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::FROUND, MVT::f32, Legal);
131 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
133 setOperationAction(ISD::FREM, MVT::f32, Custom);
134 setOperationAction(ISD::FREM, MVT::f64, Custom);
136 // Lower floating point store/load to integer store/load to reduce the number
137 // of patterns in tablegen.
138 setOperationAction(ISD::STORE, MVT::f32, Promote);
139 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
141 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
142 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
144 setOperationAction(ISD::STORE, MVT::i64, Promote);
145 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
147 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
148 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
150 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
151 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
153 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
154 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
156 setOperationAction(ISD::STORE, MVT::f64, Promote);
157 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
159 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
160 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
162 // Custom lowering of vector stores is required for local address space
164 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
165 // XXX: Native v2i32 local address space stores are possible, but not
166 // currently implemented.
167 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
169 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
170 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
171 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
173 // XXX: This can be change to Custom, once ExpandVectorStores can
174 // handle 64-bit stores.
175 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
177 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
178 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
179 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
180 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
181 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
184 setOperationAction(ISD::LOAD, MVT::f32, Promote);
185 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
187 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
188 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
190 setOperationAction(ISD::LOAD, MVT::i64, Promote);
191 AddPromotedToType(ISD::LOAD, MVT::i64, 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 for (MVT VT : MVT::integer_vector_valuetypes()) {
220 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Expand);
221 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Expand);
222 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i8, Expand);
223 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Expand);
224 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Expand);
225 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i8, Expand);
226 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Expand);
227 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Expand);
228 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v2i16, Expand);
229 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Expand);
230 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Expand);
231 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::v4i16, Expand);
234 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
236 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
237 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
238 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
239 setOperationAction(ISD::FRINT, MVT::f64, Custom);
240 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
243 if (!Subtarget->hasBFI()) {
244 // fcopysign can be done in a single instruction with BFI.
245 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
246 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
249 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
251 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
252 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
253 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
254 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
256 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
257 for (MVT VT : ScalarIntVTs) {
258 setOperationAction(ISD::SREM, VT, Expand);
259 setOperationAction(ISD::SDIV, VT, Expand);
261 // GPU does not have divrem function for signed or unsigned.
262 setOperationAction(ISD::SDIVREM, VT, Custom);
263 setOperationAction(ISD::UDIVREM, VT, Custom);
265 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
266 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
267 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
269 setOperationAction(ISD::BSWAP, VT, Expand);
270 setOperationAction(ISD::CTTZ, VT, Expand);
271 setOperationAction(ISD::CTLZ, VT, Expand);
274 if (!Subtarget->hasBCNT(32))
275 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
277 if (!Subtarget->hasBCNT(64))
278 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
280 // The hardware supports 32-bit ROTR, but not ROTL.
281 setOperationAction(ISD::ROTL, MVT::i32, Expand);
282 setOperationAction(ISD::ROTL, MVT::i64, Expand);
283 setOperationAction(ISD::ROTR, MVT::i64, Expand);
285 setOperationAction(ISD::MUL, MVT::i64, Expand);
286 setOperationAction(ISD::MULHU, MVT::i64, Expand);
287 setOperationAction(ISD::MULHS, MVT::i64, Expand);
288 setOperationAction(ISD::UDIV, MVT::i32, Expand);
289 setOperationAction(ISD::UREM, MVT::i32, Expand);
290 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
291 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
292 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
293 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
294 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
296 if (!Subtarget->hasFFBH())
297 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
299 if (!Subtarget->hasFFBL())
300 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
302 static const MVT::SimpleValueType VectorIntTypes[] = {
303 MVT::v2i32, MVT::v4i32
306 for (MVT VT : VectorIntTypes) {
307 // Expand the following operations for the current type by default.
308 setOperationAction(ISD::ADD, VT, Expand);
309 setOperationAction(ISD::AND, VT, Expand);
310 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
311 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
312 setOperationAction(ISD::MUL, VT, Expand);
313 setOperationAction(ISD::OR, VT, Expand);
314 setOperationAction(ISD::SHL, VT, Expand);
315 setOperationAction(ISD::SRA, VT, Expand);
316 setOperationAction(ISD::SRL, VT, Expand);
317 setOperationAction(ISD::ROTL, VT, Expand);
318 setOperationAction(ISD::ROTR, VT, Expand);
319 setOperationAction(ISD::SUB, VT, Expand);
320 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
321 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
322 setOperationAction(ISD::SDIV, VT, Expand);
323 setOperationAction(ISD::UDIV, VT, Expand);
324 setOperationAction(ISD::SREM, VT, Expand);
325 setOperationAction(ISD::UREM, VT, Expand);
326 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
327 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
328 setOperationAction(ISD::SDIVREM, VT, Custom);
329 setOperationAction(ISD::UDIVREM, VT, Custom);
330 setOperationAction(ISD::ADDC, VT, Expand);
331 setOperationAction(ISD::SUBC, VT, Expand);
332 setOperationAction(ISD::ADDE, VT, Expand);
333 setOperationAction(ISD::SUBE, VT, Expand);
334 setOperationAction(ISD::SELECT, VT, Expand);
335 setOperationAction(ISD::VSELECT, VT, Expand);
336 setOperationAction(ISD::SELECT_CC, VT, Expand);
337 setOperationAction(ISD::XOR, VT, Expand);
338 setOperationAction(ISD::BSWAP, VT, Expand);
339 setOperationAction(ISD::CTPOP, VT, Expand);
340 setOperationAction(ISD::CTTZ, VT, Expand);
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
342 setOperationAction(ISD::CTLZ, VT, Expand);
343 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
344 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
347 static const MVT::SimpleValueType FloatVectorTypes[] = {
348 MVT::v2f32, MVT::v4f32
351 for (MVT VT : FloatVectorTypes) {
352 setOperationAction(ISD::FABS, VT, Expand);
353 setOperationAction(ISD::FMINNUM, VT, Expand);
354 setOperationAction(ISD::FMAXNUM, VT, Expand);
355 setOperationAction(ISD::FADD, VT, Expand);
356 setOperationAction(ISD::FCEIL, VT, Expand);
357 setOperationAction(ISD::FCOS, VT, Expand);
358 setOperationAction(ISD::FDIV, VT, Expand);
359 setOperationAction(ISD::FEXP2, VT, Expand);
360 setOperationAction(ISD::FLOG2, VT, Expand);
361 setOperationAction(ISD::FREM, VT, Expand);
362 setOperationAction(ISD::FPOW, VT, Expand);
363 setOperationAction(ISD::FFLOOR, VT, Expand);
364 setOperationAction(ISD::FTRUNC, VT, Expand);
365 setOperationAction(ISD::FMUL, VT, Expand);
366 setOperationAction(ISD::FMA, VT, Expand);
367 setOperationAction(ISD::FRINT, VT, Expand);
368 setOperationAction(ISD::FNEARBYINT, VT, Expand);
369 setOperationAction(ISD::FSQRT, VT, Expand);
370 setOperationAction(ISD::FSIN, VT, Expand);
371 setOperationAction(ISD::FSUB, VT, Expand);
372 setOperationAction(ISD::FNEG, VT, Expand);
373 setOperationAction(ISD::SELECT, VT, Expand);
374 setOperationAction(ISD::VSELECT, VT, Expand);
375 setOperationAction(ISD::SELECT_CC, VT, Expand);
376 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
377 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
380 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
381 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
383 setTargetDAGCombine(ISD::MUL);
384 setTargetDAGCombine(ISD::SELECT);
385 setTargetDAGCombine(ISD::SELECT_CC);
386 setTargetDAGCombine(ISD::STORE);
388 setBooleanContents(ZeroOrNegativeOneBooleanContent);
389 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
391 setSchedulingPreference(Sched::RegPressure);
392 setJumpIsExpensive(true);
394 // SI at least has hardware support for floating point exceptions, but no way
395 // of using or handling them is implemented. They are also optional in OpenCL
397 setHasFloatingPointExceptions(false);
399 setSelectIsExpensive(false);
400 PredictableSelectIsExpensive = false;
402 // There are no integer divide instructions, and these expand to a pretty
403 // large sequence of instructions.
404 setIntDivIsCheap(false);
405 setPow2SDivIsCheap(false);
406 setFsqrtIsCheap(true);
408 // FIXME: Need to really handle these.
409 MaxStoresPerMemcpy = 4096;
410 MaxStoresPerMemmove = 4096;
411 MaxStoresPerMemset = 4096;
414 //===----------------------------------------------------------------------===//
415 // Target Information
416 //===----------------------------------------------------------------------===//
418 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
422 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
426 // The backend supports 32 and 64 bit floating point immediates.
427 // FIXME: Why are we reporting vectors of FP immediates as legal?
428 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
429 EVT ScalarVT = VT.getScalarType();
430 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
433 // We don't want to shrink f64 / f32 constants.
434 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
435 EVT ScalarVT = VT.getScalarType();
436 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
439 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
443 unsigned NewSize = NewVT.getStoreSizeInBits();
445 // If we are reducing to a 32-bit load, this is always better.
449 EVT OldVT = N->getValueType(0);
450 unsigned OldSize = OldVT.getStoreSizeInBits();
452 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
453 // extloads, so doing one requires using a buffer_load. In cases where we
454 // still couldn't use a scalar load, using the wider load shouldn't really
457 // If the old size already had to be an extload, there's no harm in continuing
458 // to reduce the width.
459 return (OldSize < 32);
462 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
464 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
467 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
468 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
470 return ((LScalarSize <= CastScalarSize) ||
471 (CastScalarSize >= 32) ||
475 // SI+ has instructions for cttz / ctlz for 32-bit values. This is probably also
476 // profitable with the expansion for 64-bit since it's generally good to
478 // FIXME: These should really have the size as a parameter.
479 bool AMDGPUTargetLowering::isCheapToSpeculateCttz() const {
483 bool AMDGPUTargetLowering::isCheapToSpeculateCtlz() const {
487 //===---------------------------------------------------------------------===//
489 //===---------------------------------------------------------------------===//
491 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
492 assert(VT.isFloatingPoint());
493 return VT == MVT::f32 || VT == MVT::f64;
496 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
497 assert(VT.isFloatingPoint());
498 return VT == MVT::f32 || VT == MVT::f64;
501 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
502 // Truncate is just accessing a subregister.
503 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
506 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
507 // Truncate is just accessing a subregister.
508 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
509 (Dest->getPrimitiveSizeInBits() % 32 == 0);
512 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
513 const DataLayout *DL = getDataLayout();
514 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
515 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
517 return SrcSize == 32 && DestSize == 64;
520 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
521 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
522 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
523 // this will enable reducing 64-bit operations the 32-bit, which is always
525 return Src == MVT::i32 && Dest == MVT::i64;
528 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
529 return isZExtFree(Val.getValueType(), VT2);
532 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
533 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
534 // limited number of native 64-bit operations. Shrinking an operation to fit
535 // in a single 32-bit register should always be helpful. As currently used,
536 // this is much less general than the name suggests, and is only used in
537 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
538 // not profitable, and may actually be harmful.
539 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
542 //===---------------------------------------------------------------------===//
543 // TargetLowering Callbacks
544 //===---------------------------------------------------------------------===//
546 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
547 const SmallVectorImpl<ISD::InputArg> &Ins) const {
549 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
552 SDValue AMDGPUTargetLowering::LowerReturn(
554 CallingConv::ID CallConv,
556 const SmallVectorImpl<ISD::OutputArg> &Outs,
557 const SmallVectorImpl<SDValue> &OutVals,
558 SDLoc DL, SelectionDAG &DAG) const {
559 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
562 //===---------------------------------------------------------------------===//
563 // Target specific lowering
564 //===---------------------------------------------------------------------===//
566 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
567 SmallVectorImpl<SDValue> &InVals) const {
568 SDValue Callee = CLI.Callee;
569 SelectionDAG &DAG = CLI.DAG;
571 const Function &Fn = *DAG.getMachineFunction().getFunction();
573 StringRef FuncName("<unknown>");
575 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
576 FuncName = G->getSymbol();
577 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
578 FuncName = G->getGlobal()->getName();
580 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
581 DAG.getContext()->diagnose(NoCalls);
585 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
586 SelectionDAG &DAG) const {
587 switch (Op.getOpcode()) {
589 Op.getNode()->dump();
590 llvm_unreachable("Custom lowering code for this"
591 "instruction is not implemented yet!");
593 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
594 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
595 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
596 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
597 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
598 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
599 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
600 case ISD::FREM: return LowerFREM(Op, DAG);
601 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
602 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
603 case ISD::FRINT: return LowerFRINT(Op, DAG);
604 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
605 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
606 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
607 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
608 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
609 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
614 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
615 SmallVectorImpl<SDValue> &Results,
616 SelectionDAG &DAG) const {
617 switch (N->getOpcode()) {
618 case ISD::SIGN_EXTEND_INREG:
619 // Different parts of legalization seem to interpret which type of
620 // sign_extend_inreg is the one to check for custom lowering. The extended
621 // from type is what really matters, but some places check for custom
622 // lowering of the result type. This results in trying to use
623 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
624 // nothing here and let the illegal result integer be handled normally.
627 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
631 Results.push_back(SDValue(Node, 0));
632 Results.push_back(SDValue(Node, 1));
633 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
635 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
639 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
640 if (Lowered.getNode())
641 Results.push_back(Lowered);
649 // FIXME: This implements accesses to initialized globals in the constant
650 // address space by copying them to private and accessing that. It does not
651 // properly handle illegal types or vectors. The private vector loads are not
652 // scalarized, and the illegal scalars hit an assertion. This technique will not
653 // work well with large initializers, and this should eventually be
654 // removed. Initialized globals should be placed into a data section that the
655 // runtime will load into a buffer before the kernel is executed. Uses of the
656 // global need to be replaced with a pointer loaded from an implicit kernel
657 // argument into this buffer holding the copy of the data, which will remove the
658 // need for any of this.
659 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
660 const GlobalValue *GV,
661 const SDValue &InitPtr,
663 SelectionDAG &DAG) const {
664 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
666 Type *InitTy = Init->getType();
668 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
669 EVT VT = EVT::getEVT(InitTy);
670 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
671 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
672 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
673 TD->getPrefTypeAlignment(InitTy));
676 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
677 EVT VT = EVT::getEVT(CFP->getType());
678 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
679 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
680 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
681 TD->getPrefTypeAlignment(CFP->getType()));
684 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
685 const StructLayout *SL = TD->getStructLayout(ST);
687 EVT PtrVT = InitPtr.getValueType();
688 SmallVector<SDValue, 8> Chains;
690 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
691 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
692 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
694 Constant *Elt = Init->getAggregateElement(I);
695 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
698 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
701 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
702 EVT PtrVT = InitPtr.getValueType();
704 unsigned NumElements;
705 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
706 NumElements = AT->getNumElements();
707 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
708 NumElements = VT->getNumElements();
710 llvm_unreachable("Unexpected type");
712 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
713 SmallVector<SDValue, 8> Chains;
714 for (unsigned i = 0; i < NumElements; ++i) {
715 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
716 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
718 Constant *Elt = Init->getAggregateElement(i);
719 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
722 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
725 if (isa<UndefValue>(Init)) {
726 EVT VT = EVT::getEVT(InitTy);
727 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
728 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
729 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
730 TD->getPrefTypeAlignment(InitTy));
734 llvm_unreachable("Unhandled constant initializer");
737 static bool hasDefinedInitializer(const GlobalValue *GV) {
738 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
739 if (!GVar || !GVar->hasInitializer())
742 if (isa<UndefValue>(GVar->getInitializer()))
748 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
750 SelectionDAG &DAG) const {
752 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
753 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
754 const GlobalValue *GV = G->getGlobal();
756 switch (G->getAddressSpace()) {
757 case AMDGPUAS::LOCAL_ADDRESS: {
758 // XXX: What does the value of G->getOffset() mean?
759 assert(G->getOffset() == 0 &&
760 "Do not know what to do with an non-zero offset");
762 // TODO: We could emit code to handle the initialization somewhere.
763 if (hasDefinedInitializer(GV))
767 if (MFI->LocalMemoryObjects.count(GV) == 0) {
768 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
769 Offset = MFI->LDSSize;
770 MFI->LocalMemoryObjects[GV] = Offset;
771 // XXX: Account for alignment?
772 MFI->LDSSize += Size;
774 Offset = MFI->LocalMemoryObjects[GV];
777 return DAG.getConstant(Offset, getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
779 case AMDGPUAS::CONSTANT_ADDRESS: {
780 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
781 Type *EltType = GV->getType()->getElementType();
782 unsigned Size = TD->getTypeAllocSize(EltType);
783 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
785 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
786 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
788 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
789 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
791 const GlobalVariable *Var = cast<GlobalVariable>(GV);
792 if (!Var->hasInitializer()) {
793 // This has no use, but bugpoint will hit it.
794 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
797 const Constant *Init = Var->getInitializer();
798 SmallVector<SDNode*, 8> WorkList;
800 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
801 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
802 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
804 WorkList.push_back(*I);
806 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
807 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
808 E = WorkList.end(); I != E; ++I) {
809 SmallVector<SDValue, 8> Ops;
810 Ops.push_back(Chain);
811 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
812 Ops.push_back((*I)->getOperand(i));
814 DAG.UpdateNodeOperands(*I, Ops);
816 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
820 const Function &Fn = *DAG.getMachineFunction().getFunction();
821 DiagnosticInfoUnsupported BadInit(Fn,
822 "initializer for address space");
823 DAG.getContext()->diagnose(BadInit);
827 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
828 SelectionDAG &DAG) const {
829 SmallVector<SDValue, 8> Args;
830 SDValue A = Op.getOperand(0);
831 SDValue B = Op.getOperand(1);
833 DAG.ExtractVectorElements(A, Args);
834 DAG.ExtractVectorElements(B, Args);
836 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
839 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
840 SelectionDAG &DAG) const {
842 SmallVector<SDValue, 8> Args;
843 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
844 EVT VT = Op.getValueType();
845 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
846 VT.getVectorNumElements());
848 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
851 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
852 SelectionDAG &DAG) const {
854 MachineFunction &MF = DAG.getMachineFunction();
855 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
856 getTargetMachine().getSubtargetImpl()->getFrameLowering());
858 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
860 unsigned FrameIndex = FIN->getIndex();
861 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
862 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
866 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
867 SelectionDAG &DAG) const {
868 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
870 EVT VT = Op.getValueType();
872 switch (IntrinsicID) {
874 case AMDGPUIntrinsic::AMDGPU_abs:
875 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
876 return LowerIntrinsicIABS(Op, DAG);
877 case AMDGPUIntrinsic::AMDGPU_lrp:
878 return LowerIntrinsicLRP(Op, DAG);
879 case AMDGPUIntrinsic::AMDGPU_fract:
880 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
881 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
883 case AMDGPUIntrinsic::AMDGPU_clamp:
884 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
885 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
886 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
888 case Intrinsic::AMDGPU_div_scale: {
889 // 3rd parameter required to be a constant.
890 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
892 return DAG.getUNDEF(VT);
894 // Translate to the operands expected by the machine instruction. The
895 // first parameter must be the same as the first instruction.
896 SDValue Numerator = Op.getOperand(1);
897 SDValue Denominator = Op.getOperand(2);
899 // Note this order is opposite of the machine instruction's operations,
900 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
901 // intrinsic has the numerator as the first operand to match a normal
902 // division operation.
904 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
906 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
907 Denominator, Numerator);
910 case Intrinsic::AMDGPU_div_fmas:
911 // FIXME: Dropping bool parameter. Work is needed to support the implicit
913 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
914 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
916 case Intrinsic::AMDGPU_div_fixup:
917 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
918 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
920 case Intrinsic::AMDGPU_trig_preop:
921 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
922 Op.getOperand(1), Op.getOperand(2));
924 case Intrinsic::AMDGPU_rcp:
925 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
927 case Intrinsic::AMDGPU_rsq:
928 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
930 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
931 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
933 case Intrinsic::AMDGPU_rsq_clamped:
934 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
935 Type *Type = VT.getTypeForEVT(*DAG.getContext());
936 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
937 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
939 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
940 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
941 DAG.getConstantFP(Max, VT));
942 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
943 DAG.getConstantFP(Min, VT));
945 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
948 case Intrinsic::AMDGPU_ldexp:
949 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
952 case AMDGPUIntrinsic::AMDGPU_imax:
953 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
955 case AMDGPUIntrinsic::AMDGPU_umax:
956 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
958 case AMDGPUIntrinsic::AMDGPU_imin:
959 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
961 case AMDGPUIntrinsic::AMDGPU_umin:
962 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
965 case AMDGPUIntrinsic::AMDGPU_umul24:
966 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
967 Op.getOperand(1), Op.getOperand(2));
969 case AMDGPUIntrinsic::AMDGPU_imul24:
970 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
971 Op.getOperand(1), Op.getOperand(2));
973 case AMDGPUIntrinsic::AMDGPU_umad24:
974 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
975 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
977 case AMDGPUIntrinsic::AMDGPU_imad24:
978 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
979 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
981 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
982 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
984 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
985 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
987 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
988 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
990 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
991 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
993 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
994 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
999 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
1000 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
1005 case AMDGPUIntrinsic::AMDGPU_bfi:
1006 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
1011 case AMDGPUIntrinsic::AMDGPU_bfm:
1012 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1016 case AMDGPUIntrinsic::AMDGPU_brev:
1017 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
1019 case Intrinsic::AMDGPU_class:
1020 return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
1021 Op.getOperand(1), Op.getOperand(2));
1023 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1024 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1026 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1027 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1028 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1029 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1033 ///IABS(a) = SMAX(sub(0, a), a)
1034 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1035 SelectionDAG &DAG) const {
1037 EVT VT = Op.getValueType();
1038 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1041 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1044 /// Linear Interpolation
1045 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
1046 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1047 SelectionDAG &DAG) const {
1049 EVT VT = Op.getValueType();
1050 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1051 DAG.getConstantFP(1.0f, MVT::f32),
1053 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1055 return DAG.getNode(ISD::FADD, DL, VT,
1056 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1060 /// \brief Generate Min/Max node
1061 SDValue AMDGPUTargetLowering::CombineFMinMaxLegacy(SDLoc DL,
1068 DAGCombinerInfo &DCI) const {
1069 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1072 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1075 SelectionDAG &DAG = DCI.DAG;
1076 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1085 case ISD::SETFALSE2:
1094 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1095 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1101 // Ordered. Assume ordered for undefined.
1103 // Only do this after legalization to avoid interfering with other combines
1104 // which might occur.
1105 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1106 !DCI.isCalledByLegalizer())
1109 // We need to permute the operands to get the correct NaN behavior. The
1110 // selected operand is the second one based on the failing compare with NaN,
1111 // so permute it based on the compare type the hardware uses.
1113 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1114 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1119 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1120 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1126 if (DCI.getDAGCombineLevel() < AfterLegalizeDAG &&
1127 !DCI.isCalledByLegalizer())
1131 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1132 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1134 case ISD::SETCC_INVALID:
1135 llvm_unreachable("Invalid setcc condcode!");
1140 /// \brief Generate Min/Max node
1141 SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1148 SelectionDAG &DAG) const {
1149 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1152 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1156 unsigned Opc = (LHS == True) ? AMDGPUISD::UMIN : AMDGPUISD::UMAX;
1157 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1161 unsigned Opc = (LHS == True) ? AMDGPUISD::SMIN : AMDGPUISD::SMAX;
1162 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1166 unsigned Opc = (LHS == True) ? AMDGPUISD::SMAX : AMDGPUISD::SMIN;
1167 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1171 unsigned Opc = (LHS == True) ? AMDGPUISD::UMAX : AMDGPUISD::UMIN;
1172 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1179 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1180 SelectionDAG &DAG) const {
1181 LoadSDNode *Load = cast<LoadSDNode>(Op);
1182 EVT MemVT = Load->getMemoryVT();
1183 EVT MemEltVT = MemVT.getVectorElementType();
1185 EVT LoadVT = Op.getValueType();
1186 EVT EltVT = LoadVT.getVectorElementType();
1187 EVT PtrVT = Load->getBasePtr().getValueType();
1189 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1190 SmallVector<SDValue, 8> Loads;
1191 SmallVector<SDValue, 8> Chains;
1194 unsigned MemEltSize = MemEltVT.getStoreSize();
1195 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1197 for (unsigned i = 0; i < NumElts; ++i) {
1198 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1199 DAG.getConstant(i * MemEltSize, PtrVT));
1202 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1203 Load->getChain(), Ptr,
1204 SrcValue.getWithOffset(i * MemEltSize),
1205 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1206 Load->isInvariant(), Load->getAlignment());
1207 Loads.push_back(NewLoad.getValue(0));
1208 Chains.push_back(NewLoad.getValue(1));
1212 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1213 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1216 return DAG.getMergeValues(Ops, SL);
1219 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1220 SelectionDAG &DAG) const {
1221 EVT VT = Op.getValueType();
1223 // If this is a 2 element vector, we really want to scalarize and not create
1224 // weird 1 element vectors.
1225 if (VT.getVectorNumElements() == 2)
1226 return ScalarizeVectorLoad(Op, DAG);
1228 LoadSDNode *Load = cast<LoadSDNode>(Op);
1229 SDValue BasePtr = Load->getBasePtr();
1230 EVT PtrVT = BasePtr.getValueType();
1231 EVT MemVT = Load->getMemoryVT();
1233 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1236 EVT LoMemVT, HiMemVT;
1239 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1240 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1241 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1243 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1244 Load->getChain(), BasePtr,
1246 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1247 Load->isInvariant(), Load->getAlignment());
1249 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1250 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1253 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1254 Load->getChain(), HiPtr,
1255 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1256 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1257 Load->isInvariant(), Load->getAlignment());
1260 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1261 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1262 LoLoad.getValue(1), HiLoad.getValue(1))
1265 return DAG.getMergeValues(Ops, SL);
1268 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1269 SelectionDAG &DAG) const {
1270 StoreSDNode *Store = cast<StoreSDNode>(Op);
1271 EVT MemVT = Store->getMemoryVT();
1272 unsigned MemBits = MemVT.getSizeInBits();
1274 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1275 // truncating store into an i32 store.
1276 // XXX: We could also handle optimize other vector bitwidths.
1277 if (!MemVT.isVector() || MemBits > 32) {
1282 SDValue Value = Store->getValue();
1283 EVT VT = Value.getValueType();
1284 EVT ElemVT = VT.getVectorElementType();
1285 SDValue Ptr = Store->getBasePtr();
1286 EVT MemEltVT = MemVT.getVectorElementType();
1287 unsigned MemEltBits = MemEltVT.getSizeInBits();
1288 unsigned MemNumElements = MemVT.getVectorNumElements();
1289 unsigned PackedSize = MemVT.getStoreSizeInBits();
1290 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1292 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1294 SDValue PackedValue;
1295 for (unsigned i = 0; i < MemNumElements; ++i) {
1296 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1297 DAG.getConstant(i, MVT::i32));
1298 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1299 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1301 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1302 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1307 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1311 if (PackedSize < 32) {
1312 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1313 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1314 Store->getMemOperand()->getPointerInfo(),
1316 Store->isNonTemporal(), Store->isVolatile(),
1317 Store->getAlignment());
1320 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1321 Store->getMemOperand()->getPointerInfo(),
1322 Store->isVolatile(), Store->isNonTemporal(),
1323 Store->getAlignment());
1326 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1327 SelectionDAG &DAG) const {
1328 StoreSDNode *Store = cast<StoreSDNode>(Op);
1329 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1330 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1331 EVT PtrVT = Store->getBasePtr().getValueType();
1332 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1335 SmallVector<SDValue, 8> Chains;
1337 unsigned EltSize = MemEltVT.getStoreSize();
1338 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1340 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1341 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1343 DAG.getConstant(i, MVT::i32));
1345 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), PtrVT);
1346 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1348 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1349 SrcValue.getWithOffset(i * EltSize),
1350 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1351 Store->getAlignment());
1352 Chains.push_back(NewStore);
1355 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1358 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1359 SelectionDAG &DAG) const {
1360 StoreSDNode *Store = cast<StoreSDNode>(Op);
1361 SDValue Val = Store->getValue();
1362 EVT VT = Val.getValueType();
1364 // If this is a 2 element vector, we really want to scalarize and not create
1365 // weird 1 element vectors.
1366 if (VT.getVectorNumElements() == 2)
1367 return ScalarizeVectorStore(Op, DAG);
1369 EVT MemVT = Store->getMemoryVT();
1370 SDValue Chain = Store->getChain();
1371 SDValue BasePtr = Store->getBasePtr();
1375 EVT LoMemVT, HiMemVT;
1378 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1379 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1380 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1382 EVT PtrVT = BasePtr.getValueType();
1383 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1384 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1386 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1388 = DAG.getTruncStore(Chain, SL, Lo,
1392 Store->isNonTemporal(),
1393 Store->isVolatile(),
1394 Store->getAlignment());
1396 = DAG.getTruncStore(Chain, SL, Hi,
1398 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1400 Store->isNonTemporal(),
1401 Store->isVolatile(),
1402 Store->getAlignment());
1404 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1408 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1410 LoadSDNode *Load = cast<LoadSDNode>(Op);
1411 ISD::LoadExtType ExtType = Load->getExtensionType();
1412 EVT VT = Op.getValueType();
1413 EVT MemVT = Load->getMemoryVT();
1415 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1416 // We can do the extload to 32-bits, and then need to separately extend to
1419 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1423 Load->getMemOperand());
1426 DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
1427 ExtLoad32.getValue(1)
1430 return DAG.getMergeValues(Ops, DL);
1433 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1434 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1435 // FIXME: Copied from PPC
1436 // First, load into 32 bits, then truncate to 1 bit.
1438 SDValue Chain = Load->getChain();
1439 SDValue BasePtr = Load->getBasePtr();
1440 MachineMemOperand *MMO = Load->getMemOperand();
1442 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1443 BasePtr, MVT::i8, MMO);
1446 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1450 return DAG.getMergeValues(Ops, DL);
1453 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1454 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1455 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1459 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1460 DAG.getConstant(2, MVT::i32));
1461 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1462 Load->getChain(), Ptr,
1463 DAG.getTargetConstant(0, MVT::i32),
1465 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1467 DAG.getConstant(0x3, MVT::i32));
1468 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1469 DAG.getConstant(3, MVT::i32));
1471 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1473 EVT MemEltVT = MemVT.getScalarType();
1474 if (ExtType == ISD::SEXTLOAD) {
1475 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1478 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1482 return DAG.getMergeValues(Ops, DL);
1486 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1490 return DAG.getMergeValues(Ops, DL);
1493 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1495 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1496 if (Result.getNode()) {
1500 StoreSDNode *Store = cast<StoreSDNode>(Op);
1501 SDValue Chain = Store->getChain();
1502 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1503 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1504 Store->getValue().getValueType().isVector()) {
1505 return ScalarizeVectorStore(Op, DAG);
1508 EVT MemVT = Store->getMemoryVT();
1509 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1510 MemVT.bitsLT(MVT::i32)) {
1512 if (Store->getMemoryVT() == MVT::i8) {
1514 } else if (Store->getMemoryVT() == MVT::i16) {
1517 SDValue BasePtr = Store->getBasePtr();
1518 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1519 DAG.getConstant(2, MVT::i32));
1520 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1521 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1523 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1524 DAG.getConstant(0x3, MVT::i32));
1526 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1527 DAG.getConstant(3, MVT::i32));
1529 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1532 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1534 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1535 MaskedValue, ShiftAmt);
1537 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1539 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1540 DAG.getConstant(0xffffffff, MVT::i32));
1541 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1543 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1544 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1545 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1550 // This is a shortcut for integer division because we have fast i32<->f32
1551 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1552 // float is enough to accurately represent up to a 24-bit integer.
1553 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1555 EVT VT = Op.getValueType();
1556 SDValue LHS = Op.getOperand(0);
1557 SDValue RHS = Op.getOperand(1);
1558 MVT IntVT = MVT::i32;
1559 MVT FltVT = MVT::f32;
1561 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1562 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1564 if (VT.isVector()) {
1565 unsigned NElts = VT.getVectorNumElements();
1566 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1567 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1570 unsigned BitSize = VT.getScalarType().getSizeInBits();
1572 SDValue jq = DAG.getConstant(1, IntVT);
1575 // char|short jq = ia ^ ib;
1576 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1578 // jq = jq >> (bitsize - 2)
1579 jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
1582 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
1585 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1588 // int ia = (int)LHS;
1590 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1592 // int ib, (int)RHS;
1594 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1596 // float fa = (float)ia;
1597 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1599 // float fb = (float)ib;
1600 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1602 // float fq = native_divide(fa, fb);
1603 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1604 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1607 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1609 // float fqneg = -fq;
1610 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1612 // float fr = mad(fqneg, fb, fa);
1613 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1614 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1616 // int iq = (int)fq;
1617 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1620 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1623 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1625 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1627 // int cv = fr >= fb;
1628 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1630 // jq = (cv ? jq : 0);
1631 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
1633 // dst = trunc/extend to legal type
1634 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1637 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1639 // Rem needs compensation, it's easier to recompute it
1640 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1641 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1647 return DAG.getMergeValues(Res, DL);
1650 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1652 SmallVectorImpl<SDValue> &Results) const {
1653 assert(Op.getValueType() == MVT::i64);
1656 EVT VT = Op.getValueType();
1657 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1659 SDValue one = DAG.getConstant(1, HalfVT);
1660 SDValue zero = DAG.getConstant(0, HalfVT);
1663 SDValue LHS = Op.getOperand(0);
1664 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1665 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1667 SDValue RHS = Op.getOperand(1);
1668 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1669 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1671 // Get Speculative values
1672 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1673 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1675 SDValue REM_Hi = zero;
1676 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1678 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1679 SDValue DIV_Lo = zero;
1681 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1683 for (unsigned i = 0; i < halfBitWidth; ++i) {
1684 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
1685 // Get Value of high bit
1687 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
1688 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
1690 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1691 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1694 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
1695 DAG.getConstant(halfBitWidth - 1, HalfVT));
1696 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
1697 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
1699 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
1700 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
1703 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1705 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
1706 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1708 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1712 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1714 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1715 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
1716 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
1719 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1720 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1721 Results.push_back(DIV);
1722 Results.push_back(REM);
1725 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1726 SelectionDAG &DAG) const {
1728 EVT VT = Op.getValueType();
1730 if (VT == MVT::i64) {
1731 SmallVector<SDValue, 2> Results;
1732 LowerUDIVREM64(Op, DAG, Results);
1733 return DAG.getMergeValues(Results, DL);
1736 SDValue Num = Op.getOperand(0);
1737 SDValue Den = Op.getOperand(1);
1739 if (VT == MVT::i32) {
1740 if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
1741 DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
1742 // TODO: We technically could do this for i64, but shouldn't that just be
1743 // handled by something generally reducing 64-bit division on 32-bit
1744 // values to 32-bit?
1745 return LowerDIVREM24(Op, DAG, false);
1749 // RCP = URECIP(Den) = 2^32 / Den + e
1750 // e is rounding error.
1751 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1753 // RCP_LO = mul(RCP, Den) */
1754 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1756 // RCP_HI = mulhu (RCP, Den) */
1757 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1759 // NEG_RCP_LO = -RCP_LO
1760 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1763 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1764 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1767 // Calculate the rounding error from the URECIP instruction
1768 // E = mulhu(ABS_RCP_LO, RCP)
1769 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1771 // RCP_A_E = RCP + E
1772 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1774 // RCP_S_E = RCP - E
1775 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1777 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1778 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1781 // Quotient = mulhu(Tmp0, Num)
1782 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1784 // Num_S_Remainder = Quotient * Den
1785 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1787 // Remainder = Num - Num_S_Remainder
1788 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1790 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1791 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1792 DAG.getConstant(-1, VT),
1793 DAG.getConstant(0, VT),
1795 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1796 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1798 DAG.getConstant(-1, VT),
1799 DAG.getConstant(0, VT),
1801 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1802 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1805 // Calculate Division result:
1807 // Quotient_A_One = Quotient + 1
1808 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1809 DAG.getConstant(1, VT));
1811 // Quotient_S_One = Quotient - 1
1812 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1813 DAG.getConstant(1, VT));
1815 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1816 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1817 Quotient, Quotient_A_One, ISD::SETEQ);
1819 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1820 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1821 Quotient_S_One, Div, ISD::SETEQ);
1823 // Calculate Rem result:
1825 // Remainder_S_Den = Remainder - Den
1826 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1828 // Remainder_A_Den = Remainder + Den
1829 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1831 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1832 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1833 Remainder, Remainder_S_Den, ISD::SETEQ);
1835 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1836 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1837 Remainder_A_Den, Rem, ISD::SETEQ);
1842 return DAG.getMergeValues(Ops, DL);
1845 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1846 SelectionDAG &DAG) const {
1848 EVT VT = Op.getValueType();
1850 SDValue LHS = Op.getOperand(0);
1851 SDValue RHS = Op.getOperand(1);
1853 if (VT == MVT::i32) {
1854 if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
1855 DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
1856 // TODO: We technically could do this for i64, but shouldn't that just be
1857 // handled by something generally reducing 64-bit division on 32-bit
1858 // values to 32-bit?
1859 return LowerDIVREM24(Op, DAG, true);
1863 SDValue Zero = DAG.getConstant(0, VT);
1864 SDValue NegOne = DAG.getConstant(-1, VT);
1866 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1867 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1868 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1869 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1871 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1872 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1874 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1875 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1877 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1878 SDValue Rem = Div.getValue(1);
1880 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1881 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1883 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1884 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1890 return DAG.getMergeValues(Res, DL);
1893 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1894 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1896 EVT VT = Op.getValueType();
1897 SDValue X = Op.getOperand(0);
1898 SDValue Y = Op.getOperand(1);
1900 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1901 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1902 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1904 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1907 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1909 SDValue Src = Op.getOperand(0);
1911 // result = trunc(src)
1912 // if (src > 0.0 && src != result)
1915 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1917 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1918 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1920 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1922 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1923 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1924 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1926 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1927 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1930 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1932 SDValue Src = Op.getOperand(0);
1934 assert(Op.getValueType() == MVT::f64);
1936 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1937 const SDValue One = DAG.getConstant(1, MVT::i32);
1939 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1941 // Extract the upper half, since this is where we will find the sign and
1943 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1945 const unsigned FractBits = 52;
1946 const unsigned ExpBits = 11;
1948 // Extract the exponent.
1949 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1951 DAG.getConstant(FractBits - 32, MVT::i32),
1952 DAG.getConstant(ExpBits, MVT::i32));
1953 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1954 DAG.getConstant(1023, MVT::i32));
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 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
2019 SDValue Src = Op.getOperand(0);
2021 // result = trunc(src);
2022 // if (src < 0.0 && src != result)
2025 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2027 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
2028 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
2030 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
2032 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
2033 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
2034 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
2036 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
2037 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2040 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2041 bool Signed) const {
2043 SDValue Src = Op.getOperand(0);
2045 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2047 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2048 DAG.getConstant(0, MVT::i32));
2049 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2050 DAG.getConstant(1, MVT::i32));
2052 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2055 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2057 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2058 DAG.getConstant(32, MVT::i32));
2060 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2063 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2064 SelectionDAG &DAG) const {
2065 SDValue S0 = Op.getOperand(0);
2066 if (S0.getValueType() != MVT::i64)
2069 EVT DestVT = Op.getValueType();
2070 if (DestVT == MVT::f64)
2071 return LowerINT_TO_FP64(Op, DAG, false);
2073 assert(DestVT == MVT::f32);
2077 // f32 uint_to_fp i64
2078 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2079 DAG.getConstant(0, MVT::i32));
2080 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2081 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2082 DAG.getConstant(1, MVT::i32));
2083 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2084 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2085 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
2086 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2089 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2090 SelectionDAG &DAG) const {
2091 SDValue Src = Op.getOperand(0);
2092 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2093 return LowerINT_TO_FP64(Op, DAG, true);
2098 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2099 bool Signed) const {
2102 SDValue Src = Op.getOperand(0);
2104 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2107 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), MVT::f64);
2109 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), MVT::f64);
2111 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2113 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2116 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2118 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2119 MVT::i32, FloorMul);
2120 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2122 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2124 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2127 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2128 SelectionDAG &DAG) const {
2129 SDValue Src = Op.getOperand(0);
2131 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2132 return LowerFP64_TO_INT(Op, DAG, true);
2137 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2138 SelectionDAG &DAG) const {
2139 SDValue Src = Op.getOperand(0);
2141 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2142 return LowerFP64_TO_INT(Op, DAG, false);
2147 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2148 SelectionDAG &DAG) const {
2149 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2150 MVT VT = Op.getSimpleValueType();
2151 MVT ScalarVT = VT.getScalarType();
2156 SDValue Src = Op.getOperand(0);
2159 // TODO: Don't scalarize on Evergreen?
2160 unsigned NElts = VT.getVectorNumElements();
2161 SmallVector<SDValue, 8> Args;
2162 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2164 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2165 for (unsigned I = 0; I < NElts; ++I)
2166 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2168 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2171 //===----------------------------------------------------------------------===//
2172 // Custom DAG optimizations
2173 //===----------------------------------------------------------------------===//
2175 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2176 APInt KnownZero, KnownOne;
2177 EVT VT = Op.getValueType();
2178 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2180 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2183 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2184 EVT VT = Op.getValueType();
2186 // In order for this to be a signed 24-bit value, bit 23, must
2188 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2189 // as unsigned 24-bit values.
2190 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2193 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2195 SelectionDAG &DAG = DCI.DAG;
2196 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2197 EVT VT = Op.getValueType();
2199 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2200 APInt KnownZero, KnownOne;
2201 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2202 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2203 DCI.CommitTargetLoweringOpt(TLO);
2206 template <typename IntTy>
2207 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2208 uint32_t Offset, uint32_t Width) {
2209 if (Width + Offset < 32) {
2210 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2211 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2212 return DAG.getConstant(Result, MVT::i32);
2215 return DAG.getConstant(Src0 >> Offset, MVT::i32);
2218 static bool usesAllNormalStores(SDNode *LoadVal) {
2219 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2220 if (!ISD::isNormalStore(*I))
2227 // If we have a copy of an illegal type, replace it with a load / store of an
2228 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2229 // instructions emitted when handling extloads and truncstores. Ideally we could
2230 // recognize the pack / unpack pattern to eliminate it.
2231 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2232 DAGCombinerInfo &DCI) const {
2233 if (!DCI.isBeforeLegalize())
2236 StoreSDNode *SN = cast<StoreSDNode>(N);
2237 SDValue Value = SN->getValue();
2238 EVT VT = Value.getValueType();
2240 if (isTypeLegal(VT) || SN->isVolatile() ||
2241 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2244 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2245 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2248 EVT MemVT = LoadVal->getMemoryVT();
2251 SelectionDAG &DAG = DCI.DAG;
2252 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2254 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2256 LoadVal->getChain(),
2257 LoadVal->getBasePtr(),
2258 LoadVal->getOffset(),
2260 LoadVal->getMemOperand());
2262 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2263 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2265 return DAG.getStore(SN->getChain(), SL, NewLoad,
2266 SN->getBasePtr(), SN->getMemOperand());
2269 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2270 DAGCombinerInfo &DCI) const {
2271 EVT VT = N->getValueType(0);
2273 if (VT.isVector() || VT.getSizeInBits() > 32)
2276 SelectionDAG &DAG = DCI.DAG;
2279 SDValue N0 = N->getOperand(0);
2280 SDValue N1 = N->getOperand(1);
2283 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2284 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2285 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2286 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2287 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2288 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2289 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2290 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2295 // We need to use sext even for MUL_U24, because MUL_U24 is used
2296 // for signed multiply of 8 and 16-bit types.
2297 return DAG.getSExtOrTrunc(Mul, DL, VT);
2300 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2301 DAGCombinerInfo &DCI) const {
2302 SelectionDAG &DAG = DCI.DAG;
2305 switch(N->getOpcode()) {
2308 return performMulCombine(N, DCI);
2309 case AMDGPUISD::MUL_I24:
2310 case AMDGPUISD::MUL_U24: {
2311 SDValue N0 = N->getOperand(0);
2312 SDValue N1 = N->getOperand(1);
2313 simplifyI24(N0, DCI);
2314 simplifyI24(N1, DCI);
2318 SDValue Cond = N->getOperand(0);
2319 if (Cond.getOpcode() == ISD::SETCC && Cond.hasOneUse()) {
2321 EVT VT = N->getValueType(0);
2322 SDValue LHS = Cond.getOperand(0);
2323 SDValue RHS = Cond.getOperand(1);
2324 SDValue CC = Cond.getOperand(2);
2326 SDValue True = N->getOperand(1);
2327 SDValue False = N->getOperand(2);
2330 return CombineFMinMaxLegacy(DL, VT, LHS, RHS, True, False, CC, DCI);
2332 // TODO: Implement min / max Evergreen instructions.
2333 if (VT == MVT::i32 &&
2334 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2335 return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2341 case AMDGPUISD::BFE_I32:
2342 case AMDGPUISD::BFE_U32: {
2343 assert(!N->getValueType(0).isVector() &&
2344 "Vector handling of BFE not implemented");
2345 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2349 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2351 return DAG.getConstant(0, MVT::i32);
2353 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2357 SDValue BitsFrom = N->getOperand(0);
2358 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2360 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2362 if (OffsetVal == 0) {
2363 // This is already sign / zero extended, so try to fold away extra BFEs.
2364 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2366 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2367 if (OpSignBits >= SignBits)
2370 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2372 // This is a sign_extend_inreg. Replace it to take advantage of existing
2373 // DAG Combines. If not eliminated, we will match back to BFE during
2376 // TODO: The sext_inreg of extended types ends, although we can could
2377 // handle them in a single BFE.
2378 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2379 DAG.getValueType(SmallVT));
2382 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2385 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2387 return constantFoldBFE<int32_t>(DAG,
2388 CVal->getSExtValue(),
2393 return constantFoldBFE<uint32_t>(DAG,
2394 CVal->getZExtValue(),
2399 if ((OffsetVal + WidthVal) >= 32) {
2400 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2401 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2402 BitsFrom, ShiftVal);
2405 if (BitsFrom.hasOneUse()) {
2406 APInt Demanded = APInt::getBitsSet(32,
2408 OffsetVal + WidthVal);
2410 APInt KnownZero, KnownOne;
2411 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2412 !DCI.isBeforeLegalizeOps());
2413 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2414 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2415 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2416 KnownZero, KnownOne, TLO)) {
2417 DCI.CommitTargetLoweringOpt(TLO);
2425 return performStoreCombine(N, DCI);
2430 //===----------------------------------------------------------------------===//
2432 //===----------------------------------------------------------------------===//
2434 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2437 const SmallVectorImpl<ISD::InputArg> &Ins,
2438 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2440 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2441 if (Ins[i].ArgVT == Ins[i].VT) {
2442 OrigIns.push_back(Ins[i]);
2447 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2448 // Vector has been split into scalars.
2449 VT = Ins[i].ArgVT.getVectorElementType();
2450 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2451 Ins[i].ArgVT.getVectorElementType() !=
2452 Ins[i].VT.getVectorElementType()) {
2453 // Vector elements have been promoted
2456 // Vector has been spilt into smaller vectors.
2460 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2461 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2462 OrigIns.push_back(Arg);
2466 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2467 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2468 return CFP->isExactlyValue(1.0);
2470 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2471 return C->isAllOnesValue();
2476 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2477 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2478 return CFP->getValueAPF().isZero();
2480 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2481 return C->isNullValue();
2486 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2487 const TargetRegisterClass *RC,
2488 unsigned Reg, EVT VT) const {
2489 MachineFunction &MF = DAG.getMachineFunction();
2490 MachineRegisterInfo &MRI = MF.getRegInfo();
2491 unsigned VirtualRegister;
2492 if (!MRI.isLiveIn(Reg)) {
2493 VirtualRegister = MRI.createVirtualRegister(RC);
2494 MRI.addLiveIn(Reg, VirtualRegister);
2496 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2498 return DAG.getRegister(VirtualRegister, VT);
2501 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2503 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2505 default: return nullptr;
2507 NODE_NAME_CASE(CALL);
2508 NODE_NAME_CASE(UMUL);
2509 NODE_NAME_CASE(RET_FLAG);
2510 NODE_NAME_CASE(BRANCH_COND);
2513 NODE_NAME_CASE(DWORDADDR)
2514 NODE_NAME_CASE(FRACT)
2515 NODE_NAME_CASE(CLAMP)
2517 NODE_NAME_CASE(FMAX_LEGACY)
2518 NODE_NAME_CASE(SMAX)
2519 NODE_NAME_CASE(UMAX)
2520 NODE_NAME_CASE(FMIN_LEGACY)
2521 NODE_NAME_CASE(SMIN)
2522 NODE_NAME_CASE(UMIN)
2523 NODE_NAME_CASE(FMAX3)
2524 NODE_NAME_CASE(SMAX3)
2525 NODE_NAME_CASE(UMAX3)
2526 NODE_NAME_CASE(FMIN3)
2527 NODE_NAME_CASE(SMIN3)
2528 NODE_NAME_CASE(UMIN3)
2529 NODE_NAME_CASE(URECIP)
2530 NODE_NAME_CASE(DIV_SCALE)
2531 NODE_NAME_CASE(DIV_FMAS)
2532 NODE_NAME_CASE(DIV_FIXUP)
2533 NODE_NAME_CASE(TRIG_PREOP)
2536 NODE_NAME_CASE(RSQ_LEGACY)
2537 NODE_NAME_CASE(RSQ_CLAMPED)
2538 NODE_NAME_CASE(LDEXP)
2539 NODE_NAME_CASE(FP_CLASS)
2540 NODE_NAME_CASE(DOT4)
2541 NODE_NAME_CASE(BFE_U32)
2542 NODE_NAME_CASE(BFE_I32)
2545 NODE_NAME_CASE(BREV)
2546 NODE_NAME_CASE(MUL_U24)
2547 NODE_NAME_CASE(MUL_I24)
2548 NODE_NAME_CASE(MAD_U24)
2549 NODE_NAME_CASE(MAD_I24)
2550 NODE_NAME_CASE(EXPORT)
2551 NODE_NAME_CASE(CONST_ADDRESS)
2552 NODE_NAME_CASE(REGISTER_LOAD)
2553 NODE_NAME_CASE(REGISTER_STORE)
2554 NODE_NAME_CASE(LOAD_CONSTANT)
2555 NODE_NAME_CASE(LOAD_INPUT)
2556 NODE_NAME_CASE(SAMPLE)
2557 NODE_NAME_CASE(SAMPLEB)
2558 NODE_NAME_CASE(SAMPLED)
2559 NODE_NAME_CASE(SAMPLEL)
2560 NODE_NAME_CASE(CVT_F32_UBYTE0)
2561 NODE_NAME_CASE(CVT_F32_UBYTE1)
2562 NODE_NAME_CASE(CVT_F32_UBYTE2)
2563 NODE_NAME_CASE(CVT_F32_UBYTE3)
2564 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2565 NODE_NAME_CASE(CONST_DATA_PTR)
2566 NODE_NAME_CASE(STORE_MSKOR)
2567 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2571 SDValue AMDGPUTargetLowering::getRsqrtEstimate(SDValue Operand,
2572 DAGCombinerInfo &DCI,
2573 unsigned &RefinementSteps,
2574 bool &UseOneConstNR) const {
2575 SelectionDAG &DAG = DCI.DAG;
2576 EVT VT = Operand.getValueType();
2578 if (VT == MVT::f32) {
2579 RefinementSteps = 0;
2580 return DAG.getNode(AMDGPUISD::RSQ, SDLoc(Operand), VT, Operand);
2583 // TODO: There is also f64 rsq instruction, but the documentation is less
2584 // clear on its precision.
2589 SDValue AMDGPUTargetLowering::getRecipEstimate(SDValue Operand,
2590 DAGCombinerInfo &DCI,
2591 unsigned &RefinementSteps) const {
2592 SelectionDAG &DAG = DCI.DAG;
2593 EVT VT = Operand.getValueType();
2595 if (VT == MVT::f32) {
2596 // Reciprocal, < 1 ulp error.
2598 // This reciprocal approximation converges to < 0.5 ulp error with one
2599 // newton rhapson performed with two fused multiple adds (FMAs).
2601 RefinementSteps = 0;
2602 return DAG.getNode(AMDGPUISD::RCP, SDLoc(Operand), VT, Operand);
2605 // TODO: There is also f64 rcp instruction, but the documentation is less
2606 // clear on its precision.
2611 static void computeKnownBitsForMinMax(const SDValue Op0,
2615 const SelectionDAG &DAG,
2617 APInt Op0Zero, Op0One;
2618 APInt Op1Zero, Op1One;
2619 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2620 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2622 KnownZero = Op0Zero & Op1Zero;
2623 KnownOne = Op0One & Op1One;
2626 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2630 const SelectionDAG &DAG,
2631 unsigned Depth) const {
2633 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2637 unsigned Opc = Op.getOpcode();
2642 case ISD::INTRINSIC_WO_CHAIN: {
2643 // FIXME: The intrinsic should just use the node.
2644 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2645 case AMDGPUIntrinsic::AMDGPU_imax:
2646 case AMDGPUIntrinsic::AMDGPU_umax:
2647 case AMDGPUIntrinsic::AMDGPU_imin:
2648 case AMDGPUIntrinsic::AMDGPU_umin:
2649 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2650 KnownZero, KnownOne, DAG, Depth);
2658 case AMDGPUISD::SMAX:
2659 case AMDGPUISD::UMAX:
2660 case AMDGPUISD::SMIN:
2661 case AMDGPUISD::UMIN:
2662 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2663 KnownZero, KnownOne, DAG, Depth);
2666 case AMDGPUISD::BFE_I32:
2667 case AMDGPUISD::BFE_U32: {
2668 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2672 unsigned BitWidth = 32;
2673 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2675 if (Opc == AMDGPUISD::BFE_U32)
2676 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2683 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2685 const SelectionDAG &DAG,
2686 unsigned Depth) const {
2687 switch (Op.getOpcode()) {
2688 case AMDGPUISD::BFE_I32: {
2689 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2693 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2694 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2695 if (!Offset || !Offset->isNullValue())
2698 // TODO: Could probably figure something out with non-0 offsets.
2699 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2700 return std::max(SignBits, Op0SignBits);
2703 case AMDGPUISD::BFE_U32: {
2704 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2705 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;