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 "AMDGPUFrameLowering.h"
19 #include "AMDGPUIntrinsicInfo.h"
20 #include "AMDGPURegisterInfo.h"
21 #include "AMDGPUSubtarget.h"
22 #include "R600MachineFunctionInfo.h"
23 #include "SIMachineFunctionInfo.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/DiagnosticPrinter.h"
37 /// Diagnostic information for unimplemented or unsupported feature reporting.
38 class DiagnosticInfoUnsupported : public DiagnosticInfo {
40 const Twine &Description;
45 static int getKindID() {
47 KindID = llvm::getNextAvailablePluginDiagnosticKind();
52 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
53 DiagnosticSeverity Severity = DS_Error)
54 : DiagnosticInfo(getKindID(), Severity),
58 const Function &getFunction() const { return Fn; }
59 const Twine &getDescription() const { return Description; }
61 void print(DiagnosticPrinter &DP) const override {
62 DP << "unsupported " << getDescription() << " in " << Fn.getName();
65 static bool classof(const DiagnosticInfo *DI) {
66 return DI->getKind() == getKindID();
70 int DiagnosticInfoUnsupported::KindID = 0;
74 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
75 CCValAssign::LocInfo LocInfo,
76 ISD::ArgFlagsTy ArgFlags, CCState &State) {
77 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
78 ArgFlags.getOrigAlign());
79 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
84 #include "AMDGPUGenCallingConv.inc"
86 // Find a larger type to do a load / store of a vector with.
87 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
88 unsigned StoreSize = VT.getStoreSizeInBits();
90 return EVT::getIntegerVT(Ctx, StoreSize);
92 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
93 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
96 // Type for a vector that will be loaded to.
97 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
98 unsigned StoreSize = VT.getStoreSizeInBits();
100 return EVT::getIntegerVT(Ctx, 32);
102 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
105 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
108 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
110 setOperationAction(ISD::Constant, MVT::i32, Legal);
111 setOperationAction(ISD::Constant, MVT::i64, Legal);
112 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
113 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
115 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
116 setOperationAction(ISD::BRIND, MVT::Other, Expand);
118 // We need to custom lower some of the intrinsics
119 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
121 // Library functions. These default to Expand, but we have instructions
123 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
124 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
125 setOperationAction(ISD::FPOW, MVT::f32, Legal);
126 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
127 setOperationAction(ISD::FABS, MVT::f32, Legal);
128 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
129 setOperationAction(ISD::FRINT, MVT::f32, Legal);
130 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
132 setOperationAction(ISD::FROUND, MVT::f32, Custom);
133 setOperationAction(ISD::FROUND, MVT::f64, Custom);
135 setOperationAction(ISD::FREM, MVT::f32, Custom);
136 setOperationAction(ISD::FREM, MVT::f64, Custom);
138 // Lower floating point store/load to integer store/load to reduce the number
139 // of patterns in tablegen.
140 setOperationAction(ISD::STORE, MVT::f32, Promote);
141 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
143 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
144 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
146 setOperationAction(ISD::STORE, MVT::i64, Promote);
147 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
149 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
150 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
152 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
153 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
155 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
156 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
158 setOperationAction(ISD::STORE, MVT::f64, Promote);
159 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
161 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
162 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
164 // Custom lowering of vector stores is required for local address space
166 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
167 // XXX: Native v2i32 local address space stores are possible, but not
168 // currently implemented.
169 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
171 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
172 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
173 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
175 // XXX: This can be change to Custom, once ExpandVectorStores can
176 // handle 64-bit stores.
177 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
179 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
180 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
181 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
182 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
183 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
186 setOperationAction(ISD::LOAD, MVT::f32, Promote);
187 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
189 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
190 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
192 setOperationAction(ISD::LOAD, MVT::i64, Promote);
193 AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
195 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
196 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
198 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
199 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
201 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
202 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
204 setOperationAction(ISD::LOAD, MVT::f64, Promote);
205 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
207 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
208 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
210 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
211 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
212 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
213 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
214 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
215 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
216 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
217 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
218 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
219 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
221 // There are no 64-bit extloads. These should be done as a 32-bit extload and
222 // an extension to 64-bit.
223 for (MVT VT : MVT::integer_valuetypes()) {
224 setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
225 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
226 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
229 for (MVT VT : MVT::integer_vector_valuetypes()) {
230 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
231 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
232 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
233 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
234 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
235 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
236 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
237 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
238 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
239 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
240 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
241 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
244 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
246 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
247 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
248 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
249 setOperationAction(ISD::FRINT, MVT::f64, Custom);
250 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
253 if (!Subtarget->hasBFI()) {
254 // fcopysign can be done in a single instruction with BFI.
255 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
256 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
259 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
261 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
262 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
263 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
264 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
266 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
267 for (MVT VT : ScalarIntVTs) {
268 setOperationAction(ISD::SREM, VT, Expand);
269 setOperationAction(ISD::SDIV, VT, Expand);
271 // GPU does not have divrem function for signed or unsigned.
272 setOperationAction(ISD::SDIVREM, VT, Custom);
273 setOperationAction(ISD::UDIVREM, VT, Custom);
275 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
276 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
277 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
279 setOperationAction(ISD::BSWAP, VT, Expand);
280 setOperationAction(ISD::CTTZ, VT, Expand);
281 setOperationAction(ISD::CTLZ, VT, Expand);
284 if (!Subtarget->hasBCNT(32))
285 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
287 if (!Subtarget->hasBCNT(64))
288 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
290 // The hardware supports 32-bit ROTR, but not ROTL.
291 setOperationAction(ISD::ROTL, MVT::i32, Expand);
292 setOperationAction(ISD::ROTL, MVT::i64, Expand);
293 setOperationAction(ISD::ROTR, MVT::i64, Expand);
295 setOperationAction(ISD::MUL, MVT::i64, Expand);
296 setOperationAction(ISD::MULHU, MVT::i64, Expand);
297 setOperationAction(ISD::MULHS, MVT::i64, Expand);
298 setOperationAction(ISD::UDIV, MVT::i32, Expand);
299 setOperationAction(ISD::UREM, MVT::i32, Expand);
300 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
301 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
302 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
303 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
304 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
306 if (!Subtarget->hasFFBH())
307 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
309 if (!Subtarget->hasFFBL())
310 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
312 static const MVT::SimpleValueType VectorIntTypes[] = {
313 MVT::v2i32, MVT::v4i32
316 for (MVT VT : VectorIntTypes) {
317 // Expand the following operations for the current type by default.
318 setOperationAction(ISD::ADD, VT, Expand);
319 setOperationAction(ISD::AND, VT, Expand);
320 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
321 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
322 setOperationAction(ISD::MUL, VT, Expand);
323 setOperationAction(ISD::OR, VT, Expand);
324 setOperationAction(ISD::SHL, VT, Expand);
325 setOperationAction(ISD::SRA, VT, Expand);
326 setOperationAction(ISD::SRL, VT, Expand);
327 setOperationAction(ISD::ROTL, VT, Expand);
328 setOperationAction(ISD::ROTR, VT, Expand);
329 setOperationAction(ISD::SUB, VT, Expand);
330 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
331 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
332 setOperationAction(ISD::SDIV, VT, Expand);
333 setOperationAction(ISD::UDIV, VT, Expand);
334 setOperationAction(ISD::SREM, VT, Expand);
335 setOperationAction(ISD::UREM, VT, Expand);
336 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
337 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
338 setOperationAction(ISD::SDIVREM, VT, Custom);
339 setOperationAction(ISD::UDIVREM, VT, Custom);
340 setOperationAction(ISD::ADDC, VT, Expand);
341 setOperationAction(ISD::SUBC, VT, Expand);
342 setOperationAction(ISD::ADDE, VT, Expand);
343 setOperationAction(ISD::SUBE, VT, Expand);
344 setOperationAction(ISD::SELECT, VT, Expand);
345 setOperationAction(ISD::VSELECT, VT, Expand);
346 setOperationAction(ISD::SELECT_CC, VT, Expand);
347 setOperationAction(ISD::XOR, VT, Expand);
348 setOperationAction(ISD::BSWAP, VT, Expand);
349 setOperationAction(ISD::CTPOP, VT, Expand);
350 setOperationAction(ISD::CTTZ, VT, Expand);
351 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
352 setOperationAction(ISD::CTLZ, VT, Expand);
353 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
354 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
357 static const MVT::SimpleValueType FloatVectorTypes[] = {
358 MVT::v2f32, MVT::v4f32
361 for (MVT VT : FloatVectorTypes) {
362 setOperationAction(ISD::FABS, VT, Expand);
363 setOperationAction(ISD::FMINNUM, VT, Expand);
364 setOperationAction(ISD::FMAXNUM, VT, Expand);
365 setOperationAction(ISD::FADD, VT, Expand);
366 setOperationAction(ISD::FCEIL, VT, Expand);
367 setOperationAction(ISD::FCOS, VT, Expand);
368 setOperationAction(ISD::FDIV, VT, Expand);
369 setOperationAction(ISD::FEXP2, VT, Expand);
370 setOperationAction(ISD::FLOG2, VT, Expand);
371 setOperationAction(ISD::FREM, VT, Expand);
372 setOperationAction(ISD::FPOW, VT, Expand);
373 setOperationAction(ISD::FFLOOR, VT, Expand);
374 setOperationAction(ISD::FTRUNC, VT, Expand);
375 setOperationAction(ISD::FMUL, VT, Expand);
376 setOperationAction(ISD::FMA, VT, Expand);
377 setOperationAction(ISD::FRINT, VT, Expand);
378 setOperationAction(ISD::FNEARBYINT, VT, Expand);
379 setOperationAction(ISD::FSQRT, VT, Expand);
380 setOperationAction(ISD::FSIN, VT, Expand);
381 setOperationAction(ISD::FSUB, VT, Expand);
382 setOperationAction(ISD::FNEG, VT, Expand);
383 setOperationAction(ISD::SELECT, VT, Expand);
384 setOperationAction(ISD::VSELECT, VT, Expand);
385 setOperationAction(ISD::SELECT_CC, VT, Expand);
386 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
387 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
390 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
391 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
393 setTargetDAGCombine(ISD::MUL);
394 setTargetDAGCombine(ISD::SELECT);
395 setTargetDAGCombine(ISD::SELECT_CC);
396 setTargetDAGCombine(ISD::STORE);
398 setBooleanContents(ZeroOrNegativeOneBooleanContent);
399 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
401 setSchedulingPreference(Sched::RegPressure);
402 setJumpIsExpensive(true);
404 // SI at least has hardware support for floating point exceptions, but no way
405 // of using or handling them is implemented. They are also optional in OpenCL
407 setHasFloatingPointExceptions(false);
409 setSelectIsExpensive(false);
410 PredictableSelectIsExpensive = false;
412 // There are no integer divide instructions, and these expand to a pretty
413 // large sequence of instructions.
414 setIntDivIsCheap(false);
415 setPow2SDivIsCheap(false);
416 setFsqrtIsCheap(true);
418 // FIXME: Need to really handle these.
419 MaxStoresPerMemcpy = 4096;
420 MaxStoresPerMemmove = 4096;
421 MaxStoresPerMemset = 4096;
424 //===----------------------------------------------------------------------===//
425 // Target Information
426 //===----------------------------------------------------------------------===//
428 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
432 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
436 // The backend supports 32 and 64 bit floating point immediates.
437 // FIXME: Why are we reporting vectors of FP immediates as legal?
438 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
439 EVT ScalarVT = VT.getScalarType();
440 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
443 // We don't want to shrink f64 / f32 constants.
444 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
445 EVT ScalarVT = VT.getScalarType();
446 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
449 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
453 unsigned NewSize = NewVT.getStoreSizeInBits();
455 // If we are reducing to a 32-bit load, this is always better.
459 EVT OldVT = N->getValueType(0);
460 unsigned OldSize = OldVT.getStoreSizeInBits();
462 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
463 // extloads, so doing one requires using a buffer_load. In cases where we
464 // still couldn't use a scalar load, using the wider load shouldn't really
467 // If the old size already had to be an extload, there's no harm in continuing
468 // to reduce the width.
469 return (OldSize < 32);
472 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
474 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
477 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
478 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
480 return ((LScalarSize <= CastScalarSize) ||
481 (CastScalarSize >= 32) ||
485 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
486 // profitable with the expansion for 64-bit since it's generally good to
488 // FIXME: These should really have the size as a parameter.
489 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
493 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
497 //===---------------------------------------------------------------------===//
499 //===---------------------------------------------------------------------===//
501 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
502 assert(VT.isFloatingPoint());
503 return VT == MVT::f32 || VT == MVT::f64;
506 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
507 assert(VT.isFloatingPoint());
508 return VT == MVT::f32 || VT == MVT::f64;
511 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
512 // Truncate is just accessing a subregister.
513 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
516 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
517 // Truncate is just accessing a subregister.
518 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
519 (Dest->getPrimitiveSizeInBits() % 32 == 0);
522 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
523 const DataLayout *DL = getDataLayout();
524 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
525 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
527 return SrcSize == 32 && DestSize == 64;
530 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
531 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
532 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
533 // this will enable reducing 64-bit operations the 32-bit, which is always
535 return Src == MVT::i32 && Dest == MVT::i64;
538 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
539 return isZExtFree(Val.getValueType(), VT2);
542 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
543 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
544 // limited number of native 64-bit operations. Shrinking an operation to fit
545 // in a single 32-bit register should always be helpful. As currently used,
546 // this is much less general than the name suggests, and is only used in
547 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
548 // not profitable, and may actually be harmful.
549 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
552 //===---------------------------------------------------------------------===//
553 // TargetLowering Callbacks
554 //===---------------------------------------------------------------------===//
556 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
557 const SmallVectorImpl<ISD::InputArg> &Ins) const {
559 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
562 SDValue AMDGPUTargetLowering::LowerReturn(
564 CallingConv::ID CallConv,
566 const SmallVectorImpl<ISD::OutputArg> &Outs,
567 const SmallVectorImpl<SDValue> &OutVals,
568 SDLoc DL, SelectionDAG &DAG) const {
569 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
572 //===---------------------------------------------------------------------===//
573 // Target specific lowering
574 //===---------------------------------------------------------------------===//
576 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
577 SmallVectorImpl<SDValue> &InVals) const {
578 SDValue Callee = CLI.Callee;
579 SelectionDAG &DAG = CLI.DAG;
581 const Function &Fn = *DAG.getMachineFunction().getFunction();
583 StringRef FuncName("<unknown>");
585 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
586 FuncName = G->getSymbol();
587 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
588 FuncName = G->getGlobal()->getName();
590 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
591 DAG.getContext()->diagnose(NoCalls);
595 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
596 SelectionDAG &DAG) const {
597 switch (Op.getOpcode()) {
599 Op.getNode()->dump();
600 llvm_unreachable("Custom lowering code for this"
601 "instruction is not implemented yet!");
603 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
604 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
605 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
606 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
607 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
608 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
609 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
610 case ISD::FREM: return LowerFREM(Op, DAG);
611 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
612 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
613 case ISD::FRINT: return LowerFRINT(Op, DAG);
614 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
615 case ISD::FROUND: return LowerFROUND(Op, DAG);
616 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
617 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
618 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
619 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
620 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
625 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
626 SmallVectorImpl<SDValue> &Results,
627 SelectionDAG &DAG) const {
628 switch (N->getOpcode()) {
629 case ISD::SIGN_EXTEND_INREG:
630 // Different parts of legalization seem to interpret which type of
631 // sign_extend_inreg is the one to check for custom lowering. The extended
632 // from type is what really matters, but some places check for custom
633 // lowering of the result type. This results in trying to use
634 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
635 // nothing here and let the illegal result integer be handled normally.
638 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
642 Results.push_back(SDValue(Node, 0));
643 Results.push_back(SDValue(Node, 1));
644 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
646 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
650 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
651 if (Lowered.getNode())
652 Results.push_back(Lowered);
660 // FIXME: This implements accesses to initialized globals in the constant
661 // address space by copying them to private and accessing that. It does not
662 // properly handle illegal types or vectors. The private vector loads are not
663 // scalarized, and the illegal scalars hit an assertion. This technique will not
664 // work well with large initializers, and this should eventually be
665 // removed. Initialized globals should be placed into a data section that the
666 // runtime will load into a buffer before the kernel is executed. Uses of the
667 // global need to be replaced with a pointer loaded from an implicit kernel
668 // argument into this buffer holding the copy of the data, which will remove the
669 // need for any of this.
670 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
671 const GlobalValue *GV,
672 const SDValue &InitPtr,
674 SelectionDAG &DAG) const {
675 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
677 Type *InitTy = Init->getType();
679 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
680 EVT VT = EVT::getEVT(InitTy);
681 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
682 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
683 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
684 TD->getPrefTypeAlignment(InitTy));
687 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
688 EVT VT = EVT::getEVT(CFP->getType());
689 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
690 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
691 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
692 TD->getPrefTypeAlignment(CFP->getType()));
695 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
696 const StructLayout *SL = TD->getStructLayout(ST);
698 EVT PtrVT = InitPtr.getValueType();
699 SmallVector<SDValue, 8> Chains;
701 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
702 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
703 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
705 Constant *Elt = Init->getAggregateElement(I);
706 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
709 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
712 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
713 EVT PtrVT = InitPtr.getValueType();
715 unsigned NumElements;
716 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
717 NumElements = AT->getNumElements();
718 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
719 NumElements = VT->getNumElements();
721 llvm_unreachable("Unexpected type");
723 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
724 SmallVector<SDValue, 8> Chains;
725 for (unsigned i = 0; i < NumElements; ++i) {
726 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
727 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
729 Constant *Elt = Init->getAggregateElement(i);
730 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
733 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
736 if (isa<UndefValue>(Init)) {
737 EVT VT = EVT::getEVT(InitTy);
738 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
739 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
740 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
741 TD->getPrefTypeAlignment(InitTy));
745 llvm_unreachable("Unhandled constant initializer");
748 static bool hasDefinedInitializer(const GlobalValue *GV) {
749 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
750 if (!GVar || !GVar->hasInitializer())
753 if (isa<UndefValue>(GVar->getInitializer()))
759 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
761 SelectionDAG &DAG) const {
763 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
764 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
765 const GlobalValue *GV = G->getGlobal();
767 switch (G->getAddressSpace()) {
768 case AMDGPUAS::LOCAL_ADDRESS: {
769 // XXX: What does the value of G->getOffset() mean?
770 assert(G->getOffset() == 0 &&
771 "Do not know what to do with an non-zero offset");
773 // TODO: We could emit code to handle the initialization somewhere.
774 if (hasDefinedInitializer(GV))
778 if (MFI->LocalMemoryObjects.count(GV) == 0) {
779 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
780 Offset = MFI->LDSSize;
781 MFI->LocalMemoryObjects[GV] = Offset;
782 // XXX: Account for alignment?
783 MFI->LDSSize += Size;
785 Offset = MFI->LocalMemoryObjects[GV];
788 return DAG.getConstant(Offset, getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
790 case AMDGPUAS::CONSTANT_ADDRESS: {
791 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
792 Type *EltType = GV->getType()->getElementType();
793 unsigned Size = TD->getTypeAllocSize(EltType);
794 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
796 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
797 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
799 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
800 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
802 const GlobalVariable *Var = cast<GlobalVariable>(GV);
803 if (!Var->hasInitializer()) {
804 // This has no use, but bugpoint will hit it.
805 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
808 const Constant *Init = Var->getInitializer();
809 SmallVector<SDNode*, 8> WorkList;
811 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
812 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
813 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
815 WorkList.push_back(*I);
817 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
818 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
819 E = WorkList.end(); I != E; ++I) {
820 SmallVector<SDValue, 8> Ops;
821 Ops.push_back(Chain);
822 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
823 Ops.push_back((*I)->getOperand(i));
825 DAG.UpdateNodeOperands(*I, Ops);
827 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
831 const Function &Fn = *DAG.getMachineFunction().getFunction();
832 DiagnosticInfoUnsupported BadInit(Fn,
833 "initializer for address space");
834 DAG.getContext()->diagnose(BadInit);
838 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
839 SelectionDAG &DAG) const {
840 SmallVector<SDValue, 8> Args;
841 SDValue A = Op.getOperand(0);
842 SDValue B = Op.getOperand(1);
844 DAG.ExtractVectorElements(A, Args);
845 DAG.ExtractVectorElements(B, Args);
847 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
850 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
851 SelectionDAG &DAG) const {
853 SmallVector<SDValue, 8> Args;
854 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
855 EVT VT = Op.getValueType();
856 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
857 VT.getVectorNumElements());
859 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
862 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
863 SelectionDAG &DAG) const {
865 MachineFunction &MF = DAG.getMachineFunction();
866 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
867 getTargetMachine().getSubtargetImpl()->getFrameLowering());
869 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
871 unsigned FrameIndex = FIN->getIndex();
872 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
873 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
877 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
878 SelectionDAG &DAG) const {
879 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
881 EVT VT = Op.getValueType();
883 switch (IntrinsicID) {
885 case AMDGPUIntrinsic::AMDGPU_abs:
886 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
887 return LowerIntrinsicIABS(Op, DAG);
888 case AMDGPUIntrinsic::AMDGPU_lrp:
889 return LowerIntrinsicLRP(Op, DAG);
890 case AMDGPUIntrinsic::AMDGPU_fract:
891 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
892 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
894 case AMDGPUIntrinsic::AMDGPU_clamp:
895 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
896 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
897 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
899 case Intrinsic::AMDGPU_div_scale: {
900 // 3rd parameter required to be a constant.
901 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
903 return DAG.getUNDEF(VT);
905 // Translate to the operands expected by the machine instruction. The
906 // first parameter must be the same as the first instruction.
907 SDValue Numerator = Op.getOperand(1);
908 SDValue Denominator = Op.getOperand(2);
910 // Note this order is opposite of the machine instruction's operations,
911 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
912 // intrinsic has the numerator as the first operand to match a normal
913 // division operation.
915 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
917 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
918 Denominator, Numerator);
921 case Intrinsic::AMDGPU_div_fmas:
922 // FIXME: Dropping bool parameter. Work is needed to support the implicit
924 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
925 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
927 case Intrinsic::AMDGPU_div_fixup:
928 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
929 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
931 case Intrinsic::AMDGPU_trig_preop:
932 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
933 Op.getOperand(1), Op.getOperand(2));
935 case Intrinsic::AMDGPU_rcp:
936 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
938 case Intrinsic::AMDGPU_rsq:
939 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
941 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
942 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
944 case Intrinsic::AMDGPU_rsq_clamped:
945 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
946 Type *Type = VT.getTypeForEVT(*DAG.getContext());
947 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
948 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
950 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
951 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
952 DAG.getConstantFP(Max, VT));
953 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
954 DAG.getConstantFP(Min, VT));
956 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
959 case Intrinsic::AMDGPU_ldexp:
960 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
963 case AMDGPUIntrinsic::AMDGPU_imax:
964 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
966 case AMDGPUIntrinsic::AMDGPU_umax:
967 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
969 case AMDGPUIntrinsic::AMDGPU_imin:
970 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
972 case AMDGPUIntrinsic::AMDGPU_umin:
973 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
976 case AMDGPUIntrinsic::AMDGPU_umul24:
977 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
978 Op.getOperand(1), Op.getOperand(2));
980 case AMDGPUIntrinsic::AMDGPU_imul24:
981 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
982 Op.getOperand(1), Op.getOperand(2));
984 case AMDGPUIntrinsic::AMDGPU_umad24:
985 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
986 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
988 case AMDGPUIntrinsic::AMDGPU_imad24:
989 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
990 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
992 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
993 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
995 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
996 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
998 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
999 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
1001 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
1002 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
1004 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
1005 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
1010 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
1011 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
1016 case AMDGPUIntrinsic::AMDGPU_bfi:
1017 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
1022 case AMDGPUIntrinsic::AMDGPU_bfm:
1023 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1027 case AMDGPUIntrinsic::AMDGPU_brev:
1028 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
1030 case Intrinsic::AMDGPU_class:
1031 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
1032 Op.getOperand(1), Op.getOperand(2));
1034 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1035 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1037 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1038 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1039 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1040 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1044 ///IABS(a) = SMAX(sub(0, a), a)
1045 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1046 SelectionDAG &DAG) const {
1048 EVT VT = Op.getValueType();
1049 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1052 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1055 /// Linear Interpolation
1056 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
1057 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1058 SelectionDAG &DAG) const {
1060 EVT VT = Op.getValueType();
1061 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1062 DAG.getConstantFP(1.0f, MVT::f32),
1064 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1066 return DAG.getNode(ISD::FADD, DL, VT,
1067 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1071 /// \brief Generate Min/Max node
1072 SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
1079 DAGCombinerInfo &DCI) const {
1080 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1083 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1086 SelectionDAG &DAG = DCI.DAG;
1087 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1096 case ISD::SETFALSE2:
1105 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1106 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1112 // Ordered. Assume ordered for undefined.
1114 // Only do this after legalization to avoid interfering with other combines
1115 // which might occur.
1116 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1117 !DCI.isCalledByLegalizer())
1120 // We need to permute the operands to get the correct NaN behavior. The
1121 // selected operand is the second one based on the failing compare with NaN,
1122 // so permute it based on the compare type the hardware uses.
1124 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1125 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1130 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1131 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1137 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1138 !DCI.isCalledByLegalizer())
1142 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1143 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1145 case ISD::SETCC_INVALID:
1146 llvm_unreachable("Invalid setcc condcode!");
1151 /// \brief Generate Min/Max node
1152 SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1159 SelectionDAG &DAG) const {
1160 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1163 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1167 unsigned Opc = (LHS == True) ? AMDGPUISD::UMIN : AMDGPUISD::UMAX;
1168 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1172 unsigned Opc = (LHS == True) ? AMDGPUISD::SMIN : AMDGPUISD::SMAX;
1173 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1177 unsigned Opc = (LHS == True) ? AMDGPUISD::SMAX : AMDGPUISD::SMIN;
1178 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1182 unsigned Opc = (LHS == True) ? AMDGPUISD::UMAX : AMDGPUISD::UMIN;
1183 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1190 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1191 SelectionDAG &DAG) const {
1192 LoadSDNode *Load = cast<LoadSDNode>(Op);
1193 EVT MemVT = Load->getMemoryVT();
1194 EVT MemEltVT = MemVT.getVectorElementType();
1196 EVT LoadVT = Op.getValueType();
1197 EVT EltVT = LoadVT.getVectorElementType();
1198 EVT PtrVT = Load->getBasePtr().getValueType();
1200 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1201 SmallVector<SDValue, 8> Loads;
1202 SmallVector<SDValue, 8> Chains;
1205 unsigned MemEltSize = MemEltVT.getStoreSize();
1206 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1208 for (unsigned i = 0; i < NumElts; ++i) {
1209 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1210 DAG.getConstant(i * MemEltSize, PtrVT));
1213 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1214 Load->getChain(), Ptr,
1215 SrcValue.getWithOffset(i * MemEltSize),
1216 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1217 Load->isInvariant(), Load->getAlignment());
1218 Loads.push_back(NewLoad.getValue(0));
1219 Chains.push_back(NewLoad.getValue(1));
1223 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1224 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1227 return DAG.getMergeValues(Ops, SL);
1230 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1231 SelectionDAG &DAG) const {
1232 EVT VT = Op.getValueType();
1234 // If this is a 2 element vector, we really want to scalarize and not create
1235 // weird 1 element vectors.
1236 if (VT.getVectorNumElements() == 2)
1237 return ScalarizeVectorLoad(Op, DAG);
1239 LoadSDNode *Load = cast<LoadSDNode>(Op);
1240 SDValue BasePtr = Load->getBasePtr();
1241 EVT PtrVT = BasePtr.getValueType();
1242 EVT MemVT = Load->getMemoryVT();
1244 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1247 EVT LoMemVT, HiMemVT;
1250 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1251 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1252 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1254 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1255 Load->getChain(), BasePtr,
1257 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1258 Load->isInvariant(), Load->getAlignment());
1260 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1261 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1264 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1265 Load->getChain(), HiPtr,
1266 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1267 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1268 Load->isInvariant(), Load->getAlignment());
1271 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1272 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1273 LoLoad.getValue(1), HiLoad.getValue(1))
1276 return DAG.getMergeValues(Ops, SL);
1279 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1280 SelectionDAG &DAG) const {
1281 StoreSDNode *Store = cast<StoreSDNode>(Op);
1282 EVT MemVT = Store->getMemoryVT();
1283 unsigned MemBits = MemVT.getSizeInBits();
1285 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1286 // truncating store into an i32 store.
1287 // XXX: We could also handle optimize other vector bitwidths.
1288 if (!MemVT.isVector() || MemBits > 32) {
1293 SDValue Value = Store->getValue();
1294 EVT VT = Value.getValueType();
1295 EVT ElemVT = VT.getVectorElementType();
1296 SDValue Ptr = Store->getBasePtr();
1297 EVT MemEltVT = MemVT.getVectorElementType();
1298 unsigned MemEltBits = MemEltVT.getSizeInBits();
1299 unsigned MemNumElements = MemVT.getVectorNumElements();
1300 unsigned PackedSize = MemVT.getStoreSizeInBits();
1301 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1303 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1305 SDValue PackedValue;
1306 for (unsigned i = 0; i < MemNumElements; ++i) {
1307 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1308 DAG.getConstant(i, MVT::i32));
1309 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1310 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1312 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1313 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1318 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1322 if (PackedSize < 32) {
1323 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1324 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1325 Store->getMemOperand()->getPointerInfo(),
1327 Store->isNonTemporal(), Store->isVolatile(),
1328 Store->getAlignment());
1331 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1332 Store->getMemOperand()->getPointerInfo(),
1333 Store->isVolatile(), Store->isNonTemporal(),
1334 Store->getAlignment());
1337 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1338 SelectionDAG &DAG) const {
1339 StoreSDNode *Store = cast<StoreSDNode>(Op);
1340 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1341 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1342 EVT PtrVT = Store->getBasePtr().getValueType();
1343 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1346 SmallVector<SDValue, 8> Chains;
1348 unsigned EltSize = MemEltVT.getStoreSize();
1349 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1351 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1352 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1354 DAG.getConstant(i, MVT::i32));
1356 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), PtrVT);
1357 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1359 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1360 SrcValue.getWithOffset(i * EltSize),
1361 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1362 Store->getAlignment());
1363 Chains.push_back(NewStore);
1366 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1369 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1370 SelectionDAG &DAG) const {
1371 StoreSDNode *Store = cast<StoreSDNode>(Op);
1372 SDValue Val = Store->getValue();
1373 EVT VT = Val.getValueType();
1375 // If this is a 2 element vector, we really want to scalarize and not create
1376 // weird 1 element vectors.
1377 if (VT.getVectorNumElements() == 2)
1378 return ScalarizeVectorStore(Op, DAG);
1380 EVT MemVT = Store->getMemoryVT();
1381 SDValue Chain = Store->getChain();
1382 SDValue BasePtr = Store->getBasePtr();
1386 EVT LoMemVT, HiMemVT;
1389 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1390 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1391 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1393 EVT PtrVT = BasePtr.getValueType();
1394 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1395 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1397 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1399 = DAG.getTruncStore(Chain, SL, Lo,
1403 Store->isNonTemporal(),
1404 Store->isVolatile(),
1405 Store->getAlignment());
1407 = DAG.getTruncStore(Chain, SL, Hi,
1409 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1411 Store->isNonTemporal(),
1412 Store->isVolatile(),
1413 Store->getAlignment());
1415 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1419 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1421 LoadSDNode *Load = cast<LoadSDNode>(Op);
1422 ISD::LoadExtType ExtType = Load->getExtensionType();
1423 EVT VT = Op.getValueType();
1424 EVT MemVT = Load->getMemoryVT();
1426 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1427 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1428 // FIXME: Copied from PPC
1429 // First, load into 32 bits, then truncate to 1 bit.
1431 SDValue Chain = Load->getChain();
1432 SDValue BasePtr = Load->getBasePtr();
1433 MachineMemOperand *MMO = Load->getMemOperand();
1435 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1436 BasePtr, MVT::i8, MMO);
1439 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1443 return DAG.getMergeValues(Ops, DL);
1446 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1447 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1448 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1452 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1453 DAG.getConstant(2, MVT::i32));
1454 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1455 Load->getChain(), Ptr,
1456 DAG.getTargetConstant(0, MVT::i32),
1458 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1460 DAG.getConstant(0x3, MVT::i32));
1461 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1462 DAG.getConstant(3, MVT::i32));
1464 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1466 EVT MemEltVT = MemVT.getScalarType();
1467 if (ExtType == ISD::SEXTLOAD) {
1468 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1471 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1475 return DAG.getMergeValues(Ops, DL);
1479 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1483 return DAG.getMergeValues(Ops, DL);
1486 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1488 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1489 if (Result.getNode()) {
1493 StoreSDNode *Store = cast<StoreSDNode>(Op);
1494 SDValue Chain = Store->getChain();
1495 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1496 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1497 Store->getValue().getValueType().isVector()) {
1498 return ScalarizeVectorStore(Op, DAG);
1501 EVT MemVT = Store->getMemoryVT();
1502 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1503 MemVT.bitsLT(MVT::i32)) {
1505 if (Store->getMemoryVT() == MVT::i8) {
1507 } else if (Store->getMemoryVT() == MVT::i16) {
1510 SDValue BasePtr = Store->getBasePtr();
1511 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1512 DAG.getConstant(2, MVT::i32));
1513 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1514 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1516 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1517 DAG.getConstant(0x3, MVT::i32));
1519 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1520 DAG.getConstant(3, MVT::i32));
1522 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1525 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1527 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1528 MaskedValue, ShiftAmt);
1530 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1532 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1533 DAG.getConstant(0xffffffff, MVT::i32));
1534 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1536 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1537 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1538 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1543 // This is a shortcut for integer division because we have fast i32<->f32
1544 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1545 // float is enough to accurately represent up to a 24-bit integer.
1546 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1548 EVT VT = Op.getValueType();
1549 SDValue LHS = Op.getOperand(0);
1550 SDValue RHS = Op.getOperand(1);
1551 MVT IntVT = MVT::i32;
1552 MVT FltVT = MVT::f32;
1554 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1555 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1557 if (VT.isVector()) {
1558 unsigned NElts = VT.getVectorNumElements();
1559 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1560 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1563 unsigned BitSize = VT.getScalarType().getSizeInBits();
1565 SDValue jq = DAG.getConstant(1, IntVT);
1568 // char|short jq = ia ^ ib;
1569 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1571 // jq = jq >> (bitsize - 2)
1572 jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
1575 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
1578 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1581 // int ia = (int)LHS;
1583 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1585 // int ib, (int)RHS;
1587 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1589 // float fa = (float)ia;
1590 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1592 // float fb = (float)ib;
1593 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1595 // float fq = native_divide(fa, fb);
1596 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1597 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1600 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1602 // float fqneg = -fq;
1603 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1605 // float fr = mad(fqneg, fb, fa);
1606 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1607 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1609 // int iq = (int)fq;
1610 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1613 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1616 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1618 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1620 // int cv = fr >= fb;
1621 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1623 // jq = (cv ? jq : 0);
1624 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
1626 // dst = trunc/extend to legal type
1627 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1630 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1632 // Rem needs compensation, it's easier to recompute it
1633 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1634 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1640 return DAG.getMergeValues(Res, DL);
1643 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1645 SmallVectorImpl<SDValue> &Results) const {
1646 assert(Op.getValueType() == MVT::i64);
1649 EVT VT = Op.getValueType();
1650 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1652 SDValue one = DAG.getConstant(1, HalfVT);
1653 SDValue zero = DAG.getConstant(0, HalfVT);
1656 SDValue LHS = Op.getOperand(0);
1657 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1658 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1660 SDValue RHS = Op.getOperand(1);
1661 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1662 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1664 // Get Speculative values
1665 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1666 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1668 SDValue REM_Hi = zero;
1669 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1671 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1672 SDValue DIV_Lo = zero;
1674 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1676 for (unsigned i = 0; i < halfBitWidth; ++i) {
1677 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
1678 // Get Value of high bit
1680 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
1681 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
1683 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1684 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1687 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
1688 DAG.getConstant(halfBitWidth - 1, HalfVT));
1689 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
1690 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
1692 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
1693 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
1696 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1698 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
1699 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1701 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1705 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1707 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1708 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
1709 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
1712 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1713 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1714 Results.push_back(DIV);
1715 Results.push_back(REM);
1718 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1719 SelectionDAG &DAG) const {
1721 EVT VT = Op.getValueType();
1723 if (VT == MVT::i64) {
1724 SmallVector<SDValue, 2> Results;
1725 LowerUDIVREM64(Op, DAG, Results);
1726 return DAG.getMergeValues(Results, DL);
1729 SDValue Num = Op.getOperand(0);
1730 SDValue Den = Op.getOperand(1);
1732 if (VT == MVT::i32) {
1733 if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
1734 DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
1735 // TODO: We technically could do this for i64, but shouldn't that just be
1736 // handled by something generally reducing 64-bit division on 32-bit
1737 // values to 32-bit?
1738 return LowerDIVREM24(Op, DAG, false);
1742 // RCP = URECIP(Den) = 2^32 / Den + e
1743 // e is rounding error.
1744 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1746 // RCP_LO = mul(RCP, Den) */
1747 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1749 // RCP_HI = mulhu (RCP, Den) */
1750 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1752 // NEG_RCP_LO = -RCP_LO
1753 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1756 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1757 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1760 // Calculate the rounding error from the URECIP instruction
1761 // E = mulhu(ABS_RCP_LO, RCP)
1762 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1764 // RCP_A_E = RCP + E
1765 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1767 // RCP_S_E = RCP - E
1768 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1770 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1771 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1774 // Quotient = mulhu(Tmp0, Num)
1775 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1777 // Num_S_Remainder = Quotient * Den
1778 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1780 // Remainder = Num - Num_S_Remainder
1781 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1783 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1784 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1785 DAG.getConstant(-1, VT),
1786 DAG.getConstant(0, VT),
1788 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1789 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1791 DAG.getConstant(-1, VT),
1792 DAG.getConstant(0, VT),
1794 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1795 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1798 // Calculate Division result:
1800 // Quotient_A_One = Quotient + 1
1801 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1802 DAG.getConstant(1, VT));
1804 // Quotient_S_One = Quotient - 1
1805 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1806 DAG.getConstant(1, VT));
1808 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1809 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1810 Quotient, Quotient_A_One, ISD::SETEQ);
1812 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1813 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1814 Quotient_S_One, Div, ISD::SETEQ);
1816 // Calculate Rem result:
1818 // Remainder_S_Den = Remainder - Den
1819 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1821 // Remainder_A_Den = Remainder + Den
1822 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1824 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1825 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1826 Remainder, Remainder_S_Den, ISD::SETEQ);
1828 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1829 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1830 Remainder_A_Den, Rem, ISD::SETEQ);
1835 return DAG.getMergeValues(Ops, DL);
1838 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1839 SelectionDAG &DAG) const {
1841 EVT VT = Op.getValueType();
1843 SDValue LHS = Op.getOperand(0);
1844 SDValue RHS = Op.getOperand(1);
1846 if (VT == MVT::i32) {
1847 if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
1848 DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
1849 // TODO: We technically could do this for i64, but shouldn't that just be
1850 // handled by something generally reducing 64-bit division on 32-bit
1851 // values to 32-bit?
1852 return LowerDIVREM24(Op, DAG, true);
1856 SDValue Zero = DAG.getConstant(0, VT);
1857 SDValue NegOne = DAG.getConstant(-1, VT);
1859 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1860 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1861 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1862 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1864 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1865 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1867 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1868 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1870 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1871 SDValue Rem = Div.getValue(1);
1873 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1874 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1876 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1877 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1883 return DAG.getMergeValues(Res, DL);
1886 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1887 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1889 EVT VT = Op.getValueType();
1890 SDValue X = Op.getOperand(0);
1891 SDValue Y = Op.getOperand(1);
1893 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1894 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1895 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1897 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1900 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1902 SDValue Src = Op.getOperand(0);
1904 // result = trunc(src)
1905 // if (src > 0.0 && src != result)
1908 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1910 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1911 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1913 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1915 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1916 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1917 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1919 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1920 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1923 static SDValue extractF64Exponent(SDValue Hi, SDLoc SL, SelectionDAG &DAG) {
1924 const unsigned FractBits = 52;
1925 const unsigned ExpBits = 11;
1927 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1929 DAG.getConstant(FractBits - 32, MVT::i32),
1930 DAG.getConstant(ExpBits, MVT::i32));
1931 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1932 DAG.getConstant(1023, MVT::i32));
1937 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1939 SDValue Src = Op.getOperand(0);
1941 assert(Op.getValueType() == MVT::f64);
1943 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1944 const SDValue One = DAG.getConstant(1, MVT::i32);
1946 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1948 // Extract the upper half, since this is where we will find the sign and
1950 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1952 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
1954 const unsigned FractBits = 52;
1956 // Extract the sign bit.
1957 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1958 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1960 // Extend back to to 64-bits.
1961 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1963 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1965 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1966 const SDValue FractMask
1967 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1969 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1970 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1971 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1973 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1975 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1977 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1978 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1980 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1981 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1983 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1986 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1988 SDValue Src = Op.getOperand(0);
1990 assert(Op.getValueType() == MVT::f64);
1992 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1993 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1994 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1996 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1997 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1999 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2001 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
2002 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
2004 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
2005 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2007 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2010 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2011 // FNEARBYINT and FRINT are the same, except in their handling of FP
2012 // exceptions. Those aren't really meaningful for us, and OpenCL only has
2013 // rint, so just treat them as equivalent.
2014 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2017 // XXX - May require not supporting f32 denormals?
2018 SDValue AMDGPUTargetLowering::LowerFROUND32(SDValue Op, SelectionDAG &DAG) const {
2020 SDValue X = Op.getOperand(0);
2022 SDValue T = DAG.getNode(ISD::FTRUNC, SL, MVT::f32, X);
2024 SDValue Diff = DAG.getNode(ISD::FSUB, SL, MVT::f32, X, T);
2026 SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, MVT::f32, Diff);
2028 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f32);
2029 const SDValue One = DAG.getConstantFP(1.0, MVT::f32);
2030 const SDValue Half = DAG.getConstantFP(0.5, MVT::f32);
2032 SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f32, One, X);
2034 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
2036 SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2038 SDValue Sel = DAG.getNode(ISD::SELECT, SL, MVT::f32, Cmp, SignOne, Zero);
2040 return DAG.getNode(ISD::FADD, SL, MVT::f32, T, Sel);
2043 SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
2045 SDValue X = Op.getOperand(0);
2047 SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
2049 const SDValue Zero = DAG.getConstant(0, MVT::i32);
2050 const SDValue One = DAG.getConstant(1, MVT::i32);
2051 const SDValue NegOne = DAG.getConstant(-1, MVT::i32);
2052 const SDValue FiftyOne = DAG.getConstant(51, MVT::i32);
2053 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
2056 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
2058 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
2060 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2062 const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), MVT::i64);
2064 SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
2065 SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
2066 DAG.getConstant(INT64_C(0x0008000000000000), MVT::i64),
2069 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
2070 SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
2071 DAG.getConstant(0, MVT::i64), Tmp0,
2074 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
2075 D, DAG.getConstant(0, MVT::i64));
2076 SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
2078 K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
2079 K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
2081 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2082 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2083 SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
2085 SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
2087 DAG.getConstantFP(1.0, MVT::f64),
2088 DAG.getConstantFP(0.0, MVT::f64));
2090 SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
2092 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
2093 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
2098 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2099 EVT VT = Op.getValueType();
2102 return LowerFROUND32(Op, DAG);
2105 return LowerFROUND64(Op, DAG);
2107 llvm_unreachable("unhandled type");
2110 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2112 SDValue Src = Op.getOperand(0);
2114 // result = trunc(src);
2115 // if (src < 0.0 && src != result)
2118 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2120 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
2121 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
2123 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
2125 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2126 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2127 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2129 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2130 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2133 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2134 bool Signed) const {
2136 SDValue Src = Op.getOperand(0);
2138 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2140 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2141 DAG.getConstant(0, MVT::i32));
2142 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2143 DAG.getConstant(1, MVT::i32));
2145 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2148 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2150 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2151 DAG.getConstant(32, MVT::i32));
2153 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2156 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2157 SelectionDAG &DAG) const {
2158 SDValue S0 = Op.getOperand(0);
2159 if (S0.getValueType() != MVT::i64)
2162 EVT DestVT = Op.getValueType();
2163 if (DestVT == MVT::f64)
2164 return LowerINT_TO_FP64(Op, DAG, false);
2166 assert(DestVT == MVT::f32);
2170 // f32 uint_to_fp i64
2171 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2172 DAG.getConstant(0, MVT::i32));
2173 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2174 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2175 DAG.getConstant(1, MVT::i32));
2176 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2177 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2178 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
2179 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2182 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2183 SelectionDAG &DAG) const {
2184 SDValue Src = Op.getOperand(0);
2185 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2186 return LowerINT_TO_FP64(Op, DAG, true);
2191 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2192 bool Signed) const {
2195 SDValue Src = Op.getOperand(0);
2197 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2200 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), MVT::f64);
2202 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), MVT::f64);
2204 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2206 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2209 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2211 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2212 MVT::i32, FloorMul);
2213 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2215 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2217 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2220 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2221 SelectionDAG &DAG) const {
2222 SDValue Src = Op.getOperand(0);
2224 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2225 return LowerFP64_TO_INT(Op, DAG, true);
2230 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2231 SelectionDAG &DAG) const {
2232 SDValue Src = Op.getOperand(0);
2234 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2235 return LowerFP64_TO_INT(Op, DAG, false);
2240 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2241 SelectionDAG &DAG) const {
2242 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2243 MVT VT = Op.getSimpleValueType();
2244 MVT ScalarVT = VT.getScalarType();
2249 SDValue Src = Op.getOperand(0);
2252 // TODO: Don't scalarize on Evergreen?
2253 unsigned NElts = VT.getVectorNumElements();
2254 SmallVector<SDValue, 8> Args;
2255 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2257 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2258 for (unsigned I = 0; I < NElts; ++I)
2259 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2261 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2264 //===----------------------------------------------------------------------===//
2265 // Custom DAG optimizations
2266 //===----------------------------------------------------------------------===//
2268 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2269 APInt KnownZero, KnownOne;
2270 EVT VT = Op.getValueType();
2271 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2273 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2276 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2277 EVT VT = Op.getValueType();
2279 // In order for this to be a signed 24-bit value, bit 23, must
2281 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2282 // as unsigned 24-bit values.
2283 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2286 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2288 SelectionDAG &DAG = DCI.DAG;
2289 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2290 EVT VT = Op.getValueType();
2292 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2293 APInt KnownZero, KnownOne;
2294 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2295 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2296 DCI.CommitTargetLoweringOpt(TLO);
2299 template <typename IntTy>
2300 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2301 uint32_t Offset, uint32_t Width) {
2302 if (Width + Offset < 32) {
2303 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2304 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2305 return DAG.getConstant(Result, MVT::i32);
2308 return DAG.getConstant(Src0 >> Offset, MVT::i32);
2311 static bool usesAllNormalStores(SDNode *LoadVal) {
2312 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2313 if (!ISD::isNormalStore(*I))
2320 // If we have a copy of an illegal type, replace it with a load / store of an
2321 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2322 // instructions emitted when handling extloads and truncstores. Ideally we could
2323 // recognize the pack / unpack pattern to eliminate it.
2324 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2325 DAGCombinerInfo &DCI) const {
2326 if (!DCI.isBeforeLegalize())
2329 StoreSDNode *SN = cast<StoreSDNode>(N);
2330 SDValue Value = SN->getValue();
2331 EVT VT = Value.getValueType();
2333 if (isTypeLegal(VT) || SN->isVolatile() ||
2334 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2337 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2338 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2341 EVT MemVT = LoadVal->getMemoryVT();
2344 SelectionDAG &DAG = DCI.DAG;
2345 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2347 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2349 LoadVal->getChain(),
2350 LoadVal->getBasePtr(),
2351 LoadVal->getOffset(),
2353 LoadVal->getMemOperand());
2355 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2356 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2358 return DAG.getStore(SN->getChain(), SL, NewLoad,
2359 SN->getBasePtr(), SN->getMemOperand());
2362 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2363 DAGCombinerInfo &DCI) const {
2364 EVT VT = N->getValueType(0);
2366 if (VT.isVector() || VT.getSizeInBits() > 32)
2369 SelectionDAG &DAG = DCI.DAG;
2372 SDValue N0 = N->getOperand(0);
2373 SDValue N1 = N->getOperand(1);
2376 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2377 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2378 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2379 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2380 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2381 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2382 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2383 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2388 // We need to use sext even for MUL_U24, because MUL_U24 is used
2389 // for signed multiply of 8 and 16-bit types.
2390 return DAG.getSExtOrTrunc(Mul, DL, VT);
2393 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2394 DAGCombinerInfo &DCI) const {
2395 SelectionDAG &DAG = DCI.DAG;
2398 switch(N->getOpcode()) {
2401 return performMulCombine(N, DCI);
2402 case AMDGPUISD::MUL_I24:
2403 case AMDGPUISD::MUL_U24: {
2404 SDValue N0 = N->getOperand(0);
2405 SDValue N1 = N->getOperand(1);
2406 simplifyI24(N0, DCI);
2407 simplifyI24(N1, DCI);
2411 SDValue Cond = N->getOperand(0);
2412 if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
2414 EVT VT = N->getValueType(0);
2415 SDValue LHS = Cond.getOperand(0);
2416 SDValue RHS = Cond.getOperand(1);
2417 SDValue CC = Cond.getOperand(2);
2419 SDValue True = N->getOperand(1);
2420 SDValue False = N->getOperand(2);
2423 return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
2425 // TODO: Implement min / max Evergreen instructions.
2426 if (VT == MVT::i32 &&
2427 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2428 return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2434 case AMDGPUISD::BFE_I32:
2435 case AMDGPUISD::BFE_U32: {
2436 assert(!N->getValueType(0).isVector() &&
2437 "Vector handling of BFE not implemented");
2438 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2442 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2444 return DAG.getConstant(0, MVT::i32);
2446 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2450 SDValue BitsFrom = N->getOperand(0);
2451 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2453 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2455 if (OffsetVal == 0) {
2456 // This is already sign / zero extended, so try to fold away extra BFEs.
2457 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2459 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2460 if (OpSignBits >= SignBits)
2463 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2465 // This is a sign_extend_inreg. Replace it to take advantage of existing
2466 // DAG Combines. If not eliminated, we will match back to BFE during
2469 // TODO: The sext_inreg of extended types ends, although we can could
2470 // handle them in a single BFE.
2471 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2472 DAG.getValueType(SmallVT));
2475 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2478 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2480 return constantFoldBFE<int32_t>(DAG,
2481 CVal->getSExtValue(),
2486 return constantFoldBFE<uint32_t>(DAG,
2487 CVal->getZExtValue(),
2492 if ((OffsetVal + WidthVal) >= 32) {
2493 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2494 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2495 BitsFrom, ShiftVal);
2498 if (BitsFrom.hasOneUse()) {
2499 APInt Demanded = APInt::getBitsSet(32,
2501 OffsetVal + WidthVal);
2503 APInt KnownZero, KnownOne;
2504 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2505 !DCI.isBeforeLegalizeOps());
2506 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2507 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2508 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2509 KnownZero, KnownOne, TLO)) {
2510 DCI.CommitTargetLoweringOpt(TLO);
2518 return performStoreCombine(N, DCI);
2523 //===----------------------------------------------------------------------===//
2525 //===----------------------------------------------------------------------===//
2527 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2530 const SmallVectorImpl<ISD::InputArg> &Ins,
2531 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2533 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2534 if (Ins[i].ArgVT == Ins[i].VT) {
2535 OrigIns.push_back(Ins[i]);
2540 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2541 // Vector has been split into scalars.
2542 VT = Ins[i].ArgVT.getVectorElementType();
2543 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2544 Ins[i].ArgVT.getVectorElementType() !=
2545 Ins[i].VT.getVectorElementType()) {
2546 // Vector elements have been promoted
2549 // Vector has been spilt into smaller vectors.
2553 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2554 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2555 OrigIns.push_back(Arg);
2559 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2560 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2561 return CFP->isExactlyValue(1.0);
2563 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2564 return C->isAllOnesValue();
2569 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2570 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2571 return CFP->getValueAPF().isZero();
2573 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2574 return C->isNullValue();
2579 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2580 const TargetRegisterClass *RC,
2581 unsigned Reg, EVT VT) const {
2582 MachineFunction &MF = DAG.getMachineFunction();
2583 MachineRegisterInfo &MRI = MF.getRegInfo();
2584 unsigned VirtualRegister;
2585 if (!MRI.isLiveIn(Reg)) {
2586 VirtualRegister = MRI.createVirtualRegister(RC);
2587 MRI.addLiveIn(Reg, VirtualRegister);
2589 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2591 return DAG.getRegister(VirtualRegister, VT);
2594 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2596 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2598 default: return nullptr;
2600 NODE_NAME_CASE(CALL);
2601 NODE_NAME_CASE(UMUL);
2602 NODE_NAME_CASE(RET_FLAG);
2603 NODE_NAME_CASE(BRANCH_COND);
2606 NODE_NAME_CASE(DWORDADDR)
2607 NODE_NAME_CASE(FRACT)
2608 NODE_NAME_CASE(CLAMP)
2610 NODE_NAME_CASE(FMAX_LEGACY)
2611 NODE_NAME_CASE(SMAX)
2612 NODE_NAME_CASE(UMAX)
2613 NODE_NAME_CASE(FMIN_LEGACY)
2614 NODE_NAME_CASE(SMIN)
2615 NODE_NAME_CASE(UMIN)
2616 NODE_NAME_CASE(FMAX3)
2617 NODE_NAME_CASE(SMAX3)
2618 NODE_NAME_CASE(UMAX3)
2619 NODE_NAME_CASE(FMIN3)
2620 NODE_NAME_CASE(SMIN3)
2621 NODE_NAME_CASE(UMIN3)
2622 NODE_NAME_CASE(URECIP)
2623 NODE_NAME_CASE(DIV_SCALE)
2624 NODE_NAME_CASE(DIV_FMAS)
2625 NODE_NAME_CASE(DIV_FIXUP)
2626 NODE_NAME_CASE(TRIG_PREOP)
2629 NODE_NAME_CASE(RSQ_LEGACY)
2630 NODE_NAME_CASE(RSQ_CLAMPED)
2631 NODE_NAME_CASE(LDEXP)
2632 NODE_NAME_CASE(FP_CLASS)
2633 NODE_NAME_CASE(DOT4)
2634 NODE_NAME_CASE(BFE_U32)
2635 NODE_NAME_CASE(BFE_I32)
2638 NODE_NAME_CASE(BREV)
2639 NODE_NAME_CASE(MUL_U24)
2640 NODE_NAME_CASE(MUL_I24)
2641 NODE_NAME_CASE(MAD_U24)
2642 NODE_NAME_CASE(MAD_I24)
2643 NODE_NAME_CASE(EXPORT)
2644 NODE_NAME_CASE(CONST_ADDRESS)
2645 NODE_NAME_CASE(REGISTER_LOAD)
2646 NODE_NAME_CASE(REGISTER_STORE)
2647 NODE_NAME_CASE(LOAD_CONSTANT)
2648 NODE_NAME_CASE(LOAD_INPUT)
2649 NODE_NAME_CASE(SAMPLE)
2650 NODE_NAME_CASE(SAMPLEB)
2651 NODE_NAME_CASE(SAMPLED)
2652 NODE_NAME_CASE(SAMPLEL)
2653 NODE_NAME_CASE(CVT_F32_UBYTE0)
2654 NODE_NAME_CASE(CVT_F32_UBYTE1)
2655 NODE_NAME_CASE(CVT_F32_UBYTE2)
2656 NODE_NAME_CASE(CVT_F32_UBYTE3)
2657 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2658 NODE_NAME_CASE(CONST_DATA_PTR)
2659 NODE_NAME_CASE(STORE_MSKOR)
2660 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2664 SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
2665 DAGCombinerInfo &DCI,
2666 unsigned &RefinementSteps,
2667 bool &UseOneConstNR) const {
2668 SelectionDAG &DAG = DCI.DAG;
2669 EVT VT = Operand.getValueType();
2671 if (VT == MVT::f32) {
2672 RefinementSteps = 0;
2673 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
2676 // TODO: There is also f64 rsq instruction, but the documentation is less
2677 // clear on its precision.
2682 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
2683 DAGCombinerInfo &DCI,
2684 unsigned &RefinementSteps) const {
2685 SelectionDAG &DAG = DCI.DAG;
2686 EVT VT = Operand.getValueType();
2688 if (VT == MVT::f32) {
2689 // Reciprocal, < 1 ulp error.
2691 // This reciprocal approximation converges to < 0.5 ulp error with one
2692 // newton rhapson performed with two fused multiple adds (FMAs).
2694 RefinementSteps = 0;
2695 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
2698 // TODO: There is also f64 rcp instruction, but the documentation is less
2699 // clear on its precision.
2704 static void computeKnownBitsForMinMax(const SDValue Op0,
2708 const SelectionDAG &DAG,
2710 APInt Op0Zero, Op0One;
2711 APInt Op1Zero, Op1One;
2712 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2713 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2715 KnownZero = Op0Zero & Op1Zero;
2716 KnownOne = Op0One & Op1One;
2719 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2723 const SelectionDAG &DAG,
2724 unsigned Depth) const {
2726 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2730 unsigned Opc = Op.getOpcode();
2735 case ISD::INTRINSIC_WO_CHAIN: {
2736 // FIXME: The intrinsic should just use the node.
2737 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2738 case AMDGPUIntrinsic::AMDGPU_imax:
2739 case AMDGPUIntrinsic::AMDGPU_umax:
2740 case AMDGPUIntrinsic::AMDGPU_imin:
2741 case AMDGPUIntrinsic::AMDGPU_umin:
2742 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2743 KnownZero, KnownOne, DAG, Depth);
2751 case AMDGPUISD::SMAX:
2752 case AMDGPUISD::UMAX:
2753 case AMDGPUISD::SMIN:
2754 case AMDGPUISD::UMIN:
2755 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2756 KnownZero, KnownOne, DAG, Depth);
2759 case AMDGPUISD::BFE_I32:
2760 case AMDGPUISD::BFE_U32: {
2761 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2765 unsigned BitWidth = 32;
2766 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2768 if (Opc == AMDGPUISD::BFE_U32)
2769 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2776 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2778 const SelectionDAG &DAG,
2779 unsigned Depth) const {
2780 switch (Op.getOpcode()) {
2781 case AMDGPUISD::BFE_I32: {
2782 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2786 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2787 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2788 if (!Offset || !Offset->isNullValue())
2791 // TODO: Could probably figure something out with non-0 offsets.
2792 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2793 return std::max(SignBits, Op0SignBits);
2796 case AMDGPUISD::BFE_U32: {
2797 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2798 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;