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::i64, Promote);
143 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
145 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
146 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
148 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
149 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
151 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
152 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
154 setOperationAction(ISD::STORE, MVT::f64, Promote);
155 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
157 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
158 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
160 // Custom lowering of vector stores is required for local address space
162 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
163 // XXX: Native v2i32 local address space stores are possible, but not
164 // currently implemented.
165 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
167 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
168 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
169 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
171 // XXX: This can be change to Custom, once ExpandVectorStores can
172 // handle 64-bit stores.
173 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
175 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
176 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
177 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
178 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
179 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
182 setOperationAction(ISD::LOAD, MVT::f32, Promote);
183 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
185 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
186 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
188 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
189 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
191 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
192 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
194 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
195 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
197 setOperationAction(ISD::LOAD, MVT::f64, Promote);
198 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
200 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
201 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
203 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
204 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
205 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
206 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
207 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
208 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
209 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
210 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
211 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
212 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
214 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
215 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
216 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
217 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
218 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
220 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
221 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
223 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
224 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
225 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
227 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
229 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
230 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
231 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
232 setOperationAction(ISD::FRINT, MVT::f64, Custom);
233 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
236 if (!Subtarget->hasBFI()) {
237 // fcopysign can be done in a single instruction with BFI.
238 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
239 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
242 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
243 for (MVT VT : ScalarIntVTs) {
244 setOperationAction(ISD::SREM, VT, Expand);
245 setOperationAction(ISD::SDIV, VT, Expand);
247 // GPU does not have divrem function for signed or unsigned.
248 setOperationAction(ISD::SDIVREM, VT, Custom);
249 setOperationAction(ISD::UDIVREM, VT, Custom);
251 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
252 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
253 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
255 setOperationAction(ISD::BSWAP, VT, Expand);
256 setOperationAction(ISD::CTTZ, VT, Expand);
257 setOperationAction(ISD::CTLZ, VT, Expand);
260 if (!Subtarget->hasBCNT(32))
261 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
263 if (!Subtarget->hasBCNT(64))
264 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
266 // The hardware supports 32-bit ROTR, but not ROTL.
267 setOperationAction(ISD::ROTL, MVT::i32, Expand);
268 setOperationAction(ISD::ROTL, MVT::i64, Expand);
269 setOperationAction(ISD::ROTR, MVT::i64, Expand);
271 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
272 setOperationAction(ISD::MUL, MVT::i64, Expand);
273 setOperationAction(ISD::MULHU, MVT::i64, Expand);
274 setOperationAction(ISD::MULHS, MVT::i64, Expand);
275 setOperationAction(ISD::UDIV, MVT::i32, Expand);
276 setOperationAction(ISD::UREM, MVT::i32, Expand);
277 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
278 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
280 static const MVT::SimpleValueType VectorIntTypes[] = {
281 MVT::v2i32, MVT::v4i32
284 for (MVT VT : VectorIntTypes) {
285 // Expand the following operations for the current type by default.
286 setOperationAction(ISD::ADD, VT, Expand);
287 setOperationAction(ISD::AND, VT, Expand);
288 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
289 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
290 setOperationAction(ISD::MUL, VT, Expand);
291 setOperationAction(ISD::OR, VT, Expand);
292 setOperationAction(ISD::SHL, VT, Expand);
293 setOperationAction(ISD::SRA, VT, Expand);
294 setOperationAction(ISD::SRL, VT, Expand);
295 setOperationAction(ISD::ROTL, VT, Expand);
296 setOperationAction(ISD::ROTR, VT, Expand);
297 setOperationAction(ISD::SUB, VT, Expand);
298 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
299 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
300 // TODO: Implement custom UREM / SREM routines.
301 setOperationAction(ISD::SDIV, VT, Expand);
302 setOperationAction(ISD::UDIV, VT, Expand);
303 setOperationAction(ISD::SREM, VT, Expand);
304 setOperationAction(ISD::UREM, VT, Expand);
305 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
306 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
307 setOperationAction(ISD::SDIVREM, VT, Custom);
308 setOperationAction(ISD::UDIVREM, VT, Custom);
309 setOperationAction(ISD::ADDC, VT, Expand);
310 setOperationAction(ISD::SUBC, VT, Expand);
311 setOperationAction(ISD::ADDE, VT, Expand);
312 setOperationAction(ISD::SUBE, VT, Expand);
313 setOperationAction(ISD::SELECT, VT, Expand);
314 setOperationAction(ISD::VSELECT, VT, Expand);
315 setOperationAction(ISD::SELECT_CC, VT, Expand);
316 setOperationAction(ISD::XOR, VT, Expand);
317 setOperationAction(ISD::BSWAP, VT, Expand);
318 setOperationAction(ISD::CTPOP, VT, Expand);
319 setOperationAction(ISD::CTTZ, VT, Expand);
320 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
321 setOperationAction(ISD::CTLZ, VT, Expand);
322 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
323 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
326 static const MVT::SimpleValueType FloatVectorTypes[] = {
327 MVT::v2f32, MVT::v4f32
330 for (MVT VT : FloatVectorTypes) {
331 setOperationAction(ISD::FABS, VT, Expand);
332 setOperationAction(ISD::FADD, VT, Expand);
333 setOperationAction(ISD::FCEIL, VT, Expand);
334 setOperationAction(ISD::FCOS, VT, Expand);
335 setOperationAction(ISD::FDIV, VT, Expand);
336 setOperationAction(ISD::FEXP2, VT, Expand);
337 setOperationAction(ISD::FLOG2, VT, Expand);
338 setOperationAction(ISD::FPOW, VT, Expand);
339 setOperationAction(ISD::FFLOOR, VT, Expand);
340 setOperationAction(ISD::FTRUNC, VT, Expand);
341 setOperationAction(ISD::FMUL, VT, Expand);
342 setOperationAction(ISD::FMA, VT, Expand);
343 setOperationAction(ISD::FRINT, VT, Expand);
344 setOperationAction(ISD::FNEARBYINT, VT, Expand);
345 setOperationAction(ISD::FSQRT, VT, Expand);
346 setOperationAction(ISD::FSIN, VT, Expand);
347 setOperationAction(ISD::FSUB, VT, Expand);
348 setOperationAction(ISD::FNEG, VT, Expand);
349 setOperationAction(ISD::SELECT, VT, Expand);
350 setOperationAction(ISD::VSELECT, VT, Expand);
351 setOperationAction(ISD::SELECT_CC, VT, Expand);
352 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
353 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
356 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
357 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
359 setTargetDAGCombine(ISD::MUL);
360 setTargetDAGCombine(ISD::SELECT_CC);
362 setSchedulingPreference(Sched::RegPressure);
363 setJumpIsExpensive(true);
365 setSelectIsExpensive(false);
366 PredictableSelectIsExpensive = false;
368 // There are no integer divide instructions, and these expand to a pretty
369 // large sequence of instructions.
370 setIntDivIsCheap(false);
371 setPow2DivIsCheap(false);
373 // TODO: Investigate this when 64-bit divides are implemented.
374 addBypassSlowDiv(64, 32);
376 // FIXME: Need to really handle these.
377 MaxStoresPerMemcpy = 4096;
378 MaxStoresPerMemmove = 4096;
379 MaxStoresPerMemset = 4096;
382 //===----------------------------------------------------------------------===//
383 // Target Information
384 //===----------------------------------------------------------------------===//
386 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
390 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
394 // The backend supports 32 and 64 bit floating point immediates.
395 // FIXME: Why are we reporting vectors of FP immediates as legal?
396 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
397 EVT ScalarVT = VT.getScalarType();
398 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
401 // We don't want to shrink f64 / f32 constants.
402 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
403 EVT ScalarVT = VT.getScalarType();
404 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
407 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
409 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
412 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
413 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
415 return ((LScalarSize <= CastScalarSize) ||
416 (CastScalarSize >= 32) ||
420 //===---------------------------------------------------------------------===//
422 //===---------------------------------------------------------------------===//
424 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
425 assert(VT.isFloatingPoint());
426 return VT == MVT::f32;
429 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
430 assert(VT.isFloatingPoint());
431 return VT == MVT::f32;
434 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
435 // Truncate is just accessing a subregister.
436 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
439 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
440 // Truncate is just accessing a subregister.
441 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
442 (Dest->getPrimitiveSizeInBits() % 32 == 0);
445 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
446 const DataLayout *DL = getDataLayout();
447 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
448 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
450 return SrcSize == 32 && DestSize == 64;
453 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
454 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
455 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
456 // this will enable reducing 64-bit operations the 32-bit, which is always
458 return Src == MVT::i32 && Dest == MVT::i64;
461 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
462 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
463 // limited number of native 64-bit operations. Shrinking an operation to fit
464 // in a single 32-bit register should always be helpful. As currently used,
465 // this is much less general than the name suggests, and is only used in
466 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
467 // not profitable, and may actually be harmful.
468 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
471 //===---------------------------------------------------------------------===//
472 // TargetLowering Callbacks
473 //===---------------------------------------------------------------------===//
475 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
476 const SmallVectorImpl<ISD::InputArg> &Ins) const {
478 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
481 SDValue AMDGPUTargetLowering::LowerReturn(
483 CallingConv::ID CallConv,
485 const SmallVectorImpl<ISD::OutputArg> &Outs,
486 const SmallVectorImpl<SDValue> &OutVals,
487 SDLoc DL, SelectionDAG &DAG) const {
488 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
491 //===---------------------------------------------------------------------===//
492 // Target specific lowering
493 //===---------------------------------------------------------------------===//
495 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
496 SmallVectorImpl<SDValue> &InVals) const {
497 SDValue Callee = CLI.Callee;
498 SelectionDAG &DAG = CLI.DAG;
500 const Function &Fn = *DAG.getMachineFunction().getFunction();
502 StringRef FuncName("<unknown>");
504 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
505 FuncName = G->getSymbol();
506 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
507 FuncName = G->getGlobal()->getName();
509 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
510 DAG.getContext()->diagnose(NoCalls);
514 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
515 SelectionDAG &DAG) const {
516 switch (Op.getOpcode()) {
518 Op.getNode()->dump();
519 llvm_unreachable("Custom lowering code for this"
520 "instruction is not implemented yet!");
522 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
523 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
524 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
525 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
526 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
527 case ISD::SDIV: return LowerSDIV(Op, DAG);
528 case ISD::SREM: return LowerSREM(Op, DAG);
529 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
530 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
531 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
532 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
533 case ISD::FRINT: return LowerFRINT(Op, DAG);
534 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
535 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
536 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
541 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
542 SmallVectorImpl<SDValue> &Results,
543 SelectionDAG &DAG) const {
544 switch (N->getOpcode()) {
545 case ISD::SIGN_EXTEND_INREG:
546 // Different parts of legalization seem to interpret which type of
547 // sign_extend_inreg is the one to check for custom lowering. The extended
548 // from type is what really matters, but some places check for custom
549 // lowering of the result type. This results in trying to use
550 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
551 // nothing here and let the illegal result integer be handled normally.
558 // FIXME: This implements accesses to initialized globals in the constant
559 // address space by copying them to private and accessing that. It does not
560 // properly handle illegal types or vectors. The private vector loads are not
561 // scalarized, and the illegal scalars hit an assertion. This technique will not
562 // work well with large initializers, and this should eventually be
563 // removed. Initialized globals should be placed into a data section that the
564 // runtime will load into a buffer before the kernel is executed. Uses of the
565 // global need to be replaced with a pointer loaded from an implicit kernel
566 // argument into this buffer holding the copy of the data, which will remove the
567 // need for any of this.
568 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
569 const GlobalValue *GV,
570 const SDValue &InitPtr,
572 SelectionDAG &DAG) const {
573 const DataLayout *TD = getTargetMachine().getDataLayout();
575 Type *InitTy = Init->getType();
577 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
578 EVT VT = EVT::getEVT(InitTy);
579 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
580 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
581 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
582 TD->getPrefTypeAlignment(InitTy));
585 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
586 EVT VT = EVT::getEVT(CFP->getType());
587 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
588 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
589 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
590 TD->getPrefTypeAlignment(CFP->getType()));
593 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
594 const StructLayout *SL = TD->getStructLayout(ST);
596 EVT PtrVT = InitPtr.getValueType();
597 SmallVector<SDValue, 8> Chains;
599 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
600 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
601 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
603 Constant *Elt = Init->getAggregateElement(I);
604 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
607 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
610 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
611 EVT PtrVT = InitPtr.getValueType();
613 unsigned NumElements;
614 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
615 NumElements = AT->getNumElements();
616 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
617 NumElements = VT->getNumElements();
619 llvm_unreachable("Unexpected type");
621 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
622 SmallVector<SDValue, 8> Chains;
623 for (unsigned i = 0; i < NumElements; ++i) {
624 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
625 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
627 Constant *Elt = Init->getAggregateElement(i);
628 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
631 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
634 if (isa<UndefValue>(Init)) {
635 EVT VT = EVT::getEVT(InitTy);
636 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
637 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
638 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
639 TD->getPrefTypeAlignment(InitTy));
643 llvm_unreachable("Unhandled constant initializer");
646 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
648 SelectionDAG &DAG) const {
650 const DataLayout *TD = getTargetMachine().getDataLayout();
651 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
652 const GlobalValue *GV = G->getGlobal();
654 switch (G->getAddressSpace()) {
655 default: llvm_unreachable("Global Address lowering not implemented for this "
657 case AMDGPUAS::LOCAL_ADDRESS: {
658 // XXX: What does the value of G->getOffset() mean?
659 assert(G->getOffset() == 0 &&
660 "Do not know what to do with an non-zero offset");
663 if (MFI->LocalMemoryObjects.count(GV) == 0) {
664 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
665 Offset = MFI->LDSSize;
666 MFI->LocalMemoryObjects[GV] = Offset;
667 // XXX: Account for alignment?
668 MFI->LDSSize += Size;
670 Offset = MFI->LocalMemoryObjects[GV];
673 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
675 case AMDGPUAS::CONSTANT_ADDRESS: {
676 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
677 Type *EltType = GV->getType()->getElementType();
678 unsigned Size = TD->getTypeAllocSize(EltType);
679 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
681 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
682 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
684 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
685 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
687 const GlobalVariable *Var = cast<GlobalVariable>(GV);
688 if (!Var->hasInitializer()) {
689 // This has no use, but bugpoint will hit it.
690 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
693 const Constant *Init = Var->getInitializer();
694 SmallVector<SDNode*, 8> WorkList;
696 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
697 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
698 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
700 WorkList.push_back(*I);
702 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
703 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
704 E = WorkList.end(); I != E; ++I) {
705 SmallVector<SDValue, 8> Ops;
706 Ops.push_back(Chain);
707 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
708 Ops.push_back((*I)->getOperand(i));
710 DAG.UpdateNodeOperands(*I, Ops);
712 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
717 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
718 SelectionDAG &DAG) const {
719 SmallVector<SDValue, 8> Args;
720 SDValue A = Op.getOperand(0);
721 SDValue B = Op.getOperand(1);
723 DAG.ExtractVectorElements(A, Args);
724 DAG.ExtractVectorElements(B, Args);
726 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
729 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
730 SelectionDAG &DAG) const {
732 SmallVector<SDValue, 8> Args;
733 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
734 EVT VT = Op.getValueType();
735 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
736 VT.getVectorNumElements());
738 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
741 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
742 SelectionDAG &DAG) const {
744 MachineFunction &MF = DAG.getMachineFunction();
745 const AMDGPUFrameLowering *TFL =
746 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
748 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
750 unsigned FrameIndex = FIN->getIndex();
751 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
752 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
756 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
757 SelectionDAG &DAG) const {
758 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
760 EVT VT = Op.getValueType();
762 switch (IntrinsicID) {
764 case AMDGPUIntrinsic::AMDGPU_abs:
765 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
766 return LowerIntrinsicIABS(Op, DAG);
767 case AMDGPUIntrinsic::AMDGPU_lrp:
768 return LowerIntrinsicLRP(Op, DAG);
769 case AMDGPUIntrinsic::AMDGPU_fract:
770 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
771 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
773 case AMDGPUIntrinsic::AMDGPU_clamp:
774 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
775 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
776 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
778 case Intrinsic::AMDGPU_div_scale: {
779 // 3rd parameter required to be a constant.
780 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
782 return DAG.getUNDEF(VT);
784 // Translate to the operands expected by the machine instruction. The
785 // first parameter must be the same as the first instruction.
786 SDValue Numerator = Op.getOperand(1);
787 SDValue Denominator = Op.getOperand(2);
788 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
790 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
791 Src0, Denominator, Numerator);
794 case Intrinsic::AMDGPU_div_fmas:
795 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
796 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
798 case Intrinsic::AMDGPU_div_fixup:
799 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
800 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
802 case Intrinsic::AMDGPU_trig_preop:
803 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
804 Op.getOperand(1), Op.getOperand(2));
806 case Intrinsic::AMDGPU_rcp:
807 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
809 case Intrinsic::AMDGPU_rsq:
810 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
812 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
813 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
815 case Intrinsic::AMDGPU_rsq_clamped:
816 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
818 case AMDGPUIntrinsic::AMDGPU_imax:
819 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
821 case AMDGPUIntrinsic::AMDGPU_umax:
822 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
824 case AMDGPUIntrinsic::AMDGPU_imin:
825 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
827 case AMDGPUIntrinsic::AMDGPU_umin:
828 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
831 case AMDGPUIntrinsic::AMDGPU_umul24:
832 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
833 Op.getOperand(1), Op.getOperand(2));
835 case AMDGPUIntrinsic::AMDGPU_imul24:
836 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
837 Op.getOperand(1), Op.getOperand(2));
839 case AMDGPUIntrinsic::AMDGPU_umad24:
840 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
841 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
843 case AMDGPUIntrinsic::AMDGPU_imad24:
844 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
845 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
847 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
848 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
850 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
851 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
853 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
854 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
856 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
857 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
859 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
860 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
865 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
866 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
871 case AMDGPUIntrinsic::AMDGPU_bfi:
872 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
877 case AMDGPUIntrinsic::AMDGPU_bfm:
878 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
882 case AMDGPUIntrinsic::AMDGPU_brev:
883 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
885 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
886 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
888 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
889 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
890 case AMDGPUIntrinsic::AMDGPU_trunc:
891 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
895 ///IABS(a) = SMAX(sub(0, a), a)
896 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
897 SelectionDAG &DAG) const {
899 EVT VT = Op.getValueType();
900 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
903 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
906 /// Linear Interpolation
907 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
908 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
909 SelectionDAG &DAG) const {
911 EVT VT = Op.getValueType();
912 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
913 DAG.getConstantFP(1.0f, MVT::f32),
915 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
917 return DAG.getNode(ISD::FADD, DL, VT,
918 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
922 /// \brief Generate Min/Max node
923 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
924 SelectionDAG &DAG) const {
926 EVT VT = N->getValueType(0);
928 SDValue LHS = N->getOperand(0);
929 SDValue RHS = N->getOperand(1);
930 SDValue True = N->getOperand(2);
931 SDValue False = N->getOperand(3);
932 SDValue CC = N->getOperand(4);
934 if (VT != MVT::f32 ||
935 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
939 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
953 llvm_unreachable("Operation should already be optimised!");
960 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
961 return DAG.getNode(Opc, DL, VT, LHS, RHS);
969 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
970 return DAG.getNode(Opc, DL, VT, LHS, RHS);
972 case ISD::SETCC_INVALID:
973 llvm_unreachable("Invalid setcc condcode!");
978 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
979 SelectionDAG &DAG) const {
980 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
981 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
982 EVT EltVT = Op.getValueType().getVectorElementType();
983 EVT PtrVT = Load->getBasePtr().getValueType();
984 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
985 SmallVector<SDValue, 8> Loads;
988 for (unsigned i = 0, e = NumElts; i != e; ++i) {
989 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
990 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
991 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
992 Load->getChain(), Ptr,
993 MachinePointerInfo(Load->getMemOperand()->getValue()),
994 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
995 Load->getAlignment()));
997 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
1000 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1001 SelectionDAG &DAG) const {
1002 StoreSDNode *Store = cast<StoreSDNode>(Op);
1003 EVT MemVT = Store->getMemoryVT();
1004 unsigned MemBits = MemVT.getSizeInBits();
1006 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1007 // truncating store into an i32 store.
1008 // XXX: We could also handle optimize other vector bitwidths.
1009 if (!MemVT.isVector() || MemBits > 32) {
1014 SDValue Value = Store->getValue();
1015 EVT VT = Value.getValueType();
1016 EVT ElemVT = VT.getVectorElementType();
1017 SDValue Ptr = Store->getBasePtr();
1018 EVT MemEltVT = MemVT.getVectorElementType();
1019 unsigned MemEltBits = MemEltVT.getSizeInBits();
1020 unsigned MemNumElements = MemVT.getVectorNumElements();
1021 unsigned PackedSize = MemVT.getStoreSizeInBits();
1022 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1024 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1026 SDValue PackedValue;
1027 for (unsigned i = 0; i < MemNumElements; ++i) {
1028 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1029 DAG.getConstant(i, MVT::i32));
1030 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1031 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1033 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1034 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1039 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1043 if (PackedSize < 32) {
1044 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1045 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1046 Store->getMemOperand()->getPointerInfo(),
1048 Store->isNonTemporal(), Store->isVolatile(),
1049 Store->getAlignment());
1052 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1053 Store->getMemOperand()->getPointerInfo(),
1054 Store->isVolatile(), Store->isNonTemporal(),
1055 Store->getAlignment());
1058 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1059 SelectionDAG &DAG) const {
1060 StoreSDNode *Store = cast<StoreSDNode>(Op);
1061 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1062 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1063 EVT PtrVT = Store->getBasePtr().getValueType();
1064 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1067 SmallVector<SDValue, 8> Chains;
1069 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1070 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1071 Store->getValue(), DAG.getConstant(i, MVT::i32));
1072 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1073 Store->getBasePtr(),
1074 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1076 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1077 MachinePointerInfo(Store->getMemOperand()->getValue()),
1078 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1079 Store->getAlignment()));
1081 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1084 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1086 LoadSDNode *Load = cast<LoadSDNode>(Op);
1087 ISD::LoadExtType ExtType = Load->getExtensionType();
1088 EVT VT = Op.getValueType();
1089 EVT MemVT = Load->getMemoryVT();
1091 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1092 // We can do the extload to 32-bits, and then need to separately extend to
1095 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1099 Load->getMemOperand());
1100 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1103 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1104 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1105 // FIXME: Copied from PPC
1106 // First, load into 32 bits, then truncate to 1 bit.
1108 SDValue Chain = Load->getChain();
1109 SDValue BasePtr = Load->getBasePtr();
1110 MachineMemOperand *MMO = Load->getMemOperand();
1112 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1113 BasePtr, MVT::i8, MMO);
1114 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1117 // Lower loads constant address space global variable loads
1118 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1119 isa<GlobalVariable>(
1120 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1122 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1123 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1124 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1125 DAG.getConstant(2, MVT::i32));
1126 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1127 Load->getChain(), Ptr,
1128 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1131 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1132 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1136 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1137 DAG.getConstant(2, MVT::i32));
1138 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1139 Load->getChain(), Ptr,
1140 DAG.getTargetConstant(0, MVT::i32),
1142 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1144 DAG.getConstant(0x3, MVT::i32));
1145 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1146 DAG.getConstant(3, MVT::i32));
1148 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1150 EVT MemEltVT = MemVT.getScalarType();
1151 if (ExtType == ISD::SEXTLOAD) {
1152 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1153 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1156 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1159 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1161 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1162 if (Result.getNode()) {
1166 StoreSDNode *Store = cast<StoreSDNode>(Op);
1167 SDValue Chain = Store->getChain();
1168 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1169 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1170 Store->getValue().getValueType().isVector()) {
1171 return SplitVectorStore(Op, DAG);
1174 EVT MemVT = Store->getMemoryVT();
1175 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1176 MemVT.bitsLT(MVT::i32)) {
1178 if (Store->getMemoryVT() == MVT::i8) {
1180 } else if (Store->getMemoryVT() == MVT::i16) {
1183 SDValue BasePtr = Store->getBasePtr();
1184 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1185 DAG.getConstant(2, MVT::i32));
1186 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1187 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1189 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1190 DAG.getConstant(0x3, MVT::i32));
1192 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1193 DAG.getConstant(3, MVT::i32));
1195 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1198 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1200 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1201 MaskedValue, ShiftAmt);
1203 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1205 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1206 DAG.getConstant(0xffffffff, MVT::i32));
1207 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1209 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1210 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1211 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1216 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1218 EVT OVT = Op.getValueType();
1219 SDValue LHS = Op.getOperand(0);
1220 SDValue RHS = Op.getOperand(1);
1223 if (!OVT.isVector()) {
1226 } else if (OVT.getVectorNumElements() == 2) {
1229 } else if (OVT.getVectorNumElements() == 4) {
1233 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1234 // char|short jq = ia ^ ib;
1235 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1237 // jq = jq >> (bitsize - 2)
1238 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1241 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1244 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1246 // int ia = (int)LHS;
1247 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1249 // int ib, (int)RHS;
1250 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1252 // float fa = (float)ia;
1253 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1255 // float fb = (float)ib;
1256 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1258 // float fq = native_divide(fa, fb);
1259 SDValue fq = DAG.getNode(ISD::FMUL, DL, FLTTY,
1260 fa, DAG.getNode(AMDGPUISD::RCP, DL, FLTTY, fb));
1263 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1265 // float fqneg = -fq;
1266 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1268 // float fr = mad(fqneg, fb, fa);
1269 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1270 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1272 // int iq = (int)fq;
1273 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1276 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1279 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1281 // int cv = fr >= fb;
1283 if (INTTY == MVT::i32) {
1284 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1286 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1288 // jq = (cv ? jq : 0);
1289 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1290 DAG.getConstant(0, OVT));
1292 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1293 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1297 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1299 EVT OVT = Op.getValueType();
1300 SDValue LHS = Op.getOperand(0);
1301 SDValue RHS = Op.getOperand(1);
1302 // The LowerSDIV32 function generates equivalent to the following IL.
1312 // ixor r10, r10, r11
1314 // ixor DST, r0, r10
1323 SDValue r10 = DAG.getSelectCC(DL,
1324 r0, DAG.getConstant(0, OVT),
1325 DAG.getConstant(-1, OVT),
1326 DAG.getConstant(0, OVT),
1330 SDValue r11 = DAG.getSelectCC(DL,
1331 r1, DAG.getConstant(0, OVT),
1332 DAG.getConstant(-1, OVT),
1333 DAG.getConstant(0, OVT),
1337 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1340 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1343 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1346 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1349 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1351 // ixor r10, r10, r11
1352 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1355 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1357 // ixor DST, r0, r10
1358 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1362 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1363 return SDValue(Op.getNode(), 0);
1366 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1367 EVT OVT = Op.getValueType().getScalarType();
1369 if (OVT == MVT::i64)
1370 return LowerSDIV64(Op, DAG);
1372 if (OVT.getScalarType() == MVT::i32)
1373 return LowerSDIV32(Op, DAG);
1375 if (OVT == MVT::i16 || OVT == MVT::i8) {
1376 // FIXME: We should be checking for the masked bits. This isn't reached
1377 // because i8 and i16 are not legal types.
1378 return LowerSDIV24(Op, DAG);
1381 return SDValue(Op.getNode(), 0);
1384 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1386 EVT OVT = Op.getValueType();
1387 SDValue LHS = Op.getOperand(0);
1388 SDValue RHS = Op.getOperand(1);
1389 // The LowerSREM32 function generates equivalent to the following IL.
1399 // umul r20, r20, r1
1402 // ixor DST, r0, r10
1411 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1414 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1417 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1420 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1423 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1426 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1429 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1431 // umul r20, r20, r1
1432 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1435 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1438 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1440 // ixor DST, r0, r10
1441 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1445 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1446 return SDValue(Op.getNode(), 0);
1449 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1450 EVT OVT = Op.getValueType();
1452 if (OVT.getScalarType() == MVT::i64)
1453 return LowerSREM64(Op, DAG);
1455 if (OVT.getScalarType() == MVT::i32)
1456 return LowerSREM32(Op, DAG);
1458 return SDValue(Op.getNode(), 0);
1461 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1462 SelectionDAG &DAG) const {
1464 EVT VT = Op.getValueType();
1466 SDValue Num = Op.getOperand(0);
1467 SDValue Den = Op.getOperand(1);
1469 // RCP = URECIP(Den) = 2^32 / Den + e
1470 // e is rounding error.
1471 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1473 // RCP_LO = umulo(RCP, Den) */
1474 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1476 // RCP_HI = mulhu (RCP, Den) */
1477 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1479 // NEG_RCP_LO = -RCP_LO
1480 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1483 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1484 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1487 // Calculate the rounding error from the URECIP instruction
1488 // E = mulhu(ABS_RCP_LO, RCP)
1489 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1491 // RCP_A_E = RCP + E
1492 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1494 // RCP_S_E = RCP - E
1495 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1497 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1498 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1501 // Quotient = mulhu(Tmp0, Num)
1502 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1504 // Num_S_Remainder = Quotient * Den
1505 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1507 // Remainder = Num - Num_S_Remainder
1508 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1510 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1511 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1512 DAG.getConstant(-1, VT),
1513 DAG.getConstant(0, VT),
1515 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1516 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1518 DAG.getConstant(-1, VT),
1519 DAG.getConstant(0, VT),
1521 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1522 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1525 // Calculate Division result:
1527 // Quotient_A_One = Quotient + 1
1528 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1529 DAG.getConstant(1, VT));
1531 // Quotient_S_One = Quotient - 1
1532 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1533 DAG.getConstant(1, VT));
1535 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1536 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1537 Quotient, Quotient_A_One, ISD::SETEQ);
1539 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1540 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1541 Quotient_S_One, Div, ISD::SETEQ);
1543 // Calculate Rem result:
1545 // Remainder_S_Den = Remainder - Den
1546 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1548 // Remainder_A_Den = Remainder + Den
1549 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1551 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1552 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1553 Remainder, Remainder_S_Den, ISD::SETEQ);
1555 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1556 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1557 Remainder_A_Den, Rem, ISD::SETEQ);
1562 return DAG.getMergeValues(Ops, DL);
1565 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1566 SelectionDAG &DAG) const {
1568 EVT VT = Op.getValueType();
1570 SDValue Zero = DAG.getConstant(0, VT);
1571 SDValue NegOne = DAG.getConstant(-1, VT);
1573 SDValue LHS = Op.getOperand(0);
1574 SDValue RHS = Op.getOperand(1);
1576 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1577 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1578 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1579 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1581 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1582 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1584 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1585 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1587 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1588 SDValue Rem = Div.getValue(1);
1590 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1591 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1593 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1594 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1600 return DAG.getMergeValues(Res, DL);
1603 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1605 SDValue Src = Op.getOperand(0);
1607 // result = trunc(src)
1608 // if (src > 0.0 && src != result)
1611 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1613 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1614 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1616 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1618 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1619 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1620 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1622 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1623 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1626 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1628 SDValue Src = Op.getOperand(0);
1630 assert(Op.getValueType() == MVT::f64);
1632 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1633 const SDValue One = DAG.getConstant(1, MVT::i32);
1635 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1637 // Extract the upper half, since this is where we will find the sign and
1639 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1641 const unsigned FractBits = 52;
1642 const unsigned ExpBits = 11;
1644 // Extract the exponent.
1645 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1647 DAG.getConstant(FractBits - 32, MVT::i32),
1648 DAG.getConstant(ExpBits, MVT::i32));
1649 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1650 DAG.getConstant(1023, MVT::i32));
1652 // Extract the sign bit.
1653 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1654 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1656 // Extend back to to 64-bits.
1657 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1659 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1661 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1662 const SDValue FractMask
1663 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1665 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1666 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1667 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1669 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1671 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1673 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1674 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1676 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1677 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1679 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1682 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1684 SDValue Src = Op.getOperand(0);
1686 assert(Op.getValueType() == MVT::f64);
1688 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1689 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1690 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1692 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1693 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1695 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1697 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1698 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1700 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1701 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1703 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1706 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1707 // FNEARBYINT and FRINT are the same, except in their handling of FP
1708 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1709 // rint, so just treat them as equivalent.
1710 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1713 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1715 SDValue Src = Op.getOperand(0);
1717 // result = trunc(src);
1718 // if (src < 0.0 && src != result)
1721 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1723 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1724 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1726 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1728 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1729 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1730 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1732 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1733 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1736 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1737 SelectionDAG &DAG) const {
1738 SDValue S0 = Op.getOperand(0);
1740 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1743 // f32 uint_to_fp i64
1744 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1745 DAG.getConstant(0, MVT::i32));
1746 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1747 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1748 DAG.getConstant(1, MVT::i32));
1749 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1750 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1751 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1752 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1755 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1757 SelectionDAG &DAG) const {
1758 MVT VT = Op.getSimpleValueType();
1760 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1761 // Shift left by 'Shift' bits.
1762 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1763 // Signed shift Right by 'Shift' bits.
1764 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1767 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1768 SelectionDAG &DAG) const {
1769 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1770 MVT VT = Op.getSimpleValueType();
1771 MVT ScalarVT = VT.getScalarType();
1776 SDValue Src = Op.getOperand(0);
1779 // TODO: Don't scalarize on Evergreen?
1780 unsigned NElts = VT.getVectorNumElements();
1781 SmallVector<SDValue, 8> Args;
1782 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1784 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1785 for (unsigned I = 0; I < NElts; ++I)
1786 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1788 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1791 //===----------------------------------------------------------------------===//
1792 // Custom DAG optimizations
1793 //===----------------------------------------------------------------------===//
1795 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1796 APInt KnownZero, KnownOne;
1797 EVT VT = Op.getValueType();
1798 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1800 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1803 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1804 EVT VT = Op.getValueType();
1806 // In order for this to be a signed 24-bit value, bit 23, must
1808 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1809 // as unsigned 24-bit values.
1810 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1813 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1815 SelectionDAG &DAG = DCI.DAG;
1816 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1817 EVT VT = Op.getValueType();
1819 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1820 APInt KnownZero, KnownOne;
1821 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1822 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1823 DCI.CommitTargetLoweringOpt(TLO);
1826 template <typename IntTy>
1827 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1828 uint32_t Offset, uint32_t Width) {
1829 if (Width + Offset < 32) {
1830 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1831 return DAG.getConstant(Result, MVT::i32);
1834 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1837 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1838 DAGCombinerInfo &DCI) const {
1839 SelectionDAG &DAG = DCI.DAG;
1842 switch(N->getOpcode()) {
1845 EVT VT = N->getValueType(0);
1846 SDValue N0 = N->getOperand(0);
1847 SDValue N1 = N->getOperand(1);
1850 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1851 if (VT.isVector() || VT.getSizeInBits() > 32)
1854 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1855 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1856 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1857 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1858 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1859 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1860 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1861 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1866 // We need to use sext even for MUL_U24, because MUL_U24 is used
1867 // for signed multiply of 8 and 16-bit types.
1868 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1872 case AMDGPUISD::MUL_I24:
1873 case AMDGPUISD::MUL_U24: {
1874 SDValue N0 = N->getOperand(0);
1875 SDValue N1 = N->getOperand(1);
1876 simplifyI24(N0, DCI);
1877 simplifyI24(N1, DCI);
1880 case ISD::SELECT_CC: {
1881 return CombineMinMax(N, DAG);
1883 case AMDGPUISD::BFE_I32:
1884 case AMDGPUISD::BFE_U32: {
1885 assert(!N->getValueType(0).isVector() &&
1886 "Vector handling of BFE not implemented");
1887 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1891 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1893 return DAG.getConstant(0, MVT::i32);
1895 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1899 SDValue BitsFrom = N->getOperand(0);
1900 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1902 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1904 if (OffsetVal == 0) {
1905 // This is already sign / zero extended, so try to fold away extra BFEs.
1906 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1908 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1909 if (OpSignBits >= SignBits)
1912 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1914 // This is a sign_extend_inreg. Replace it to take advantage of existing
1915 // DAG Combines. If not eliminated, we will match back to BFE during
1918 // TODO: The sext_inreg of extended types ends, although we can could
1919 // handle them in a single BFE.
1920 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1921 DAG.getValueType(SmallVT));
1924 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1927 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1929 return constantFoldBFE<int32_t>(DAG,
1930 Val->getSExtValue(),
1935 return constantFoldBFE<uint32_t>(DAG,
1936 Val->getZExtValue(),
1941 APInt Demanded = APInt::getBitsSet(32,
1943 OffsetVal + WidthVal);
1945 if ((OffsetVal + WidthVal) >= 32) {
1946 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1947 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1948 BitsFrom, ShiftVal);
1951 APInt KnownZero, KnownOne;
1952 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1953 !DCI.isBeforeLegalizeOps());
1954 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1955 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1956 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1957 DCI.CommitTargetLoweringOpt(TLO);
1966 //===----------------------------------------------------------------------===//
1968 //===----------------------------------------------------------------------===//
1970 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1973 const SmallVectorImpl<ISD::InputArg> &Ins,
1974 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1976 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1977 if (Ins[i].ArgVT == Ins[i].VT) {
1978 OrigIns.push_back(Ins[i]);
1983 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1984 // Vector has been split into scalars.
1985 VT = Ins[i].ArgVT.getVectorElementType();
1986 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1987 Ins[i].ArgVT.getVectorElementType() !=
1988 Ins[i].VT.getVectorElementType()) {
1989 // Vector elements have been promoted
1992 // Vector has been spilt into smaller vectors.
1996 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1997 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1998 OrigIns.push_back(Arg);
2002 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2003 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2004 return CFP->isExactlyValue(1.0);
2006 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2007 return C->isAllOnesValue();
2012 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2013 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2014 return CFP->getValueAPF().isZero();
2016 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2017 return C->isNullValue();
2022 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2023 const TargetRegisterClass *RC,
2024 unsigned Reg, EVT VT) const {
2025 MachineFunction &MF = DAG.getMachineFunction();
2026 MachineRegisterInfo &MRI = MF.getRegInfo();
2027 unsigned VirtualRegister;
2028 if (!MRI.isLiveIn(Reg)) {
2029 VirtualRegister = MRI.createVirtualRegister(RC);
2030 MRI.addLiveIn(Reg, VirtualRegister);
2032 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2034 return DAG.getRegister(VirtualRegister, VT);
2037 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2039 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2041 default: return nullptr;
2043 NODE_NAME_CASE(CALL);
2044 NODE_NAME_CASE(UMUL);
2045 NODE_NAME_CASE(RET_FLAG);
2046 NODE_NAME_CASE(BRANCH_COND);
2049 NODE_NAME_CASE(DWORDADDR)
2050 NODE_NAME_CASE(FRACT)
2051 NODE_NAME_CASE(CLAMP)
2052 NODE_NAME_CASE(FMAX)
2053 NODE_NAME_CASE(SMAX)
2054 NODE_NAME_CASE(UMAX)
2055 NODE_NAME_CASE(FMIN)
2056 NODE_NAME_CASE(SMIN)
2057 NODE_NAME_CASE(UMIN)
2058 NODE_NAME_CASE(URECIP)
2059 NODE_NAME_CASE(DIV_SCALE)
2060 NODE_NAME_CASE(DIV_FMAS)
2061 NODE_NAME_CASE(DIV_FIXUP)
2062 NODE_NAME_CASE(TRIG_PREOP)
2065 NODE_NAME_CASE(RSQ_LEGACY)
2066 NODE_NAME_CASE(RSQ_CLAMPED)
2067 NODE_NAME_CASE(DOT4)
2068 NODE_NAME_CASE(BFE_U32)
2069 NODE_NAME_CASE(BFE_I32)
2072 NODE_NAME_CASE(BREV)
2073 NODE_NAME_CASE(MUL_U24)
2074 NODE_NAME_CASE(MUL_I24)
2075 NODE_NAME_CASE(MAD_U24)
2076 NODE_NAME_CASE(MAD_I24)
2077 NODE_NAME_CASE(EXPORT)
2078 NODE_NAME_CASE(CONST_ADDRESS)
2079 NODE_NAME_CASE(REGISTER_LOAD)
2080 NODE_NAME_CASE(REGISTER_STORE)
2081 NODE_NAME_CASE(LOAD_CONSTANT)
2082 NODE_NAME_CASE(LOAD_INPUT)
2083 NODE_NAME_CASE(SAMPLE)
2084 NODE_NAME_CASE(SAMPLEB)
2085 NODE_NAME_CASE(SAMPLED)
2086 NODE_NAME_CASE(SAMPLEL)
2087 NODE_NAME_CASE(CVT_F32_UBYTE0)
2088 NODE_NAME_CASE(CVT_F32_UBYTE1)
2089 NODE_NAME_CASE(CVT_F32_UBYTE2)
2090 NODE_NAME_CASE(CVT_F32_UBYTE3)
2091 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2092 NODE_NAME_CASE(STORE_MSKOR)
2093 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2097 static void computeKnownBitsForMinMax(const SDValue Op0,
2101 const SelectionDAG &DAG,
2103 APInt Op0Zero, Op0One;
2104 APInt Op1Zero, Op1One;
2105 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2106 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2108 KnownZero = Op0Zero & Op1Zero;
2109 KnownOne = Op0One & Op1One;
2112 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2116 const SelectionDAG &DAG,
2117 unsigned Depth) const {
2119 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2123 unsigned Opc = Op.getOpcode();
2128 case ISD::INTRINSIC_WO_CHAIN: {
2129 // FIXME: The intrinsic should just use the node.
2130 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2131 case AMDGPUIntrinsic::AMDGPU_imax:
2132 case AMDGPUIntrinsic::AMDGPU_umax:
2133 case AMDGPUIntrinsic::AMDGPU_imin:
2134 case AMDGPUIntrinsic::AMDGPU_umin:
2135 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2136 KnownZero, KnownOne, DAG, Depth);
2144 case AMDGPUISD::SMAX:
2145 case AMDGPUISD::UMAX:
2146 case AMDGPUISD::SMIN:
2147 case AMDGPUISD::UMIN:
2148 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2149 KnownZero, KnownOne, DAG, Depth);
2152 case AMDGPUISD::BFE_I32:
2153 case AMDGPUISD::BFE_U32: {
2154 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2158 unsigned BitWidth = 32;
2159 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2161 KnownZero = APInt::getAllOnesValue(BitWidth);
2162 KnownOne = APInt::getNullValue(BitWidth);
2166 // FIXME: This could do a lot more. If offset is 0, should be the same as
2167 // sign_extend_inreg implementation, but that involves duplicating it.
2168 if (Opc == AMDGPUISD::BFE_I32)
2169 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2171 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2178 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2180 const SelectionDAG &DAG,
2181 unsigned Depth) const {
2182 switch (Op.getOpcode()) {
2183 case AMDGPUISD::BFE_I32: {
2184 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2188 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2189 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2190 if (!Offset || !Offset->isNullValue())
2193 // TODO: Could probably figure something out with non-0 offsets.
2194 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2195 return std::max(SignBits, Op0SignBits);
2198 case AMDGPUISD::BFE_U32: {
2199 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2200 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;