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,
106 const AMDGPUSubtarget &STI)
107 : TargetLowering(TM), Subtarget(&STI) {
108 setOperationAction(ISD::Constant, MVT::i32, Legal);
109 setOperationAction(ISD::Constant, MVT::i64, Legal);
110 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
111 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
113 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
114 setOperationAction(ISD::BRIND, MVT::Other, Expand);
116 // We need to custom lower some of the intrinsics
117 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
119 // Library functions. These default to Expand, but we have instructions
121 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
122 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
123 setOperationAction(ISD::FPOW, MVT::f32, Legal);
124 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
125 setOperationAction(ISD::FABS, MVT::f32, Legal);
126 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
127 setOperationAction(ISD::FRINT, MVT::f32, Legal);
128 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
129 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
130 setOperationAction(ISD::FMAXNUM, 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 // v_mad_f32 does not support denormals according to some sources.
139 if (!Subtarget->hasFP32Denormals())
140 setOperationAction(ISD::FMAD, MVT::f32, Legal);
142 // Expand to fneg + fadd.
143 setOperationAction(ISD::FSUB, MVT::f64, Expand);
145 // Lower floating point store/load to integer store/load to reduce the number
146 // of patterns in tablegen.
147 setOperationAction(ISD::STORE, MVT::f32, Promote);
148 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
150 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
151 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
153 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
154 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
156 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
157 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
159 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
160 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
162 setOperationAction(ISD::STORE, MVT::f64, Promote);
163 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
165 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
166 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
168 // Custom lowering of vector stores is required for local address space
170 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
172 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
173 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
174 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
176 // XXX: This can be change to Custom, once ExpandVectorStores can
177 // handle 64-bit stores.
178 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
180 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
181 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
182 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
183 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
184 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
187 setOperationAction(ISD::LOAD, MVT::f32, Promote);
188 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
190 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
191 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
193 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
194 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
196 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
197 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
199 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
200 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
202 setOperationAction(ISD::LOAD, MVT::f64, Promote);
203 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
205 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
206 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
208 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
209 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
210 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
211 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
212 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
213 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
214 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
215 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
216 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
217 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
219 // There are no 64-bit extloads. These should be done as a 32-bit extload and
220 // an extension to 64-bit.
221 for (MVT VT : MVT::integer_valuetypes()) {
222 setLoadExtAction(ISD::EXTLOAD, MVT::i64, VT, Expand);
223 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, VT, Expand);
224 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, VT, Expand);
227 for (MVT VT : MVT::integer_vector_valuetypes()) {
228 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
229 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
230 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
231 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
232 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
233 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
234 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
235 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
236 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
237 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
238 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
239 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
242 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
244 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
245 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
246 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
247 setOperationAction(ISD::FRINT, MVT::f64, Custom);
248 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
251 if (!Subtarget->hasBFI()) {
252 // fcopysign can be done in a single instruction with BFI.
253 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
254 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
257 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
259 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
260 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
261 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
262 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
264 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
265 for (MVT VT : ScalarIntVTs) {
266 setOperationAction(ISD::SREM, VT, Expand);
267 setOperationAction(ISD::SDIV, VT, Expand);
269 // GPU does not have divrem function for signed or unsigned.
270 setOperationAction(ISD::SDIVREM, VT, Custom);
271 setOperationAction(ISD::UDIVREM, VT, Custom);
273 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
274 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
275 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
277 setOperationAction(ISD::BSWAP, VT, Expand);
278 setOperationAction(ISD::CTTZ, VT, Expand);
279 setOperationAction(ISD::CTLZ, VT, Expand);
282 if (!Subtarget->hasBCNT(32))
283 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
285 if (!Subtarget->hasBCNT(64))
286 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
288 // The hardware supports 32-bit ROTR, but not ROTL.
289 setOperationAction(ISD::ROTL, MVT::i32, Expand);
290 setOperationAction(ISD::ROTL, MVT::i64, Expand);
291 setOperationAction(ISD::ROTR, MVT::i64, Expand);
293 setOperationAction(ISD::MUL, MVT::i64, Expand);
294 setOperationAction(ISD::MULHU, MVT::i64, Expand);
295 setOperationAction(ISD::MULHS, MVT::i64, Expand);
296 setOperationAction(ISD::UDIV, MVT::i32, Expand);
297 setOperationAction(ISD::UREM, MVT::i32, Expand);
298 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
299 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
300 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
301 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
302 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
304 if (!Subtarget->hasFFBH())
305 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
307 if (!Subtarget->hasFFBL())
308 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
310 static const MVT::SimpleValueType VectorIntTypes[] = {
311 MVT::v2i32, MVT::v4i32
314 for (MVT VT : VectorIntTypes) {
315 // Expand the following operations for the current type by default.
316 setOperationAction(ISD::ADD, VT, Expand);
317 setOperationAction(ISD::AND, VT, Expand);
318 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
319 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
320 setOperationAction(ISD::MUL, VT, Expand);
321 setOperationAction(ISD::OR, VT, Expand);
322 setOperationAction(ISD::SHL, VT, Expand);
323 setOperationAction(ISD::SRA, VT, Expand);
324 setOperationAction(ISD::SRL, VT, Expand);
325 setOperationAction(ISD::ROTL, VT, Expand);
326 setOperationAction(ISD::ROTR, VT, Expand);
327 setOperationAction(ISD::SUB, VT, Expand);
328 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
329 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
330 setOperationAction(ISD::SDIV, VT, Expand);
331 setOperationAction(ISD::UDIV, VT, Expand);
332 setOperationAction(ISD::SREM, VT, Expand);
333 setOperationAction(ISD::UREM, VT, Expand);
334 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
335 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
336 setOperationAction(ISD::SDIVREM, VT, Custom);
337 setOperationAction(ISD::UDIVREM, VT, Custom);
338 setOperationAction(ISD::ADDC, VT, Expand);
339 setOperationAction(ISD::SUBC, VT, Expand);
340 setOperationAction(ISD::ADDE, VT, Expand);
341 setOperationAction(ISD::SUBE, VT, Expand);
342 setOperationAction(ISD::SELECT, VT, Expand);
343 setOperationAction(ISD::VSELECT, VT, Expand);
344 setOperationAction(ISD::SELECT_CC, VT, Expand);
345 setOperationAction(ISD::XOR, VT, Expand);
346 setOperationAction(ISD::BSWAP, VT, Expand);
347 setOperationAction(ISD::CTPOP, VT, Expand);
348 setOperationAction(ISD::CTTZ, VT, Expand);
349 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
350 setOperationAction(ISD::CTLZ, VT, Expand);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
352 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
355 static const MVT::SimpleValueType FloatVectorTypes[] = {
356 MVT::v2f32, MVT::v4f32
359 for (MVT VT : FloatVectorTypes) {
360 setOperationAction(ISD::FABS, VT, Expand);
361 setOperationAction(ISD::FMINNUM, VT, Expand);
362 setOperationAction(ISD::FMAXNUM, VT, Expand);
363 setOperationAction(ISD::FADD, VT, Expand);
364 setOperationAction(ISD::FCEIL, VT, Expand);
365 setOperationAction(ISD::FCOS, VT, Expand);
366 setOperationAction(ISD::FDIV, VT, Expand);
367 setOperationAction(ISD::FEXP2, VT, Expand);
368 setOperationAction(ISD::FLOG2, VT, Expand);
369 setOperationAction(ISD::FREM, VT, Expand);
370 setOperationAction(ISD::FPOW, VT, Expand);
371 setOperationAction(ISD::FFLOOR, VT, Expand);
372 setOperationAction(ISD::FTRUNC, VT, Expand);
373 setOperationAction(ISD::FMUL, VT, Expand);
374 setOperationAction(ISD::FMA, VT, Expand);
375 setOperationAction(ISD::FRINT, VT, Expand);
376 setOperationAction(ISD::FNEARBYINT, VT, Expand);
377 setOperationAction(ISD::FSQRT, VT, Expand);
378 setOperationAction(ISD::FSIN, VT, Expand);
379 setOperationAction(ISD::FSUB, VT, Expand);
380 setOperationAction(ISD::FNEG, VT, Expand);
381 setOperationAction(ISD::SELECT, VT, Expand);
382 setOperationAction(ISD::VSELECT, VT, Expand);
383 setOperationAction(ISD::SELECT_CC, VT, Expand);
384 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
385 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
388 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
389 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
391 setTargetDAGCombine(ISD::MUL);
392 setTargetDAGCombine(ISD::SELECT);
393 setTargetDAGCombine(ISD::SELECT_CC);
394 setTargetDAGCombine(ISD::STORE);
396 setTargetDAGCombine(ISD::FADD);
397 setTargetDAGCombine(ISD::FSUB);
399 setBooleanContents(ZeroOrNegativeOneBooleanContent);
400 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
402 setSchedulingPreference(Sched::RegPressure);
403 setJumpIsExpensive(true);
405 // SI at least has hardware support for floating point exceptions, but no way
406 // of using or handling them is implemented. They are also optional in OpenCL
408 setHasFloatingPointExceptions(false);
410 setSelectIsExpensive(false);
411 PredictableSelectIsExpensive = false;
413 // There are no integer divide instructions, and these expand to a pretty
414 // large sequence of instructions.
415 setIntDivIsCheap(false);
416 setPow2SDivIsCheap(false);
417 setFsqrtIsCheap(true);
419 // FIXME: Need to really handle these.
420 MaxStoresPerMemcpy = 4096;
421 MaxStoresPerMemmove = 4096;
422 MaxStoresPerMemset = 4096;
425 //===----------------------------------------------------------------------===//
426 // Target Information
427 //===----------------------------------------------------------------------===//
429 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
433 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
437 // The backend supports 32 and 64 bit floating point immediates.
438 // FIXME: Why are we reporting vectors of FP immediates as legal?
439 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
440 EVT ScalarVT = VT.getScalarType();
441 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
444 // We don't want to shrink f64 / f32 constants.
445 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
446 EVT ScalarVT = VT.getScalarType();
447 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
450 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
454 unsigned NewSize = NewVT.getStoreSizeInBits();
456 // If we are reducing to a 32-bit load, this is always better.
460 EVT OldVT = N->getValueType(0);
461 unsigned OldSize = OldVT.getStoreSizeInBits();
463 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
464 // extloads, so doing one requires using a buffer_load. In cases where we
465 // still couldn't use a scalar load, using the wider load shouldn't really
468 // If the old size already had to be an extload, there's no harm in continuing
469 // to reduce the width.
470 return (OldSize < 32);
473 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
475 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
478 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
479 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
481 return ((LScalarSize <= CastScalarSize) ||
482 (CastScalarSize >= 32) ||
486 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
487 // profitable with the expansion for 64-bit since it's generally good to
489 // FIXME: These should really have the size as a parameter.
490 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
494 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
498 //===---------------------------------------------------------------------===//
500 //===---------------------------------------------------------------------===//
502 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
503 assert(VT.isFloatingPoint());
504 return VT == MVT::f32 || VT == MVT::f64;
507 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
508 assert(VT.isFloatingPoint());
509 return VT == MVT::f32 || VT == MVT::f64;
512 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
513 // Truncate is just accessing a subregister.
514 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
517 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
518 // Truncate is just accessing a subregister.
519 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
520 (Dest->getPrimitiveSizeInBits() % 32 == 0);
523 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
524 const DataLayout *DL = getDataLayout();
525 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
526 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
528 return SrcSize == 32 && DestSize == 64;
531 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
532 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
533 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
534 // this will enable reducing 64-bit operations the 32-bit, which is always
536 return Src == MVT::i32 && Dest == MVT::i64;
539 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
540 return isZExtFree(Val.getValueType(), VT2);
543 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
544 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
545 // limited number of native 64-bit operations. Shrinking an operation to fit
546 // in a single 32-bit register should always be helpful. As currently used,
547 // this is much less general than the name suggests, and is only used in
548 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
549 // not profitable, and may actually be harmful.
550 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
553 //===---------------------------------------------------------------------===//
554 // TargetLowering Callbacks
555 //===---------------------------------------------------------------------===//
557 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
558 const SmallVectorImpl<ISD::InputArg> &Ins) const {
560 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
563 SDValue AMDGPUTargetLowering::LowerReturn(
565 CallingConv::ID CallConv,
567 const SmallVectorImpl<ISD::OutputArg> &Outs,
568 const SmallVectorImpl<SDValue> &OutVals,
569 SDLoc DL, SelectionDAG &DAG) const {
570 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
573 //===---------------------------------------------------------------------===//
574 // Target specific lowering
575 //===---------------------------------------------------------------------===//
577 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
578 SmallVectorImpl<SDValue> &InVals) const {
579 SDValue Callee = CLI.Callee;
580 SelectionDAG &DAG = CLI.DAG;
582 const Function &Fn = *DAG.getMachineFunction().getFunction();
584 StringRef FuncName("<unknown>");
586 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
587 FuncName = G->getSymbol();
588 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
589 FuncName = G->getGlobal()->getName();
591 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
592 DAG.getContext()->diagnose(NoCalls);
596 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
597 SelectionDAG &DAG) const {
598 switch (Op.getOpcode()) {
600 Op.getNode()->dump();
601 llvm_unreachable("Custom lowering code for this"
602 "instruction is not implemented yet!");
604 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
605 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
606 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
607 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
608 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
609 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
610 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
611 case ISD::FREM: return LowerFREM(Op, DAG);
612 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
613 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
614 case ISD::FRINT: return LowerFRINT(Op, DAG);
615 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
616 case ISD::FROUND: return LowerFROUND(Op, DAG);
617 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
618 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
619 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
620 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
621 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
626 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
627 SmallVectorImpl<SDValue> &Results,
628 SelectionDAG &DAG) const {
629 switch (N->getOpcode()) {
630 case ISD::SIGN_EXTEND_INREG:
631 // Different parts of legalization seem to interpret which type of
632 // sign_extend_inreg is the one to check for custom lowering. The extended
633 // from type is what really matters, but some places check for custom
634 // lowering of the result type. This results in trying to use
635 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
636 // nothing here and let the illegal result integer be handled normally.
639 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
643 Results.push_back(SDValue(Node, 0));
644 Results.push_back(SDValue(Node, 1));
645 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
647 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
651 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
652 if (Lowered.getNode())
653 Results.push_back(Lowered);
661 // FIXME: This implements accesses to initialized globals in the constant
662 // address space by copying them to private and accessing that. It does not
663 // properly handle illegal types or vectors. The private vector loads are not
664 // scalarized, and the illegal scalars hit an assertion. This technique will not
665 // work well with large initializers, and this should eventually be
666 // removed. Initialized globals should be placed into a data section that the
667 // runtime will load into a buffer before the kernel is executed. Uses of the
668 // global need to be replaced with a pointer loaded from an implicit kernel
669 // argument into this buffer holding the copy of the data, which will remove the
670 // need for any of this.
671 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
672 const GlobalValue *GV,
673 const SDValue &InitPtr,
675 SelectionDAG &DAG) const {
676 const DataLayout *TD = getDataLayout();
678 Type *InitTy = Init->getType();
680 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
681 EVT VT = EVT::getEVT(InitTy);
682 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
683 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, DL, VT), InitPtr,
684 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
685 TD->getPrefTypeAlignment(InitTy));
688 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
689 EVT VT = EVT::getEVT(CFP->getType());
690 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
691 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, DL, VT), InitPtr,
692 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
693 TD->getPrefTypeAlignment(CFP->getType()));
696 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
697 const StructLayout *SL = TD->getStructLayout(ST);
699 EVT PtrVT = InitPtr.getValueType();
700 SmallVector<SDValue, 8> Chains;
702 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
703 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), DL, PtrVT);
704 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
706 Constant *Elt = Init->getAggregateElement(I);
707 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
710 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
713 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
714 EVT PtrVT = InitPtr.getValueType();
716 unsigned NumElements;
717 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
718 NumElements = AT->getNumElements();
719 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
720 NumElements = VT->getNumElements();
722 llvm_unreachable("Unexpected type");
724 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
725 SmallVector<SDValue, 8> Chains;
726 for (unsigned i = 0; i < NumElements; ++i) {
727 SDValue Offset = DAG.getConstant(i * EltSize, DL, PtrVT);
728 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
730 Constant *Elt = Init->getAggregateElement(i);
731 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
734 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
737 if (isa<UndefValue>(Init)) {
738 EVT VT = EVT::getEVT(InitTy);
739 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
740 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
741 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
742 TD->getPrefTypeAlignment(InitTy));
746 llvm_unreachable("Unhandled constant initializer");
749 static bool hasDefinedInitializer(const GlobalValue *GV) {
750 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
751 if (!GVar || !GVar->hasInitializer())
754 if (isa<UndefValue>(GVar->getInitializer()))
760 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
762 SelectionDAG &DAG) const {
764 const DataLayout *TD = getDataLayout();
765 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
766 const GlobalValue *GV = G->getGlobal();
768 switch (G->getAddressSpace()) {
769 case AMDGPUAS::LOCAL_ADDRESS: {
770 // XXX: What does the value of G->getOffset() mean?
771 assert(G->getOffset() == 0 &&
772 "Do not know what to do with an non-zero offset");
774 // TODO: We could emit code to handle the initialization somewhere.
775 if (hasDefinedInitializer(GV))
779 if (MFI->LocalMemoryObjects.count(GV) == 0) {
780 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
781 Offset = MFI->LDSSize;
782 MFI->LocalMemoryObjects[GV] = Offset;
783 // XXX: Account for alignment?
784 MFI->LDSSize += Size;
786 Offset = MFI->LocalMemoryObjects[GV];
789 return DAG.getConstant(Offset, SDLoc(Op),
790 getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
792 case AMDGPUAS::CONSTANT_ADDRESS: {
793 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
794 Type *EltType = GV->getType()->getElementType();
795 unsigned Size = TD->getTypeAllocSize(EltType);
796 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
798 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
799 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
801 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
802 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
804 const GlobalVariable *Var = cast<GlobalVariable>(GV);
805 if (!Var->hasInitializer()) {
806 // This has no use, but bugpoint will hit it.
807 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
810 const Constant *Init = Var->getInitializer();
811 SmallVector<SDNode*, 8> WorkList;
813 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
814 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
815 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
817 WorkList.push_back(*I);
819 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
820 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
821 E = WorkList.end(); I != E; ++I) {
822 SmallVector<SDValue, 8> Ops;
823 Ops.push_back(Chain);
824 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
825 Ops.push_back((*I)->getOperand(i));
827 DAG.UpdateNodeOperands(*I, Ops);
829 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
833 const Function &Fn = *DAG.getMachineFunction().getFunction();
834 DiagnosticInfoUnsupported BadInit(Fn,
835 "initializer for address space");
836 DAG.getContext()->diagnose(BadInit);
840 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
841 SelectionDAG &DAG) const {
842 SmallVector<SDValue, 8> Args;
844 for (const SDUse &U : Op->ops())
845 DAG.ExtractVectorElements(U.get(), 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 = Subtarget->getFrameLowering();
868 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
870 unsigned FrameIndex = FIN->getIndex();
871 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
872 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), SDLoc(Op),
876 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
877 SelectionDAG &DAG) const {
878 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
880 EVT VT = Op.getValueType();
882 switch (IntrinsicID) {
884 case AMDGPUIntrinsic::AMDGPU_abs:
885 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
886 return LowerIntrinsicIABS(Op, DAG);
887 case AMDGPUIntrinsic::AMDGPU_lrp:
888 return LowerIntrinsicLRP(Op, DAG);
890 case AMDGPUIntrinsic::AMDGPU_clamp:
891 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
892 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
893 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
895 case Intrinsic::AMDGPU_div_scale: {
896 // 3rd parameter required to be a constant.
897 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
899 return DAG.getUNDEF(VT);
901 // Translate to the operands expected by the machine instruction. The
902 // first parameter must be the same as the first instruction.
903 SDValue Numerator = Op.getOperand(1);
904 SDValue Denominator = Op.getOperand(2);
906 // Note this order is opposite of the machine instruction's operations,
907 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
908 // intrinsic has the numerator as the first operand to match a normal
909 // division operation.
911 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
913 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
914 Denominator, Numerator);
917 case Intrinsic::AMDGPU_div_fmas:
918 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
919 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3),
922 case Intrinsic::AMDGPU_div_fixup:
923 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
924 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
926 case Intrinsic::AMDGPU_trig_preop:
927 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
928 Op.getOperand(1), Op.getOperand(2));
930 case Intrinsic::AMDGPU_rcp:
931 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
933 case Intrinsic::AMDGPU_rsq:
934 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
936 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
937 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
939 case Intrinsic::AMDGPU_rsq_clamped:
940 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
941 Type *Type = VT.getTypeForEVT(*DAG.getContext());
942 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
943 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
945 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
946 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
947 DAG.getConstantFP(Max, DL, VT));
948 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
949 DAG.getConstantFP(Min, DL, VT));
951 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
954 case Intrinsic::AMDGPU_ldexp:
955 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
958 case AMDGPUIntrinsic::AMDGPU_imax:
959 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
961 case AMDGPUIntrinsic::AMDGPU_umax:
962 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
964 case AMDGPUIntrinsic::AMDGPU_imin:
965 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
967 case AMDGPUIntrinsic::AMDGPU_umin:
968 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
971 case AMDGPUIntrinsic::AMDGPU_umul24:
972 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
973 Op.getOperand(1), Op.getOperand(2));
975 case AMDGPUIntrinsic::AMDGPU_imul24:
976 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
977 Op.getOperand(1), Op.getOperand(2));
979 case AMDGPUIntrinsic::AMDGPU_umad24:
980 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
981 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
983 case AMDGPUIntrinsic::AMDGPU_imad24:
984 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
985 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
987 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
988 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
990 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
991 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
993 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
994 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
996 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
997 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
999 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
1000 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
1005 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
1006 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
1011 case AMDGPUIntrinsic::AMDGPU_bfi:
1012 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
1017 case AMDGPUIntrinsic::AMDGPU_bfm:
1018 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1022 case AMDGPUIntrinsic::AMDGPU_brev:
1023 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
1025 case Intrinsic::AMDGPU_class:
1026 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
1027 Op.getOperand(1), Op.getOperand(2));
1029 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1030 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1032 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1033 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1034 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1035 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1039 ///IABS(a) = SMAX(sub(0, a), a)
1040 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1041 SelectionDAG &DAG) const {
1043 EVT VT = Op.getValueType();
1044 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1047 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1050 /// Linear Interpolation
1051 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
1052 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1053 SelectionDAG &DAG) const {
1055 EVT VT = Op.getValueType();
1056 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1057 DAG.getConstantFP(1.0f, DL, MVT::f32),
1059 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1061 return DAG.getNode(ISD::FADD, DL, VT,
1062 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1066 /// \brief Generate Min/Max node
1067 SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
1074 DAGCombinerInfo &DCI) const {
1075 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1078 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1081 SelectionDAG &DAG = DCI.DAG;
1082 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1091 case ISD::SETFALSE2:
1100 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1101 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1107 // Ordered. Assume ordered for undefined.
1109 // Only do this after legalization to avoid interfering with other combines
1110 // which might occur.
1111 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1112 !DCI.isCalledByLegalizer())
1115 // We need to permute the operands to get the correct NaN behavior. The
1116 // selected operand is the second one based on the failing compare with NaN,
1117 // so permute it based on the compare type the hardware uses.
1119 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1120 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1125 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1126 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1132 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1133 !DCI.isCalledByLegalizer())
1137 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1138 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1140 case ISD::SETCC_INVALID:
1141 llvm_unreachable("Invalid setcc condcode!");
1146 /// \brief Generate Min/Max node
1147 SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1154 SelectionDAG &DAG) const {
1155 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1158 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1162 unsigned Opc = (LHS == True) ? AMDGPUISD::UMIN : AMDGPUISD::UMAX;
1163 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1167 unsigned Opc = (LHS == True) ? AMDGPUISD::SMIN : AMDGPUISD::SMAX;
1168 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1172 unsigned Opc = (LHS == True) ? AMDGPUISD::SMAX : AMDGPUISD::SMIN;
1173 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1177 unsigned Opc = (LHS == True) ? AMDGPUISD::UMAX : AMDGPUISD::UMIN;
1178 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1185 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1186 SelectionDAG &DAG) const {
1187 LoadSDNode *Load = cast<LoadSDNode>(Op);
1188 EVT MemVT = Load->getMemoryVT();
1189 EVT MemEltVT = MemVT.getVectorElementType();
1191 EVT LoadVT = Op.getValueType();
1192 EVT EltVT = LoadVT.getVectorElementType();
1193 EVT PtrVT = Load->getBasePtr().getValueType();
1195 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1196 SmallVector<SDValue, 8> Loads;
1197 SmallVector<SDValue, 8> Chains;
1200 unsigned MemEltSize = MemEltVT.getStoreSize();
1201 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1203 for (unsigned i = 0; i < NumElts; ++i) {
1204 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1205 DAG.getConstant(i * MemEltSize, SL, PtrVT));
1208 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1209 Load->getChain(), Ptr,
1210 SrcValue.getWithOffset(i * MemEltSize),
1211 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1212 Load->isInvariant(), Load->getAlignment());
1213 Loads.push_back(NewLoad.getValue(0));
1214 Chains.push_back(NewLoad.getValue(1));
1218 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1219 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1222 return DAG.getMergeValues(Ops, SL);
1225 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1226 SelectionDAG &DAG) const {
1227 EVT VT = Op.getValueType();
1229 // If this is a 2 element vector, we really want to scalarize and not create
1230 // weird 1 element vectors.
1231 if (VT.getVectorNumElements() == 2)
1232 return ScalarizeVectorLoad(Op, DAG);
1234 LoadSDNode *Load = cast<LoadSDNode>(Op);
1235 SDValue BasePtr = Load->getBasePtr();
1236 EVT PtrVT = BasePtr.getValueType();
1237 EVT MemVT = Load->getMemoryVT();
1239 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1242 EVT LoMemVT, HiMemVT;
1245 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1246 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1247 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1249 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1250 Load->getChain(), BasePtr,
1252 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1253 Load->isInvariant(), Load->getAlignment());
1255 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1256 DAG.getConstant(LoMemVT.getStoreSize(), SL,
1260 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1261 Load->getChain(), HiPtr,
1262 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1263 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1264 Load->isInvariant(), Load->getAlignment());
1267 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1268 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1269 LoLoad.getValue(1), HiLoad.getValue(1))
1272 return DAG.getMergeValues(Ops, SL);
1275 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1276 SelectionDAG &DAG) const {
1277 StoreSDNode *Store = cast<StoreSDNode>(Op);
1278 EVT MemVT = Store->getMemoryVT();
1279 unsigned MemBits = MemVT.getSizeInBits();
1281 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1282 // truncating store into an i32 store.
1283 // XXX: We could also handle optimize other vector bitwidths.
1284 if (!MemVT.isVector() || MemBits > 32) {
1289 SDValue Value = Store->getValue();
1290 EVT VT = Value.getValueType();
1291 EVT ElemVT = VT.getVectorElementType();
1292 SDValue Ptr = Store->getBasePtr();
1293 EVT MemEltVT = MemVT.getVectorElementType();
1294 unsigned MemEltBits = MemEltVT.getSizeInBits();
1295 unsigned MemNumElements = MemVT.getVectorNumElements();
1296 unsigned PackedSize = MemVT.getStoreSizeInBits();
1297 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, DL, MVT::i32);
1299 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1301 SDValue PackedValue;
1302 for (unsigned i = 0; i < MemNumElements; ++i) {
1303 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1304 DAG.getConstant(i, DL, MVT::i32));
1305 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1306 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1308 SDValue Shift = DAG.getConstant(MemEltBits * i, DL, MVT::i32);
1309 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1314 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1318 if (PackedSize < 32) {
1319 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1320 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1321 Store->getMemOperand()->getPointerInfo(),
1323 Store->isNonTemporal(), Store->isVolatile(),
1324 Store->getAlignment());
1327 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1328 Store->getMemOperand()->getPointerInfo(),
1329 Store->isVolatile(), Store->isNonTemporal(),
1330 Store->getAlignment());
1333 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1334 SelectionDAG &DAG) const {
1335 StoreSDNode *Store = cast<StoreSDNode>(Op);
1336 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1337 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1338 EVT PtrVT = Store->getBasePtr().getValueType();
1339 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1342 SmallVector<SDValue, 8> Chains;
1344 unsigned EltSize = MemEltVT.getStoreSize();
1345 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1347 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1348 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1350 DAG.getConstant(i, SL, MVT::i32));
1352 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), SL, PtrVT);
1353 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1355 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1356 SrcValue.getWithOffset(i * EltSize),
1357 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1358 Store->getAlignment());
1359 Chains.push_back(NewStore);
1362 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1365 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1366 SelectionDAG &DAG) const {
1367 StoreSDNode *Store = cast<StoreSDNode>(Op);
1368 SDValue Val = Store->getValue();
1369 EVT VT = Val.getValueType();
1371 // If this is a 2 element vector, we really want to scalarize and not create
1372 // weird 1 element vectors.
1373 if (VT.getVectorNumElements() == 2)
1374 return ScalarizeVectorStore(Op, DAG);
1376 EVT MemVT = Store->getMemoryVT();
1377 SDValue Chain = Store->getChain();
1378 SDValue BasePtr = Store->getBasePtr();
1382 EVT LoMemVT, HiMemVT;
1385 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1386 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1387 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1389 EVT PtrVT = BasePtr.getValueType();
1390 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1391 DAG.getConstant(LoMemVT.getStoreSize(), SL,
1394 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1396 = DAG.getTruncStore(Chain, SL, Lo,
1400 Store->isNonTemporal(),
1401 Store->isVolatile(),
1402 Store->getAlignment());
1404 = DAG.getTruncStore(Chain, SL, Hi,
1406 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1408 Store->isNonTemporal(),
1409 Store->isVolatile(),
1410 Store->getAlignment());
1412 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1416 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1418 LoadSDNode *Load = cast<LoadSDNode>(Op);
1419 ISD::LoadExtType ExtType = Load->getExtensionType();
1420 EVT VT = Op.getValueType();
1421 EVT MemVT = Load->getMemoryVT();
1423 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1424 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1425 // FIXME: Copied from PPC
1426 // First, load into 32 bits, then truncate to 1 bit.
1428 SDValue Chain = Load->getChain();
1429 SDValue BasePtr = Load->getBasePtr();
1430 MachineMemOperand *MMO = Load->getMemOperand();
1432 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1433 BasePtr, MVT::i8, MMO);
1436 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1440 return DAG.getMergeValues(Ops, DL);
1443 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1444 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1445 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1449 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1450 DAG.getConstant(2, DL, MVT::i32));
1451 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1452 Load->getChain(), Ptr,
1453 DAG.getTargetConstant(0, DL, MVT::i32),
1455 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1457 DAG.getConstant(0x3, DL, MVT::i32));
1458 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1459 DAG.getConstant(3, DL, MVT::i32));
1461 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1463 EVT MemEltVT = MemVT.getScalarType();
1464 if (ExtType == ISD::SEXTLOAD) {
1465 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1468 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1472 return DAG.getMergeValues(Ops, DL);
1476 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1480 return DAG.getMergeValues(Ops, DL);
1483 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1485 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1486 if (Result.getNode()) {
1490 StoreSDNode *Store = cast<StoreSDNode>(Op);
1491 SDValue Chain = Store->getChain();
1492 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1493 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1494 Store->getValue().getValueType().isVector()) {
1495 return ScalarizeVectorStore(Op, DAG);
1498 EVT MemVT = Store->getMemoryVT();
1499 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1500 MemVT.bitsLT(MVT::i32)) {
1502 if (Store->getMemoryVT() == MVT::i8) {
1504 } else if (Store->getMemoryVT() == MVT::i16) {
1507 SDValue BasePtr = Store->getBasePtr();
1508 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1509 DAG.getConstant(2, DL, MVT::i32));
1510 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1512 DAG.getTargetConstant(0, DL, MVT::i32));
1514 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1515 DAG.getConstant(0x3, DL, MVT::i32));
1517 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1518 DAG.getConstant(3, DL, MVT::i32));
1520 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1523 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1525 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1526 MaskedValue, ShiftAmt);
1528 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32,
1529 DAG.getConstant(Mask, DL, MVT::i32),
1531 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1532 DAG.getConstant(0xffffffff, DL, MVT::i32));
1533 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1535 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1536 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1538 DAG.getTargetConstant(0, DL, 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, DL, 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,
1573 DAG.getConstant(BitSize - 2, DL, VT));
1576 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, DL, VT));
1579 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1582 // int ia = (int)LHS;
1584 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1586 // int ib, (int)RHS;
1588 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1590 // float fa = (float)ia;
1591 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1593 // float fb = (float)ib;
1594 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1596 // float fq = native_divide(fa, fb);
1597 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1598 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1601 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1603 // float fqneg = -fq;
1604 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1606 // float fr = mad(fqneg, fb, fa);
1607 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1608 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1610 // int iq = (int)fq;
1611 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1614 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1617 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1619 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1621 // int cv = fr >= fb;
1622 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1624 // jq = (cv ? jq : 0);
1625 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, DL, VT));
1627 // dst = trunc/extend to legal type
1628 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1631 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1633 // Rem needs compensation, it's easier to recompute it
1634 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1635 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1641 return DAG.getMergeValues(Res, DL);
1644 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1646 SmallVectorImpl<SDValue> &Results) const {
1647 assert(Op.getValueType() == MVT::i64);
1650 EVT VT = Op.getValueType();
1651 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1653 SDValue one = DAG.getConstant(1, DL, HalfVT);
1654 SDValue zero = DAG.getConstant(0, DL, HalfVT);
1657 SDValue LHS = Op.getOperand(0);
1658 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1659 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1661 SDValue RHS = Op.getOperand(1);
1662 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1663 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1665 if (VT == MVT::i64 &&
1666 DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
1667 DAG.MaskedValueIsZero(LHS, APInt::getHighBitsSet(64, 32))) {
1669 SDValue Res = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1672 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(0), zero);
1673 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, Res.getValue(1), zero);
1674 Results.push_back(DIV);
1675 Results.push_back(REM);
1679 // Get Speculative values
1680 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1681 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1683 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1684 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, zero);
1686 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1687 SDValue DIV_Lo = zero;
1689 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1691 for (unsigned i = 0; i < halfBitWidth; ++i) {
1692 const unsigned bitPos = halfBitWidth - i - 1;
1693 SDValue POS = DAG.getConstant(bitPos, DL, HalfVT);
1694 // Get value of high bit
1695 SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1696 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1697 HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
1700 REM = DAG.getNode(ISD::SHL, DL, VT, REM, DAG.getConstant(1, DL, VT));
1702 REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
1704 SDValue BIT = DAG.getConstant(1 << bitPos, DL, HalfVT);
1705 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1707 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1710 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1711 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1714 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1715 Results.push_back(DIV);
1716 Results.push_back(REM);
1719 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1720 SelectionDAG &DAG) const {
1722 EVT VT = Op.getValueType();
1724 if (VT == MVT::i64) {
1725 SmallVector<SDValue, 2> Results;
1726 LowerUDIVREM64(Op, DAG, Results);
1727 return DAG.getMergeValues(Results, DL);
1730 SDValue Num = Op.getOperand(0);
1731 SDValue Den = Op.getOperand(1);
1733 if (VT == MVT::i32) {
1734 if (DAG.MaskedValueIsZero(Num, APInt::getHighBitsSet(32, 8)) &&
1735 DAG.MaskedValueIsZero(Den, APInt::getHighBitsSet(32, 8))) {
1736 // TODO: We technically could do this for i64, but shouldn't that just be
1737 // handled by something generally reducing 64-bit division on 32-bit
1738 // values to 32-bit?
1739 return LowerDIVREM24(Op, DAG, false);
1743 // RCP = URECIP(Den) = 2^32 / Den + e
1744 // e is rounding error.
1745 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1747 // RCP_LO = mul(RCP, Den) */
1748 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1750 // RCP_HI = mulhu (RCP, Den) */
1751 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1753 // NEG_RCP_LO = -RCP_LO
1754 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
1757 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1758 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1761 // Calculate the rounding error from the URECIP instruction
1762 // E = mulhu(ABS_RCP_LO, RCP)
1763 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1765 // RCP_A_E = RCP + E
1766 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1768 // RCP_S_E = RCP - E
1769 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1771 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1772 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, DL, VT),
1775 // Quotient = mulhu(Tmp0, Num)
1776 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1778 // Num_S_Remainder = Quotient * Den
1779 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1781 // Remainder = Num - Num_S_Remainder
1782 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1784 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1785 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1786 DAG.getConstant(-1, DL, VT),
1787 DAG.getConstant(0, DL, VT),
1789 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1790 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1792 DAG.getConstant(-1, DL, VT),
1793 DAG.getConstant(0, DL, VT),
1795 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1796 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1799 // Calculate Division result:
1801 // Quotient_A_One = Quotient + 1
1802 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1803 DAG.getConstant(1, DL, VT));
1805 // Quotient_S_One = Quotient - 1
1806 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1807 DAG.getConstant(1, DL, VT));
1809 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1810 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1811 Quotient, Quotient_A_One, ISD::SETEQ);
1813 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1814 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1815 Quotient_S_One, Div, ISD::SETEQ);
1817 // Calculate Rem result:
1819 // Remainder_S_Den = Remainder - Den
1820 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1822 // Remainder_A_Den = Remainder + Den
1823 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1825 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1826 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, DL, VT),
1827 Remainder, Remainder_S_Den, ISD::SETEQ);
1829 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1830 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, DL, VT),
1831 Remainder_A_Den, Rem, ISD::SETEQ);
1836 return DAG.getMergeValues(Ops, DL);
1839 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1840 SelectionDAG &DAG) const {
1842 EVT VT = Op.getValueType();
1844 SDValue LHS = Op.getOperand(0);
1845 SDValue RHS = Op.getOperand(1);
1847 SDValue Zero = DAG.getConstant(0, DL, VT);
1848 SDValue NegOne = DAG.getConstant(-1, DL, VT);
1850 if (VT == MVT::i32 &&
1851 DAG.ComputeNumSignBits(LHS) > 8 &&
1852 DAG.ComputeNumSignBits(RHS) > 8) {
1853 return LowerDIVREM24(Op, DAG, true);
1855 if (VT == MVT::i64 &&
1856 DAG.ComputeNumSignBits(LHS) > 32 &&
1857 DAG.ComputeNumSignBits(RHS) > 32) {
1858 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1861 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, Zero);
1862 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, Zero);
1863 SDValue DIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(HalfVT, HalfVT),
1866 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(0)),
1867 DAG.getNode(ISD::SIGN_EXTEND, DL, VT, DIVREM.getValue(1))
1869 return DAG.getMergeValues(Res, DL);
1872 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1873 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1874 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1875 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1877 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1878 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1880 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1881 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1883 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1884 SDValue Rem = Div.getValue(1);
1886 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1887 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1889 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1890 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1896 return DAG.getMergeValues(Res, DL);
1899 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1900 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1902 EVT VT = Op.getValueType();
1903 SDValue X = Op.getOperand(0);
1904 SDValue Y = Op.getOperand(1);
1906 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1907 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1908 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1910 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1913 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1915 SDValue Src = Op.getOperand(0);
1917 // result = trunc(src)
1918 // if (src > 0.0 && src != result)
1921 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1923 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
1924 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f64);
1926 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1928 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1929 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1930 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1932 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1933 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1936 static SDValue extractF64Exponent(SDValue Hi, SDLoc SL, SelectionDAG &DAG) {
1937 const unsigned FractBits = 52;
1938 const unsigned ExpBits = 11;
1940 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1942 DAG.getConstant(FractBits - 32, SL, MVT::i32),
1943 DAG.getConstant(ExpBits, SL, MVT::i32));
1944 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1945 DAG.getConstant(1023, SL, MVT::i32));
1950 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1952 SDValue Src = Op.getOperand(0);
1954 assert(Op.getValueType() == MVT::f64);
1956 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
1957 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
1959 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1961 // Extract the upper half, since this is where we will find the sign and
1963 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1965 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
1967 const unsigned FractBits = 52;
1969 // Extract the sign bit.
1970 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, SL, MVT::i32);
1971 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1973 // Extend back to to 64-bits.
1974 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1976 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1978 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1979 const SDValue FractMask
1980 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, SL, MVT::i64);
1982 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1983 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1984 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1986 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1988 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, SL, MVT::i32);
1990 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1991 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1993 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1994 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1996 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1999 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
2001 SDValue Src = Op.getOperand(0);
2003 assert(Op.getValueType() == MVT::f64);
2005 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
2006 SDValue C1 = DAG.getConstantFP(C1Val, SL, MVT::f64);
2007 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
2009 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
2010 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
2012 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
2014 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
2015 SDValue C2 = DAG.getConstantFP(C2Val, SL, MVT::f64);
2017 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
2018 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
2020 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
2023 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
2024 // FNEARBYINT and FRINT are the same, except in their handling of FP
2025 // exceptions. Those aren't really meaningful for us, and OpenCL only has
2026 // rint, so just treat them as equivalent.
2027 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
2030 // XXX - May require not supporting f32 denormals?
2031 SDValue AMDGPUTargetLowering::LowerFROUND32(SDValue Op, SelectionDAG &DAG) const {
2033 SDValue X = Op.getOperand(0);
2035 SDValue T = DAG.getNode(ISD::FTRUNC, SL, MVT::f32, X);
2037 SDValue Diff = DAG.getNode(ISD::FSUB, SL, MVT::f32, X, T);
2039 SDValue AbsDiff = DAG.getNode(ISD::FABS, SL, MVT::f32, Diff);
2041 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f32);
2042 const SDValue One = DAG.getConstantFP(1.0, SL, MVT::f32);
2043 const SDValue Half = DAG.getConstantFP(0.5, SL, MVT::f32);
2045 SDValue SignOne = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f32, One, X);
2047 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f32);
2049 SDValue Cmp = DAG.getSetCC(SL, SetCCVT, AbsDiff, Half, ISD::SETOGE);
2051 SDValue Sel = DAG.getNode(ISD::SELECT, SL, MVT::f32, Cmp, SignOne, Zero);
2053 return DAG.getNode(ISD::FADD, SL, MVT::f32, T, Sel);
2056 SDValue AMDGPUTargetLowering::LowerFROUND64(SDValue Op, SelectionDAG &DAG) const {
2058 SDValue X = Op.getOperand(0);
2060 SDValue L = DAG.getNode(ISD::BITCAST, SL, MVT::i64, X);
2062 const SDValue Zero = DAG.getConstant(0, SL, MVT::i32);
2063 const SDValue One = DAG.getConstant(1, SL, MVT::i32);
2064 const SDValue NegOne = DAG.getConstant(-1, SL, MVT::i32);
2065 const SDValue FiftyOne = DAG.getConstant(51, SL, MVT::i32);
2066 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
2069 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, X);
2071 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC, One);
2073 SDValue Exp = extractF64Exponent(Hi, SL, DAG);
2075 const SDValue Mask = DAG.getConstant(INT64_C(0x000fffffffffffff), SL,
2078 SDValue M = DAG.getNode(ISD::SRA, SL, MVT::i64, Mask, Exp);
2079 SDValue D = DAG.getNode(ISD::SRA, SL, MVT::i64,
2080 DAG.getConstant(INT64_C(0x0008000000000000), SL,
2084 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, L, M);
2085 SDValue Tmp1 = DAG.getSetCC(SL, SetCCVT,
2086 DAG.getConstant(0, SL, MVT::i64), Tmp0,
2089 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, Tmp1,
2090 D, DAG.getConstant(0, SL, MVT::i64));
2091 SDValue K = DAG.getNode(ISD::ADD, SL, MVT::i64, L, Tmp2);
2093 K = DAG.getNode(ISD::AND, SL, MVT::i64, K, DAG.getNOT(SL, M, MVT::i64));
2094 K = DAG.getNode(ISD::BITCAST, SL, MVT::f64, K);
2096 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
2097 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
2098 SDValue ExpEqNegOne = DAG.getSetCC(SL, SetCCVT, NegOne, Exp, ISD::SETEQ);
2100 SDValue Mag = DAG.getNode(ISD::SELECT, SL, MVT::f64,
2102 DAG.getConstantFP(1.0, SL, MVT::f64),
2103 DAG.getConstantFP(0.0, SL, MVT::f64));
2105 SDValue S = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, Mag, X);
2107 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpLt0, S, K);
2108 K = DAG.getNode(ISD::SELECT, SL, MVT::f64, ExpGt51, X, K);
2113 SDValue AMDGPUTargetLowering::LowerFROUND(SDValue Op, SelectionDAG &DAG) const {
2114 EVT VT = Op.getValueType();
2117 return LowerFROUND32(Op, DAG);
2120 return LowerFROUND64(Op, DAG);
2122 llvm_unreachable("unhandled type");
2125 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2127 SDValue Src = Op.getOperand(0);
2129 // result = trunc(src);
2130 // if (src < 0.0 && src != result)
2133 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2135 const SDValue Zero = DAG.getConstantFP(0.0, SL, MVT::f64);
2136 const SDValue NegOne = DAG.getConstantFP(-1.0, SL, MVT::f64);
2138 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
2140 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2141 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2142 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2144 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2145 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2148 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2149 bool Signed) const {
2151 SDValue Src = Op.getOperand(0);
2153 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2155 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2156 DAG.getConstant(0, SL, MVT::i32));
2157 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2158 DAG.getConstant(1, SL, MVT::i32));
2160 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2163 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2165 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2166 DAG.getConstant(32, SL, MVT::i32));
2168 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2171 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2172 SelectionDAG &DAG) const {
2173 SDValue S0 = Op.getOperand(0);
2174 if (S0.getValueType() != MVT::i64)
2177 EVT DestVT = Op.getValueType();
2178 if (DestVT == MVT::f64)
2179 return LowerINT_TO_FP64(Op, DAG, false);
2181 assert(DestVT == MVT::f32);
2185 // f32 uint_to_fp i64
2186 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2187 DAG.getConstant(0, DL, MVT::i32));
2188 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2189 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2190 DAG.getConstant(1, DL, MVT::i32));
2191 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2192 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2193 DAG.getConstantFP(4294967296.0f, DL, MVT::f32)); // 2^32
2194 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2197 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2198 SelectionDAG &DAG) const {
2199 SDValue Src = Op.getOperand(0);
2200 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2201 return LowerINT_TO_FP64(Op, DAG, true);
2206 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2207 bool Signed) const {
2210 SDValue Src = Op.getOperand(0);
2212 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2214 SDValue K0 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), SL,
2216 SDValue K1 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), SL,
2219 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2221 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2224 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2226 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2227 MVT::i32, FloorMul);
2228 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2230 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2232 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2235 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2236 SelectionDAG &DAG) const {
2237 SDValue Src = Op.getOperand(0);
2239 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2240 return LowerFP64_TO_INT(Op, DAG, true);
2245 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2246 SelectionDAG &DAG) const {
2247 SDValue Src = Op.getOperand(0);
2249 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2250 return LowerFP64_TO_INT(Op, DAG, false);
2255 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2256 SelectionDAG &DAG) const {
2257 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2258 MVT VT = Op.getSimpleValueType();
2259 MVT ScalarVT = VT.getScalarType();
2264 SDValue Src = Op.getOperand(0);
2267 // TODO: Don't scalarize on Evergreen?
2268 unsigned NElts = VT.getVectorNumElements();
2269 SmallVector<SDValue, 8> Args;
2270 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2272 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2273 for (unsigned I = 0; I < NElts; ++I)
2274 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2276 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2279 //===----------------------------------------------------------------------===//
2280 // Custom DAG optimizations
2281 //===----------------------------------------------------------------------===//
2283 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2284 APInt KnownZero, KnownOne;
2285 EVT VT = Op.getValueType();
2286 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2288 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2291 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2292 EVT VT = Op.getValueType();
2294 // In order for this to be a signed 24-bit value, bit 23, must
2296 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2297 // as unsigned 24-bit values.
2298 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2301 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2303 SelectionDAG &DAG = DCI.DAG;
2304 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2305 EVT VT = Op.getValueType();
2307 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2308 APInt KnownZero, KnownOne;
2309 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2310 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2311 DCI.CommitTargetLoweringOpt(TLO);
2314 template <typename IntTy>
2315 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2316 uint32_t Offset, uint32_t Width, SDLoc DL) {
2317 if (Width + Offset < 32) {
2318 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2319 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2320 return DAG.getConstant(Result, DL, MVT::i32);
2323 return DAG.getConstant(Src0 >> Offset, DL, MVT::i32);
2326 static bool usesAllNormalStores(SDNode *LoadVal) {
2327 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2328 if (!ISD::isNormalStore(*I))
2335 // If we have a copy of an illegal type, replace it with a load / store of an
2336 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2337 // instructions emitted when handling extloads and truncstores. Ideally we could
2338 // recognize the pack / unpack pattern to eliminate it.
2339 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2340 DAGCombinerInfo &DCI) const {
2341 if (!DCI.isBeforeLegalize())
2344 StoreSDNode *SN = cast<StoreSDNode>(N);
2345 SDValue Value = SN->getValue();
2346 EVT VT = Value.getValueType();
2348 if (isTypeLegal(VT) || SN->isVolatile() ||
2349 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2352 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2353 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2356 EVT MemVT = LoadVal->getMemoryVT();
2359 SelectionDAG &DAG = DCI.DAG;
2360 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2362 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2364 LoadVal->getChain(),
2365 LoadVal->getBasePtr(),
2366 LoadVal->getOffset(),
2368 LoadVal->getMemOperand());
2370 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2371 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2373 return DAG.getStore(SN->getChain(), SL, NewLoad,
2374 SN->getBasePtr(), SN->getMemOperand());
2377 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2378 DAGCombinerInfo &DCI) const {
2379 EVT VT = N->getValueType(0);
2381 if (VT.isVector() || VT.getSizeInBits() > 32)
2384 SelectionDAG &DAG = DCI.DAG;
2387 SDValue N0 = N->getOperand(0);
2388 SDValue N1 = N->getOperand(1);
2391 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2392 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2393 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2394 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2395 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2396 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2397 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2398 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2403 // We need to use sext even for MUL_U24, because MUL_U24 is used
2404 // for signed multiply of 8 and 16-bit types.
2405 return DAG.getSExtOrTrunc(Mul, DL, VT);
2408 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2409 DAGCombinerInfo &DCI) const {
2410 SelectionDAG &DAG = DCI.DAG;
2413 switch(N->getOpcode()) {
2416 return performMulCombine(N, DCI);
2417 case AMDGPUISD::MUL_I24:
2418 case AMDGPUISD::MUL_U24: {
2419 SDValue N0 = N->getOperand(0);
2420 SDValue N1 = N->getOperand(1);
2421 simplifyI24(N0, DCI);
2422 simplifyI24(N1, DCI);
2426 SDValue Cond = N->getOperand(0);
2427 if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
2428 EVT VT = N->getValueType(0);
2429 SDValue LHS = Cond.getOperand(0);
2430 SDValue RHS = Cond.getOperand(1);
2431 SDValue CC = Cond.getOperand(2);
2433 SDValue True = N->getOperand(1);
2434 SDValue False = N->getOperand(2);
2437 return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
2439 // TODO: Implement min / max Evergreen instructions.
2440 if (VT == MVT::i32 &&
2441 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2442 return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2448 case AMDGPUISD::BFE_I32:
2449 case AMDGPUISD::BFE_U32: {
2450 assert(!N->getValueType(0).isVector() &&
2451 "Vector handling of BFE not implemented");
2452 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2456 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2458 return DAG.getConstant(0, DL, MVT::i32);
2460 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2464 SDValue BitsFrom = N->getOperand(0);
2465 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2467 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2469 if (OffsetVal == 0) {
2470 // This is already sign / zero extended, so try to fold away extra BFEs.
2471 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2473 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2474 if (OpSignBits >= SignBits)
2477 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2479 // This is a sign_extend_inreg. Replace it to take advantage of existing
2480 // DAG Combines. If not eliminated, we will match back to BFE during
2483 // TODO: The sext_inreg of extended types ends, although we can could
2484 // handle them in a single BFE.
2485 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2486 DAG.getValueType(SmallVT));
2489 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2492 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2494 return constantFoldBFE<int32_t>(DAG,
2495 CVal->getSExtValue(),
2501 return constantFoldBFE<uint32_t>(DAG,
2502 CVal->getZExtValue(),
2508 if ((OffsetVal + WidthVal) >= 32) {
2509 SDValue ShiftVal = DAG.getConstant(OffsetVal, DL, MVT::i32);
2510 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2511 BitsFrom, ShiftVal);
2514 if (BitsFrom.hasOneUse()) {
2515 APInt Demanded = APInt::getBitsSet(32,
2517 OffsetVal + WidthVal);
2519 APInt KnownZero, KnownOne;
2520 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2521 !DCI.isBeforeLegalizeOps());
2522 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2523 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2524 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2525 KnownZero, KnownOne, TLO)) {
2526 DCI.CommitTargetLoweringOpt(TLO);
2534 return performStoreCombine(N, DCI);
2539 //===----------------------------------------------------------------------===//
2541 //===----------------------------------------------------------------------===//
2543 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2546 const SmallVectorImpl<ISD::InputArg> &Ins,
2547 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2549 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2550 if (Ins[i].ArgVT == Ins[i].VT) {
2551 OrigIns.push_back(Ins[i]);
2556 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2557 // Vector has been split into scalars.
2558 VT = Ins[i].ArgVT.getVectorElementType();
2559 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2560 Ins[i].ArgVT.getVectorElementType() !=
2561 Ins[i].VT.getVectorElementType()) {
2562 // Vector elements have been promoted
2565 // Vector has been spilt into smaller vectors.
2569 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2570 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2571 OrigIns.push_back(Arg);
2575 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2576 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2577 return CFP->isExactlyValue(1.0);
2579 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2580 return C->isAllOnesValue();
2585 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2586 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2587 return CFP->getValueAPF().isZero();
2589 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2590 return C->isNullValue();
2595 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2596 const TargetRegisterClass *RC,
2597 unsigned Reg, EVT VT) const {
2598 MachineFunction &MF = DAG.getMachineFunction();
2599 MachineRegisterInfo &MRI = MF.getRegInfo();
2600 unsigned VirtualRegister;
2601 if (!MRI.isLiveIn(Reg)) {
2602 VirtualRegister = MRI.createVirtualRegister(RC);
2603 MRI.addLiveIn(Reg, VirtualRegister);
2605 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2607 return DAG.getRegister(VirtualRegister, VT);
2610 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2612 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2613 switch ((AMDGPUISD::NodeType)Opcode) {
2614 case AMDGPUISD::FIRST_NUMBER: break;
2616 NODE_NAME_CASE(CALL);
2617 NODE_NAME_CASE(UMUL);
2618 NODE_NAME_CASE(RET_FLAG);
2619 NODE_NAME_CASE(BRANCH_COND);
2622 NODE_NAME_CASE(DWORDADDR)
2623 NODE_NAME_CASE(FRACT)
2624 NODE_NAME_CASE(CLAMP)
2625 NODE_NAME_CASE(COS_HW)
2626 NODE_NAME_CASE(SIN_HW)
2627 NODE_NAME_CASE(FMAX_LEGACY)
2628 NODE_NAME_CASE(SMAX)
2629 NODE_NAME_CASE(UMAX)
2630 NODE_NAME_CASE(FMIN_LEGACY)
2631 NODE_NAME_CASE(SMIN)
2632 NODE_NAME_CASE(UMIN)
2633 NODE_NAME_CASE(FMAX3)
2634 NODE_NAME_CASE(SMAX3)
2635 NODE_NAME_CASE(UMAX3)
2636 NODE_NAME_CASE(FMIN3)
2637 NODE_NAME_CASE(SMIN3)
2638 NODE_NAME_CASE(UMIN3)
2639 NODE_NAME_CASE(URECIP)
2640 NODE_NAME_CASE(DIV_SCALE)
2641 NODE_NAME_CASE(DIV_FMAS)
2642 NODE_NAME_CASE(DIV_FIXUP)
2643 NODE_NAME_CASE(TRIG_PREOP)
2646 NODE_NAME_CASE(RSQ_LEGACY)
2647 NODE_NAME_CASE(RSQ_CLAMPED)
2648 NODE_NAME_CASE(LDEXP)
2649 NODE_NAME_CASE(FP_CLASS)
2650 NODE_NAME_CASE(DOT4)
2651 NODE_NAME_CASE(CARRY)
2652 NODE_NAME_CASE(BORROW)
2653 NODE_NAME_CASE(BFE_U32)
2654 NODE_NAME_CASE(BFE_I32)
2657 NODE_NAME_CASE(BREV)
2658 NODE_NAME_CASE(MUL_U24)
2659 NODE_NAME_CASE(MUL_I24)
2660 NODE_NAME_CASE(MAD_U24)
2661 NODE_NAME_CASE(MAD_I24)
2662 NODE_NAME_CASE(TEXTURE_FETCH)
2663 NODE_NAME_CASE(EXPORT)
2664 NODE_NAME_CASE(CONST_ADDRESS)
2665 NODE_NAME_CASE(REGISTER_LOAD)
2666 NODE_NAME_CASE(REGISTER_STORE)
2667 NODE_NAME_CASE(LOAD_CONSTANT)
2668 NODE_NAME_CASE(LOAD_INPUT)
2669 NODE_NAME_CASE(SAMPLE)
2670 NODE_NAME_CASE(SAMPLEB)
2671 NODE_NAME_CASE(SAMPLED)
2672 NODE_NAME_CASE(SAMPLEL)
2673 NODE_NAME_CASE(CVT_F32_UBYTE0)
2674 NODE_NAME_CASE(CVT_F32_UBYTE1)
2675 NODE_NAME_CASE(CVT_F32_UBYTE2)
2676 NODE_NAME_CASE(CVT_F32_UBYTE3)
2677 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2678 NODE_NAME_CASE(CONST_DATA_PTR)
2679 case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
2680 NODE_NAME_CASE(SENDMSG)
2681 NODE_NAME_CASE(STORE_MSKOR)
2682 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2683 case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
2688 SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
2689 DAGCombinerInfo &DCI,
2690 unsigned &RefinementSteps,
2691 bool &UseOneConstNR) const {
2692 SelectionDAG &DAG = DCI.DAG;
2693 EVT VT = Operand.getValueType();
2695 if (VT == MVT::f32) {
2696 RefinementSteps = 0;
2697 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
2700 // TODO: There is also f64 rsq instruction, but the documentation is less
2701 // clear on its precision.
2706 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
2707 DAGCombinerInfo &DCI,
2708 unsigned &RefinementSteps) const {
2709 SelectionDAG &DAG = DCI.DAG;
2710 EVT VT = Operand.getValueType();
2712 if (VT == MVT::f32) {
2713 // Reciprocal, < 1 ulp error.
2715 // This reciprocal approximation converges to < 0.5 ulp error with one
2716 // newton rhapson performed with two fused multiple adds (FMAs).
2718 RefinementSteps = 0;
2719 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
2722 // TODO: There is also f64 rcp instruction, but the documentation is less
2723 // clear on its precision.
2728 static void computeKnownBitsForMinMax(const SDValue Op0,
2732 const SelectionDAG &DAG,
2734 APInt Op0Zero, Op0One;
2735 APInt Op1Zero, Op1One;
2736 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2737 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2739 KnownZero = Op0Zero & Op1Zero;
2740 KnownOne = Op0One & Op1One;
2743 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2747 const SelectionDAG &DAG,
2748 unsigned Depth) const {
2750 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2754 unsigned Opc = Op.getOpcode();
2759 case ISD::INTRINSIC_WO_CHAIN: {
2760 // FIXME: The intrinsic should just use the node.
2761 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2762 case AMDGPUIntrinsic::AMDGPU_imax:
2763 case AMDGPUIntrinsic::AMDGPU_umax:
2764 case AMDGPUIntrinsic::AMDGPU_imin:
2765 case AMDGPUIntrinsic::AMDGPU_umin:
2766 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2767 KnownZero, KnownOne, DAG, Depth);
2775 case AMDGPUISD::SMAX:
2776 case AMDGPUISD::UMAX:
2777 case AMDGPUISD::SMIN:
2778 case AMDGPUISD::UMIN:
2779 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2780 KnownZero, KnownOne, DAG, Depth);
2783 case AMDGPUISD::CARRY:
2784 case AMDGPUISD::BORROW: {
2785 KnownZero = APInt::getHighBitsSet(32, 31);
2789 case AMDGPUISD::BFE_I32:
2790 case AMDGPUISD::BFE_U32: {
2791 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2795 unsigned BitWidth = 32;
2796 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2798 if (Opc == AMDGPUISD::BFE_U32)
2799 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2806 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2808 const SelectionDAG &DAG,
2809 unsigned Depth) const {
2810 switch (Op.getOpcode()) {
2811 case AMDGPUISD::BFE_I32: {
2812 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2816 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2817 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2818 if (!Offset || !Offset->isNullValue())
2821 // TODO: Could probably figure something out with non-0 offsets.
2822 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2823 return std::max(SignBits, Op0SignBits);
2826 case AMDGPUISD::BFE_U32: {
2827 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2828 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
2831 case AMDGPUISD::CARRY:
2832 case AMDGPUISD::BORROW: