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/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
38 /// Diagnostic information for unimplemented or unsupported feature reporting.
39 class DiagnosticInfoUnsupported : public DiagnosticInfo {
41 const Twine &Description;
46 static int getKindID() {
48 KindID = llvm::getNextAvailablePluginDiagnosticKind();
53 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
54 DiagnosticSeverity Severity = DS_Error)
55 : DiagnosticInfo(getKindID(), Severity),
59 const Function &getFunction() const { return Fn; }
60 const Twine &getDescription() const { return Description; }
62 void print(DiagnosticPrinter &DP) const override {
63 DP << "unsupported " << getDescription() << " in " << Fn.getName();
66 static bool classof(const DiagnosticInfo *DI) {
67 return DI->getKind() == getKindID();
71 int DiagnosticInfoUnsupported::KindID = 0;
75 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
76 CCValAssign::LocInfo LocInfo,
77 ISD::ArgFlagsTy ArgFlags, CCState &State) {
78 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
79 ArgFlags.getOrigAlign());
80 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
85 #include "AMDGPUGenCallingConv.inc"
87 // Find a larger type to do a load / store of a vector with.
88 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
89 unsigned StoreSize = VT.getStoreSizeInBits();
91 return EVT::getIntegerVT(Ctx, StoreSize);
93 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
94 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
97 // Type for a vector that will be loaded to.
98 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
99 unsigned StoreSize = VT.getStoreSizeInBits();
101 return EVT::getIntegerVT(Ctx, 32);
103 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
106 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
107 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
109 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
111 setOperationAction(ISD::Constant, MVT::i32, Legal);
112 setOperationAction(ISD::Constant, MVT::i64, Legal);
113 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
114 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
116 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
117 setOperationAction(ISD::BRIND, MVT::Other, Expand);
119 // We need to custom lower some of the intrinsics
120 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
122 // Library functions. These default to Expand, but we have instructions
124 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
125 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
126 setOperationAction(ISD::FPOW, MVT::f32, Legal);
127 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
128 setOperationAction(ISD::FABS, MVT::f32, Legal);
129 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
130 setOperationAction(ISD::FRINT, MVT::f32, Legal);
131 setOperationAction(ISD::FROUND, MVT::f32, Legal);
132 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
134 // Lower floating point store/load to integer store/load to reduce the number
135 // of patterns in tablegen.
136 setOperationAction(ISD::STORE, MVT::f32, Promote);
137 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
139 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
140 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
142 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
143 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
145 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
146 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
148 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
149 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
151 setOperationAction(ISD::STORE, MVT::f64, Promote);
152 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
154 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
155 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
157 // Custom lowering of vector stores is required for local address space
159 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
160 // XXX: Native v2i32 local address space stores are possible, but not
161 // currently implemented.
162 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
164 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
165 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
166 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
168 // XXX: This can be change to Custom, once ExpandVectorStores can
169 // handle 64-bit stores.
170 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
172 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
173 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
174 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
175 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
176 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
179 setOperationAction(ISD::LOAD, MVT::f32, Promote);
180 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
182 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
183 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
185 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
186 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
188 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
189 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
191 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
192 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
194 setOperationAction(ISD::LOAD, MVT::f64, Promote);
195 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
197 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
198 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
200 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
201 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
202 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
203 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
204 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
205 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
206 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
207 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
208 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
209 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
211 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
212 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
213 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
214 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
216 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
217 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
218 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
220 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
221 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
224 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
226 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
227 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
228 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
229 setOperationAction(ISD::FRINT, MVT::f64, Custom);
230 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
233 if (!Subtarget->hasBFI()) {
234 // fcopysign can be done in a single instruction with BFI.
235 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
236 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
239 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
240 for (MVT VT : ScalarIntVTs) {
241 setOperationAction(ISD::SREM, VT, Expand);
242 setOperationAction(ISD::SDIV, VT, Expand);
244 // GPU does not have divrem function for signed or unsigned.
245 setOperationAction(ISD::SDIVREM, VT, Custom);
246 setOperationAction(ISD::UDIVREM, VT, Custom);
248 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
249 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
250 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
252 setOperationAction(ISD::BSWAP, VT, Expand);
253 setOperationAction(ISD::CTTZ, VT, Expand);
254 setOperationAction(ISD::CTLZ, VT, Expand);
257 if (!Subtarget->hasBCNT(32))
258 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
260 if (!Subtarget->hasBCNT(64))
261 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
263 // The hardware supports 32-bit ROTR, but not ROTL.
264 setOperationAction(ISD::ROTL, MVT::i32, Expand);
265 setOperationAction(ISD::ROTL, MVT::i64, Expand);
266 setOperationAction(ISD::ROTR, MVT::i64, Expand);
268 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
269 setOperationAction(ISD::MUL, MVT::i64, Expand);
270 setOperationAction(ISD::MULHU, MVT::i64, Expand);
271 setOperationAction(ISD::MULHS, MVT::i64, Expand);
272 setOperationAction(ISD::UDIV, MVT::i32, Expand);
273 setOperationAction(ISD::UREM, MVT::i32, Expand);
274 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
275 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
277 static const MVT::SimpleValueType VectorIntTypes[] = {
278 MVT::v2i32, MVT::v4i32
281 for (MVT VT : VectorIntTypes) {
282 // Expand the following operations for the current type by default.
283 setOperationAction(ISD::ADD, VT, Expand);
284 setOperationAction(ISD::AND, VT, Expand);
285 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
286 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
287 setOperationAction(ISD::MUL, VT, Expand);
288 setOperationAction(ISD::OR, VT, Expand);
289 setOperationAction(ISD::SHL, VT, Expand);
290 setOperationAction(ISD::SRA, VT, Expand);
291 setOperationAction(ISD::SRL, VT, Expand);
292 setOperationAction(ISD::ROTL, VT, Expand);
293 setOperationAction(ISD::ROTR, VT, Expand);
294 setOperationAction(ISD::SUB, VT, Expand);
295 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
296 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
297 // TODO: Implement custom UREM / SREM routines.
298 setOperationAction(ISD::SDIV, VT, Expand);
299 setOperationAction(ISD::UDIV, VT, Expand);
300 setOperationAction(ISD::SREM, VT, Expand);
301 setOperationAction(ISD::UREM, VT, Expand);
302 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
303 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
304 setOperationAction(ISD::SDIVREM, VT, Custom);
305 setOperationAction(ISD::UDIVREM, VT, Custom);
306 setOperationAction(ISD::ADDC, VT, Expand);
307 setOperationAction(ISD::SUBC, VT, Expand);
308 setOperationAction(ISD::ADDE, VT, Expand);
309 setOperationAction(ISD::SUBE, VT, Expand);
310 setOperationAction(ISD::SELECT, VT, Expand);
311 setOperationAction(ISD::VSELECT, VT, Expand);
312 setOperationAction(ISD::SELECT_CC, VT, Expand);
313 setOperationAction(ISD::XOR, VT, Expand);
314 setOperationAction(ISD::BSWAP, VT, Expand);
315 setOperationAction(ISD::CTPOP, VT, Expand);
316 setOperationAction(ISD::CTTZ, VT, Expand);
317 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
318 setOperationAction(ISD::CTLZ, VT, Expand);
319 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
320 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
323 static const MVT::SimpleValueType FloatVectorTypes[] = {
324 MVT::v2f32, MVT::v4f32
327 for (MVT VT : FloatVectorTypes) {
328 setOperationAction(ISD::FABS, VT, Expand);
329 setOperationAction(ISD::FADD, VT, Expand);
330 setOperationAction(ISD::FCEIL, VT, Expand);
331 setOperationAction(ISD::FCOS, VT, Expand);
332 setOperationAction(ISD::FDIV, VT, Expand);
333 setOperationAction(ISD::FEXP2, VT, Expand);
334 setOperationAction(ISD::FLOG2, VT, Expand);
335 setOperationAction(ISD::FPOW, VT, Expand);
336 setOperationAction(ISD::FFLOOR, VT, Expand);
337 setOperationAction(ISD::FTRUNC, VT, Expand);
338 setOperationAction(ISD::FMUL, VT, Expand);
339 setOperationAction(ISD::FRINT, VT, Expand);
340 setOperationAction(ISD::FNEARBYINT, VT, Expand);
341 setOperationAction(ISD::FSQRT, VT, Expand);
342 setOperationAction(ISD::FSIN, VT, Expand);
343 setOperationAction(ISD::FSUB, VT, Expand);
344 setOperationAction(ISD::FNEG, VT, Expand);
345 setOperationAction(ISD::SELECT, VT, Expand);
346 setOperationAction(ISD::VSELECT, VT, Expand);
347 setOperationAction(ISD::SELECT_CC, VT, Expand);
348 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
349 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
352 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
353 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
355 setTargetDAGCombine(ISD::MUL);
356 setTargetDAGCombine(ISD::SELECT_CC);
358 setSchedulingPreference(Sched::RegPressure);
359 setJumpIsExpensive(true);
361 setSelectIsExpensive(false);
362 PredictableSelectIsExpensive = false;
364 // There are no integer divide instructions, and these expand to a pretty
365 // large sequence of instructions.
366 setIntDivIsCheap(false);
367 setPow2DivIsCheap(false);
369 // TODO: Investigate this when 64-bit divides are implemented.
370 addBypassSlowDiv(64, 32);
372 // FIXME: Need to really handle these.
373 MaxStoresPerMemcpy = 4096;
374 MaxStoresPerMemmove = 4096;
375 MaxStoresPerMemset = 4096;
378 //===----------------------------------------------------------------------===//
379 // Target Information
380 //===----------------------------------------------------------------------===//
382 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
386 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
390 // The backend supports 32 and 64 bit floating point immediates.
391 // FIXME: Why are we reporting vectors of FP immediates as legal?
392 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
393 EVT ScalarVT = VT.getScalarType();
394 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
397 // We don't want to shrink f64 / f32 constants.
398 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
399 EVT ScalarVT = VT.getScalarType();
400 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
403 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
405 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
408 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
409 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
411 return ((LScalarSize <= CastScalarSize) ||
412 (CastScalarSize >= 32) ||
416 //===---------------------------------------------------------------------===//
418 //===---------------------------------------------------------------------===//
420 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
421 assert(VT.isFloatingPoint());
422 return VT == MVT::f32;
425 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
426 assert(VT.isFloatingPoint());
427 return VT == MVT::f32;
430 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
431 // Truncate is just accessing a subregister.
432 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
435 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
436 // Truncate is just accessing a subregister.
437 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
438 (Dest->getPrimitiveSizeInBits() % 32 == 0);
441 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
442 const DataLayout *DL = getDataLayout();
443 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
444 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
446 return SrcSize == 32 && DestSize == 64;
449 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
450 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
451 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
452 // this will enable reducing 64-bit operations the 32-bit, which is always
454 return Src == MVT::i32 && Dest == MVT::i64;
457 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
458 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
459 // limited number of native 64-bit operations. Shrinking an operation to fit
460 // in a single 32-bit register should always be helpful. As currently used,
461 // this is much less general than the name suggests, and is only used in
462 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
463 // not profitable, and may actually be harmful.
464 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
467 //===---------------------------------------------------------------------===//
468 // TargetLowering Callbacks
469 //===---------------------------------------------------------------------===//
471 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
472 const SmallVectorImpl<ISD::InputArg> &Ins) const {
474 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
477 SDValue AMDGPUTargetLowering::LowerReturn(
479 CallingConv::ID CallConv,
481 const SmallVectorImpl<ISD::OutputArg> &Outs,
482 const SmallVectorImpl<SDValue> &OutVals,
483 SDLoc DL, SelectionDAG &DAG) const {
484 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
487 //===---------------------------------------------------------------------===//
488 // Target specific lowering
489 //===---------------------------------------------------------------------===//
491 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
492 SmallVectorImpl<SDValue> &InVals) const {
493 SDValue Callee = CLI.Callee;
494 SelectionDAG &DAG = CLI.DAG;
496 const Function &Fn = *DAG.getMachineFunction().getFunction();
498 StringRef FuncName("<unknown>");
500 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
501 FuncName = G->getSymbol();
502 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
503 FuncName = G->getGlobal()->getName();
505 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
506 DAG.getContext()->diagnose(NoCalls);
510 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
511 SelectionDAG &DAG) const {
512 switch (Op.getOpcode()) {
514 Op.getNode()->dump();
515 llvm_unreachable("Custom lowering code for this"
516 "instruction is not implemented yet!");
518 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
519 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
520 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
521 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
522 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
523 case ISD::SDIV: return LowerSDIV(Op, DAG);
524 case ISD::SREM: return LowerSREM(Op, DAG);
525 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
526 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
527 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
528 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
529 case ISD::FRINT: return LowerFRINT(Op, DAG);
530 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
531 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
532 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
537 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
538 SmallVectorImpl<SDValue> &Results,
539 SelectionDAG &DAG) const {
540 switch (N->getOpcode()) {
541 case ISD::SIGN_EXTEND_INREG:
542 // Different parts of legalization seem to interpret which type of
543 // sign_extend_inreg is the one to check for custom lowering. The extended
544 // from type is what really matters, but some places check for custom
545 // lowering of the result type. This results in trying to use
546 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
547 // nothing here and let the illegal result integer be handled normally.
554 // FIXME: This implements accesses to initialized globals in the constant
555 // address space by copying them to private and accessing that. It does not
556 // properly handle illegal types or vectors. The private vector loads are not
557 // scalarized, and the illegal scalars hit an assertion. This technique will not
558 // work well with large initializers, and this should eventually be
559 // removed. Initialized globals should be placed into a data section that the
560 // runtime will load into a buffer before the kernel is executed. Uses of the
561 // global need to be replaced with a pointer loaded from an implicit kernel
562 // argument into this buffer holding the copy of the data, which will remove the
563 // need for any of this.
564 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
565 const GlobalValue *GV,
566 const SDValue &InitPtr,
568 SelectionDAG &DAG) const {
569 const DataLayout *TD = getTargetMachine().getDataLayout();
571 Type *InitTy = Init->getType();
573 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
574 EVT VT = EVT::getEVT(InitTy);
575 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
576 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
577 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
578 TD->getPrefTypeAlignment(InitTy));
581 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
582 EVT VT = EVT::getEVT(CFP->getType());
583 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
584 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
585 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
586 TD->getPrefTypeAlignment(CFP->getType()));
589 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
590 const StructLayout *SL = TD->getStructLayout(ST);
592 EVT PtrVT = InitPtr.getValueType();
593 SmallVector<SDValue, 8> Chains;
595 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
596 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
597 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
599 Constant *Elt = Init->getAggregateElement(I);
600 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
603 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
606 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
607 EVT PtrVT = InitPtr.getValueType();
609 unsigned NumElements;
610 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
611 NumElements = AT->getNumElements();
612 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
613 NumElements = VT->getNumElements();
615 llvm_unreachable("Unexpected type");
617 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
618 SmallVector<SDValue, 8> Chains;
619 for (unsigned i = 0; i < NumElements; ++i) {
620 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
621 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
623 Constant *Elt = Init->getAggregateElement(i);
624 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
627 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
630 if (isa<UndefValue>(Init)) {
631 EVT VT = EVT::getEVT(InitTy);
632 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
633 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
634 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
635 TD->getPrefTypeAlignment(InitTy));
639 llvm_unreachable("Unhandled constant initializer");
642 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
644 SelectionDAG &DAG) const {
646 const DataLayout *TD = getTargetMachine().getDataLayout();
647 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
648 const GlobalValue *GV = G->getGlobal();
650 switch (G->getAddressSpace()) {
651 default: llvm_unreachable("Global Address lowering not implemented for this "
653 case AMDGPUAS::LOCAL_ADDRESS: {
654 // XXX: What does the value of G->getOffset() mean?
655 assert(G->getOffset() == 0 &&
656 "Do not know what to do with an non-zero offset");
659 if (MFI->LocalMemoryObjects.count(GV) == 0) {
660 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
661 Offset = MFI->LDSSize;
662 MFI->LocalMemoryObjects[GV] = Offset;
663 // XXX: Account for alignment?
664 MFI->LDSSize += Size;
666 Offset = MFI->LocalMemoryObjects[GV];
669 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
671 case AMDGPUAS::CONSTANT_ADDRESS: {
672 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
673 Type *EltType = GV->getType()->getElementType();
674 unsigned Size = TD->getTypeAllocSize(EltType);
675 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
677 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
678 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
680 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
681 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
683 const GlobalVariable *Var = cast<GlobalVariable>(GV);
684 if (!Var->hasInitializer()) {
685 // This has no use, but bugpoint will hit it.
686 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
689 const Constant *Init = Var->getInitializer();
690 SmallVector<SDNode*, 8> WorkList;
692 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
693 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
694 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
696 WorkList.push_back(*I);
698 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
699 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
700 E = WorkList.end(); I != E; ++I) {
701 SmallVector<SDValue, 8> Ops;
702 Ops.push_back(Chain);
703 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
704 Ops.push_back((*I)->getOperand(i));
706 DAG.UpdateNodeOperands(*I, Ops);
708 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
713 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
714 SelectionDAG &DAG) const {
715 SmallVector<SDValue, 8> Args;
716 SDValue A = Op.getOperand(0);
717 SDValue B = Op.getOperand(1);
719 DAG.ExtractVectorElements(A, Args);
720 DAG.ExtractVectorElements(B, Args);
722 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
725 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
726 SelectionDAG &DAG) const {
728 SmallVector<SDValue, 8> Args;
729 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
730 EVT VT = Op.getValueType();
731 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
732 VT.getVectorNumElements());
734 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
737 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
738 SelectionDAG &DAG) const {
740 MachineFunction &MF = DAG.getMachineFunction();
741 const AMDGPUFrameLowering *TFL =
742 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
744 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
746 unsigned FrameIndex = FIN->getIndex();
747 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
748 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
752 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
753 SelectionDAG &DAG) const {
754 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
756 EVT VT = Op.getValueType();
758 switch (IntrinsicID) {
760 case AMDGPUIntrinsic::AMDGPU_abs:
761 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
762 return LowerIntrinsicIABS(Op, DAG);
763 case AMDGPUIntrinsic::AMDGPU_lrp:
764 return LowerIntrinsicLRP(Op, DAG);
765 case AMDGPUIntrinsic::AMDGPU_fract:
766 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
767 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
769 case AMDGPUIntrinsic::AMDGPU_clamp:
770 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
771 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
772 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
774 case Intrinsic::AMDGPU_div_scale: {
775 // 3rd parameter required to be a constant.
776 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
778 return DAG.getUNDEF(VT);
780 // Translate to the operands expected by the machine instruction. The
781 // first parameter must be the same as the first instruction.
782 SDValue Numerator = Op.getOperand(1);
783 SDValue Denominator = Op.getOperand(2);
784 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
786 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
787 Src0, Denominator, Numerator);
790 case Intrinsic::AMDGPU_div_fmas:
791 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
792 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
794 case Intrinsic::AMDGPU_div_fixup:
795 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
796 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
798 case Intrinsic::AMDGPU_trig_preop:
799 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
800 Op.getOperand(1), Op.getOperand(2));
802 case Intrinsic::AMDGPU_rcp:
803 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
805 case Intrinsic::AMDGPU_rsq:
806 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
808 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
809 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
811 case Intrinsic::AMDGPU_rsq_clamped:
812 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
814 case AMDGPUIntrinsic::AMDGPU_imax:
815 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
817 case AMDGPUIntrinsic::AMDGPU_umax:
818 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
820 case AMDGPUIntrinsic::AMDGPU_imin:
821 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
823 case AMDGPUIntrinsic::AMDGPU_umin:
824 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
827 case AMDGPUIntrinsic::AMDGPU_umul24:
828 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
829 Op.getOperand(1), Op.getOperand(2));
831 case AMDGPUIntrinsic::AMDGPU_imul24:
832 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
833 Op.getOperand(1), Op.getOperand(2));
835 case AMDGPUIntrinsic::AMDGPU_umad24:
836 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
837 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
839 case AMDGPUIntrinsic::AMDGPU_imad24:
840 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
841 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
843 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
844 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
846 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
847 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
849 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
850 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
852 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
853 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
855 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
856 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
861 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
862 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
867 case AMDGPUIntrinsic::AMDGPU_bfi:
868 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
873 case AMDGPUIntrinsic::AMDGPU_bfm:
874 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
878 case AMDGPUIntrinsic::AMDGPU_brev:
879 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
881 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
882 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
884 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
885 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
886 case AMDGPUIntrinsic::AMDGPU_trunc:
887 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
891 ///IABS(a) = SMAX(sub(0, a), a)
892 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
893 SelectionDAG &DAG) const {
895 EVT VT = Op.getValueType();
896 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
899 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
902 /// Linear Interpolation
903 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
904 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
905 SelectionDAG &DAG) const {
907 EVT VT = Op.getValueType();
908 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
909 DAG.getConstantFP(1.0f, MVT::f32),
911 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
913 return DAG.getNode(ISD::FADD, DL, VT,
914 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
918 /// \brief Generate Min/Max node
919 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
920 SelectionDAG &DAG) const {
922 EVT VT = N->getValueType(0);
924 SDValue LHS = N->getOperand(0);
925 SDValue RHS = N->getOperand(1);
926 SDValue True = N->getOperand(2);
927 SDValue False = N->getOperand(3);
928 SDValue CC = N->getOperand(4);
930 if (VT != MVT::f32 ||
931 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
935 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
949 llvm_unreachable("Operation should already be optimised!");
956 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
957 return DAG.getNode(Opc, DL, VT, LHS, RHS);
965 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
966 return DAG.getNode(Opc, DL, VT, LHS, RHS);
968 case ISD::SETCC_INVALID:
969 llvm_unreachable("Invalid setcc condcode!");
974 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
975 SelectionDAG &DAG) const {
976 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
977 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
978 EVT EltVT = Op.getValueType().getVectorElementType();
979 EVT PtrVT = Load->getBasePtr().getValueType();
980 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
981 SmallVector<SDValue, 8> Loads;
984 for (unsigned i = 0, e = NumElts; i != e; ++i) {
985 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
986 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
987 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
988 Load->getChain(), Ptr,
989 MachinePointerInfo(Load->getMemOperand()->getValue()),
990 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
991 Load->getAlignment()));
993 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
996 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
997 SelectionDAG &DAG) const {
998 StoreSDNode *Store = cast<StoreSDNode>(Op);
999 EVT MemVT = Store->getMemoryVT();
1000 unsigned MemBits = MemVT.getSizeInBits();
1002 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1003 // truncating store into an i32 store.
1004 // XXX: We could also handle optimize other vector bitwidths.
1005 if (!MemVT.isVector() || MemBits > 32) {
1010 SDValue Value = Store->getValue();
1011 EVT VT = Value.getValueType();
1012 EVT ElemVT = VT.getVectorElementType();
1013 SDValue Ptr = Store->getBasePtr();
1014 EVT MemEltVT = MemVT.getVectorElementType();
1015 unsigned MemEltBits = MemEltVT.getSizeInBits();
1016 unsigned MemNumElements = MemVT.getVectorNumElements();
1017 unsigned PackedSize = MemVT.getStoreSizeInBits();
1018 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1020 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1022 SDValue PackedValue;
1023 for (unsigned i = 0; i < MemNumElements; ++i) {
1024 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1025 DAG.getConstant(i, MVT::i32));
1026 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1027 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1029 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1030 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1035 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1039 if (PackedSize < 32) {
1040 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1041 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1042 Store->getMemOperand()->getPointerInfo(),
1044 Store->isNonTemporal(), Store->isVolatile(),
1045 Store->getAlignment());
1048 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1049 Store->getMemOperand()->getPointerInfo(),
1050 Store->isVolatile(), Store->isNonTemporal(),
1051 Store->getAlignment());
1054 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1055 SelectionDAG &DAG) const {
1056 StoreSDNode *Store = cast<StoreSDNode>(Op);
1057 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1058 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1059 EVT PtrVT = Store->getBasePtr().getValueType();
1060 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1063 SmallVector<SDValue, 8> Chains;
1065 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1066 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1067 Store->getValue(), DAG.getConstant(i, MVT::i32));
1068 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1069 Store->getBasePtr(),
1070 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1072 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1073 MachinePointerInfo(Store->getMemOperand()->getValue()),
1074 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1075 Store->getAlignment()));
1077 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1080 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1082 LoadSDNode *Load = cast<LoadSDNode>(Op);
1083 ISD::LoadExtType ExtType = Load->getExtensionType();
1084 EVT VT = Op.getValueType();
1085 EVT MemVT = Load->getMemoryVT();
1087 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1088 // We can do the extload to 32-bits, and then need to separately extend to
1091 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1095 Load->getMemOperand());
1096 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1099 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1100 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1101 // FIXME: Copied from PPC
1102 // First, load into 32 bits, then truncate to 1 bit.
1104 SDValue Chain = Load->getChain();
1105 SDValue BasePtr = Load->getBasePtr();
1106 MachineMemOperand *MMO = Load->getMemOperand();
1108 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1109 BasePtr, MVT::i8, MMO);
1110 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1113 // Lower loads constant address space global variable loads
1114 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1115 isa<GlobalVariable>(
1116 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1118 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1119 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1120 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1121 DAG.getConstant(2, MVT::i32));
1122 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1123 Load->getChain(), Ptr,
1124 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1127 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1128 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1132 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1133 DAG.getConstant(2, MVT::i32));
1134 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1135 Load->getChain(), Ptr,
1136 DAG.getTargetConstant(0, MVT::i32),
1138 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1140 DAG.getConstant(0x3, MVT::i32));
1141 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1142 DAG.getConstant(3, MVT::i32));
1144 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1146 EVT MemEltVT = MemVT.getScalarType();
1147 if (ExtType == ISD::SEXTLOAD) {
1148 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1149 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1152 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1155 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1157 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1158 if (Result.getNode()) {
1162 StoreSDNode *Store = cast<StoreSDNode>(Op);
1163 SDValue Chain = Store->getChain();
1164 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1165 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1166 Store->getValue().getValueType().isVector()) {
1167 return SplitVectorStore(Op, DAG);
1170 EVT MemVT = Store->getMemoryVT();
1171 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1172 MemVT.bitsLT(MVT::i32)) {
1174 if (Store->getMemoryVT() == MVT::i8) {
1176 } else if (Store->getMemoryVT() == MVT::i16) {
1179 SDValue BasePtr = Store->getBasePtr();
1180 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1181 DAG.getConstant(2, MVT::i32));
1182 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1183 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1185 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1186 DAG.getConstant(0x3, MVT::i32));
1188 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1189 DAG.getConstant(3, MVT::i32));
1191 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1194 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1196 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1197 MaskedValue, ShiftAmt);
1199 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1201 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1202 DAG.getConstant(0xffffffff, MVT::i32));
1203 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1205 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1206 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1207 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1212 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1214 EVT OVT = Op.getValueType();
1215 SDValue LHS = Op.getOperand(0);
1216 SDValue RHS = Op.getOperand(1);
1219 if (!OVT.isVector()) {
1222 } else if (OVT.getVectorNumElements() == 2) {
1225 } else if (OVT.getVectorNumElements() == 4) {
1229 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1230 // char|short jq = ia ^ ib;
1231 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1233 // jq = jq >> (bitsize - 2)
1234 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1237 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1240 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1242 // int ia = (int)LHS;
1243 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1245 // int ib, (int)RHS;
1246 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1248 // float fa = (float)ia;
1249 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1251 // float fb = (float)ib;
1252 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1254 // float fq = native_divide(fa, fb);
1255 SDValue fq = DAG.getNode(ISD::FMUL, DL, FLTTY,
1256 fa, DAG.getNode(AMDGPUISD::RCP, DL, FLTTY, fb));
1259 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1261 // float fqneg = -fq;
1262 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1264 // float fr = mad(fqneg, fb, fa);
1265 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1266 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1268 // int iq = (int)fq;
1269 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1272 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1275 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1277 // int cv = fr >= fb;
1279 if (INTTY == MVT::i32) {
1280 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1282 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1284 // jq = (cv ? jq : 0);
1285 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1286 DAG.getConstant(0, OVT));
1288 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1289 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1293 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1295 EVT OVT = Op.getValueType();
1296 SDValue LHS = Op.getOperand(0);
1297 SDValue RHS = Op.getOperand(1);
1298 // The LowerSDIV32 function generates equivalent to the following IL.
1308 // ixor r10, r10, r11
1310 // ixor DST, r0, r10
1319 SDValue r10 = DAG.getSelectCC(DL,
1320 r0, DAG.getConstant(0, OVT),
1321 DAG.getConstant(-1, OVT),
1322 DAG.getConstant(0, OVT),
1326 SDValue r11 = DAG.getSelectCC(DL,
1327 r1, DAG.getConstant(0, OVT),
1328 DAG.getConstant(-1, OVT),
1329 DAG.getConstant(0, OVT),
1333 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1336 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1339 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1342 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1345 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1347 // ixor r10, r10, r11
1348 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1351 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1353 // ixor DST, r0, r10
1354 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1358 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1359 return SDValue(Op.getNode(), 0);
1362 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1363 EVT OVT = Op.getValueType().getScalarType();
1365 if (OVT == MVT::i64)
1366 return LowerSDIV64(Op, DAG);
1368 if (OVT.getScalarType() == MVT::i32)
1369 return LowerSDIV32(Op, DAG);
1371 if (OVT == MVT::i16 || OVT == MVT::i8) {
1372 // FIXME: We should be checking for the masked bits. This isn't reached
1373 // because i8 and i16 are not legal types.
1374 return LowerSDIV24(Op, DAG);
1377 return SDValue(Op.getNode(), 0);
1380 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1382 EVT OVT = Op.getValueType();
1383 SDValue LHS = Op.getOperand(0);
1384 SDValue RHS = Op.getOperand(1);
1385 // The LowerSREM32 function generates equivalent to the following IL.
1395 // umul r20, r20, r1
1398 // ixor DST, r0, r10
1407 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1410 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1413 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1416 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1419 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1422 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1425 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1427 // umul r20, r20, r1
1428 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1431 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1434 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1436 // ixor DST, r0, r10
1437 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1441 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1442 return SDValue(Op.getNode(), 0);
1445 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1446 EVT OVT = Op.getValueType();
1448 if (OVT.getScalarType() == MVT::i64)
1449 return LowerSREM64(Op, DAG);
1451 if (OVT.getScalarType() == MVT::i32)
1452 return LowerSREM32(Op, DAG);
1454 return SDValue(Op.getNode(), 0);
1457 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1458 SelectionDAG &DAG) const {
1460 EVT VT = Op.getValueType();
1462 SDValue Num = Op.getOperand(0);
1463 SDValue Den = Op.getOperand(1);
1465 // RCP = URECIP(Den) = 2^32 / Den + e
1466 // e is rounding error.
1467 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1469 // RCP_LO = umulo(RCP, Den) */
1470 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1472 // RCP_HI = mulhu (RCP, Den) */
1473 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1475 // NEG_RCP_LO = -RCP_LO
1476 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1479 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1480 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1483 // Calculate the rounding error from the URECIP instruction
1484 // E = mulhu(ABS_RCP_LO, RCP)
1485 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1487 // RCP_A_E = RCP + E
1488 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1490 // RCP_S_E = RCP - E
1491 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1493 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1494 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1497 // Quotient = mulhu(Tmp0, Num)
1498 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1500 // Num_S_Remainder = Quotient * Den
1501 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1503 // Remainder = Num - Num_S_Remainder
1504 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1506 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1507 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1508 DAG.getConstant(-1, VT),
1509 DAG.getConstant(0, VT),
1511 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1512 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1514 DAG.getConstant(-1, VT),
1515 DAG.getConstant(0, VT),
1517 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1518 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1521 // Calculate Division result:
1523 // Quotient_A_One = Quotient + 1
1524 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1525 DAG.getConstant(1, VT));
1527 // Quotient_S_One = Quotient - 1
1528 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1529 DAG.getConstant(1, VT));
1531 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1532 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1533 Quotient, Quotient_A_One, ISD::SETEQ);
1535 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1536 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1537 Quotient_S_One, Div, ISD::SETEQ);
1539 // Calculate Rem result:
1541 // Remainder_S_Den = Remainder - Den
1542 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1544 // Remainder_A_Den = Remainder + Den
1545 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1547 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1548 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1549 Remainder, Remainder_S_Den, ISD::SETEQ);
1551 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1552 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1553 Remainder_A_Den, Rem, ISD::SETEQ);
1558 return DAG.getMergeValues(Ops, DL);
1561 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1562 SelectionDAG &DAG) const {
1564 EVT VT = Op.getValueType();
1566 SDValue Zero = DAG.getConstant(0, VT);
1567 SDValue NegOne = DAG.getConstant(-1, VT);
1569 SDValue LHS = Op.getOperand(0);
1570 SDValue RHS = Op.getOperand(1);
1572 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1573 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1574 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1575 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1577 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1578 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1580 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1581 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1583 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1584 SDValue Rem = Div.getValue(1);
1586 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1587 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1589 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1590 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1596 return DAG.getMergeValues(Res, DL);
1599 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1601 SDValue Src = Op.getOperand(0);
1603 // result = trunc(src)
1604 // if (src > 0.0 && src != result)
1607 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1609 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1610 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1612 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1614 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1615 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1616 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1618 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1619 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1622 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1624 SDValue Src = Op.getOperand(0);
1626 assert(Op.getValueType() == MVT::f64);
1628 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1629 const SDValue One = DAG.getConstant(1, MVT::i32);
1631 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1633 // Extract the upper half, since this is where we will find the sign and
1635 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1637 const unsigned FractBits = 52;
1638 const unsigned ExpBits = 11;
1640 // Extract the exponent.
1641 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1643 DAG.getConstant(FractBits - 32, MVT::i32),
1644 DAG.getConstant(ExpBits, MVT::i32));
1645 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1646 DAG.getConstant(1023, MVT::i32));
1648 // Extract the sign bit.
1649 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1650 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1652 // Extend back to to 64-bits.
1653 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1655 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1657 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1658 const SDValue FractMask
1659 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1661 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1662 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1663 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1665 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1667 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1669 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1670 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1672 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1673 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1675 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1678 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1680 SDValue Src = Op.getOperand(0);
1682 assert(Op.getValueType() == MVT::f64);
1684 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1685 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1686 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1688 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1689 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1691 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1693 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1694 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1696 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1697 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1699 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1702 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1703 // FNEARBYINT and FRINT are the same, except in their handling of FP
1704 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1705 // rint, so just treat them as equivalent.
1706 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1709 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1711 SDValue Src = Op.getOperand(0);
1713 // result = trunc(src);
1714 // if (src < 0.0 && src != result)
1717 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1719 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1720 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1722 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1724 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1725 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1726 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1728 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1729 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1732 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1733 SelectionDAG &DAG) const {
1734 SDValue S0 = Op.getOperand(0);
1736 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1739 // f32 uint_to_fp i64
1740 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1741 DAG.getConstant(0, MVT::i32));
1742 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1743 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1744 DAG.getConstant(1, MVT::i32));
1745 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1746 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1747 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1748 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1751 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1753 SelectionDAG &DAG) const {
1754 MVT VT = Op.getSimpleValueType();
1756 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1757 // Shift left by 'Shift' bits.
1758 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1759 // Signed shift Right by 'Shift' bits.
1760 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1763 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1764 SelectionDAG &DAG) const {
1765 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1766 MVT VT = Op.getSimpleValueType();
1767 MVT ScalarVT = VT.getScalarType();
1772 SDValue Src = Op.getOperand(0);
1775 // TODO: Don't scalarize on Evergreen?
1776 unsigned NElts = VT.getVectorNumElements();
1777 SmallVector<SDValue, 8> Args;
1778 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1780 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1781 for (unsigned I = 0; I < NElts; ++I)
1782 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1784 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1787 //===----------------------------------------------------------------------===//
1788 // Custom DAG optimizations
1789 //===----------------------------------------------------------------------===//
1791 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1792 APInt KnownZero, KnownOne;
1793 EVT VT = Op.getValueType();
1794 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1796 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1799 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1800 EVT VT = Op.getValueType();
1802 // In order for this to be a signed 24-bit value, bit 23, must
1804 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1805 // as unsigned 24-bit values.
1806 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1809 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1811 SelectionDAG &DAG = DCI.DAG;
1812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1813 EVT VT = Op.getValueType();
1815 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1816 APInt KnownZero, KnownOne;
1817 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1818 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1819 DCI.CommitTargetLoweringOpt(TLO);
1822 template <typename IntTy>
1823 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1824 uint32_t Offset, uint32_t Width) {
1825 if (Width + Offset < 32) {
1826 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1827 return DAG.getConstant(Result, MVT::i32);
1830 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1833 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1834 DAGCombinerInfo &DCI) const {
1835 SelectionDAG &DAG = DCI.DAG;
1838 switch(N->getOpcode()) {
1841 EVT VT = N->getValueType(0);
1842 SDValue N0 = N->getOperand(0);
1843 SDValue N1 = N->getOperand(1);
1846 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1847 if (VT.isVector() || VT.getSizeInBits() > 32)
1850 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1851 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1852 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1853 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1854 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1855 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1856 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1857 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1862 // We need to use sext even for MUL_U24, because MUL_U24 is used
1863 // for signed multiply of 8 and 16-bit types.
1864 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1868 case AMDGPUISD::MUL_I24:
1869 case AMDGPUISD::MUL_U24: {
1870 SDValue N0 = N->getOperand(0);
1871 SDValue N1 = N->getOperand(1);
1872 simplifyI24(N0, DCI);
1873 simplifyI24(N1, DCI);
1876 case ISD::SELECT_CC: {
1877 return CombineMinMax(N, DAG);
1879 case AMDGPUISD::BFE_I32:
1880 case AMDGPUISD::BFE_U32: {
1881 assert(!N->getValueType(0).isVector() &&
1882 "Vector handling of BFE not implemented");
1883 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1887 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1889 return DAG.getConstant(0, MVT::i32);
1891 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1895 SDValue BitsFrom = N->getOperand(0);
1896 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1898 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1900 if (OffsetVal == 0) {
1901 // This is already sign / zero extended, so try to fold away extra BFEs.
1902 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1904 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1905 if (OpSignBits >= SignBits)
1908 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1910 // This is a sign_extend_inreg. Replace it to take advantage of existing
1911 // DAG Combines. If not eliminated, we will match back to BFE during
1914 // TODO: The sext_inreg of extended types ends, although we can could
1915 // handle them in a single BFE.
1916 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1917 DAG.getValueType(SmallVT));
1920 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1923 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1925 return constantFoldBFE<int32_t>(DAG,
1926 Val->getSExtValue(),
1931 return constantFoldBFE<uint32_t>(DAG,
1932 Val->getZExtValue(),
1937 APInt Demanded = APInt::getBitsSet(32,
1939 OffsetVal + WidthVal);
1941 if ((OffsetVal + WidthVal) >= 32) {
1942 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1943 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1944 BitsFrom, ShiftVal);
1947 APInt KnownZero, KnownOne;
1948 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1949 !DCI.isBeforeLegalizeOps());
1950 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1951 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1952 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1953 DCI.CommitTargetLoweringOpt(TLO);
1962 //===----------------------------------------------------------------------===//
1964 //===----------------------------------------------------------------------===//
1966 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1969 const SmallVectorImpl<ISD::InputArg> &Ins,
1970 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1972 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1973 if (Ins[i].ArgVT == Ins[i].VT) {
1974 OrigIns.push_back(Ins[i]);
1979 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1980 // Vector has been split into scalars.
1981 VT = Ins[i].ArgVT.getVectorElementType();
1982 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1983 Ins[i].ArgVT.getVectorElementType() !=
1984 Ins[i].VT.getVectorElementType()) {
1985 // Vector elements have been promoted
1988 // Vector has been spilt into smaller vectors.
1992 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1993 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1994 OrigIns.push_back(Arg);
1998 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1999 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2000 return CFP->isExactlyValue(1.0);
2002 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2003 return C->isAllOnesValue();
2008 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2009 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2010 return CFP->getValueAPF().isZero();
2012 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2013 return C->isNullValue();
2018 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2019 const TargetRegisterClass *RC,
2020 unsigned Reg, EVT VT) const {
2021 MachineFunction &MF = DAG.getMachineFunction();
2022 MachineRegisterInfo &MRI = MF.getRegInfo();
2023 unsigned VirtualRegister;
2024 if (!MRI.isLiveIn(Reg)) {
2025 VirtualRegister = MRI.createVirtualRegister(RC);
2026 MRI.addLiveIn(Reg, VirtualRegister);
2028 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2030 return DAG.getRegister(VirtualRegister, VT);
2033 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2035 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2037 default: return nullptr;
2039 NODE_NAME_CASE(CALL);
2040 NODE_NAME_CASE(UMUL);
2041 NODE_NAME_CASE(RET_FLAG);
2042 NODE_NAME_CASE(BRANCH_COND);
2045 NODE_NAME_CASE(DWORDADDR)
2046 NODE_NAME_CASE(FRACT)
2047 NODE_NAME_CASE(CLAMP)
2048 NODE_NAME_CASE(FMAX)
2049 NODE_NAME_CASE(SMAX)
2050 NODE_NAME_CASE(UMAX)
2051 NODE_NAME_CASE(FMIN)
2052 NODE_NAME_CASE(SMIN)
2053 NODE_NAME_CASE(UMIN)
2054 NODE_NAME_CASE(URECIP)
2055 NODE_NAME_CASE(DIV_SCALE)
2056 NODE_NAME_CASE(DIV_FMAS)
2057 NODE_NAME_CASE(DIV_FIXUP)
2058 NODE_NAME_CASE(TRIG_PREOP)
2061 NODE_NAME_CASE(RSQ_LEGACY)
2062 NODE_NAME_CASE(RSQ_CLAMPED)
2063 NODE_NAME_CASE(DOT4)
2064 NODE_NAME_CASE(BFE_U32)
2065 NODE_NAME_CASE(BFE_I32)
2068 NODE_NAME_CASE(BREV)
2069 NODE_NAME_CASE(MUL_U24)
2070 NODE_NAME_CASE(MUL_I24)
2071 NODE_NAME_CASE(MAD_U24)
2072 NODE_NAME_CASE(MAD_I24)
2073 NODE_NAME_CASE(EXPORT)
2074 NODE_NAME_CASE(CONST_ADDRESS)
2075 NODE_NAME_CASE(REGISTER_LOAD)
2076 NODE_NAME_CASE(REGISTER_STORE)
2077 NODE_NAME_CASE(LOAD_CONSTANT)
2078 NODE_NAME_CASE(LOAD_INPUT)
2079 NODE_NAME_CASE(SAMPLE)
2080 NODE_NAME_CASE(SAMPLEB)
2081 NODE_NAME_CASE(SAMPLED)
2082 NODE_NAME_CASE(SAMPLEL)
2083 NODE_NAME_CASE(CVT_F32_UBYTE0)
2084 NODE_NAME_CASE(CVT_F32_UBYTE1)
2085 NODE_NAME_CASE(CVT_F32_UBYTE2)
2086 NODE_NAME_CASE(CVT_F32_UBYTE3)
2087 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2088 NODE_NAME_CASE(STORE_MSKOR)
2089 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2093 static void computeKnownBitsForMinMax(const SDValue Op0,
2097 const SelectionDAG &DAG,
2099 APInt Op0Zero, Op0One;
2100 APInt Op1Zero, Op1One;
2101 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2102 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2104 KnownZero = Op0Zero & Op1Zero;
2105 KnownOne = Op0One & Op1One;
2108 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2112 const SelectionDAG &DAG,
2113 unsigned Depth) const {
2115 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2119 unsigned Opc = Op.getOpcode();
2124 case ISD::INTRINSIC_WO_CHAIN: {
2125 // FIXME: The intrinsic should just use the node.
2126 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2127 case AMDGPUIntrinsic::AMDGPU_imax:
2128 case AMDGPUIntrinsic::AMDGPU_umax:
2129 case AMDGPUIntrinsic::AMDGPU_imin:
2130 case AMDGPUIntrinsic::AMDGPU_umin:
2131 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2132 KnownZero, KnownOne, DAG, Depth);
2140 case AMDGPUISD::SMAX:
2141 case AMDGPUISD::UMAX:
2142 case AMDGPUISD::SMIN:
2143 case AMDGPUISD::UMIN:
2144 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2145 KnownZero, KnownOne, DAG, Depth);
2148 case AMDGPUISD::BFE_I32:
2149 case AMDGPUISD::BFE_U32: {
2150 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2154 unsigned BitWidth = 32;
2155 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2157 KnownZero = APInt::getAllOnesValue(BitWidth);
2158 KnownOne = APInt::getNullValue(BitWidth);
2162 // FIXME: This could do a lot more. If offset is 0, should be the same as
2163 // sign_extend_inreg implementation, but that involves duplicating it.
2164 if (Opc == AMDGPUISD::BFE_I32)
2165 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2167 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2174 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2176 const SelectionDAG &DAG,
2177 unsigned Depth) const {
2178 switch (Op.getOpcode()) {
2179 case AMDGPUISD::BFE_I32: {
2180 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2184 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2185 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2186 if (!Offset || !Offset->isNullValue())
2189 // TODO: Could probably figure something out with non-0 offsets.
2190 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2191 return std::max(SignBits, Op0SignBits);
2194 case AMDGPUISD::BFE_U32: {
2195 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2196 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;