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 "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDILIntrinsicInfo.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 // Initialize target lowering borrowed from AMDIL
114 // We need to custom lower some of the intrinsics
115 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
117 // Library functions. These default to Expand, but we have instructions
119 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
120 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
121 setOperationAction(ISD::FPOW, MVT::f32, Legal);
122 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
123 setOperationAction(ISD::FABS, MVT::f32, Legal);
124 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
125 setOperationAction(ISD::FRINT, MVT::f32, Legal);
126 setOperationAction(ISD::FROUND, MVT::f32, Legal);
127 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
129 // Lower floating point store/load to integer store/load to reduce the number
130 // of patterns in tablegen.
131 setOperationAction(ISD::STORE, MVT::f32, Promote);
132 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
134 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
135 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
137 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
138 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
140 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
141 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
143 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
144 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
146 setOperationAction(ISD::STORE, MVT::f64, Promote);
147 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
149 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
150 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
152 // Custom lowering of vector stores is required for local address space
154 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
155 // XXX: Native v2i32 local address space stores are possible, but not
156 // currently implemented.
157 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
159 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
160 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
161 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
163 // XXX: This can be change to Custom, once ExpandVectorStores can
164 // handle 64-bit stores.
165 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
167 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
168 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
169 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
170 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
171 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
174 setOperationAction(ISD::LOAD, MVT::f32, Promote);
175 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
177 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
178 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
180 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
181 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
183 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
184 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
186 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
187 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
189 setOperationAction(ISD::LOAD, MVT::f64, Promote);
190 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
192 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
193 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
195 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
196 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
197 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
198 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
199 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
200 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
201 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
202 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
203 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
204 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
206 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
207 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
208 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
209 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
210 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
211 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
212 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
213 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
214 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
215 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
217 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
219 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
221 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
222 setOperationAction(ISD::FRINT, MVT::f64, Custom);
225 if (!Subtarget->hasBFI()) {
226 // fcopysign can be done in a single instruction with BFI.
227 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
228 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
231 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
232 for (MVT VT : ScalarIntVTs) {
233 setOperationAction(ISD::SREM, VT, Expand);
234 setOperationAction(ISD::SDIV, VT, Custom);
236 // GPU does not have divrem function for signed or unsigned.
237 setOperationAction(ISD::SDIVREM, VT, Expand);
238 setOperationAction(ISD::UDIVREM, VT, Custom);
240 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
241 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
242 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
244 setOperationAction(ISD::BSWAP, VT, Expand);
245 setOperationAction(ISD::CTTZ, VT, Expand);
246 setOperationAction(ISD::CTLZ, VT, Expand);
249 if (!Subtarget->hasBCNT(32))
250 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
252 if (!Subtarget->hasBCNT(64))
253 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
255 // The hardware supports 32-bit ROTR, but not ROTL.
256 setOperationAction(ISD::ROTL, MVT::i32, Expand);
257 setOperationAction(ISD::ROTL, MVT::i64, Expand);
258 setOperationAction(ISD::ROTR, MVT::i64, Expand);
260 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
261 setOperationAction(ISD::MUL, MVT::i64, Expand);
262 setOperationAction(ISD::MULHU, MVT::i64, Expand);
263 setOperationAction(ISD::MULHS, MVT::i64, Expand);
264 setOperationAction(ISD::SUB, MVT::i64, Expand);
265 setOperationAction(ISD::UDIV, MVT::i32, Expand);
266 setOperationAction(ISD::UREM, MVT::i32, Expand);
267 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
268 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
270 static const MVT::SimpleValueType VectorIntTypes[] = {
271 MVT::v2i32, MVT::v4i32
274 for (MVT VT : VectorIntTypes) {
275 // Expand the following operations for the current type by default.
276 setOperationAction(ISD::ADD, VT, Expand);
277 setOperationAction(ISD::AND, VT, Expand);
278 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
279 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
280 setOperationAction(ISD::MUL, VT, Expand);
281 setOperationAction(ISD::OR, VT, Expand);
282 setOperationAction(ISD::SHL, VT, Expand);
283 setOperationAction(ISD::SRA, VT, Expand);
284 setOperationAction(ISD::SRL, VT, Expand);
285 setOperationAction(ISD::ROTL, VT, Expand);
286 setOperationAction(ISD::ROTR, VT, Expand);
287 setOperationAction(ISD::SUB, VT, Expand);
288 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
289 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
290 // TODO: Implement custom UREM / SREM routines.
291 setOperationAction(ISD::SDIV, VT, Custom);
292 setOperationAction(ISD::UDIV, VT, Expand);
293 setOperationAction(ISD::SREM, VT, Expand);
294 setOperationAction(ISD::UREM, VT, Expand);
295 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
296 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
297 setOperationAction(ISD::SDIVREM, VT, Expand);
298 setOperationAction(ISD::UDIVREM, VT, Custom);
299 setOperationAction(ISD::SELECT, VT, Expand);
300 setOperationAction(ISD::VSELECT, VT, Expand);
301 setOperationAction(ISD::XOR, VT, Expand);
302 setOperationAction(ISD::BSWAP, VT, Expand);
303 setOperationAction(ISD::CTPOP, VT, Expand);
304 setOperationAction(ISD::CTTZ, VT, Expand);
305 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
306 setOperationAction(ISD::CTLZ, VT, Expand);
307 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
310 static const MVT::SimpleValueType FloatVectorTypes[] = {
311 MVT::v2f32, MVT::v4f32
314 for (MVT VT : FloatVectorTypes) {
315 setOperationAction(ISD::FABS, VT, Expand);
316 setOperationAction(ISD::FADD, VT, Expand);
317 setOperationAction(ISD::FCOS, VT, Expand);
318 setOperationAction(ISD::FDIV, VT, Expand);
319 setOperationAction(ISD::FPOW, VT, Expand);
320 setOperationAction(ISD::FFLOOR, VT, Expand);
321 setOperationAction(ISD::FTRUNC, VT, Expand);
322 setOperationAction(ISD::FMUL, VT, Expand);
323 setOperationAction(ISD::FRINT, VT, Expand);
324 setOperationAction(ISD::FSQRT, VT, Expand);
325 setOperationAction(ISD::FSIN, VT, Expand);
326 setOperationAction(ISD::FSUB, VT, Expand);
327 setOperationAction(ISD::FNEG, VT, Expand);
328 setOperationAction(ISD::SELECT, VT, Expand);
329 setOperationAction(ISD::VSELECT, VT, Expand);
330 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
333 setTargetDAGCombine(ISD::MUL);
334 setTargetDAGCombine(ISD::SELECT_CC);
336 setSchedulingPreference(Sched::RegPressure);
337 setJumpIsExpensive(true);
339 // There are no integer divide instructions, and these expand to a pretty
340 // large sequence of instructions.
341 setIntDivIsCheap(false);
343 // TODO: Investigate this when 64-bit divides are implemented.
344 addBypassSlowDiv(64, 32);
346 // FIXME: Need to really handle these.
347 MaxStoresPerMemcpy = 4096;
348 MaxStoresPerMemmove = 4096;
349 MaxStoresPerMemset = 4096;
352 //===----------------------------------------------------------------------===//
353 // Target Information
354 //===----------------------------------------------------------------------===//
356 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
360 // The backend supports 32 and 64 bit floating point immediates.
361 // FIXME: Why are we reporting vectors of FP immediates as legal?
362 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
363 EVT ScalarVT = VT.getScalarType();
364 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
367 // We don't want to shrink f64 / f32 constants.
368 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
369 EVT ScalarVT = VT.getScalarType();
370 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
373 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
375 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
378 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
379 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
381 return ((LScalarSize <= CastScalarSize) ||
382 (CastScalarSize >= 32) ||
386 //===---------------------------------------------------------------------===//
388 //===---------------------------------------------------------------------===//
390 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
391 assert(VT.isFloatingPoint());
392 return VT == MVT::f32;
395 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
396 assert(VT.isFloatingPoint());
397 return VT == MVT::f32;
400 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
401 // Truncate is just accessing a subregister.
402 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
405 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
406 // Truncate is just accessing a subregister.
407 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
408 (Dest->getPrimitiveSizeInBits() % 32 == 0);
411 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
412 const DataLayout *DL = getDataLayout();
413 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
414 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
416 return SrcSize == 32 && DestSize == 64;
419 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
420 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
421 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
422 // this will enable reducing 64-bit operations the 32-bit, which is always
424 return Src == MVT::i32 && Dest == MVT::i64;
427 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
428 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
429 // limited number of native 64-bit operations. Shrinking an operation to fit
430 // in a single 32-bit register should always be helpful. As currently used,
431 // this is much less general than the name suggests, and is only used in
432 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
433 // not profitable, and may actually be harmful.
434 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
437 //===---------------------------------------------------------------------===//
438 // TargetLowering Callbacks
439 //===---------------------------------------------------------------------===//
441 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
442 const SmallVectorImpl<ISD::InputArg> &Ins) const {
444 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
447 SDValue AMDGPUTargetLowering::LowerReturn(
449 CallingConv::ID CallConv,
451 const SmallVectorImpl<ISD::OutputArg> &Outs,
452 const SmallVectorImpl<SDValue> &OutVals,
453 SDLoc DL, SelectionDAG &DAG) const {
454 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
457 //===---------------------------------------------------------------------===//
458 // Target specific lowering
459 //===---------------------------------------------------------------------===//
461 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
462 SmallVectorImpl<SDValue> &InVals) const {
463 SDValue Callee = CLI.Callee;
464 SelectionDAG &DAG = CLI.DAG;
466 const Function &Fn = *DAG.getMachineFunction().getFunction();
468 StringRef FuncName("<unknown>");
470 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
471 FuncName = G->getSymbol();
472 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
473 FuncName = G->getGlobal()->getName();
475 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
476 DAG.getContext()->diagnose(NoCalls);
480 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
481 SelectionDAG &DAG) const {
482 switch (Op.getOpcode()) {
484 Op.getNode()->dump();
485 llvm_unreachable("Custom lowering code for this"
486 "instruction is not implemented yet!");
488 // AMDGPU DAG lowering.
489 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
490 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
491 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
492 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
493 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
494 case ISD::SDIV: return LowerSDIV(Op, DAG);
495 case ISD::SREM: return LowerSREM(Op, DAG);
496 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
497 case ISD::FRINT: return LowerFRINT(Op, DAG);
498 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
500 // AMDIL DAG lowering.
501 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
506 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
507 SmallVectorImpl<SDValue> &Results,
508 SelectionDAG &DAG) const {
509 switch (N->getOpcode()) {
510 case ISD::SIGN_EXTEND_INREG:
511 // Different parts of legalization seem to interpret which type of
512 // sign_extend_inreg is the one to check for custom lowering. The extended
513 // from type is what really matters, but some places check for custom
514 // lowering of the result type. This results in trying to use
515 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
516 // nothing here and let the illegal result integer be handled normally.
519 SDValue Op = SDValue(N, 0);
521 EVT VT = Op.getValueType();
522 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
523 N->getOperand(0), N->getOperand(1));
524 Results.push_back(UDIVREM);
528 SDValue Op = SDValue(N, 0);
530 EVT VT = Op.getValueType();
531 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
532 N->getOperand(0), N->getOperand(1));
533 Results.push_back(UDIVREM.getValue(1));
537 SDValue Op = SDValue(N, 0);
539 EVT VT = Op.getValueType();
540 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
542 SDValue one = DAG.getConstant(1, HalfVT);
543 SDValue zero = DAG.getConstant(0, HalfVT);
546 SDValue LHS = N->getOperand(0);
547 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
548 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
550 SDValue RHS = N->getOperand(1);
551 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
552 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
554 // Get Speculative values
555 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
556 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
558 SDValue REM_Hi = zero;
559 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
561 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
562 SDValue DIV_Lo = zero;
564 const unsigned halfBitWidth = HalfVT.getSizeInBits();
566 for (unsigned i = 0; i < halfBitWidth; ++i) {
567 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
568 // Get Value of high bit
570 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
571 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
573 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
574 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
577 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
578 DAG.getConstant(halfBitWidth - 1, HalfVT));
579 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
580 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
582 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
583 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
586 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
588 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
589 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
591 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
595 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
597 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
598 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
599 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
602 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
603 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
604 Results.push_back(DIV);
605 Results.push_back(REM);
613 // FIXME: This implements accesses to initialized globals in the constant
614 // address space by copying them to private and accessing that. It does not
615 // properly handle illegal types or vectors. The private vector loads are not
616 // scalarized, and the illegal scalars hit an assertion. This technique will not
617 // work well with large initializers, and this should eventually be
618 // removed. Initialized globals should be placed into a data section that the
619 // runtime will load into a buffer before the kernel is executed. Uses of the
620 // global need to be replaced with a pointer loaded from an implicit kernel
621 // argument into this buffer holding the copy of the data, which will remove the
622 // need for any of this.
623 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
624 const GlobalValue *GV,
625 const SDValue &InitPtr,
627 SelectionDAG &DAG) const {
628 const DataLayout *TD = getTargetMachine().getDataLayout();
630 Type *InitTy = Init->getType();
632 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
633 EVT VT = EVT::getEVT(InitTy);
634 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
635 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
636 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
637 TD->getPrefTypeAlignment(InitTy));
640 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
641 EVT VT = EVT::getEVT(CFP->getType());
642 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
643 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
644 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
645 TD->getPrefTypeAlignment(CFP->getType()));
648 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
649 const StructLayout *SL = TD->getStructLayout(ST);
651 EVT PtrVT = InitPtr.getValueType();
652 SmallVector<SDValue, 8> Chains;
654 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
655 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
656 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
658 Constant *Elt = Init->getAggregateElement(I);
659 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
662 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
665 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
666 EVT PtrVT = InitPtr.getValueType();
668 unsigned NumElements;
669 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
670 NumElements = AT->getNumElements();
671 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
672 NumElements = VT->getNumElements();
674 llvm_unreachable("Unexpected type");
676 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
677 SmallVector<SDValue, 8> Chains;
678 for (unsigned i = 0; i < NumElements; ++i) {
679 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
680 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
682 Constant *Elt = Init->getAggregateElement(i);
683 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
686 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
689 if (isa<UndefValue>(Init)) {
690 EVT VT = EVT::getEVT(InitTy);
691 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
692 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
693 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
694 TD->getPrefTypeAlignment(InitTy));
698 llvm_unreachable("Unhandled constant initializer");
701 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
703 SelectionDAG &DAG) const {
705 const DataLayout *TD = getTargetMachine().getDataLayout();
706 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
707 const GlobalValue *GV = G->getGlobal();
709 switch (G->getAddressSpace()) {
710 default: llvm_unreachable("Global Address lowering not implemented for this "
712 case AMDGPUAS::LOCAL_ADDRESS: {
713 // XXX: What does the value of G->getOffset() mean?
714 assert(G->getOffset() == 0 &&
715 "Do not know what to do with an non-zero offset");
718 if (MFI->LocalMemoryObjects.count(GV) == 0) {
719 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
720 Offset = MFI->LDSSize;
721 MFI->LocalMemoryObjects[GV] = Offset;
722 // XXX: Account for alignment?
723 MFI->LDSSize += Size;
725 Offset = MFI->LocalMemoryObjects[GV];
728 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
730 case AMDGPUAS::CONSTANT_ADDRESS: {
731 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
732 Type *EltType = GV->getType()->getElementType();
733 unsigned Size = TD->getTypeAllocSize(EltType);
734 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
736 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
737 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
739 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
740 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
742 const GlobalVariable *Var = cast<GlobalVariable>(GV);
743 if (!Var->hasInitializer()) {
744 // This has no use, but bugpoint will hit it.
745 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
748 const Constant *Init = Var->getInitializer();
749 SmallVector<SDNode*, 8> WorkList;
751 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
752 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
753 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
755 WorkList.push_back(*I);
757 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
758 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
759 E = WorkList.end(); I != E; ++I) {
760 SmallVector<SDValue, 8> Ops;
761 Ops.push_back(Chain);
762 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
763 Ops.push_back((*I)->getOperand(i));
765 DAG.UpdateNodeOperands(*I, Ops);
767 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
772 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
773 SelectionDAG &DAG) const {
774 SmallVector<SDValue, 8> Args;
775 SDValue A = Op.getOperand(0);
776 SDValue B = Op.getOperand(1);
778 DAG.ExtractVectorElements(A, Args);
779 DAG.ExtractVectorElements(B, Args);
781 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
784 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
785 SelectionDAG &DAG) const {
787 SmallVector<SDValue, 8> Args;
788 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
789 EVT VT = Op.getValueType();
790 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
791 VT.getVectorNumElements());
793 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
796 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
797 SelectionDAG &DAG) const {
799 MachineFunction &MF = DAG.getMachineFunction();
800 const AMDGPUFrameLowering *TFL =
801 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
803 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
805 unsigned FrameIndex = FIN->getIndex();
806 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
807 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
811 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
812 SelectionDAG &DAG) const {
813 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
815 EVT VT = Op.getValueType();
817 switch (IntrinsicID) {
819 case AMDGPUIntrinsic::AMDGPU_abs:
820 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
821 return LowerIntrinsicIABS(Op, DAG);
822 case AMDGPUIntrinsic::AMDGPU_lrp:
823 return LowerIntrinsicLRP(Op, DAG);
824 case AMDGPUIntrinsic::AMDGPU_fract:
825 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
826 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
828 case AMDGPUIntrinsic::AMDGPU_clamp:
829 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
830 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
831 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
833 case AMDGPUIntrinsic::AMDGPU_imax:
834 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
836 case AMDGPUIntrinsic::AMDGPU_umax:
837 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
839 case AMDGPUIntrinsic::AMDGPU_imin:
840 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
842 case AMDGPUIntrinsic::AMDGPU_umin:
843 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
846 case AMDGPUIntrinsic::AMDGPU_umul24:
847 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
848 Op.getOperand(1), Op.getOperand(2));
850 case AMDGPUIntrinsic::AMDGPU_imul24:
851 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
852 Op.getOperand(1), Op.getOperand(2));
854 case AMDGPUIntrinsic::AMDGPU_umad24:
855 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
856 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
858 case AMDGPUIntrinsic::AMDGPU_imad24:
859 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
860 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
862 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
863 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
865 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
866 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
868 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
869 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
871 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
872 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
874 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
875 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
880 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
881 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
886 case AMDGPUIntrinsic::AMDGPU_bfi:
887 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
892 case AMDGPUIntrinsic::AMDGPU_bfm:
893 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
897 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
898 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
900 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
901 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
905 ///IABS(a) = SMAX(sub(0, a), a)
906 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
907 SelectionDAG &DAG) const {
909 EVT VT = Op.getValueType();
910 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
913 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
916 /// Linear Interpolation
917 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
918 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
919 SelectionDAG &DAG) const {
921 EVT VT = Op.getValueType();
922 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
923 DAG.getConstantFP(1.0f, MVT::f32),
925 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
927 return DAG.getNode(ISD::FADD, DL, VT,
928 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
932 /// \brief Generate Min/Max node
933 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
934 SelectionDAG &DAG) const {
936 EVT VT = N->getValueType(0);
938 SDValue LHS = N->getOperand(0);
939 SDValue RHS = N->getOperand(1);
940 SDValue True = N->getOperand(2);
941 SDValue False = N->getOperand(3);
942 SDValue CC = N->getOperand(4);
944 if (VT != MVT::f32 ||
945 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
949 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
963 llvm_unreachable("Operation should already be optimised!");
970 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
971 return DAG.getNode(Opc, DL, VT, LHS, RHS);
979 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
980 return DAG.getNode(Opc, DL, VT, LHS, RHS);
982 case ISD::SETCC_INVALID:
983 llvm_unreachable("Invalid setcc condcode!");
988 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
989 SelectionDAG &DAG) const {
990 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
991 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
992 EVT EltVT = Op.getValueType().getVectorElementType();
993 EVT PtrVT = Load->getBasePtr().getValueType();
994 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
995 SmallVector<SDValue, 8> Loads;
998 for (unsigned i = 0, e = NumElts; i != e; ++i) {
999 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1000 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
1001 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1002 Load->getChain(), Ptr,
1003 MachinePointerInfo(Load->getMemOperand()->getValue()),
1004 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1005 Load->getAlignment()));
1007 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
1010 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1011 SelectionDAG &DAG) const {
1012 StoreSDNode *Store = cast<StoreSDNode>(Op);
1013 EVT MemVT = Store->getMemoryVT();
1014 unsigned MemBits = MemVT.getSizeInBits();
1016 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1017 // truncating store into an i32 store.
1018 // XXX: We could also handle optimize other vector bitwidths.
1019 if (!MemVT.isVector() || MemBits > 32) {
1024 SDValue Value = Store->getValue();
1025 EVT VT = Value.getValueType();
1026 EVT ElemVT = VT.getVectorElementType();
1027 SDValue Ptr = Store->getBasePtr();
1028 EVT MemEltVT = MemVT.getVectorElementType();
1029 unsigned MemEltBits = MemEltVT.getSizeInBits();
1030 unsigned MemNumElements = MemVT.getVectorNumElements();
1031 unsigned PackedSize = MemVT.getStoreSizeInBits();
1032 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1034 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1036 SDValue PackedValue;
1037 for (unsigned i = 0; i < MemNumElements; ++i) {
1038 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1039 DAG.getConstant(i, MVT::i32));
1040 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1041 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1043 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1044 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1049 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1053 if (PackedSize < 32) {
1054 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1055 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1056 Store->getMemOperand()->getPointerInfo(),
1058 Store->isNonTemporal(), Store->isVolatile(),
1059 Store->getAlignment());
1062 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1063 Store->getMemOperand()->getPointerInfo(),
1064 Store->isVolatile(), Store->isNonTemporal(),
1065 Store->getAlignment());
1068 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1069 SelectionDAG &DAG) const {
1070 StoreSDNode *Store = cast<StoreSDNode>(Op);
1071 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1072 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1073 EVT PtrVT = Store->getBasePtr().getValueType();
1074 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1077 SmallVector<SDValue, 8> Chains;
1079 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1080 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1081 Store->getValue(), DAG.getConstant(i, MVT::i32));
1082 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1083 Store->getBasePtr(),
1084 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1086 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1087 MachinePointerInfo(Store->getMemOperand()->getValue()),
1088 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1089 Store->getAlignment()));
1091 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1094 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1096 LoadSDNode *Load = cast<LoadSDNode>(Op);
1097 ISD::LoadExtType ExtType = Load->getExtensionType();
1098 EVT VT = Op.getValueType();
1099 EVT MemVT = Load->getMemoryVT();
1101 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1102 // We can do the extload to 32-bits, and then need to separately extend to
1105 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1109 Load->getMemOperand());
1110 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1113 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1114 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1115 // FIXME: Copied from PPC
1116 // First, load into 32 bits, then truncate to 1 bit.
1118 SDValue Chain = Load->getChain();
1119 SDValue BasePtr = Load->getBasePtr();
1120 MachineMemOperand *MMO = Load->getMemOperand();
1122 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1123 BasePtr, MVT::i8, MMO);
1124 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1127 // Lower loads constant address space global variable loads
1128 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1129 isa<GlobalVariable>(
1130 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1132 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1133 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1134 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1135 DAG.getConstant(2, MVT::i32));
1136 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1137 Load->getChain(), Ptr,
1138 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1141 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1142 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1146 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1147 DAG.getConstant(2, MVT::i32));
1148 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1149 Load->getChain(), Ptr,
1150 DAG.getTargetConstant(0, MVT::i32),
1152 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1154 DAG.getConstant(0x3, MVT::i32));
1155 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1156 DAG.getConstant(3, MVT::i32));
1158 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1160 EVT MemEltVT = MemVT.getScalarType();
1161 if (ExtType == ISD::SEXTLOAD) {
1162 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1163 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1166 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1169 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1171 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1172 if (Result.getNode()) {
1176 StoreSDNode *Store = cast<StoreSDNode>(Op);
1177 SDValue Chain = Store->getChain();
1178 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1179 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1180 Store->getValue().getValueType().isVector()) {
1181 return SplitVectorStore(Op, DAG);
1184 EVT MemVT = Store->getMemoryVT();
1185 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1186 MemVT.bitsLT(MVT::i32)) {
1188 if (Store->getMemoryVT() == MVT::i8) {
1190 } else if (Store->getMemoryVT() == MVT::i16) {
1193 SDValue BasePtr = Store->getBasePtr();
1194 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1195 DAG.getConstant(2, MVT::i32));
1196 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1197 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1199 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1200 DAG.getConstant(0x3, MVT::i32));
1202 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1203 DAG.getConstant(3, MVT::i32));
1205 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1208 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1210 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1211 MaskedValue, ShiftAmt);
1213 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1215 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1216 DAG.getConstant(0xffffffff, MVT::i32));
1217 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1219 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1220 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1221 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1226 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1228 EVT OVT = Op.getValueType();
1229 SDValue LHS = Op.getOperand(0);
1230 SDValue RHS = Op.getOperand(1);
1233 if (!OVT.isVector()) {
1236 } else if (OVT.getVectorNumElements() == 2) {
1239 } else if (OVT.getVectorNumElements() == 4) {
1243 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1244 // char|short jq = ia ^ ib;
1245 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1247 // jq = jq >> (bitsize - 2)
1248 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1251 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1254 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1256 // int ia = (int)LHS;
1257 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1259 // int ib, (int)RHS;
1260 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1262 // float fa = (float)ia;
1263 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1265 // float fb = (float)ib;
1266 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1268 // float fq = native_divide(fa, fb);
1269 SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
1272 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1274 // float fqneg = -fq;
1275 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1277 // float fr = mad(fqneg, fb, fa);
1278 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1279 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1281 // int iq = (int)fq;
1282 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1285 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1288 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1290 // int cv = fr >= fb;
1292 if (INTTY == MVT::i32) {
1293 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1295 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1297 // jq = (cv ? jq : 0);
1298 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1299 DAG.getConstant(0, OVT));
1301 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1302 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1306 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1308 EVT OVT = Op.getValueType();
1309 SDValue LHS = Op.getOperand(0);
1310 SDValue RHS = Op.getOperand(1);
1311 // The LowerSDIV32 function generates equivalent to the following IL.
1321 // ixor r10, r10, r11
1323 // ixor DST, r0, r10
1332 SDValue r10 = DAG.getSelectCC(DL,
1333 r0, DAG.getConstant(0, OVT),
1334 DAG.getConstant(-1, OVT),
1335 DAG.getConstant(0, OVT),
1339 SDValue r11 = DAG.getSelectCC(DL,
1340 r1, DAG.getConstant(0, OVT),
1341 DAG.getConstant(-1, OVT),
1342 DAG.getConstant(0, OVT),
1346 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1349 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1352 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1355 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1358 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1360 // ixor r10, r10, r11
1361 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1364 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1366 // ixor DST, r0, r10
1367 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1371 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1372 return SDValue(Op.getNode(), 0);
1375 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1376 EVT OVT = Op.getValueType().getScalarType();
1378 if (OVT == MVT::i64)
1379 return LowerSDIV64(Op, DAG);
1381 if (OVT.getScalarType() == MVT::i32)
1382 return LowerSDIV32(Op, DAG);
1384 if (OVT == MVT::i16 || OVT == MVT::i8) {
1385 // FIXME: We should be checking for the masked bits. This isn't reached
1386 // because i8 and i16 are not legal types.
1387 return LowerSDIV24(Op, DAG);
1390 return SDValue(Op.getNode(), 0);
1393 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1395 EVT OVT = Op.getValueType();
1396 SDValue LHS = Op.getOperand(0);
1397 SDValue RHS = Op.getOperand(1);
1398 // The LowerSREM32 function generates equivalent to the following IL.
1408 // umul r20, r20, r1
1411 // ixor DST, r0, r10
1420 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1423 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1426 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1429 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1432 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1435 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1438 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1440 // umul r20, r20, r1
1441 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1444 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1447 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1449 // ixor DST, r0, r10
1450 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1454 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1455 return SDValue(Op.getNode(), 0);
1458 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1459 EVT OVT = Op.getValueType();
1461 if (OVT.getScalarType() == MVT::i64)
1462 return LowerSREM64(Op, DAG);
1464 if (OVT.getScalarType() == MVT::i32)
1465 return LowerSREM32(Op, DAG);
1467 return SDValue(Op.getNode(), 0);
1470 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1471 SelectionDAG &DAG) const {
1473 EVT VT = Op.getValueType();
1475 SDValue Num = Op.getOperand(0);
1476 SDValue Den = Op.getOperand(1);
1478 // RCP = URECIP(Den) = 2^32 / Den + e
1479 // e is rounding error.
1480 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1482 // RCP_LO = umulo(RCP, Den) */
1483 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1485 // RCP_HI = mulhu (RCP, Den) */
1486 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1488 // NEG_RCP_LO = -RCP_LO
1489 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1492 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1493 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1496 // Calculate the rounding error from the URECIP instruction
1497 // E = mulhu(ABS_RCP_LO, RCP)
1498 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1500 // RCP_A_E = RCP + E
1501 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1503 // RCP_S_E = RCP - E
1504 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1506 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1507 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1510 // Quotient = mulhu(Tmp0, Num)
1511 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1513 // Num_S_Remainder = Quotient * Den
1514 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1516 // Remainder = Num - Num_S_Remainder
1517 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1519 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1520 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1521 DAG.getConstant(-1, VT),
1522 DAG.getConstant(0, VT),
1524 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1525 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1527 DAG.getConstant(-1, VT),
1528 DAG.getConstant(0, VT),
1530 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1531 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1534 // Calculate Division result:
1536 // Quotient_A_One = Quotient + 1
1537 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1538 DAG.getConstant(1, VT));
1540 // Quotient_S_One = Quotient - 1
1541 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1542 DAG.getConstant(1, VT));
1544 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1545 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1546 Quotient, Quotient_A_One, ISD::SETEQ);
1548 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1549 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1550 Quotient_S_One, Div, ISD::SETEQ);
1552 // Calculate Rem result:
1554 // Remainder_S_Den = Remainder - Den
1555 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1557 // Remainder_A_Den = Remainder + Den
1558 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1560 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1561 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1562 Remainder, Remainder_S_Den, ISD::SETEQ);
1564 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1565 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1566 Remainder_A_Den, Rem, ISD::SETEQ);
1571 return DAG.getMergeValues(Ops, DL);
1574 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1576 SDValue Src = Op.getOperand(0);
1578 assert(Op.getValueType() == MVT::f64);
1580 SDValue C1 = DAG.getConstantFP(0x1.0p+52, MVT::f64);
1581 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1583 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1584 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1586 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1587 SDValue C2 = DAG.getConstantFP(0x1.fffffffffffffp+51, MVT::f64);
1589 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1590 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1592 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1595 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1596 SelectionDAG &DAG) const {
1597 SDValue S0 = Op.getOperand(0);
1599 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1602 // f32 uint_to_fp i64
1603 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1604 DAG.getConstant(0, MVT::i32));
1605 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1606 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1607 DAG.getConstant(1, MVT::i32));
1608 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1609 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1610 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1611 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1614 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1616 SelectionDAG &DAG) const {
1617 MVT VT = Op.getSimpleValueType();
1619 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1620 // Shift left by 'Shift' bits.
1621 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1622 // Signed shift Right by 'Shift' bits.
1623 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1626 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1627 SelectionDAG &DAG) const {
1628 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1629 MVT VT = Op.getSimpleValueType();
1630 MVT ScalarVT = VT.getScalarType();
1635 SDValue Src = Op.getOperand(0);
1638 // TODO: Don't scalarize on Evergreen?
1639 unsigned NElts = VT.getVectorNumElements();
1640 SmallVector<SDValue, 8> Args;
1641 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1643 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1644 for (unsigned I = 0; I < NElts; ++I)
1645 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1647 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1650 //===----------------------------------------------------------------------===//
1651 // Custom DAG optimizations
1652 //===----------------------------------------------------------------------===//
1654 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1655 APInt KnownZero, KnownOne;
1656 EVT VT = Op.getValueType();
1657 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1659 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1662 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1663 EVT VT = Op.getValueType();
1665 // In order for this to be a signed 24-bit value, bit 23, must
1667 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1668 // as unsigned 24-bit values.
1669 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1672 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1674 SelectionDAG &DAG = DCI.DAG;
1675 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1676 EVT VT = Op.getValueType();
1678 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1679 APInt KnownZero, KnownOne;
1680 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1681 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1682 DCI.CommitTargetLoweringOpt(TLO);
1685 template <typename IntTy>
1686 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1687 uint32_t Offset, uint32_t Width) {
1688 if (Width + Offset < 32) {
1689 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1690 return DAG.getConstant(Result, MVT::i32);
1693 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1696 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1697 DAGCombinerInfo &DCI) const {
1698 SelectionDAG &DAG = DCI.DAG;
1701 switch(N->getOpcode()) {
1704 EVT VT = N->getValueType(0);
1705 SDValue N0 = N->getOperand(0);
1706 SDValue N1 = N->getOperand(1);
1709 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1710 if (VT.isVector() || VT.getSizeInBits() > 32)
1713 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1714 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1715 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1716 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1717 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1718 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1719 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1720 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1725 // We need to use sext even for MUL_U24, because MUL_U24 is used
1726 // for signed multiply of 8 and 16-bit types.
1727 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1731 case AMDGPUISD::MUL_I24:
1732 case AMDGPUISD::MUL_U24: {
1733 SDValue N0 = N->getOperand(0);
1734 SDValue N1 = N->getOperand(1);
1735 simplifyI24(N0, DCI);
1736 simplifyI24(N1, DCI);
1739 case ISD::SELECT_CC: {
1740 return CombineMinMax(N, DAG);
1742 case AMDGPUISD::BFE_I32:
1743 case AMDGPUISD::BFE_U32: {
1744 assert(!N->getValueType(0).isVector() &&
1745 "Vector handling of BFE not implemented");
1746 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1750 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1752 return DAG.getConstant(0, MVT::i32);
1754 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1758 SDValue BitsFrom = N->getOperand(0);
1759 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1761 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1763 if (OffsetVal == 0) {
1764 // This is already sign / zero extended, so try to fold away extra BFEs.
1765 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1767 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1768 if (OpSignBits >= SignBits)
1771 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1773 // This is a sign_extend_inreg. Replace it to take advantage of existing
1774 // DAG Combines. If not eliminated, we will match back to BFE during
1777 // TODO: The sext_inreg of extended types ends, although we can could
1778 // handle them in a single BFE.
1779 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1780 DAG.getValueType(SmallVT));
1783 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1786 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1788 return constantFoldBFE<int32_t>(DAG,
1789 Val->getSExtValue(),
1794 return constantFoldBFE<uint32_t>(DAG,
1795 Val->getZExtValue(),
1800 APInt Demanded = APInt::getBitsSet(32,
1802 OffsetVal + WidthVal);
1804 if ((OffsetVal + WidthVal) >= 32) {
1805 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1806 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1807 BitsFrom, ShiftVal);
1810 APInt KnownZero, KnownOne;
1811 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1812 !DCI.isBeforeLegalizeOps());
1813 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1814 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1815 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1816 DCI.CommitTargetLoweringOpt(TLO);
1825 //===----------------------------------------------------------------------===//
1827 //===----------------------------------------------------------------------===//
1829 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1832 const SmallVectorImpl<ISD::InputArg> &Ins,
1833 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1835 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1836 if (Ins[i].ArgVT == Ins[i].VT) {
1837 OrigIns.push_back(Ins[i]);
1842 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1843 // Vector has been split into scalars.
1844 VT = Ins[i].ArgVT.getVectorElementType();
1845 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1846 Ins[i].ArgVT.getVectorElementType() !=
1847 Ins[i].VT.getVectorElementType()) {
1848 // Vector elements have been promoted
1851 // Vector has been spilt into smaller vectors.
1855 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1856 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1857 OrigIns.push_back(Arg);
1861 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1862 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1863 return CFP->isExactlyValue(1.0);
1865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1866 return C->isAllOnesValue();
1871 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1872 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1873 return CFP->getValueAPF().isZero();
1875 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1876 return C->isNullValue();
1881 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1882 const TargetRegisterClass *RC,
1883 unsigned Reg, EVT VT) const {
1884 MachineFunction &MF = DAG.getMachineFunction();
1885 MachineRegisterInfo &MRI = MF.getRegInfo();
1886 unsigned VirtualRegister;
1887 if (!MRI.isLiveIn(Reg)) {
1888 VirtualRegister = MRI.createVirtualRegister(RC);
1889 MRI.addLiveIn(Reg, VirtualRegister);
1891 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1893 return DAG.getRegister(VirtualRegister, VT);
1896 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1898 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1900 default: return nullptr;
1902 NODE_NAME_CASE(CALL);
1903 NODE_NAME_CASE(UMUL);
1904 NODE_NAME_CASE(DIV_INF);
1905 NODE_NAME_CASE(RET_FLAG);
1906 NODE_NAME_CASE(BRANCH_COND);
1909 NODE_NAME_CASE(DWORDADDR)
1910 NODE_NAME_CASE(FRACT)
1911 NODE_NAME_CASE(CLAMP)
1912 NODE_NAME_CASE(FMAX)
1913 NODE_NAME_CASE(SMAX)
1914 NODE_NAME_CASE(UMAX)
1915 NODE_NAME_CASE(FMIN)
1916 NODE_NAME_CASE(SMIN)
1917 NODE_NAME_CASE(UMIN)
1918 NODE_NAME_CASE(BFE_U32)
1919 NODE_NAME_CASE(BFE_I32)
1922 NODE_NAME_CASE(MUL_U24)
1923 NODE_NAME_CASE(MUL_I24)
1924 NODE_NAME_CASE(MAD_U24)
1925 NODE_NAME_CASE(MAD_I24)
1926 NODE_NAME_CASE(URECIP)
1927 NODE_NAME_CASE(DOT4)
1928 NODE_NAME_CASE(EXPORT)
1929 NODE_NAME_CASE(CONST_ADDRESS)
1930 NODE_NAME_CASE(REGISTER_LOAD)
1931 NODE_NAME_CASE(REGISTER_STORE)
1932 NODE_NAME_CASE(LOAD_CONSTANT)
1933 NODE_NAME_CASE(LOAD_INPUT)
1934 NODE_NAME_CASE(SAMPLE)
1935 NODE_NAME_CASE(SAMPLEB)
1936 NODE_NAME_CASE(SAMPLED)
1937 NODE_NAME_CASE(SAMPLEL)
1938 NODE_NAME_CASE(CVT_F32_UBYTE0)
1939 NODE_NAME_CASE(CVT_F32_UBYTE1)
1940 NODE_NAME_CASE(CVT_F32_UBYTE2)
1941 NODE_NAME_CASE(CVT_F32_UBYTE3)
1942 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
1943 NODE_NAME_CASE(STORE_MSKOR)
1944 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1948 static void computeKnownBitsForMinMax(const SDValue Op0,
1952 const SelectionDAG &DAG,
1954 APInt Op0Zero, Op0One;
1955 APInt Op1Zero, Op1One;
1956 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1957 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1959 KnownZero = Op0Zero & Op1Zero;
1960 KnownOne = Op0One & Op1One;
1963 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1967 const SelectionDAG &DAG,
1968 unsigned Depth) const {
1970 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1974 unsigned Opc = Op.getOpcode();
1979 case ISD::INTRINSIC_WO_CHAIN: {
1980 // FIXME: The intrinsic should just use the node.
1981 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1982 case AMDGPUIntrinsic::AMDGPU_imax:
1983 case AMDGPUIntrinsic::AMDGPU_umax:
1984 case AMDGPUIntrinsic::AMDGPU_imin:
1985 case AMDGPUIntrinsic::AMDGPU_umin:
1986 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1987 KnownZero, KnownOne, DAG, Depth);
1995 case AMDGPUISD::SMAX:
1996 case AMDGPUISD::UMAX:
1997 case AMDGPUISD::SMIN:
1998 case AMDGPUISD::UMIN:
1999 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2000 KnownZero, KnownOne, DAG, Depth);
2003 case AMDGPUISD::BFE_I32:
2004 case AMDGPUISD::BFE_U32: {
2005 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2009 unsigned BitWidth = 32;
2010 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2012 KnownZero = APInt::getAllOnesValue(BitWidth);
2013 KnownOne = APInt::getNullValue(BitWidth);
2017 // FIXME: This could do a lot more. If offset is 0, should be the same as
2018 // sign_extend_inreg implementation, but that involves duplicating it.
2019 if (Opc == AMDGPUISD::BFE_I32)
2020 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2022 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2029 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2031 const SelectionDAG &DAG,
2032 unsigned Depth) const {
2033 switch (Op.getOpcode()) {
2034 case AMDGPUISD::BFE_I32: {
2035 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2039 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2040 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2041 if (!Offset || !Offset->isNullValue())
2044 // TODO: Could probably figure something out with non-0 offsets.
2045 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2046 return std::max(SignBits, Op0SignBits);
2049 case AMDGPUISD::BFE_U32: {
2050 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2051 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;