1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUFrameLowering.h"
19 #include "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDILIntrinsicInfo.h"
22 #include "R600MachineFunctionInfo.h"
23 #include "SIMachineFunctionInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
38 /// Diagnostic information for unimplemented or unsupported feature reporting.
39 class DiagnosticInfoUnsupported : public DiagnosticInfo {
41 const Twine &Description;
46 static int getKindID() {
48 KindID = llvm::getNextAvailablePluginDiagnosticKind();
53 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
54 DiagnosticSeverity Severity = DS_Error)
55 : DiagnosticInfo(getKindID(), Severity),
59 const Function &getFunction() const { return Fn; }
60 const Twine &getDescription() const { return Description; }
62 void print(DiagnosticPrinter &DP) const override {
63 DP << "unsupported " << getDescription() << " in " << Fn.getName();
66 static bool classof(const DiagnosticInfo *DI) {
67 return DI->getKind() == getKindID();
71 int DiagnosticInfoUnsupported::KindID = 0;
75 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
76 CCValAssign::LocInfo LocInfo,
77 ISD::ArgFlagsTy ArgFlags, CCState &State) {
78 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
79 ArgFlags.getOrigAlign());
80 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
85 #include "AMDGPUGenCallingConv.inc"
87 // Find a larger type to do a load / store of a vector with.
88 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
89 unsigned StoreSize = VT.getStoreSizeInBits();
91 return EVT::getIntegerVT(Ctx, StoreSize);
93 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
94 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
97 // Type for a vector that will be loaded to.
98 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
99 unsigned StoreSize = VT.getStoreSizeInBits();
101 return EVT::getIntegerVT(Ctx, 32);
103 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
106 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
107 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
109 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
111 // Initialize target lowering borrowed from AMDIL
114 // We need to custom lower some of the intrinsics
115 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
117 // Library functions. These default to Expand, but we have instructions
119 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
120 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
121 setOperationAction(ISD::FPOW, MVT::f32, Legal);
122 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
123 setOperationAction(ISD::FABS, MVT::f32, Legal);
124 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
125 setOperationAction(ISD::FRINT, MVT::f32, Legal);
126 setOperationAction(ISD::FROUND, MVT::f32, Legal);
127 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
129 // Lower floating point store/load to integer store/load to reduce the number
130 // of patterns in tablegen.
131 setOperationAction(ISD::STORE, MVT::f32, Promote);
132 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
134 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
135 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
137 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
138 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
140 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
141 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
143 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
144 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
146 setOperationAction(ISD::STORE, MVT::f64, Promote);
147 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
149 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
150 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
152 // Custom lowering of vector stores is required for local address space
154 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
155 // XXX: Native v2i32 local address space stores are possible, but not
156 // currently implemented.
157 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
159 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
160 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
161 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
163 // XXX: This can be change to Custom, once ExpandVectorStores can
164 // handle 64-bit stores.
165 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
167 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
168 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
169 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
170 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
171 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
174 setOperationAction(ISD::LOAD, MVT::f32, Promote);
175 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
177 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
178 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
180 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
181 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
183 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
184 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
186 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
187 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
189 setOperationAction(ISD::LOAD, MVT::f64, Promote);
190 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
192 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
193 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
195 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
196 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
197 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
198 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
199 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
200 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
201 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
202 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
203 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
204 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
206 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
207 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
208 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
209 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
210 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
211 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
212 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
213 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
214 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
215 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
217 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
219 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
221 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
222 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
223 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
224 setOperationAction(ISD::FRINT, MVT::f64, Custom);
225 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
228 if (!Subtarget->hasBFI()) {
229 // fcopysign can be done in a single instruction with BFI.
230 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
231 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
234 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
235 for (MVT VT : ScalarIntVTs) {
236 setOperationAction(ISD::SREM, VT, Expand);
237 setOperationAction(ISD::SDIV, VT, Custom);
239 // GPU does not have divrem function for signed or unsigned.
240 setOperationAction(ISD::SDIVREM, VT, Expand);
241 setOperationAction(ISD::UDIVREM, VT, Custom);
243 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
244 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
245 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
247 setOperationAction(ISD::BSWAP, VT, Expand);
248 setOperationAction(ISD::CTTZ, VT, Expand);
249 setOperationAction(ISD::CTLZ, VT, Expand);
252 if (!Subtarget->hasBCNT(32))
253 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
255 if (!Subtarget->hasBCNT(64))
256 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
258 // The hardware supports 32-bit ROTR, but not ROTL.
259 setOperationAction(ISD::ROTL, MVT::i32, Expand);
260 setOperationAction(ISD::ROTL, MVT::i64, Expand);
261 setOperationAction(ISD::ROTR, MVT::i64, Expand);
263 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
264 setOperationAction(ISD::MUL, MVT::i64, Expand);
265 setOperationAction(ISD::MULHU, MVT::i64, Expand);
266 setOperationAction(ISD::MULHS, MVT::i64, Expand);
267 setOperationAction(ISD::SUB, MVT::i64, Expand);
268 setOperationAction(ISD::UDIV, MVT::i32, Expand);
269 setOperationAction(ISD::UREM, MVT::i32, Expand);
270 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
271 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
273 static const MVT::SimpleValueType VectorIntTypes[] = {
274 MVT::v2i32, MVT::v4i32
277 for (MVT VT : VectorIntTypes) {
278 // Expand the following operations for the current type by default.
279 setOperationAction(ISD::ADD, VT, Expand);
280 setOperationAction(ISD::AND, VT, Expand);
281 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
282 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
283 setOperationAction(ISD::MUL, VT, Expand);
284 setOperationAction(ISD::OR, VT, Expand);
285 setOperationAction(ISD::SHL, VT, Expand);
286 setOperationAction(ISD::SRA, VT, Expand);
287 setOperationAction(ISD::SRL, VT, Expand);
288 setOperationAction(ISD::ROTL, VT, Expand);
289 setOperationAction(ISD::ROTR, VT, Expand);
290 setOperationAction(ISD::SUB, VT, Expand);
291 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
292 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
293 // TODO: Implement custom UREM / SREM routines.
294 setOperationAction(ISD::SDIV, VT, Custom);
295 setOperationAction(ISD::UDIV, VT, Expand);
296 setOperationAction(ISD::SREM, VT, Expand);
297 setOperationAction(ISD::UREM, VT, Expand);
298 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
299 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
300 setOperationAction(ISD::SDIVREM, VT, Expand);
301 setOperationAction(ISD::UDIVREM, VT, Custom);
302 setOperationAction(ISD::SELECT, VT, Expand);
303 setOperationAction(ISD::VSELECT, VT, Expand);
304 setOperationAction(ISD::XOR, VT, Expand);
305 setOperationAction(ISD::BSWAP, VT, Expand);
306 setOperationAction(ISD::CTPOP, VT, Expand);
307 setOperationAction(ISD::CTTZ, VT, Expand);
308 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
309 setOperationAction(ISD::CTLZ, VT, Expand);
310 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
313 static const MVT::SimpleValueType FloatVectorTypes[] = {
314 MVT::v2f32, MVT::v4f32
317 for (MVT VT : FloatVectorTypes) {
318 setOperationAction(ISD::FABS, VT, Expand);
319 setOperationAction(ISD::FADD, VT, Expand);
320 setOperationAction(ISD::FCOS, VT, Expand);
321 setOperationAction(ISD::FDIV, VT, Expand);
322 setOperationAction(ISD::FPOW, VT, Expand);
323 setOperationAction(ISD::FFLOOR, VT, Expand);
324 setOperationAction(ISD::FTRUNC, VT, Expand);
325 setOperationAction(ISD::FMUL, VT, Expand);
326 setOperationAction(ISD::FRINT, VT, Expand);
327 setOperationAction(ISD::FSQRT, VT, Expand);
328 setOperationAction(ISD::FSIN, VT, Expand);
329 setOperationAction(ISD::FSUB, VT, Expand);
330 setOperationAction(ISD::FNEG, VT, Expand);
331 setOperationAction(ISD::SELECT, VT, Expand);
332 setOperationAction(ISD::VSELECT, VT, Expand);
333 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
336 setTargetDAGCombine(ISD::MUL);
337 setTargetDAGCombine(ISD::SELECT_CC);
339 setSchedulingPreference(Sched::RegPressure);
340 setJumpIsExpensive(true);
342 // There are no integer divide instructions, and these expand to a pretty
343 // large sequence of instructions.
344 setIntDivIsCheap(false);
346 // TODO: Investigate this when 64-bit divides are implemented.
347 addBypassSlowDiv(64, 32);
349 // FIXME: Need to really handle these.
350 MaxStoresPerMemcpy = 4096;
351 MaxStoresPerMemmove = 4096;
352 MaxStoresPerMemset = 4096;
355 //===----------------------------------------------------------------------===//
356 // Target Information
357 //===----------------------------------------------------------------------===//
359 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
363 // The backend supports 32 and 64 bit floating point immediates.
364 // FIXME: Why are we reporting vectors of FP immediates as legal?
365 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
366 EVT ScalarVT = VT.getScalarType();
367 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
370 // We don't want to shrink f64 / f32 constants.
371 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
372 EVT ScalarVT = VT.getScalarType();
373 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
376 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
378 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
381 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
382 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
384 return ((LScalarSize <= CastScalarSize) ||
385 (CastScalarSize >= 32) ||
389 //===---------------------------------------------------------------------===//
391 //===---------------------------------------------------------------------===//
393 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
394 assert(VT.isFloatingPoint());
395 return VT == MVT::f32;
398 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
399 assert(VT.isFloatingPoint());
400 return VT == MVT::f32;
403 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
404 // Truncate is just accessing a subregister.
405 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
408 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
409 // Truncate is just accessing a subregister.
410 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
411 (Dest->getPrimitiveSizeInBits() % 32 == 0);
414 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
415 const DataLayout *DL = getDataLayout();
416 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
417 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
419 return SrcSize == 32 && DestSize == 64;
422 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
423 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
424 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
425 // this will enable reducing 64-bit operations the 32-bit, which is always
427 return Src == MVT::i32 && Dest == MVT::i64;
430 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
431 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
432 // limited number of native 64-bit operations. Shrinking an operation to fit
433 // in a single 32-bit register should always be helpful. As currently used,
434 // this is much less general than the name suggests, and is only used in
435 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
436 // not profitable, and may actually be harmful.
437 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
440 //===---------------------------------------------------------------------===//
441 // TargetLowering Callbacks
442 //===---------------------------------------------------------------------===//
444 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
445 const SmallVectorImpl<ISD::InputArg> &Ins) const {
447 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
450 SDValue AMDGPUTargetLowering::LowerReturn(
452 CallingConv::ID CallConv,
454 const SmallVectorImpl<ISD::OutputArg> &Outs,
455 const SmallVectorImpl<SDValue> &OutVals,
456 SDLoc DL, SelectionDAG &DAG) const {
457 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
460 //===---------------------------------------------------------------------===//
461 // Target specific lowering
462 //===---------------------------------------------------------------------===//
464 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
465 SmallVectorImpl<SDValue> &InVals) const {
466 SDValue Callee = CLI.Callee;
467 SelectionDAG &DAG = CLI.DAG;
469 const Function &Fn = *DAG.getMachineFunction().getFunction();
471 StringRef FuncName("<unknown>");
473 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
474 FuncName = G->getSymbol();
475 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
476 FuncName = G->getGlobal()->getName();
478 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
479 DAG.getContext()->diagnose(NoCalls);
483 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
484 SelectionDAG &DAG) const {
485 switch (Op.getOpcode()) {
487 Op.getNode()->dump();
488 llvm_unreachable("Custom lowering code for this"
489 "instruction is not implemented yet!");
491 // AMDGPU DAG lowering.
492 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
493 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
494 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
495 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
496 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
497 case ISD::SDIV: return LowerSDIV(Op, DAG);
498 case ISD::SREM: return LowerSREM(Op, DAG);
499 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
500 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
501 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
502 case ISD::FRINT: return LowerFRINT(Op, DAG);
503 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
504 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
506 // AMDIL DAG lowering.
507 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
512 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
513 SmallVectorImpl<SDValue> &Results,
514 SelectionDAG &DAG) const {
515 switch (N->getOpcode()) {
516 case ISD::SIGN_EXTEND_INREG:
517 // Different parts of legalization seem to interpret which type of
518 // sign_extend_inreg is the one to check for custom lowering. The extended
519 // from type is what really matters, but some places check for custom
520 // lowering of the result type. This results in trying to use
521 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
522 // nothing here and let the illegal result integer be handled normally.
525 SDValue Op = SDValue(N, 0);
527 EVT VT = Op.getValueType();
528 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
529 N->getOperand(0), N->getOperand(1));
530 Results.push_back(UDIVREM);
534 SDValue Op = SDValue(N, 0);
536 EVT VT = Op.getValueType();
537 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
538 N->getOperand(0), N->getOperand(1));
539 Results.push_back(UDIVREM.getValue(1));
543 SDValue Op = SDValue(N, 0);
545 EVT VT = Op.getValueType();
546 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
548 SDValue one = DAG.getConstant(1, HalfVT);
549 SDValue zero = DAG.getConstant(0, HalfVT);
552 SDValue LHS = N->getOperand(0);
553 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
554 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
556 SDValue RHS = N->getOperand(1);
557 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
558 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
560 // Get Speculative values
561 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
562 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
564 SDValue REM_Hi = zero;
565 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
567 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
568 SDValue DIV_Lo = zero;
570 const unsigned halfBitWidth = HalfVT.getSizeInBits();
572 for (unsigned i = 0; i < halfBitWidth; ++i) {
573 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
574 // Get Value of high bit
576 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
577 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
579 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
580 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
583 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
584 DAG.getConstant(halfBitWidth - 1, HalfVT));
585 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
586 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
588 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
589 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
592 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
594 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
595 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
597 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
601 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
603 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
604 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
605 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
608 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
609 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
610 Results.push_back(DIV);
611 Results.push_back(REM);
619 // FIXME: This implements accesses to initialized globals in the constant
620 // address space by copying them to private and accessing that. It does not
621 // properly handle illegal types or vectors. The private vector loads are not
622 // scalarized, and the illegal scalars hit an assertion. This technique will not
623 // work well with large initializers, and this should eventually be
624 // removed. Initialized globals should be placed into a data section that the
625 // runtime will load into a buffer before the kernel is executed. Uses of the
626 // global need to be replaced with a pointer loaded from an implicit kernel
627 // argument into this buffer holding the copy of the data, which will remove the
628 // need for any of this.
629 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
630 const GlobalValue *GV,
631 const SDValue &InitPtr,
633 SelectionDAG &DAG) const {
634 const DataLayout *TD = getTargetMachine().getDataLayout();
636 Type *InitTy = Init->getType();
638 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
639 EVT VT = EVT::getEVT(InitTy);
640 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
641 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
642 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
643 TD->getPrefTypeAlignment(InitTy));
646 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
647 EVT VT = EVT::getEVT(CFP->getType());
648 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
649 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
650 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
651 TD->getPrefTypeAlignment(CFP->getType()));
654 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
655 const StructLayout *SL = TD->getStructLayout(ST);
657 EVT PtrVT = InitPtr.getValueType();
658 SmallVector<SDValue, 8> Chains;
660 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
661 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
662 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
664 Constant *Elt = Init->getAggregateElement(I);
665 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
668 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
671 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
672 EVT PtrVT = InitPtr.getValueType();
674 unsigned NumElements;
675 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
676 NumElements = AT->getNumElements();
677 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
678 NumElements = VT->getNumElements();
680 llvm_unreachable("Unexpected type");
682 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
683 SmallVector<SDValue, 8> Chains;
684 for (unsigned i = 0; i < NumElements; ++i) {
685 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
686 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
688 Constant *Elt = Init->getAggregateElement(i);
689 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
692 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
695 if (isa<UndefValue>(Init)) {
696 EVT VT = EVT::getEVT(InitTy);
697 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
698 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
699 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
700 TD->getPrefTypeAlignment(InitTy));
704 llvm_unreachable("Unhandled constant initializer");
707 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
709 SelectionDAG &DAG) const {
711 const DataLayout *TD = getTargetMachine().getDataLayout();
712 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
713 const GlobalValue *GV = G->getGlobal();
715 switch (G->getAddressSpace()) {
716 default: llvm_unreachable("Global Address lowering not implemented for this "
718 case AMDGPUAS::LOCAL_ADDRESS: {
719 // XXX: What does the value of G->getOffset() mean?
720 assert(G->getOffset() == 0 &&
721 "Do not know what to do with an non-zero offset");
724 if (MFI->LocalMemoryObjects.count(GV) == 0) {
725 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
726 Offset = MFI->LDSSize;
727 MFI->LocalMemoryObjects[GV] = Offset;
728 // XXX: Account for alignment?
729 MFI->LDSSize += Size;
731 Offset = MFI->LocalMemoryObjects[GV];
734 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
736 case AMDGPUAS::CONSTANT_ADDRESS: {
737 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
738 Type *EltType = GV->getType()->getElementType();
739 unsigned Size = TD->getTypeAllocSize(EltType);
740 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
742 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
743 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
745 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
746 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
748 const GlobalVariable *Var = cast<GlobalVariable>(GV);
749 if (!Var->hasInitializer()) {
750 // This has no use, but bugpoint will hit it.
751 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
754 const Constant *Init = Var->getInitializer();
755 SmallVector<SDNode*, 8> WorkList;
757 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
758 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
759 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
761 WorkList.push_back(*I);
763 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
764 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
765 E = WorkList.end(); I != E; ++I) {
766 SmallVector<SDValue, 8> Ops;
767 Ops.push_back(Chain);
768 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
769 Ops.push_back((*I)->getOperand(i));
771 DAG.UpdateNodeOperands(*I, Ops);
773 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
778 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
779 SelectionDAG &DAG) const {
780 SmallVector<SDValue, 8> Args;
781 SDValue A = Op.getOperand(0);
782 SDValue B = Op.getOperand(1);
784 DAG.ExtractVectorElements(A, Args);
785 DAG.ExtractVectorElements(B, Args);
787 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
790 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
791 SelectionDAG &DAG) const {
793 SmallVector<SDValue, 8> Args;
794 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
795 EVT VT = Op.getValueType();
796 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
797 VT.getVectorNumElements());
799 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
802 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
803 SelectionDAG &DAG) const {
805 MachineFunction &MF = DAG.getMachineFunction();
806 const AMDGPUFrameLowering *TFL =
807 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
809 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
811 unsigned FrameIndex = FIN->getIndex();
812 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
813 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
817 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
818 SelectionDAG &DAG) const {
819 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
821 EVT VT = Op.getValueType();
823 switch (IntrinsicID) {
825 case AMDGPUIntrinsic::AMDGPU_abs:
826 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
827 return LowerIntrinsicIABS(Op, DAG);
828 case AMDGPUIntrinsic::AMDGPU_lrp:
829 return LowerIntrinsicLRP(Op, DAG);
830 case AMDGPUIntrinsic::AMDGPU_fract:
831 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
832 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
834 case AMDGPUIntrinsic::AMDGPU_clamp:
835 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
836 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
837 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
839 case AMDGPUIntrinsic::AMDGPU_imax:
840 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
842 case AMDGPUIntrinsic::AMDGPU_umax:
843 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
845 case AMDGPUIntrinsic::AMDGPU_imin:
846 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
848 case AMDGPUIntrinsic::AMDGPU_umin:
849 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
852 case AMDGPUIntrinsic::AMDGPU_umul24:
853 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
854 Op.getOperand(1), Op.getOperand(2));
856 case AMDGPUIntrinsic::AMDGPU_imul24:
857 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
858 Op.getOperand(1), Op.getOperand(2));
860 case AMDGPUIntrinsic::AMDGPU_umad24:
861 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
862 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
864 case AMDGPUIntrinsic::AMDGPU_imad24:
865 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
866 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
868 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
869 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
871 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
872 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
874 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
875 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
877 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
878 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
880 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
881 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
886 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
887 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
892 case AMDGPUIntrinsic::AMDGPU_bfi:
893 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
898 case AMDGPUIntrinsic::AMDGPU_bfm:
899 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
903 case AMDGPUIntrinsic::AMDGPU_brev:
904 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
906 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
907 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
909 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
910 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
914 ///IABS(a) = SMAX(sub(0, a), a)
915 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
916 SelectionDAG &DAG) const {
918 EVT VT = Op.getValueType();
919 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
922 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
925 /// Linear Interpolation
926 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
927 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
928 SelectionDAG &DAG) const {
930 EVT VT = Op.getValueType();
931 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
932 DAG.getConstantFP(1.0f, MVT::f32),
934 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
936 return DAG.getNode(ISD::FADD, DL, VT,
937 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
941 /// \brief Generate Min/Max node
942 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
943 SelectionDAG &DAG) const {
945 EVT VT = N->getValueType(0);
947 SDValue LHS = N->getOperand(0);
948 SDValue RHS = N->getOperand(1);
949 SDValue True = N->getOperand(2);
950 SDValue False = N->getOperand(3);
951 SDValue CC = N->getOperand(4);
953 if (VT != MVT::f32 ||
954 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
958 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
972 llvm_unreachable("Operation should already be optimised!");
979 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
980 return DAG.getNode(Opc, DL, VT, LHS, RHS);
988 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
989 return DAG.getNode(Opc, DL, VT, LHS, RHS);
991 case ISD::SETCC_INVALID:
992 llvm_unreachable("Invalid setcc condcode!");
997 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
998 SelectionDAG &DAG) const {
999 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
1000 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
1001 EVT EltVT = Op.getValueType().getVectorElementType();
1002 EVT PtrVT = Load->getBasePtr().getValueType();
1003 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1004 SmallVector<SDValue, 8> Loads;
1007 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1008 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1009 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
1010 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1011 Load->getChain(), Ptr,
1012 MachinePointerInfo(Load->getMemOperand()->getValue()),
1013 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1014 Load->getAlignment()));
1016 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
1019 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1020 SelectionDAG &DAG) const {
1021 StoreSDNode *Store = cast<StoreSDNode>(Op);
1022 EVT MemVT = Store->getMemoryVT();
1023 unsigned MemBits = MemVT.getSizeInBits();
1025 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1026 // truncating store into an i32 store.
1027 // XXX: We could also handle optimize other vector bitwidths.
1028 if (!MemVT.isVector() || MemBits > 32) {
1033 SDValue Value = Store->getValue();
1034 EVT VT = Value.getValueType();
1035 EVT ElemVT = VT.getVectorElementType();
1036 SDValue Ptr = Store->getBasePtr();
1037 EVT MemEltVT = MemVT.getVectorElementType();
1038 unsigned MemEltBits = MemEltVT.getSizeInBits();
1039 unsigned MemNumElements = MemVT.getVectorNumElements();
1040 unsigned PackedSize = MemVT.getStoreSizeInBits();
1041 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1043 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1045 SDValue PackedValue;
1046 for (unsigned i = 0; i < MemNumElements; ++i) {
1047 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1048 DAG.getConstant(i, MVT::i32));
1049 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1050 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1052 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1053 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1058 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1062 if (PackedSize < 32) {
1063 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1064 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1065 Store->getMemOperand()->getPointerInfo(),
1067 Store->isNonTemporal(), Store->isVolatile(),
1068 Store->getAlignment());
1071 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1072 Store->getMemOperand()->getPointerInfo(),
1073 Store->isVolatile(), Store->isNonTemporal(),
1074 Store->getAlignment());
1077 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1078 SelectionDAG &DAG) const {
1079 StoreSDNode *Store = cast<StoreSDNode>(Op);
1080 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1081 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1082 EVT PtrVT = Store->getBasePtr().getValueType();
1083 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1086 SmallVector<SDValue, 8> Chains;
1088 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1089 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1090 Store->getValue(), DAG.getConstant(i, MVT::i32));
1091 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1092 Store->getBasePtr(),
1093 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1095 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1096 MachinePointerInfo(Store->getMemOperand()->getValue()),
1097 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1098 Store->getAlignment()));
1100 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1103 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1105 LoadSDNode *Load = cast<LoadSDNode>(Op);
1106 ISD::LoadExtType ExtType = Load->getExtensionType();
1107 EVT VT = Op.getValueType();
1108 EVT MemVT = Load->getMemoryVT();
1110 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1111 // We can do the extload to 32-bits, and then need to separately extend to
1114 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1118 Load->getMemOperand());
1119 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1122 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1123 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1124 // FIXME: Copied from PPC
1125 // First, load into 32 bits, then truncate to 1 bit.
1127 SDValue Chain = Load->getChain();
1128 SDValue BasePtr = Load->getBasePtr();
1129 MachineMemOperand *MMO = Load->getMemOperand();
1131 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1132 BasePtr, MVT::i8, MMO);
1133 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1136 // Lower loads constant address space global variable loads
1137 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1138 isa<GlobalVariable>(
1139 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1141 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1142 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1143 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1144 DAG.getConstant(2, MVT::i32));
1145 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1146 Load->getChain(), Ptr,
1147 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1150 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1151 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1155 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1156 DAG.getConstant(2, MVT::i32));
1157 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1158 Load->getChain(), Ptr,
1159 DAG.getTargetConstant(0, MVT::i32),
1161 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1163 DAG.getConstant(0x3, MVT::i32));
1164 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1165 DAG.getConstant(3, MVT::i32));
1167 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1169 EVT MemEltVT = MemVT.getScalarType();
1170 if (ExtType == ISD::SEXTLOAD) {
1171 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1172 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1175 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1178 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1180 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1181 if (Result.getNode()) {
1185 StoreSDNode *Store = cast<StoreSDNode>(Op);
1186 SDValue Chain = Store->getChain();
1187 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1188 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1189 Store->getValue().getValueType().isVector()) {
1190 return SplitVectorStore(Op, DAG);
1193 EVT MemVT = Store->getMemoryVT();
1194 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1195 MemVT.bitsLT(MVT::i32)) {
1197 if (Store->getMemoryVT() == MVT::i8) {
1199 } else if (Store->getMemoryVT() == MVT::i16) {
1202 SDValue BasePtr = Store->getBasePtr();
1203 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1204 DAG.getConstant(2, MVT::i32));
1205 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1206 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1208 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1209 DAG.getConstant(0x3, MVT::i32));
1211 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1212 DAG.getConstant(3, MVT::i32));
1214 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1217 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1219 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1220 MaskedValue, ShiftAmt);
1222 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1224 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1225 DAG.getConstant(0xffffffff, MVT::i32));
1226 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1228 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1229 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1230 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1235 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1237 EVT OVT = Op.getValueType();
1238 SDValue LHS = Op.getOperand(0);
1239 SDValue RHS = Op.getOperand(1);
1242 if (!OVT.isVector()) {
1245 } else if (OVT.getVectorNumElements() == 2) {
1248 } else if (OVT.getVectorNumElements() == 4) {
1252 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1253 // char|short jq = ia ^ ib;
1254 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1256 // jq = jq >> (bitsize - 2)
1257 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1260 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1263 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1265 // int ia = (int)LHS;
1266 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1268 // int ib, (int)RHS;
1269 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1271 // float fa = (float)ia;
1272 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1274 // float fb = (float)ib;
1275 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1277 // float fq = native_divide(fa, fb);
1278 SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
1281 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1283 // float fqneg = -fq;
1284 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1286 // float fr = mad(fqneg, fb, fa);
1287 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1288 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1290 // int iq = (int)fq;
1291 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1294 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1297 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1299 // int cv = fr >= fb;
1301 if (INTTY == MVT::i32) {
1302 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1304 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1306 // jq = (cv ? jq : 0);
1307 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1308 DAG.getConstant(0, OVT));
1310 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1311 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1315 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1317 EVT OVT = Op.getValueType();
1318 SDValue LHS = Op.getOperand(0);
1319 SDValue RHS = Op.getOperand(1);
1320 // The LowerSDIV32 function generates equivalent to the following IL.
1330 // ixor r10, r10, r11
1332 // ixor DST, r0, r10
1341 SDValue r10 = DAG.getSelectCC(DL,
1342 r0, DAG.getConstant(0, OVT),
1343 DAG.getConstant(-1, OVT),
1344 DAG.getConstant(0, OVT),
1348 SDValue r11 = DAG.getSelectCC(DL,
1349 r1, DAG.getConstant(0, OVT),
1350 DAG.getConstant(-1, OVT),
1351 DAG.getConstant(0, OVT),
1355 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1358 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1361 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1364 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1367 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1369 // ixor r10, r10, r11
1370 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1373 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1375 // ixor DST, r0, r10
1376 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1380 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1381 return SDValue(Op.getNode(), 0);
1384 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1385 EVT OVT = Op.getValueType().getScalarType();
1387 if (OVT == MVT::i64)
1388 return LowerSDIV64(Op, DAG);
1390 if (OVT.getScalarType() == MVT::i32)
1391 return LowerSDIV32(Op, DAG);
1393 if (OVT == MVT::i16 || OVT == MVT::i8) {
1394 // FIXME: We should be checking for the masked bits. This isn't reached
1395 // because i8 and i16 are not legal types.
1396 return LowerSDIV24(Op, DAG);
1399 return SDValue(Op.getNode(), 0);
1402 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1404 EVT OVT = Op.getValueType();
1405 SDValue LHS = Op.getOperand(0);
1406 SDValue RHS = Op.getOperand(1);
1407 // The LowerSREM32 function generates equivalent to the following IL.
1417 // umul r20, r20, r1
1420 // ixor DST, r0, r10
1429 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1432 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1435 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1438 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1441 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1444 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1447 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1449 // umul r20, r20, r1
1450 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1453 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1456 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1458 // ixor DST, r0, r10
1459 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1463 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1464 return SDValue(Op.getNode(), 0);
1467 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1468 EVT OVT = Op.getValueType();
1470 if (OVT.getScalarType() == MVT::i64)
1471 return LowerSREM64(Op, DAG);
1473 if (OVT.getScalarType() == MVT::i32)
1474 return LowerSREM32(Op, DAG);
1476 return SDValue(Op.getNode(), 0);
1479 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1480 SelectionDAG &DAG) const {
1482 EVT VT = Op.getValueType();
1484 SDValue Num = Op.getOperand(0);
1485 SDValue Den = Op.getOperand(1);
1487 // RCP = URECIP(Den) = 2^32 / Den + e
1488 // e is rounding error.
1489 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1491 // RCP_LO = umulo(RCP, Den) */
1492 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1494 // RCP_HI = mulhu (RCP, Den) */
1495 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1497 // NEG_RCP_LO = -RCP_LO
1498 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1501 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1502 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1505 // Calculate the rounding error from the URECIP instruction
1506 // E = mulhu(ABS_RCP_LO, RCP)
1507 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1509 // RCP_A_E = RCP + E
1510 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1512 // RCP_S_E = RCP - E
1513 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1515 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1516 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1519 // Quotient = mulhu(Tmp0, Num)
1520 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1522 // Num_S_Remainder = Quotient * Den
1523 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1525 // Remainder = Num - Num_S_Remainder
1526 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1528 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1529 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1530 DAG.getConstant(-1, VT),
1531 DAG.getConstant(0, VT),
1533 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1534 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1536 DAG.getConstant(-1, VT),
1537 DAG.getConstant(0, VT),
1539 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1540 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1543 // Calculate Division result:
1545 // Quotient_A_One = Quotient + 1
1546 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1547 DAG.getConstant(1, VT));
1549 // Quotient_S_One = Quotient - 1
1550 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1551 DAG.getConstant(1, VT));
1553 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1554 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1555 Quotient, Quotient_A_One, ISD::SETEQ);
1557 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1558 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1559 Quotient_S_One, Div, ISD::SETEQ);
1561 // Calculate Rem result:
1563 // Remainder_S_Den = Remainder - Den
1564 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1566 // Remainder_A_Den = Remainder + Den
1567 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1569 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1570 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1571 Remainder, Remainder_S_Den, ISD::SETEQ);
1573 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1574 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1575 Remainder_A_Den, Rem, ISD::SETEQ);
1580 return DAG.getMergeValues(Ops, DL);
1583 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1585 SDValue Src = Op.getOperand(0);
1587 // result = trunc(src)
1588 // if (src > 0.0 && src != result)
1591 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1593 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1594 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1596 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1598 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1599 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1600 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1602 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1603 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1606 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1608 SDValue Src = Op.getOperand(0);
1610 assert(Op.getValueType() == MVT::f64);
1612 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1613 const SDValue One = DAG.getConstant(1, MVT::i32);
1615 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1617 // Extract the upper half, since this is where we will find the sign and
1619 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1621 const unsigned FractBits = 52;
1622 const unsigned ExpBits = 11;
1624 // Extract the exponent.
1625 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1627 DAG.getConstant(FractBits - 32, MVT::i32),
1628 DAG.getConstant(ExpBits, MVT::i32));
1629 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1630 DAG.getConstant(1023, MVT::i32));
1632 // Extract the sign bit.
1633 const SDValue SignBitMask = DAG.getConstant(1ul << 31, MVT::i32);
1634 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1636 // Extend back to to 64-bits.
1637 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1639 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1641 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1642 const SDValue FractMask = DAG.getConstant((1L << FractBits) - 1, MVT::i64);
1644 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1645 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1646 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1648 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1650 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1652 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1653 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1655 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1656 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1658 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1661 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1663 SDValue Src = Op.getOperand(0);
1665 assert(Op.getValueType() == MVT::f64);
1667 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1668 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1669 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1671 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1672 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1674 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1676 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1677 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1679 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1680 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1682 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1685 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1687 SDValue Src = Op.getOperand(0);
1689 // result = trunc(src);
1690 // if (src < 0.0 && src != result)
1693 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1695 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1696 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1698 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1700 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1701 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1702 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1704 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1705 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1708 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1709 SelectionDAG &DAG) const {
1710 SDValue S0 = Op.getOperand(0);
1712 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1715 // f32 uint_to_fp i64
1716 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1717 DAG.getConstant(0, MVT::i32));
1718 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1719 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1720 DAG.getConstant(1, MVT::i32));
1721 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1722 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1723 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1724 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1727 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1729 SelectionDAG &DAG) const {
1730 MVT VT = Op.getSimpleValueType();
1732 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1733 // Shift left by 'Shift' bits.
1734 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1735 // Signed shift Right by 'Shift' bits.
1736 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1739 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1740 SelectionDAG &DAG) const {
1741 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1742 MVT VT = Op.getSimpleValueType();
1743 MVT ScalarVT = VT.getScalarType();
1748 SDValue Src = Op.getOperand(0);
1751 // TODO: Don't scalarize on Evergreen?
1752 unsigned NElts = VT.getVectorNumElements();
1753 SmallVector<SDValue, 8> Args;
1754 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1756 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1757 for (unsigned I = 0; I < NElts; ++I)
1758 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1760 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1763 //===----------------------------------------------------------------------===//
1764 // Custom DAG optimizations
1765 //===----------------------------------------------------------------------===//
1767 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1768 APInt KnownZero, KnownOne;
1769 EVT VT = Op.getValueType();
1770 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1772 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1775 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1776 EVT VT = Op.getValueType();
1778 // In order for this to be a signed 24-bit value, bit 23, must
1780 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1781 // as unsigned 24-bit values.
1782 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1785 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1787 SelectionDAG &DAG = DCI.DAG;
1788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1789 EVT VT = Op.getValueType();
1791 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1792 APInt KnownZero, KnownOne;
1793 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1794 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1795 DCI.CommitTargetLoweringOpt(TLO);
1798 template <typename IntTy>
1799 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1800 uint32_t Offset, uint32_t Width) {
1801 if (Width + Offset < 32) {
1802 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1803 return DAG.getConstant(Result, MVT::i32);
1806 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1809 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1810 DAGCombinerInfo &DCI) const {
1811 SelectionDAG &DAG = DCI.DAG;
1814 switch(N->getOpcode()) {
1817 EVT VT = N->getValueType(0);
1818 SDValue N0 = N->getOperand(0);
1819 SDValue N1 = N->getOperand(1);
1822 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1823 if (VT.isVector() || VT.getSizeInBits() > 32)
1826 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1827 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1828 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1829 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1830 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1831 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1832 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1833 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1838 // We need to use sext even for MUL_U24, because MUL_U24 is used
1839 // for signed multiply of 8 and 16-bit types.
1840 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1844 case AMDGPUISD::MUL_I24:
1845 case AMDGPUISD::MUL_U24: {
1846 SDValue N0 = N->getOperand(0);
1847 SDValue N1 = N->getOperand(1);
1848 simplifyI24(N0, DCI);
1849 simplifyI24(N1, DCI);
1852 case ISD::SELECT_CC: {
1853 return CombineMinMax(N, DAG);
1855 case AMDGPUISD::BFE_I32:
1856 case AMDGPUISD::BFE_U32: {
1857 assert(!N->getValueType(0).isVector() &&
1858 "Vector handling of BFE not implemented");
1859 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1863 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1865 return DAG.getConstant(0, MVT::i32);
1867 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1871 SDValue BitsFrom = N->getOperand(0);
1872 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1874 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1876 if (OffsetVal == 0) {
1877 // This is already sign / zero extended, so try to fold away extra BFEs.
1878 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1880 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1881 if (OpSignBits >= SignBits)
1884 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1886 // This is a sign_extend_inreg. Replace it to take advantage of existing
1887 // DAG Combines. If not eliminated, we will match back to BFE during
1890 // TODO: The sext_inreg of extended types ends, although we can could
1891 // handle them in a single BFE.
1892 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1893 DAG.getValueType(SmallVT));
1896 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1899 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1901 return constantFoldBFE<int32_t>(DAG,
1902 Val->getSExtValue(),
1907 return constantFoldBFE<uint32_t>(DAG,
1908 Val->getZExtValue(),
1913 APInt Demanded = APInt::getBitsSet(32,
1915 OffsetVal + WidthVal);
1917 if ((OffsetVal + WidthVal) >= 32) {
1918 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1919 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1920 BitsFrom, ShiftVal);
1923 APInt KnownZero, KnownOne;
1924 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1925 !DCI.isBeforeLegalizeOps());
1926 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1927 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1928 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1929 DCI.CommitTargetLoweringOpt(TLO);
1938 //===----------------------------------------------------------------------===//
1940 //===----------------------------------------------------------------------===//
1942 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1945 const SmallVectorImpl<ISD::InputArg> &Ins,
1946 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1948 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1949 if (Ins[i].ArgVT == Ins[i].VT) {
1950 OrigIns.push_back(Ins[i]);
1955 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1956 // Vector has been split into scalars.
1957 VT = Ins[i].ArgVT.getVectorElementType();
1958 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1959 Ins[i].ArgVT.getVectorElementType() !=
1960 Ins[i].VT.getVectorElementType()) {
1961 // Vector elements have been promoted
1964 // Vector has been spilt into smaller vectors.
1968 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1969 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1970 OrigIns.push_back(Arg);
1974 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1975 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1976 return CFP->isExactlyValue(1.0);
1978 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1979 return C->isAllOnesValue();
1984 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1985 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1986 return CFP->getValueAPF().isZero();
1988 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1989 return C->isNullValue();
1994 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1995 const TargetRegisterClass *RC,
1996 unsigned Reg, EVT VT) const {
1997 MachineFunction &MF = DAG.getMachineFunction();
1998 MachineRegisterInfo &MRI = MF.getRegInfo();
1999 unsigned VirtualRegister;
2000 if (!MRI.isLiveIn(Reg)) {
2001 VirtualRegister = MRI.createVirtualRegister(RC);
2002 MRI.addLiveIn(Reg, VirtualRegister);
2004 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2006 return DAG.getRegister(VirtualRegister, VT);
2009 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2011 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2013 default: return nullptr;
2015 NODE_NAME_CASE(CALL);
2016 NODE_NAME_CASE(UMUL);
2017 NODE_NAME_CASE(DIV_INF);
2018 NODE_NAME_CASE(RET_FLAG);
2019 NODE_NAME_CASE(BRANCH_COND);
2022 NODE_NAME_CASE(DWORDADDR)
2023 NODE_NAME_CASE(FRACT)
2024 NODE_NAME_CASE(CLAMP)
2025 NODE_NAME_CASE(FMAX)
2026 NODE_NAME_CASE(SMAX)
2027 NODE_NAME_CASE(UMAX)
2028 NODE_NAME_CASE(FMIN)
2029 NODE_NAME_CASE(SMIN)
2030 NODE_NAME_CASE(UMIN)
2031 NODE_NAME_CASE(BFE_U32)
2032 NODE_NAME_CASE(BFE_I32)
2035 NODE_NAME_CASE(BREV)
2036 NODE_NAME_CASE(MUL_U24)
2037 NODE_NAME_CASE(MUL_I24)
2038 NODE_NAME_CASE(MAD_U24)
2039 NODE_NAME_CASE(MAD_I24)
2040 NODE_NAME_CASE(URECIP)
2041 NODE_NAME_CASE(DOT4)
2042 NODE_NAME_CASE(EXPORT)
2043 NODE_NAME_CASE(CONST_ADDRESS)
2044 NODE_NAME_CASE(REGISTER_LOAD)
2045 NODE_NAME_CASE(REGISTER_STORE)
2046 NODE_NAME_CASE(LOAD_CONSTANT)
2047 NODE_NAME_CASE(LOAD_INPUT)
2048 NODE_NAME_CASE(SAMPLE)
2049 NODE_NAME_CASE(SAMPLEB)
2050 NODE_NAME_CASE(SAMPLED)
2051 NODE_NAME_CASE(SAMPLEL)
2052 NODE_NAME_CASE(CVT_F32_UBYTE0)
2053 NODE_NAME_CASE(CVT_F32_UBYTE1)
2054 NODE_NAME_CASE(CVT_F32_UBYTE2)
2055 NODE_NAME_CASE(CVT_F32_UBYTE3)
2056 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2057 NODE_NAME_CASE(STORE_MSKOR)
2058 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2062 static void computeKnownBitsForMinMax(const SDValue Op0,
2066 const SelectionDAG &DAG,
2068 APInt Op0Zero, Op0One;
2069 APInt Op1Zero, Op1One;
2070 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2071 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2073 KnownZero = Op0Zero & Op1Zero;
2074 KnownOne = Op0One & Op1One;
2077 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2081 const SelectionDAG &DAG,
2082 unsigned Depth) const {
2084 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2088 unsigned Opc = Op.getOpcode();
2093 case ISD::INTRINSIC_WO_CHAIN: {
2094 // FIXME: The intrinsic should just use the node.
2095 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2096 case AMDGPUIntrinsic::AMDGPU_imax:
2097 case AMDGPUIntrinsic::AMDGPU_umax:
2098 case AMDGPUIntrinsic::AMDGPU_imin:
2099 case AMDGPUIntrinsic::AMDGPU_umin:
2100 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2101 KnownZero, KnownOne, DAG, Depth);
2109 case AMDGPUISD::SMAX:
2110 case AMDGPUISD::UMAX:
2111 case AMDGPUISD::SMIN:
2112 case AMDGPUISD::UMIN:
2113 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2114 KnownZero, KnownOne, DAG, Depth);
2117 case AMDGPUISD::BFE_I32:
2118 case AMDGPUISD::BFE_U32: {
2119 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2123 unsigned BitWidth = 32;
2124 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2126 KnownZero = APInt::getAllOnesValue(BitWidth);
2127 KnownOne = APInt::getNullValue(BitWidth);
2131 // FIXME: This could do a lot more. If offset is 0, should be the same as
2132 // sign_extend_inreg implementation, but that involves duplicating it.
2133 if (Opc == AMDGPUISD::BFE_I32)
2134 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2136 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2143 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2145 const SelectionDAG &DAG,
2146 unsigned Depth) const {
2147 switch (Op.getOpcode()) {
2148 case AMDGPUISD::BFE_I32: {
2149 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2153 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2154 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2155 if (!Offset || !Offset->isNullValue())
2158 // TODO: Could probably figure something out with non-0 offsets.
2159 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2160 return std::max(SignBits, Op0SignBits);
2163 case AMDGPUISD::BFE_U32: {
2164 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2165 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;