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 // Initialize target lowering borrowed from AMDIL
114 setOperationAction(ISD::Constant, MVT::i32, Legal);
115 setOperationAction(ISD::Constant, MVT::i64, Legal);
116 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
117 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
119 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
120 setOperationAction(ISD::BRIND, MVT::Other, Expand);
122 // We need to custom lower some of the intrinsics
123 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
125 // Library functions. These default to Expand, but we have instructions
127 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
128 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
129 setOperationAction(ISD::FPOW, MVT::f32, Legal);
130 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
131 setOperationAction(ISD::FABS, MVT::f32, Legal);
132 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
133 setOperationAction(ISD::FRINT, MVT::f32, Legal);
134 setOperationAction(ISD::FROUND, MVT::f32, Legal);
135 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
137 // Lower floating point store/load to integer store/load to reduce the number
138 // of patterns in tablegen.
139 setOperationAction(ISD::STORE, MVT::f32, Promote);
140 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
142 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
143 AddPromotedToType(ISD::STORE, MVT::v2f32, 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::SELECT, VT, Expand);
310 setOperationAction(ISD::VSELECT, VT, Expand);
311 setOperationAction(ISD::SELECT_CC, VT, Expand);
312 setOperationAction(ISD::XOR, VT, Expand);
313 setOperationAction(ISD::BSWAP, VT, Expand);
314 setOperationAction(ISD::CTPOP, VT, Expand);
315 setOperationAction(ISD::CTTZ, VT, Expand);
316 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
317 setOperationAction(ISD::CTLZ, VT, Expand);
318 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
319 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
322 static const MVT::SimpleValueType FloatVectorTypes[] = {
323 MVT::v2f32, MVT::v4f32
326 for (MVT VT : FloatVectorTypes) {
327 setOperationAction(ISD::FABS, VT, Expand);
328 setOperationAction(ISD::FADD, VT, Expand);
329 setOperationAction(ISD::FCEIL, VT, Expand);
330 setOperationAction(ISD::FCOS, VT, Expand);
331 setOperationAction(ISD::FDIV, VT, Expand);
332 setOperationAction(ISD::FEXP2, VT, Expand);
333 setOperationAction(ISD::FLOG2, VT, Expand);
334 setOperationAction(ISD::FPOW, VT, Expand);
335 setOperationAction(ISD::FFLOOR, VT, Expand);
336 setOperationAction(ISD::FTRUNC, VT, Expand);
337 setOperationAction(ISD::FMUL, VT, Expand);
338 setOperationAction(ISD::FRINT, VT, Expand);
339 setOperationAction(ISD::FNEARBYINT, VT, Expand);
340 setOperationAction(ISD::FSQRT, VT, Expand);
341 setOperationAction(ISD::FSIN, VT, Expand);
342 setOperationAction(ISD::FSUB, VT, Expand);
343 setOperationAction(ISD::FNEG, VT, Expand);
344 setOperationAction(ISD::SELECT, VT, Expand);
345 setOperationAction(ISD::VSELECT, VT, Expand);
346 setOperationAction(ISD::SELECT_CC, VT, Expand);
347 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
348 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
351 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
352 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
354 setTargetDAGCombine(ISD::MUL);
355 setTargetDAGCombine(ISD::SELECT_CC);
357 setSchedulingPreference(Sched::RegPressure);
358 setJumpIsExpensive(true);
360 // There are no integer divide instructions, and these expand to a pretty
361 // large sequence of instructions.
362 setIntDivIsCheap(false);
363 setPow2DivIsCheap(false);
365 // TODO: Investigate this when 64-bit divides are implemented.
366 addBypassSlowDiv(64, 32);
368 // FIXME: Need to really handle these.
369 MaxStoresPerMemcpy = 4096;
370 MaxStoresPerMemmove = 4096;
371 MaxStoresPerMemset = 4096;
374 //===----------------------------------------------------------------------===//
375 // Target Information
376 //===----------------------------------------------------------------------===//
378 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
382 // The backend supports 32 and 64 bit floating point immediates.
383 // FIXME: Why are we reporting vectors of FP immediates as legal?
384 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
385 EVT ScalarVT = VT.getScalarType();
386 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
389 // We don't want to shrink f64 / f32 constants.
390 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
391 EVT ScalarVT = VT.getScalarType();
392 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
395 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
397 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
400 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
401 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
403 return ((LScalarSize <= CastScalarSize) ||
404 (CastScalarSize >= 32) ||
408 //===---------------------------------------------------------------------===//
410 //===---------------------------------------------------------------------===//
412 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
413 assert(VT.isFloatingPoint());
414 return VT == MVT::f32;
417 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
418 assert(VT.isFloatingPoint());
419 return VT == MVT::f32;
422 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
423 // Truncate is just accessing a subregister.
424 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
427 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
428 // Truncate is just accessing a subregister.
429 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
430 (Dest->getPrimitiveSizeInBits() % 32 == 0);
433 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
434 const DataLayout *DL = getDataLayout();
435 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
436 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
438 return SrcSize == 32 && DestSize == 64;
441 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
442 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
443 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
444 // this will enable reducing 64-bit operations the 32-bit, which is always
446 return Src == MVT::i32 && Dest == MVT::i64;
449 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
450 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
451 // limited number of native 64-bit operations. Shrinking an operation to fit
452 // in a single 32-bit register should always be helpful. As currently used,
453 // this is much less general than the name suggests, and is only used in
454 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
455 // not profitable, and may actually be harmful.
456 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
459 //===---------------------------------------------------------------------===//
460 // TargetLowering Callbacks
461 //===---------------------------------------------------------------------===//
463 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
464 const SmallVectorImpl<ISD::InputArg> &Ins) const {
466 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
469 SDValue AMDGPUTargetLowering::LowerReturn(
471 CallingConv::ID CallConv,
473 const SmallVectorImpl<ISD::OutputArg> &Outs,
474 const SmallVectorImpl<SDValue> &OutVals,
475 SDLoc DL, SelectionDAG &DAG) const {
476 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
479 //===---------------------------------------------------------------------===//
480 // Target specific lowering
481 //===---------------------------------------------------------------------===//
483 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
484 SmallVectorImpl<SDValue> &InVals) const {
485 SDValue Callee = CLI.Callee;
486 SelectionDAG &DAG = CLI.DAG;
488 const Function &Fn = *DAG.getMachineFunction().getFunction();
490 StringRef FuncName("<unknown>");
492 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
493 FuncName = G->getSymbol();
494 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
495 FuncName = G->getGlobal()->getName();
497 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
498 DAG.getContext()->diagnose(NoCalls);
502 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
503 SelectionDAG &DAG) const {
504 switch (Op.getOpcode()) {
506 Op.getNode()->dump();
507 llvm_unreachable("Custom lowering code for this"
508 "instruction is not implemented yet!");
510 // AMDGPU DAG lowering.
511 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
512 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
513 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
514 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
515 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
516 case ISD::SDIV: return LowerSDIV(Op, DAG);
517 case ISD::SREM: return LowerSREM(Op, DAG);
518 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
519 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
520 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
521 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
522 case ISD::FRINT: return LowerFRINT(Op, DAG);
523 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
524 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
525 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
527 // AMDIL DAG lowering.
528 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
533 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
534 SmallVectorImpl<SDValue> &Results,
535 SelectionDAG &DAG) const {
536 switch (N->getOpcode()) {
537 case ISD::SIGN_EXTEND_INREG:
538 // Different parts of legalization seem to interpret which type of
539 // sign_extend_inreg is the one to check for custom lowering. The extended
540 // from type is what really matters, but some places check for custom
541 // lowering of the result type. This results in trying to use
542 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
543 // nothing here and let the illegal result integer be handled normally.
550 // FIXME: This implements accesses to initialized globals in the constant
551 // address space by copying them to private and accessing that. It does not
552 // properly handle illegal types or vectors. The private vector loads are not
553 // scalarized, and the illegal scalars hit an assertion. This technique will not
554 // work well with large initializers, and this should eventually be
555 // removed. Initialized globals should be placed into a data section that the
556 // runtime will load into a buffer before the kernel is executed. Uses of the
557 // global need to be replaced with a pointer loaded from an implicit kernel
558 // argument into this buffer holding the copy of the data, which will remove the
559 // need for any of this.
560 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
561 const GlobalValue *GV,
562 const SDValue &InitPtr,
564 SelectionDAG &DAG) const {
565 const DataLayout *TD = getTargetMachine().getDataLayout();
567 Type *InitTy = Init->getType();
569 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
570 EVT VT = EVT::getEVT(InitTy);
571 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
572 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
573 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
574 TD->getPrefTypeAlignment(InitTy));
577 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
578 EVT VT = EVT::getEVT(CFP->getType());
579 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
580 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
581 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
582 TD->getPrefTypeAlignment(CFP->getType()));
585 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
586 const StructLayout *SL = TD->getStructLayout(ST);
588 EVT PtrVT = InitPtr.getValueType();
589 SmallVector<SDValue, 8> Chains;
591 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
592 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
593 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
595 Constant *Elt = Init->getAggregateElement(I);
596 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
599 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
602 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
603 EVT PtrVT = InitPtr.getValueType();
605 unsigned NumElements;
606 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
607 NumElements = AT->getNumElements();
608 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
609 NumElements = VT->getNumElements();
611 llvm_unreachable("Unexpected type");
613 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
614 SmallVector<SDValue, 8> Chains;
615 for (unsigned i = 0; i < NumElements; ++i) {
616 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
617 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
619 Constant *Elt = Init->getAggregateElement(i);
620 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
623 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
626 if (isa<UndefValue>(Init)) {
627 EVT VT = EVT::getEVT(InitTy);
628 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
629 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
630 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
631 TD->getPrefTypeAlignment(InitTy));
635 llvm_unreachable("Unhandled constant initializer");
638 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
640 SelectionDAG &DAG) const {
642 const DataLayout *TD = getTargetMachine().getDataLayout();
643 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
644 const GlobalValue *GV = G->getGlobal();
646 switch (G->getAddressSpace()) {
647 default: llvm_unreachable("Global Address lowering not implemented for this "
649 case AMDGPUAS::LOCAL_ADDRESS: {
650 // XXX: What does the value of G->getOffset() mean?
651 assert(G->getOffset() == 0 &&
652 "Do not know what to do with an non-zero offset");
655 if (MFI->LocalMemoryObjects.count(GV) == 0) {
656 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
657 Offset = MFI->LDSSize;
658 MFI->LocalMemoryObjects[GV] = Offset;
659 // XXX: Account for alignment?
660 MFI->LDSSize += Size;
662 Offset = MFI->LocalMemoryObjects[GV];
665 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
667 case AMDGPUAS::CONSTANT_ADDRESS: {
668 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
669 Type *EltType = GV->getType()->getElementType();
670 unsigned Size = TD->getTypeAllocSize(EltType);
671 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
673 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
674 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
676 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
677 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
679 const GlobalVariable *Var = cast<GlobalVariable>(GV);
680 if (!Var->hasInitializer()) {
681 // This has no use, but bugpoint will hit it.
682 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
685 const Constant *Init = Var->getInitializer();
686 SmallVector<SDNode*, 8> WorkList;
688 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
689 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
690 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
692 WorkList.push_back(*I);
694 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
695 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
696 E = WorkList.end(); I != E; ++I) {
697 SmallVector<SDValue, 8> Ops;
698 Ops.push_back(Chain);
699 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
700 Ops.push_back((*I)->getOperand(i));
702 DAG.UpdateNodeOperands(*I, Ops);
704 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
709 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
710 SelectionDAG &DAG) const {
711 SmallVector<SDValue, 8> Args;
712 SDValue A = Op.getOperand(0);
713 SDValue B = Op.getOperand(1);
715 DAG.ExtractVectorElements(A, Args);
716 DAG.ExtractVectorElements(B, Args);
718 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
721 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
722 SelectionDAG &DAG) const {
724 SmallVector<SDValue, 8> Args;
725 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
726 EVT VT = Op.getValueType();
727 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
728 VT.getVectorNumElements());
730 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
733 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
734 SelectionDAG &DAG) const {
736 MachineFunction &MF = DAG.getMachineFunction();
737 const AMDGPUFrameLowering *TFL =
738 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
740 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
742 unsigned FrameIndex = FIN->getIndex();
743 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
744 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
748 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
749 SelectionDAG &DAG) const {
750 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
752 EVT VT = Op.getValueType();
754 switch (IntrinsicID) {
756 case AMDGPUIntrinsic::AMDGPU_abs:
757 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
758 return LowerIntrinsicIABS(Op, DAG);
759 case AMDGPUIntrinsic::AMDGPU_lrp:
760 return LowerIntrinsicLRP(Op, DAG);
761 case AMDGPUIntrinsic::AMDGPU_fract:
762 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
763 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
765 case AMDGPUIntrinsic::AMDGPU_clamp:
766 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
767 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
768 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
770 case Intrinsic::AMDGPU_div_scale:
771 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
772 Op.getOperand(1), Op.getOperand(2));
774 case Intrinsic::AMDGPU_div_fmas:
775 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
776 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
778 case Intrinsic::AMDGPU_div_fixup:
779 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
780 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
782 case Intrinsic::AMDGPU_trig_preop:
783 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
784 Op.getOperand(1), Op.getOperand(2));
786 case Intrinsic::AMDGPU_rcp:
787 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
789 case Intrinsic::AMDGPU_rsq:
790 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
792 case AMDGPUIntrinsic::AMDGPU_imax:
793 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
795 case AMDGPUIntrinsic::AMDGPU_umax:
796 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
798 case AMDGPUIntrinsic::AMDGPU_imin:
799 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
801 case AMDGPUIntrinsic::AMDGPU_umin:
802 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
805 case AMDGPUIntrinsic::AMDGPU_umul24:
806 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
807 Op.getOperand(1), Op.getOperand(2));
809 case AMDGPUIntrinsic::AMDGPU_imul24:
810 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
811 Op.getOperand(1), Op.getOperand(2));
813 case AMDGPUIntrinsic::AMDGPU_umad24:
814 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
815 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
817 case AMDGPUIntrinsic::AMDGPU_imad24:
818 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
819 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
821 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
822 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
824 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
825 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
827 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
828 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
830 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
831 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
833 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
834 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
839 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
840 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
845 case AMDGPUIntrinsic::AMDGPU_bfi:
846 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
851 case AMDGPUIntrinsic::AMDGPU_bfm:
852 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
856 case AMDGPUIntrinsic::AMDGPU_brev:
857 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
859 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
860 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
862 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
863 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
864 case AMDGPUIntrinsic::AMDGPU_trunc:
865 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
869 ///IABS(a) = SMAX(sub(0, a), a)
870 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
871 SelectionDAG &DAG) const {
873 EVT VT = Op.getValueType();
874 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
877 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
880 /// Linear Interpolation
881 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
882 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
883 SelectionDAG &DAG) const {
885 EVT VT = Op.getValueType();
886 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
887 DAG.getConstantFP(1.0f, MVT::f32),
889 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
891 return DAG.getNode(ISD::FADD, DL, VT,
892 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
896 /// \brief Generate Min/Max node
897 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
898 SelectionDAG &DAG) const {
900 EVT VT = N->getValueType(0);
902 SDValue LHS = N->getOperand(0);
903 SDValue RHS = N->getOperand(1);
904 SDValue True = N->getOperand(2);
905 SDValue False = N->getOperand(3);
906 SDValue CC = N->getOperand(4);
908 if (VT != MVT::f32 ||
909 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
913 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
927 llvm_unreachable("Operation should already be optimised!");
934 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
935 return DAG.getNode(Opc, DL, VT, LHS, RHS);
943 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
944 return DAG.getNode(Opc, DL, VT, LHS, RHS);
946 case ISD::SETCC_INVALID:
947 llvm_unreachable("Invalid setcc condcode!");
952 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
953 SelectionDAG &DAG) const {
954 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
955 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
956 EVT EltVT = Op.getValueType().getVectorElementType();
957 EVT PtrVT = Load->getBasePtr().getValueType();
958 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
959 SmallVector<SDValue, 8> Loads;
962 for (unsigned i = 0, e = NumElts; i != e; ++i) {
963 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
964 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
965 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
966 Load->getChain(), Ptr,
967 MachinePointerInfo(Load->getMemOperand()->getValue()),
968 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
969 Load->getAlignment()));
971 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
974 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
975 SelectionDAG &DAG) const {
976 StoreSDNode *Store = cast<StoreSDNode>(Op);
977 EVT MemVT = Store->getMemoryVT();
978 unsigned MemBits = MemVT.getSizeInBits();
980 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
981 // truncating store into an i32 store.
982 // XXX: We could also handle optimize other vector bitwidths.
983 if (!MemVT.isVector() || MemBits > 32) {
988 SDValue Value = Store->getValue();
989 EVT VT = Value.getValueType();
990 EVT ElemVT = VT.getVectorElementType();
991 SDValue Ptr = Store->getBasePtr();
992 EVT MemEltVT = MemVT.getVectorElementType();
993 unsigned MemEltBits = MemEltVT.getSizeInBits();
994 unsigned MemNumElements = MemVT.getVectorNumElements();
995 unsigned PackedSize = MemVT.getStoreSizeInBits();
996 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
998 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1000 SDValue PackedValue;
1001 for (unsigned i = 0; i < MemNumElements; ++i) {
1002 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1003 DAG.getConstant(i, MVT::i32));
1004 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1005 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1007 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1008 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1013 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1017 if (PackedSize < 32) {
1018 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1019 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1020 Store->getMemOperand()->getPointerInfo(),
1022 Store->isNonTemporal(), Store->isVolatile(),
1023 Store->getAlignment());
1026 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1027 Store->getMemOperand()->getPointerInfo(),
1028 Store->isVolatile(), Store->isNonTemporal(),
1029 Store->getAlignment());
1032 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1033 SelectionDAG &DAG) const {
1034 StoreSDNode *Store = cast<StoreSDNode>(Op);
1035 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1036 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1037 EVT PtrVT = Store->getBasePtr().getValueType();
1038 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1041 SmallVector<SDValue, 8> Chains;
1043 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1044 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1045 Store->getValue(), DAG.getConstant(i, MVT::i32));
1046 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1047 Store->getBasePtr(),
1048 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1050 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1051 MachinePointerInfo(Store->getMemOperand()->getValue()),
1052 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1053 Store->getAlignment()));
1055 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1058 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1060 LoadSDNode *Load = cast<LoadSDNode>(Op);
1061 ISD::LoadExtType ExtType = Load->getExtensionType();
1062 EVT VT = Op.getValueType();
1063 EVT MemVT = Load->getMemoryVT();
1065 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1066 // We can do the extload to 32-bits, and then need to separately extend to
1069 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1073 Load->getMemOperand());
1074 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1077 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1078 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1079 // FIXME: Copied from PPC
1080 // First, load into 32 bits, then truncate to 1 bit.
1082 SDValue Chain = Load->getChain();
1083 SDValue BasePtr = Load->getBasePtr();
1084 MachineMemOperand *MMO = Load->getMemOperand();
1086 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1087 BasePtr, MVT::i8, MMO);
1088 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1091 // Lower loads constant address space global variable loads
1092 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1093 isa<GlobalVariable>(
1094 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1096 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1097 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1098 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1099 DAG.getConstant(2, MVT::i32));
1100 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1101 Load->getChain(), Ptr,
1102 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1105 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1106 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1110 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1111 DAG.getConstant(2, MVT::i32));
1112 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1113 Load->getChain(), Ptr,
1114 DAG.getTargetConstant(0, MVT::i32),
1116 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1118 DAG.getConstant(0x3, MVT::i32));
1119 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1120 DAG.getConstant(3, MVT::i32));
1122 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1124 EVT MemEltVT = MemVT.getScalarType();
1125 if (ExtType == ISD::SEXTLOAD) {
1126 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1127 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1130 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1133 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1135 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1136 if (Result.getNode()) {
1140 StoreSDNode *Store = cast<StoreSDNode>(Op);
1141 SDValue Chain = Store->getChain();
1142 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1143 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1144 Store->getValue().getValueType().isVector()) {
1145 return SplitVectorStore(Op, DAG);
1148 EVT MemVT = Store->getMemoryVT();
1149 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1150 MemVT.bitsLT(MVT::i32)) {
1152 if (Store->getMemoryVT() == MVT::i8) {
1154 } else if (Store->getMemoryVT() == MVT::i16) {
1157 SDValue BasePtr = Store->getBasePtr();
1158 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1159 DAG.getConstant(2, MVT::i32));
1160 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1161 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1163 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1164 DAG.getConstant(0x3, MVT::i32));
1166 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1167 DAG.getConstant(3, MVT::i32));
1169 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1172 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1174 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1175 MaskedValue, ShiftAmt);
1177 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1179 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1180 DAG.getConstant(0xffffffff, MVT::i32));
1181 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1183 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1184 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1185 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1190 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1192 EVT OVT = Op.getValueType();
1193 SDValue LHS = Op.getOperand(0);
1194 SDValue RHS = Op.getOperand(1);
1197 if (!OVT.isVector()) {
1200 } else if (OVT.getVectorNumElements() == 2) {
1203 } else if (OVT.getVectorNumElements() == 4) {
1207 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1208 // char|short jq = ia ^ ib;
1209 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1211 // jq = jq >> (bitsize - 2)
1212 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1215 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1218 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1220 // int ia = (int)LHS;
1221 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1223 // int ib, (int)RHS;
1224 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1226 // float fa = (float)ia;
1227 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1229 // float fb = (float)ib;
1230 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1232 // float fq = native_divide(fa, fb);
1233 SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
1236 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1238 // float fqneg = -fq;
1239 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1241 // float fr = mad(fqneg, fb, fa);
1242 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1243 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1245 // int iq = (int)fq;
1246 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1249 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1252 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1254 // int cv = fr >= fb;
1256 if (INTTY == MVT::i32) {
1257 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1259 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1261 // jq = (cv ? jq : 0);
1262 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1263 DAG.getConstant(0, OVT));
1265 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1266 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1270 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1272 EVT OVT = Op.getValueType();
1273 SDValue LHS = Op.getOperand(0);
1274 SDValue RHS = Op.getOperand(1);
1275 // The LowerSDIV32 function generates equivalent to the following IL.
1285 // ixor r10, r10, r11
1287 // ixor DST, r0, r10
1296 SDValue r10 = DAG.getSelectCC(DL,
1297 r0, DAG.getConstant(0, OVT),
1298 DAG.getConstant(-1, OVT),
1299 DAG.getConstant(0, OVT),
1303 SDValue r11 = DAG.getSelectCC(DL,
1304 r1, DAG.getConstant(0, OVT),
1305 DAG.getConstant(-1, OVT),
1306 DAG.getConstant(0, OVT),
1310 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1313 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1316 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1319 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1322 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1324 // ixor r10, r10, r11
1325 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1328 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1330 // ixor DST, r0, r10
1331 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1335 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1336 return SDValue(Op.getNode(), 0);
1339 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1340 EVT OVT = Op.getValueType().getScalarType();
1342 if (OVT == MVT::i64)
1343 return LowerSDIV64(Op, DAG);
1345 if (OVT.getScalarType() == MVT::i32)
1346 return LowerSDIV32(Op, DAG);
1348 if (OVT == MVT::i16 || OVT == MVT::i8) {
1349 // FIXME: We should be checking for the masked bits. This isn't reached
1350 // because i8 and i16 are not legal types.
1351 return LowerSDIV24(Op, DAG);
1354 return SDValue(Op.getNode(), 0);
1357 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1359 EVT OVT = Op.getValueType();
1360 SDValue LHS = Op.getOperand(0);
1361 SDValue RHS = Op.getOperand(1);
1362 // The LowerSREM32 function generates equivalent to the following IL.
1372 // umul r20, r20, r1
1375 // ixor DST, r0, r10
1384 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1387 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1390 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1393 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1396 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1399 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1402 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1404 // umul r20, r20, r1
1405 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1408 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1411 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1413 // ixor DST, r0, r10
1414 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1418 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1419 return SDValue(Op.getNode(), 0);
1422 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1423 EVT OVT = Op.getValueType();
1425 if (OVT.getScalarType() == MVT::i64)
1426 return LowerSREM64(Op, DAG);
1428 if (OVT.getScalarType() == MVT::i32)
1429 return LowerSREM32(Op, DAG);
1431 return SDValue(Op.getNode(), 0);
1434 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1435 SelectionDAG &DAG) const {
1437 EVT VT = Op.getValueType();
1439 SDValue Num = Op.getOperand(0);
1440 SDValue Den = Op.getOperand(1);
1442 // RCP = URECIP(Den) = 2^32 / Den + e
1443 // e is rounding error.
1444 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1446 // RCP_LO = umulo(RCP, Den) */
1447 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1449 // RCP_HI = mulhu (RCP, Den) */
1450 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1452 // NEG_RCP_LO = -RCP_LO
1453 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1456 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1457 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1460 // Calculate the rounding error from the URECIP instruction
1461 // E = mulhu(ABS_RCP_LO, RCP)
1462 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1464 // RCP_A_E = RCP + E
1465 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1467 // RCP_S_E = RCP - E
1468 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1470 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1471 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1474 // Quotient = mulhu(Tmp0, Num)
1475 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1477 // Num_S_Remainder = Quotient * Den
1478 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1480 // Remainder = Num - Num_S_Remainder
1481 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1483 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1484 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1485 DAG.getConstant(-1, VT),
1486 DAG.getConstant(0, VT),
1488 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1489 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1491 DAG.getConstant(-1, VT),
1492 DAG.getConstant(0, VT),
1494 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1495 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1498 // Calculate Division result:
1500 // Quotient_A_One = Quotient + 1
1501 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1502 DAG.getConstant(1, VT));
1504 // Quotient_S_One = Quotient - 1
1505 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1506 DAG.getConstant(1, VT));
1508 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1509 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1510 Quotient, Quotient_A_One, ISD::SETEQ);
1512 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1513 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1514 Quotient_S_One, Div, ISD::SETEQ);
1516 // Calculate Rem result:
1518 // Remainder_S_Den = Remainder - Den
1519 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1521 // Remainder_A_Den = Remainder + Den
1522 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1524 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1525 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1526 Remainder, Remainder_S_Den, ISD::SETEQ);
1528 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1529 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1530 Remainder_A_Den, Rem, ISD::SETEQ);
1535 return DAG.getMergeValues(Ops, DL);
1538 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1539 SelectionDAG &DAG) const {
1541 EVT VT = Op.getValueType();
1543 SDValue Zero = DAG.getConstant(0, VT);
1544 SDValue NegOne = DAG.getConstant(-1, VT);
1546 SDValue LHS = Op.getOperand(0);
1547 SDValue RHS = Op.getOperand(1);
1549 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1550 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1551 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1552 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1554 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1555 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1557 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1558 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1560 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1561 SDValue Rem = Div.getValue(1);
1563 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1564 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1566 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1567 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1573 return DAG.getMergeValues(Res, DL);
1576 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1578 SDValue Src = Op.getOperand(0);
1580 // result = trunc(src)
1581 // if (src > 0.0 && src != result)
1584 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1586 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1587 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1589 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1591 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1592 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1593 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1595 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1596 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1599 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1601 SDValue Src = Op.getOperand(0);
1603 assert(Op.getValueType() == MVT::f64);
1605 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1606 const SDValue One = DAG.getConstant(1, MVT::i32);
1608 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1610 // Extract the upper half, since this is where we will find the sign and
1612 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1614 const unsigned FractBits = 52;
1615 const unsigned ExpBits = 11;
1617 // Extract the exponent.
1618 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1620 DAG.getConstant(FractBits - 32, MVT::i32),
1621 DAG.getConstant(ExpBits, MVT::i32));
1622 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1623 DAG.getConstant(1023, MVT::i32));
1625 // Extract the sign bit.
1626 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1627 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1629 // Extend back to to 64-bits.
1630 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1632 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1634 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1635 const SDValue FractMask
1636 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1638 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1639 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1640 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1642 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1644 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1646 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1647 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1649 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1650 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1652 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1655 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1657 SDValue Src = Op.getOperand(0);
1659 assert(Op.getValueType() == MVT::f64);
1661 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1662 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1663 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1665 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1666 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1668 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1670 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1671 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1673 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1674 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1676 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1679 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1680 // FNEARBYINT and FRINT are the same, except in their handling of FP
1681 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1682 // rint, so just treat them as equivalent.
1683 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1686 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1688 SDValue Src = Op.getOperand(0);
1690 // result = trunc(src);
1691 // if (src < 0.0 && src != result)
1694 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1696 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1697 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1699 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1701 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1702 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1703 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1705 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1706 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1709 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1710 SelectionDAG &DAG) const {
1711 SDValue S0 = Op.getOperand(0);
1713 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1716 // f32 uint_to_fp i64
1717 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1718 DAG.getConstant(0, MVT::i32));
1719 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1720 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1721 DAG.getConstant(1, MVT::i32));
1722 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1723 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1724 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1725 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1728 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1730 SelectionDAG &DAG) const {
1731 MVT VT = Op.getSimpleValueType();
1733 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1734 // Shift left by 'Shift' bits.
1735 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1736 // Signed shift Right by 'Shift' bits.
1737 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1740 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1741 SelectionDAG &DAG) const {
1742 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1743 MVT VT = Op.getSimpleValueType();
1744 MVT ScalarVT = VT.getScalarType();
1749 SDValue Src = Op.getOperand(0);
1752 // TODO: Don't scalarize on Evergreen?
1753 unsigned NElts = VT.getVectorNumElements();
1754 SmallVector<SDValue, 8> Args;
1755 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1757 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1758 for (unsigned I = 0; I < NElts; ++I)
1759 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1761 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1764 //===----------------------------------------------------------------------===//
1765 // Custom DAG optimizations
1766 //===----------------------------------------------------------------------===//
1768 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1769 APInt KnownZero, KnownOne;
1770 EVT VT = Op.getValueType();
1771 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1773 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1776 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1777 EVT VT = Op.getValueType();
1779 // In order for this to be a signed 24-bit value, bit 23, must
1781 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1782 // as unsigned 24-bit values.
1783 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1786 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1788 SelectionDAG &DAG = DCI.DAG;
1789 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1790 EVT VT = Op.getValueType();
1792 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1793 APInt KnownZero, KnownOne;
1794 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1795 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1796 DCI.CommitTargetLoweringOpt(TLO);
1799 template <typename IntTy>
1800 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1801 uint32_t Offset, uint32_t Width) {
1802 if (Width + Offset < 32) {
1803 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1804 return DAG.getConstant(Result, MVT::i32);
1807 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1810 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1811 DAGCombinerInfo &DCI) const {
1812 SelectionDAG &DAG = DCI.DAG;
1815 switch(N->getOpcode()) {
1818 EVT VT = N->getValueType(0);
1819 SDValue N0 = N->getOperand(0);
1820 SDValue N1 = N->getOperand(1);
1823 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1824 if (VT.isVector() || VT.getSizeInBits() > 32)
1827 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1828 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1829 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1830 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1831 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1832 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1833 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1834 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1839 // We need to use sext even for MUL_U24, because MUL_U24 is used
1840 // for signed multiply of 8 and 16-bit types.
1841 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1845 case AMDGPUISD::MUL_I24:
1846 case AMDGPUISD::MUL_U24: {
1847 SDValue N0 = N->getOperand(0);
1848 SDValue N1 = N->getOperand(1);
1849 simplifyI24(N0, DCI);
1850 simplifyI24(N1, DCI);
1853 case ISD::SELECT_CC: {
1854 return CombineMinMax(N, DAG);
1856 case AMDGPUISD::BFE_I32:
1857 case AMDGPUISD::BFE_U32: {
1858 assert(!N->getValueType(0).isVector() &&
1859 "Vector handling of BFE not implemented");
1860 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1864 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1866 return DAG.getConstant(0, MVT::i32);
1868 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1872 SDValue BitsFrom = N->getOperand(0);
1873 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1875 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1877 if (OffsetVal == 0) {
1878 // This is already sign / zero extended, so try to fold away extra BFEs.
1879 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1881 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1882 if (OpSignBits >= SignBits)
1885 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1887 // This is a sign_extend_inreg. Replace it to take advantage of existing
1888 // DAG Combines. If not eliminated, we will match back to BFE during
1891 // TODO: The sext_inreg of extended types ends, although we can could
1892 // handle them in a single BFE.
1893 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1894 DAG.getValueType(SmallVT));
1897 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1900 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1902 return constantFoldBFE<int32_t>(DAG,
1903 Val->getSExtValue(),
1908 return constantFoldBFE<uint32_t>(DAG,
1909 Val->getZExtValue(),
1914 APInt Demanded = APInt::getBitsSet(32,
1916 OffsetVal + WidthVal);
1918 if ((OffsetVal + WidthVal) >= 32) {
1919 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1920 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1921 BitsFrom, ShiftVal);
1924 APInt KnownZero, KnownOne;
1925 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1926 !DCI.isBeforeLegalizeOps());
1927 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1928 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1929 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1930 DCI.CommitTargetLoweringOpt(TLO);
1939 //===----------------------------------------------------------------------===//
1941 //===----------------------------------------------------------------------===//
1943 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1946 const SmallVectorImpl<ISD::InputArg> &Ins,
1947 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1949 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1950 if (Ins[i].ArgVT == Ins[i].VT) {
1951 OrigIns.push_back(Ins[i]);
1956 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1957 // Vector has been split into scalars.
1958 VT = Ins[i].ArgVT.getVectorElementType();
1959 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1960 Ins[i].ArgVT.getVectorElementType() !=
1961 Ins[i].VT.getVectorElementType()) {
1962 // Vector elements have been promoted
1965 // Vector has been spilt into smaller vectors.
1969 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1970 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1971 OrigIns.push_back(Arg);
1975 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1976 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1977 return CFP->isExactlyValue(1.0);
1979 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1980 return C->isAllOnesValue();
1985 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1986 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1987 return CFP->getValueAPF().isZero();
1989 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1990 return C->isNullValue();
1995 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1996 const TargetRegisterClass *RC,
1997 unsigned Reg, EVT VT) const {
1998 MachineFunction &MF = DAG.getMachineFunction();
1999 MachineRegisterInfo &MRI = MF.getRegInfo();
2000 unsigned VirtualRegister;
2001 if (!MRI.isLiveIn(Reg)) {
2002 VirtualRegister = MRI.createVirtualRegister(RC);
2003 MRI.addLiveIn(Reg, VirtualRegister);
2005 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2007 return DAG.getRegister(VirtualRegister, VT);
2010 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2012 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2014 default: return nullptr;
2016 NODE_NAME_CASE(CALL);
2017 NODE_NAME_CASE(UMUL);
2018 NODE_NAME_CASE(DIV_INF);
2019 NODE_NAME_CASE(RET_FLAG);
2020 NODE_NAME_CASE(BRANCH_COND);
2023 NODE_NAME_CASE(DWORDADDR)
2024 NODE_NAME_CASE(FRACT)
2025 NODE_NAME_CASE(CLAMP)
2026 NODE_NAME_CASE(FMAX)
2027 NODE_NAME_CASE(SMAX)
2028 NODE_NAME_CASE(UMAX)
2029 NODE_NAME_CASE(FMIN)
2030 NODE_NAME_CASE(SMIN)
2031 NODE_NAME_CASE(UMIN)
2032 NODE_NAME_CASE(URECIP)
2033 NODE_NAME_CASE(DIV_SCALE)
2034 NODE_NAME_CASE(DIV_FMAS)
2035 NODE_NAME_CASE(DIV_FIXUP)
2036 NODE_NAME_CASE(TRIG_PREOP)
2039 NODE_NAME_CASE(DOT4)
2040 NODE_NAME_CASE(BFE_U32)
2041 NODE_NAME_CASE(BFE_I32)
2044 NODE_NAME_CASE(BREV)
2045 NODE_NAME_CASE(MUL_U24)
2046 NODE_NAME_CASE(MUL_I24)
2047 NODE_NAME_CASE(MAD_U24)
2048 NODE_NAME_CASE(MAD_I24)
2049 NODE_NAME_CASE(EXPORT)
2050 NODE_NAME_CASE(CONST_ADDRESS)
2051 NODE_NAME_CASE(REGISTER_LOAD)
2052 NODE_NAME_CASE(REGISTER_STORE)
2053 NODE_NAME_CASE(LOAD_CONSTANT)
2054 NODE_NAME_CASE(LOAD_INPUT)
2055 NODE_NAME_CASE(SAMPLE)
2056 NODE_NAME_CASE(SAMPLEB)
2057 NODE_NAME_CASE(SAMPLED)
2058 NODE_NAME_CASE(SAMPLEL)
2059 NODE_NAME_CASE(CVT_F32_UBYTE0)
2060 NODE_NAME_CASE(CVT_F32_UBYTE1)
2061 NODE_NAME_CASE(CVT_F32_UBYTE2)
2062 NODE_NAME_CASE(CVT_F32_UBYTE3)
2063 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2064 NODE_NAME_CASE(STORE_MSKOR)
2065 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2069 static void computeKnownBitsForMinMax(const SDValue Op0,
2073 const SelectionDAG &DAG,
2075 APInt Op0Zero, Op0One;
2076 APInt Op1Zero, Op1One;
2077 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2078 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2080 KnownZero = Op0Zero & Op1Zero;
2081 KnownOne = Op0One & Op1One;
2084 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2088 const SelectionDAG &DAG,
2089 unsigned Depth) const {
2091 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2095 unsigned Opc = Op.getOpcode();
2100 case ISD::INTRINSIC_WO_CHAIN: {
2101 // FIXME: The intrinsic should just use the node.
2102 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2103 case AMDGPUIntrinsic::AMDGPU_imax:
2104 case AMDGPUIntrinsic::AMDGPU_umax:
2105 case AMDGPUIntrinsic::AMDGPU_imin:
2106 case AMDGPUIntrinsic::AMDGPU_umin:
2107 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2108 KnownZero, KnownOne, DAG, Depth);
2116 case AMDGPUISD::SMAX:
2117 case AMDGPUISD::UMAX:
2118 case AMDGPUISD::SMIN:
2119 case AMDGPUISD::UMIN:
2120 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2121 KnownZero, KnownOne, DAG, Depth);
2124 case AMDGPUISD::BFE_I32:
2125 case AMDGPUISD::BFE_U32: {
2126 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2130 unsigned BitWidth = 32;
2131 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2133 KnownZero = APInt::getAllOnesValue(BitWidth);
2134 KnownOne = APInt::getNullValue(BitWidth);
2138 // FIXME: This could do a lot more. If offset is 0, should be the same as
2139 // sign_extend_inreg implementation, but that involves duplicating it.
2140 if (Opc == AMDGPUISD::BFE_I32)
2141 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2143 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2150 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2152 const SelectionDAG &DAG,
2153 unsigned Depth) const {
2154 switch (Op.getOpcode()) {
2155 case AMDGPUISD::BFE_I32: {
2156 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2160 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2161 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2162 if (!Offset || !Offset->isNullValue())
2165 // TODO: Could probably figure something out with non-0 offsets.
2166 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2167 return std::max(SignBits, Op0SignBits);
2170 case AMDGPUISD::BFE_U32: {
2171 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2172 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;