLHS = Cond.getOperand(0);
2815 // select (setcc x, 0, eq), -1, (ctlz_zero_undef x) -> ffbh_u32 x
2816 if (CCOpcode == ISD::SETEQ &&
2817 isCtlzOpc(RHS.getOpcode()) &&
2818 RHS.getOperand(0) == CmpLHS &&
2819 isNegativeOne(LHS)) {
2820 return getFFBH_U32(DAG, CmpLHS, SL);
2823 // select (setcc x, 0, ne), (ctlz_zero_undef x), -1 -> ffbh_u32 x
2824 if (CCOpcode == ISD::SETNE &&
2825 isCtlzOpc(LHS.getOpcode()) &&
2826 LHS.getOperand(0) == CmpLHS &&
2827 isNegativeOne(RHS)) {
2828 return getFFBH_U32(DAG, CmpLHS, SL);
2834 static SDValue distributeOpThroughSelect(TargetLowering::DAGCombinerInfo &DCI,
2840 SelectionDAG &DAG = DCI.DAG;
2841 EVT VT = N1.getValueType();
2843 SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT, Cond,
2844 N1.getOperand(0), N2.getOperand(0));
2845 DCI.AddToWorklist(NewSelect.getNode());
2846 return DAG.getNode(Op, SL, VT, NewSelect);
2849 // Pull a free FP operation out of a select so it may fold into uses.
2851 // select c, (fneg x), (fneg y) -> fneg (select c, x, y)
2852 // select c, (fneg x), k -> fneg (select c, x, (fneg k))
2854 // select c, (fabs x), (fabs y) -> fabs (select c, x, y)
2855 // select c, (fabs x), +k -> fabs (select c, x, k)
2856 static SDValue foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
2858 SelectionDAG &DAG = DCI.DAG;
2859 SDValue Cond = N.getOperand(0);
2860 SDValue LHS = N.getOperand(1);
2861 SDValue RHS = N.getOperand(2);
2863 EVT VT = N.getValueType();
2864 if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
2865 (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
2866 return distributeOpThroughSelect(DCI, LHS.getOpcode(),
2867 SDLoc(N), Cond, LHS, RHS);
2871 if (RHS.getOpcode() == ISD::FABS || RHS.getOpcode() == ISD::FNEG) {
2872 std::swap(LHS, RHS);
2876 // TODO: Support vector constants.
2877 ConstantFPSDNode *CRHS = dyn_cast<ConstantFPSDNode>(RHS);
2878 if ((LHS.getOpcode() == ISD::FNEG || LHS.getOpcode() == ISD::FABS) && CRHS) {
2880 // If one side is an fneg/fabs and the other is a constant, we can push the
2881 // fneg/fabs down. If it's an fabs, the constant needs to be non-negative.
2882 SDValue NewLHS = LHS.getOperand(0);
2883 SDValue NewRHS = RHS;
2885 // Careful: if the neg can be folded up, don't try to pull it back down.
2886 bool ShouldFoldNeg = true;
2888 if (NewLHS.hasOneUse()) {
2889 unsigned Opc = NewLHS.getOpcode();
2890 if (LHS.getOpcode() == ISD::FNEG && fnegFoldsIntoOp(Opc))
2891 ShouldFoldNeg = false;
2892 if (LHS.getOpcode() == ISD::FABS && Opc == ISD::FMUL)
2893 ShouldFoldNeg = false;
2896 if (ShouldFoldNeg) {
2897 if (LHS.getOpcode() == ISD::FNEG)
2898 NewRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
2899 else if (CRHS->isNegative())
2903 std::swap(NewLHS, NewRHS);
2905 SDValue NewSelect = DAG.getNode(ISD::SELECT, SL, VT,
2906 Cond, NewLHS, NewRHS);
2907 DCI.AddToWorklist(NewSelect.getNode());
2908 return DAG.getNode(LHS.getOpcode(), SL, VT, NewSelect);
2916 SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
2917 DAGCombinerInfo &DCI) const {
2918 if (SDValue Folded = foldFreeOpFromSelect(DCI, SDValue(N, 0)))
2921 SDValue Cond = N->getOperand(0);
2922 if (Cond.getOpcode() != ISD::SETCC)
2925 EVT VT = N->getValueType(0);
2926 SDValue LHS = Cond.getOperand(0);
2927 SDValue RHS = Cond.getOperand(1);
2928 SDValue CC = Cond.getOperand(2);
2930 SDValue True = N->getOperand(1);
2931 SDValue False = N->getOperand(2);
2933 if (Cond.hasOneUse()) { // TODO: Look for multiple select uses.
2934 SelectionDAG &DAG = DCI.DAG;
2935 if ((DAG.isConstantValueOfAnyType(True) ||
2936 DAG.isConstantValueOfAnyType(True)) &&
2937 (!DAG.isConstantValueOfAnyType(False) &&
2938 !DAG.isConstantValueOfAnyType(False))) {
2939 // Swap cmp + select pair to move constant to false input.
2940 // This will allow using VOPC cndmasks more often.
2941 // select (setcc x, y), k, x -> select (setcc y, x) x, x
2944 ISD::CondCode NewCC = getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
2945 LHS.getValueType().isInteger());
2947 SDValue NewCond = DAG.getSetCC(SL, Cond.getValueType(), LHS, RHS, NewCC);
2948 return DAG.getNode(ISD::SELECT, SL, VT, NewCond, False, True);
2951 if (VT == MVT::f32 && Subtarget->hasFminFmaxLegacy()) {
2953 = combineFMinMaxLegacy(SDLoc(N), VT, LHS, RHS, True, False, CC, DCI);
2954 // Revisit this node so we can catch min3/max3/med3 patterns.
2955 //DCI.AddToWorklist(MinMax.getNode());
2960 // There's no reason to not do this if the condition has other uses.
2961 return performCtlzCombine(SDLoc(N), Cond, True, False, DCI);
2964 static bool isConstantFPZero(SDValue N) {
2965 if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N))
2966 return C->isZero() && !C->isNegative();
2970 static unsigned inverseMinMax(unsigned Opc) {
2973 return ISD::FMINNUM;
2975 return ISD::FMAXNUM;
2976 case AMDGPUISD::FMAX_LEGACY:
2977 return AMDGPUISD::FMIN_LEGACY;
2978 case AMDGPUISD::FMIN_LEGACY:
2979 return AMDGPUISD::FMAX_LEGACY;
2981 llvm_unreachable("invalid min/max opcode");
2985 SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
2986 DAGCombinerInfo &DCI) const {
2987 SelectionDAG &DAG = DCI.DAG;
2988 SDValue N0 = N->getOperand(0);
2989 EVT VT = N->getValueType(0);
2991 unsigned Opc = N0.getOpcode();
2993 // If the input has multiple uses and we can either fold the negate down, or
2994 // the other uses cannot, give up. This both prevents unprofitable
2995 // transformations and infinite loops: we won't repeatedly try to fold around
2996 // a negate that has no 'good' form.
2997 if (N0.hasOneUse()) {
2998 // This may be able to fold into the source, but at a code size cost. Don't
2999 // fold if the fold into the user is free.
3000 if (allUsesHaveSourceMods(N, 0))
3003 if (fnegFoldsIntoOp(Opc) &&
3004 (allUsesHaveSourceMods(N) || !allUsesHaveSourceMods(N0.getNode())))
3011 if (!mayIgnoreSignedZero(N0))
3014 // (fneg (fadd x, y)) -> (fadd (fneg x), (fneg y))
3015 SDValue LHS = N0.getOperand(0);
3016 SDValue RHS = N0.getOperand(1);
3018 if (LHS.getOpcode() != ISD::FNEG)
3019 LHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3021 LHS = LHS.getOperand(0);
3023 if (RHS.getOpcode() != ISD::FNEG)
3024 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3026 RHS = RHS.getOperand(0);
3028 SDValue Res = DAG.getNode(ISD::FADD, SL, VT, LHS, RHS, N0->getFlags());
3029 if (!N0.hasOneUse())
3030 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3034 case AMDGPUISD::FMUL_LEGACY: {
3035 // (fneg (fmul x, y)) -> (fmul x, (fneg y))
3036 // (fneg (fmul_legacy x, y)) -> (fmul_legacy x, (fneg y))
3037 SDValue LHS = N0.getOperand(0);
3038 SDValue RHS = N0.getOperand(1);
3040 if (LHS.getOpcode() == ISD::FNEG)
3041 LHS = LHS.getOperand(0);
3042 else if (RHS.getOpcode() == ISD::FNEG)
3043 RHS = RHS.getOperand(0);
3045 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3047 SDValue Res = DAG.getNode(Opc, SL, VT, LHS, RHS, N0->getFlags());
3048 if (!N0.hasOneUse())
3049 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3054 if (!mayIgnoreSignedZero(N0))
3057 // (fneg (fma x, y, z)) -> (fma x, (fneg y), (fneg z))
3058 SDValue LHS = N0.getOperand(0);
3059 SDValue MHS = N0.getOperand(1);
3060 SDValue RHS = N0.getOperand(2);
3062 if (LHS.getOpcode() == ISD::FNEG)
3063 LHS = LHS.getOperand(0);
3064 else if (MHS.getOpcode() == ISD::FNEG)
3065 MHS = MHS.getOperand(0);
3067 MHS = DAG.getNode(ISD::FNEG, SL, VT, MHS);
3069 if (RHS.getOpcode() != ISD::FNEG)
3070 RHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3072 RHS = RHS.getOperand(0);
3074 SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
3075 if (!N0.hasOneUse())
3076 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3081 case AMDGPUISD::FMAX_LEGACY:
3082 case AMDGPUISD::FMIN_LEGACY: {
3083 // fneg (fmaxnum x, y) -> fminnum (fneg x), (fneg y)
3084 // fneg (fminnum x, y) -> fmaxnum (fneg x), (fneg y)
3085 // fneg (fmax_legacy x, y) -> fmin_legacy (fneg x), (fneg y)
3086 // fneg (fmin_legacy x, y) -> fmax_legacy (fneg x), (fneg y)
3088 SDValue LHS = N0.getOperand(0);
3089 SDValue RHS = N0.getOperand(1);
3091 // 0 doesn't have a negated inline immediate.
3092 // TODO: Shouldn't fold 1/2pi either, and should be generalized to other
3094 if (isConstantFPZero(RHS))
3097 SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);
3098 SDValue NegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
3099 unsigned Opposite = inverseMinMax(Opc);
3101 SDValue Res = DAG.getNode(Opposite, SL, VT, NegLHS, NegRHS, N0->getFlags());
3102 if (!N0.hasOneUse())
3103 DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
3106 case ISD::FP_EXTEND:
3109 case ISD::FNEARBYINT: // XXX - Should fround be handled?
3111 case AMDGPUISD::RCP:
3112 case AMDGPUISD::RCP_LEGACY:
3113 case AMDGPUISD::SIN_HW: {
3114 SDValue CvtSrc = N0.getOperand(0);
3115 if (CvtSrc.getOpcode() == ISD::FNEG) {
3116 // (fneg (fp_extend (fneg x))) -> (fp_extend x)
3117 // (fneg (rcp (fneg x))) -> (rcp x)
3118 return DAG.getNode(Opc, SL, VT, CvtSrc.getOperand(0));
3121 if (!N0.hasOneUse())
3124 // (fneg (fp_extend x)) -> (fp_extend (fneg x))
3125 // (fneg (rcp x)) -> (rcp (fneg x))
3126 SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3127 return DAG.getNode(Opc, SL, VT, Neg, N0->getFlags());
3129 case ISD::FP_ROUND: {
3130 SDValue CvtSrc = N0.getOperand(0);
3132 if (CvtSrc.getOpcode() == ISD::FNEG) {
3133 // (fneg (fp_round (fneg x))) -> (fp_round x)
3134 return DAG.getNode(ISD::FP_ROUND, SL, VT,
3135 CvtSrc.getOperand(0), N0.getOperand(1));
3138 if (!N0.hasOneUse())
3141 // (fneg (fp_round x)) -> (fp_round (fneg x))
3142 SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
3143 return DAG.getNode(ISD::FP_ROUND, SL, VT, Neg, N0.getOperand(1));
3145 case ISD::FP16_TO_FP: {
3146 // v_cvt_f32_f16 supports source modifiers on pre-VI targets without legal
3147 // f16, but legalization of f16 fneg ends up pulling it out of the source.
3148 // Put the fneg back as a legal source operation that can be matched later.
3151 SDValue Src = N0.getOperand(0);
3152 EVT SrcVT = Src.getValueType();
3154 // fneg (fp16_to_fp x) -> fp16_to_fp (xor x, 0x8000)
3155 SDValue IntFNeg = DAG.getNode(ISD::XOR, SL, SrcVT, Src,
3156 DAG.getConstant(0x8000, SL, SrcVT));
3157 return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFNeg);
3164 SDValue AMDGPUTargetLowering::performFAbsCombine(SDNode *N,
3165 DAGCombinerInfo &DCI) const {
3166 SelectionDAG &DAG = DCI.DAG;
3167 SDValue N0 = N->getOperand(0);
3169 if (!N0.hasOneUse())
3172 switch (N0.getOpcode()) {
3173 case ISD::FP16_TO_FP: {
3174 assert(!Subtarget->has16BitInsts() && "should only see if f16 is illegal");
3176 SDValue Src = N0.getOperand(0);
3177 EVT SrcVT = Src.getValueType();
3179 // fabs (fp16_to_fp x) -> fp16_to_fp (and x, 0x7fff)
3180 SDValue IntFAbs = DAG.getNode(ISD::AND, SL, SrcVT, Src,
3181 DAG.getConstant(0x7fff, SL, SrcVT));
3182 return DAG.getNode(ISD::FP16_TO_FP, SL, N->getValueType(0), IntFAbs);
3189 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
3190 DAGCombinerInfo &DCI) const {
3191 SelectionDAG &DAG = DCI.DAG;
3194 switch(N->getOpcode()) {
3197 case ISD::BITCAST: {
3198 EVT DestVT = N->getValueType(0);
3200 // Push casts through vector builds. This helps avoid emitting a large
3201 // number of copies when materializing floating point vector constants.
3203 // vNt1 bitcast (vNt0 (build_vector t0:x, t0:y)) =>
3204 // vnt1 = build_vector (t1 (bitcast t0:x)), (t1 (bitcast t0:y))
3205 if (DestVT.isVector()) {
3206 SDValue Src = N->getOperand(0);
3207 if (Src.getOpcode() == ISD::BUILD_VECTOR) {
3208 EVT SrcVT = Src.getValueType();
3209 unsigned NElts = DestVT.getVectorNumElements();
3211 if (SrcVT.getVectorNumElements() == NElts) {
3212 EVT DestEltVT = DestVT.getVectorElementType();
3214 SmallVector<SDValue, 8> CastedElts;
3216 for (unsigned I = 0, E = SrcVT.getVectorNumElements(); I != E; ++I) {
3217 SDValue Elt = Src.getOperand(I);
3218 CastedElts.push_back(DAG.getNode(ISD::BITCAST, DL, DestEltVT, Elt));
3221 return DAG.getBuildVector(DestVT, SL, CastedElts);
3226 if (DestVT.getSizeInBits() != 64 && !DestVT.isVector())
3229 // Fold bitcasts of constants.
3231 // v2i32 (bitcast i64:k) -> build_vector lo_32(k), hi_32(k)
3232 // TODO: Generalize and move to DAGCombiner
3233 SDValue Src = N->getOperand(0);
3234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src)) {
3235 assert(Src.getValueType() == MVT::i64);
3237 uint64_t CVal = C->getZExtValue();
3238 return DAG.getNode(ISD::BUILD_VECTOR, SL, DestVT,
3239 DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3240 DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3243 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Src)) {
3244 const APInt &Val = C->getValueAPF().bitcastToAPInt();
3246 uint64_t CVal = Val.getZExtValue();
3247 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
3248 DAG.getConstant(Lo_32(CVal), SL, MVT::i32),
3249 DAG.getConstant(Hi_32(CVal), SL, MVT::i32));
3251 return DAG.getNode(ISD::BITCAST, SL, DestVT, Vec);
3257 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3260 return performShlCombine(N, DCI);
3263 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3266 return performSrlCombine(N, DCI);
3269 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
3272 return performSraCombine(N, DCI);
3275 return performMulCombine(N, DCI);
3277 return performMulhsCombine(N, DCI);
3279 return performMulhuCombine(N, DCI);
3280 case AMDGPUISD::MUL_I24:
3281 case AMDGPUISD::MUL_U24:
3282 case AMDGPUISD::MULHI_I24:
3283 case AMDGPUISD::MULHI_U24: {
3284 // If the first call to simplify is successfull, then N may end up being
3285 // deleted, so we shouldn't call simplifyI24 again.
3286 simplifyI24(N, 0, DCI) || simplifyI24(N, 1, DCI);
3289 case AMDGPUISD::MUL_LOHI_I24:
3290 case AMDGPUISD::MUL_LOHI_U24:
3291 return performMulLoHi24Combine(N, DCI);
3293 return performSelectCombine(N, DCI);
3295 return performFNegCombine(N, DCI);
3297 return performFAbsCombine(N, DCI);
3298 case AMDGPUISD::BFE_I32:
3299 case AMDGPUISD::BFE_U32: {
3300 assert(!N->getValueType(0).isVector() &&
3301 "Vector handling of BFE not implemented");
3302 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
3306 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
3308 return DAG.getConstant(0, DL, MVT::i32);
3310 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
3314 SDValue BitsFrom = N->getOperand(0);
3315 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
3317 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
3319 if (OffsetVal == 0) {
3320 // This is already sign / zero extended, so try to fold away extra BFEs.
3321 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
3323 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
3324 if (OpSignBits >= SignBits)
3327 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
3329 // This is a sign_extend_inreg. Replace it to take advantage of existing
3330 // DAG Combines. If not eliminated, we will match back to BFE during
3333 // TODO: The sext_inreg of extended types ends, although we can could
3334 // handle them in a single BFE.
3335 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
3336 DAG.getValueType(SmallVT));
3339 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
3342 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
3344 return constantFoldBFE<int32_t>(DAG,
3345 CVal->getSExtValue(),
3351 return constantFoldBFE<uint32_t>(DAG,
3352 CVal->getZExtValue(),
3358 if ((OffsetVal + WidthVal) >= 32) {
3359 SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
3360 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
3361 BitsFrom, ShiftVal);
3364 if (BitsFrom.hasOneUse()) {
3365 APInt Demanded = APInt::getBitsSet(32,
3367 OffsetVal + WidthVal);
3370 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
3371 !DCI.isBeforeLegalizeOps());
3372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3373 if (TLI.ShrinkDemandedConstant(BitsFrom, Demanded, TLO) ||
3374 TLI.SimplifyDemandedBits(BitsFrom, Demanded, Known, TLO)) {
3375 DCI.CommitTargetLoweringOpt(TLO);
3382 return performLoadCombine(N, DCI);
3384 return performStoreCombine(N, DCI);
3385 case AMDGPUISD::CLAMP:
3386 return performClampCombine(N, DCI);
3387 case AMDGPUISD::RCP: {
3388 if (const auto *CFP = dyn_cast<ConstantFPSDNode>(N->getOperand(0))) {
3389 // XXX - Should this flush denormals?
3390 const APFloat &Val = CFP->getValueAPF();
3391 APFloat One(Val.getSemantics(), "1.0");
3392 return DAG.getConstantFP(One / Val, SDLoc(N), N->getValueType(0));
3401 //===----------------------------------------------------------------------===//
3403 //===----------------------------------------------------------------------===//
3405 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
3406 const TargetRegisterClass *RC,
3407 unsigned Reg, EVT VT) const {
3408 MachineFunction &MF = DAG.getMachineFunction();
3409 MachineRegisterInfo &MRI = MF.getRegInfo();
3410 unsigned VirtualRegister;
3411 if (!MRI.isLiveIn(Reg)) {
3412 VirtualRegister = MRI.createVirtualRegister(RC);
3413 MRI.addLiveIn(Reg, VirtualRegister);
3415 VirtualRegister = MRI.getLiveInVirtReg(Reg);
3417 return DAG.getRegister(VirtualRegister, VT);
3420 uint32_t AMDGPUTargetLowering::getImplicitParameterOffset(
3421 const AMDGPUMachineFunction *MFI, const ImplicitParameter Param) const {
3422 unsigned Alignment = Subtarget->getAlignmentForImplicitArgPtr();
3423 uint64_t ArgOffset = alignTo(MFI->getABIArgOffset(), Alignment);
3428 return ArgOffset + 4;
3430 llvm_unreachable("unexpected implicit parameter type");
3433 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
3435 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
3436 switch ((AMDGPUISD::NodeType)Opcode) {
3437 case AMDGPUISD::FIRST_NUMBER: break;
3439 NODE_NAME_CASE(UMUL);
3440 NODE_NAME_CASE(BRANCH_COND);
3444 NODE_NAME_CASE(ELSE)
3445 NODE_NAME_CASE(LOOP)
3446 NODE_NAME_CASE(CALL)
3447 NODE_NAME_CASE(TRAP)
3448 NODE_NAME_CASE(RET_FLAG)
3449 NODE_NAME_CASE(RETURN_TO_EPILOG)
3450 NODE_NAME_CASE(ENDPGM)
3451 NODE_NAME_CASE(DWORDADDR)
3452 NODE_NAME_CASE(FRACT)
3453 NODE_NAME_CASE(SETCC)
3454 NODE_NAME_CASE(SETREG)
3455 NODE_NAME_CASE(FMA_W_CHAIN)
3456 NODE_NAME_CASE(FMUL_W_CHAIN)
3457 NODE_NAME_CASE(CLAMP)
3458 NODE_NAME_CASE(COS_HW)
3459 NODE_NAME_CASE(SIN_HW)
3460 NODE_NAME_CASE(FMAX_LEGACY)
3461 NODE_NAME_CASE(FMIN_LEGACY)
3462 NODE_NAME_CASE(FMAX3)
3463 NODE_NAME_CASE(SMAX3)
3464 NODE_NAME_CASE(UMAX3)
3465 NODE_NAME_CASE(FMIN3)
3466 NODE_NAME_CASE(SMIN3)
3467 NODE_NAME_CASE(UMIN3)
3468 NODE_NAME_CASE(FMED3)
3469 NODE_NAME_CASE(SMED3)
3470 NODE_NAME_CASE(UMED3)
3471 NODE_NAME_CASE(URECIP)
3472 NODE_NAME_CASE(DIV_SCALE)
3473 NODE_NAME_CASE(DIV_FMAS)
3474 NODE_NAME_CASE(DIV_FIXUP)
3475 NODE_NAME_CASE(FMAD_FTZ)
3476 NODE_NAME_CASE(TRIG_PREOP)
3479 NODE_NAME_CASE(RCP_LEGACY)
3480 NODE_NAME_CASE(RSQ_LEGACY)
3481 NODE_NAME_CASE(FMUL_LEGACY)
3482 NODE_NAME_CASE(RSQ_CLAMP)
3483 NODE_NAME_CASE(LDEXP)
3484 NODE_NAME_CASE(FP_CLASS)
3485 NODE_NAME_CASE(DOT4)
3486 NODE_NAME_CASE(CARRY)
3487 NODE_NAME_CASE(BORROW)
3488 NODE_NAME_CASE(BFE_U32)
3489 NODE_NAME_CASE(BFE_I32)
3492 NODE_NAME_CASE(FFBH_U32)
3493 NODE_NAME_CASE(FFBH_I32)
3494 NODE_NAME_CASE(MUL_U24)
3495 NODE_NAME_CASE(MUL_I24)
3496 NODE_NAME_CASE(MULHI_U24)
3497 NODE_NAME_CASE(MULHI_I24)
3498 NODE_NAME_CASE(MUL_LOHI_U24)
3499 NODE_NAME_CASE(MUL_LOHI_I24)
3500 NODE_NAME_CASE(MAD_U24)
3501 NODE_NAME_CASE(MAD_I24)
3502 NODE_NAME_CASE(TEXTURE_FETCH)
3503 NODE_NAME_CASE(EXPORT)
3504 NODE_NAME_CASE(EXPORT_DONE)
3505 NODE_NAME_CASE(R600_EXPORT)
3506 NODE_NAME_CASE(CONST_ADDRESS)
3507 NODE_NAME_CASE(REGISTER_LOAD)
3508 NODE_NAME_CASE(REGISTER_STORE)
3509 NODE_NAME_CASE(SAMPLE)
3510 NODE_NAME_CASE(SAMPLEB)
3511 NODE_NAME_CASE(SAMPLED)
3512 NODE_NAME_CASE(SAMPLEL)
3513 NODE_NAME_CASE(CVT_F32_UBYTE0)
3514 NODE_NAME_CASE(CVT_F32_UBYTE1)
3515 NODE_NAME_CASE(CVT_F32_UBYTE2)
3516 NODE_NAME_CASE(CVT_F32_UBYTE3)
3517 NODE_NAME_CASE(CVT_PKRTZ_F16_F32)
3518 NODE_NAME_CASE(FP_TO_FP16)
3519 NODE_NAME_CASE(FP16_ZEXT)
3520 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
3521 NODE_NAME_CASE(CONST_DATA_PTR)
3522 NODE_NAME_CASE(PC_ADD_REL_OFFSET)
3523 NODE_NAME_CASE(KILL)
3524 NODE_NAME_CASE(DUMMY_CHAIN)
3525 case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
3526 NODE_NAME_CASE(INIT_EXEC)
3527 NODE_NAME_CASE(INIT_EXEC_FROM_INPUT)
3528 NODE_NAME_CASE(SENDMSG)
3529 NODE_NAME_CASE(SENDMSGHALT)
3530 NODE_NAME_CASE(INTERP_MOV)
3531 NODE_NAME_CASE(INTERP_P1)
3532 NODE_NAME_CASE(INTERP_P2)
3533 NODE_NAME_CASE(STORE_MSKOR)
3534 NODE_NAME_CASE(LOAD_CONSTANT)
3535 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
3536 NODE_NAME_CASE(ATOMIC_CMP_SWAP)
3537 NODE_NAME_CASE(ATOMIC_INC)
3538 NODE_NAME_CASE(ATOMIC_DEC)
3539 NODE_NAME_CASE(BUFFER_LOAD)
3540 NODE_NAME_CASE(BUFFER_LOAD_FORMAT)
3541 case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
3546 SDValue AMDGPUTargetLowering::getSqrtEstimate(SDValue Operand,
3547 SelectionDAG &DAG, int Enabled,
3548 int &RefinementSteps,
3549 bool &UseOneConstNR,
3550 bool Reciprocal) const {
3551 EVT VT = Operand.getValueType();
3553 if (VT == MVT::f32) {
3554 RefinementSteps = 0;
3555 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
3558 // TODO: There is also f64 rsq instruction, but the documentation is less
3559 // clear on its precision.
3564 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
3565 SelectionDAG &DAG, int Enabled,
3566 int &RefinementSteps) const {
3567 EVT VT = Operand.getValueType();
3569 if (VT == MVT::f32) {
3570 // Reciprocal, < 1 ulp error.
3572 // This reciprocal approximation converges to < 0.5 ulp error with one
3573 // newton rhapson performed with two fused multiple adds (FMAs).
3575 RefinementSteps = 0;
3576 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
3579 // TODO: There is also f64 rcp instruction, but the documentation is less
3580 // clear on its precision.
3585 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
3586 const SDValue Op, KnownBits &Known,
3587 const APInt &DemandedElts, const SelectionDAG &DAG, unsigned Depth) const {
3589 Known.resetAll(); // Don't know anything.
3592 unsigned Opc = Op.getOpcode();
3597 case AMDGPUISD::CARRY:
3598 case AMDGPUISD::BORROW: {
3599 Known.Zero = APInt::getHighBitsSet(32, 31);
3603 case AMDGPUISD::BFE_I32:
3604 case AMDGPUISD::BFE_U32: {
3605 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3609 uint32_t Width = CWidth->getZExtValue() & 0x1f;
3611 if (Opc == AMDGPUISD::BFE_U32)
3612 Known.Zero = APInt::getHighBitsSet(32, 32 - Width);
3616 case AMDGPUISD::FP_TO_FP16:
3617 case AMDGPUISD::FP16_ZEXT: {
3618 unsigned BitWidth = Known.getBitWidth();
3620 // High bits are zero.
3621 Known.Zero = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
3627 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
3628 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
3629 unsigned Depth) const {
3630 switch (Op.getOpcode()) {
3631 case AMDGPUISD::BFE_I32: {
3632 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3636 unsigned SignBits = 32 - Width->getZExtValue() + 1;
3637 if (!isNullConstant(Op.getOperand(1)))
3640 // TODO: Could probably figure something out with non-0 offsets.
3641 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
3642 return std::max(SignBits, Op0SignBits);
3645 case AMDGPUISD::BFE_U32: {
3646 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
3647 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
3650 case AMDGPUISD::CARRY:
3651 case AMDGPUISD::BORROW:
3653 case AMDGPUISD::FP_TO_FP16:
3654 case AMDGPUISD::FP16_ZEXT: