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::FCEIL, VT, Expand);
321 setOperationAction(ISD::FCOS, VT, Expand);
322 setOperationAction(ISD::FDIV, VT, Expand);
323 setOperationAction(ISD::FEXP2, VT, Expand);
324 setOperationAction(ISD::FLOG2, VT, Expand);
325 setOperationAction(ISD::FPOW, VT, Expand);
326 setOperationAction(ISD::FFLOOR, VT, Expand);
327 setOperationAction(ISD::FTRUNC, VT, Expand);
328 setOperationAction(ISD::FMUL, VT, Expand);
329 setOperationAction(ISD::FRINT, VT, Expand);
330 setOperationAction(ISD::FNEARBYINT, VT, Expand);
331 setOperationAction(ISD::FSQRT, VT, Expand);
332 setOperationAction(ISD::FSIN, VT, Expand);
333 setOperationAction(ISD::FSUB, VT, Expand);
334 setOperationAction(ISD::FNEG, VT, Expand);
335 setOperationAction(ISD::SELECT, VT, Expand);
336 setOperationAction(ISD::VSELECT, VT, Expand);
337 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
340 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
341 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
343 setTargetDAGCombine(ISD::MUL);
344 setTargetDAGCombine(ISD::SELECT_CC);
346 setSchedulingPreference(Sched::RegPressure);
347 setJumpIsExpensive(true);
349 // There are no integer divide instructions, and these expand to a pretty
350 // large sequence of instructions.
351 setIntDivIsCheap(false);
353 // TODO: Investigate this when 64-bit divides are implemented.
354 addBypassSlowDiv(64, 32);
356 // FIXME: Need to really handle these.
357 MaxStoresPerMemcpy = 4096;
358 MaxStoresPerMemmove = 4096;
359 MaxStoresPerMemset = 4096;
362 //===----------------------------------------------------------------------===//
363 // Target Information
364 //===----------------------------------------------------------------------===//
366 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
370 // The backend supports 32 and 64 bit floating point immediates.
371 // FIXME: Why are we reporting vectors of FP immediates as legal?
372 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
373 EVT ScalarVT = VT.getScalarType();
374 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
377 // We don't want to shrink f64 / f32 constants.
378 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
379 EVT ScalarVT = VT.getScalarType();
380 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
383 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
385 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
388 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
389 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
391 return ((LScalarSize <= CastScalarSize) ||
392 (CastScalarSize >= 32) ||
396 //===---------------------------------------------------------------------===//
398 //===---------------------------------------------------------------------===//
400 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
401 assert(VT.isFloatingPoint());
402 return VT == MVT::f32;
405 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
406 assert(VT.isFloatingPoint());
407 return VT == MVT::f32;
410 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
411 // Truncate is just accessing a subregister.
412 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
415 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
416 // Truncate is just accessing a subregister.
417 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
418 (Dest->getPrimitiveSizeInBits() % 32 == 0);
421 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
422 const DataLayout *DL = getDataLayout();
423 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
424 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
426 return SrcSize == 32 && DestSize == 64;
429 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
430 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
431 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
432 // this will enable reducing 64-bit operations the 32-bit, which is always
434 return Src == MVT::i32 && Dest == MVT::i64;
437 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
438 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
439 // limited number of native 64-bit operations. Shrinking an operation to fit
440 // in a single 32-bit register should always be helpful. As currently used,
441 // this is much less general than the name suggests, and is only used in
442 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
443 // not profitable, and may actually be harmful.
444 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
447 //===---------------------------------------------------------------------===//
448 // TargetLowering Callbacks
449 //===---------------------------------------------------------------------===//
451 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
452 const SmallVectorImpl<ISD::InputArg> &Ins) const {
454 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
457 SDValue AMDGPUTargetLowering::LowerReturn(
459 CallingConv::ID CallConv,
461 const SmallVectorImpl<ISD::OutputArg> &Outs,
462 const SmallVectorImpl<SDValue> &OutVals,
463 SDLoc DL, SelectionDAG &DAG) const {
464 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
467 //===---------------------------------------------------------------------===//
468 // Target specific lowering
469 //===---------------------------------------------------------------------===//
471 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
472 SmallVectorImpl<SDValue> &InVals) const {
473 SDValue Callee = CLI.Callee;
474 SelectionDAG &DAG = CLI.DAG;
476 const Function &Fn = *DAG.getMachineFunction().getFunction();
478 StringRef FuncName("<unknown>");
480 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
481 FuncName = G->getSymbol();
482 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
483 FuncName = G->getGlobal()->getName();
485 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
486 DAG.getContext()->diagnose(NoCalls);
490 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
491 SelectionDAG &DAG) const {
492 switch (Op.getOpcode()) {
494 Op.getNode()->dump();
495 llvm_unreachable("Custom lowering code for this"
496 "instruction is not implemented yet!");
498 // AMDGPU DAG lowering.
499 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
500 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
501 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
502 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
503 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
504 case ISD::SDIV: return LowerSDIV(Op, DAG);
505 case ISD::SREM: return LowerSREM(Op, DAG);
506 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
507 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
508 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
509 case ISD::FRINT: return LowerFRINT(Op, DAG);
510 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
511 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
512 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
514 // AMDIL DAG lowering.
515 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
520 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
521 SmallVectorImpl<SDValue> &Results,
522 SelectionDAG &DAG) const {
523 switch (N->getOpcode()) {
524 case ISD::SIGN_EXTEND_INREG:
525 // Different parts of legalization seem to interpret which type of
526 // sign_extend_inreg is the one to check for custom lowering. The extended
527 // from type is what really matters, but some places check for custom
528 // lowering of the result type. This results in trying to use
529 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
530 // nothing here and let the illegal result integer be handled normally.
533 SDValue Op = SDValue(N, 0);
535 EVT VT = Op.getValueType();
536 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
537 N->getOperand(0), N->getOperand(1));
538 Results.push_back(UDIVREM);
542 SDValue Op = SDValue(N, 0);
544 EVT VT = Op.getValueType();
545 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
546 N->getOperand(0), N->getOperand(1));
547 Results.push_back(UDIVREM.getValue(1));
551 SDValue Op = SDValue(N, 0);
553 EVT VT = Op.getValueType();
554 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
556 SDValue one = DAG.getConstant(1, HalfVT);
557 SDValue zero = DAG.getConstant(0, HalfVT);
560 SDValue LHS = N->getOperand(0);
561 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
562 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
564 SDValue RHS = N->getOperand(1);
565 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
566 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
568 // Get Speculative values
569 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
570 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
572 SDValue REM_Hi = zero;
573 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
575 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
576 SDValue DIV_Lo = zero;
578 const unsigned halfBitWidth = HalfVT.getSizeInBits();
580 for (unsigned i = 0; i < halfBitWidth; ++i) {
581 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
582 // Get Value of high bit
584 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
585 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
587 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
588 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
591 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
592 DAG.getConstant(halfBitWidth - 1, HalfVT));
593 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
594 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
596 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
597 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
600 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
602 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
603 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
605 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
609 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
611 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
612 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
613 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
616 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
617 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
618 Results.push_back(DIV);
619 Results.push_back(REM);
627 // FIXME: This implements accesses to initialized globals in the constant
628 // address space by copying them to private and accessing that. It does not
629 // properly handle illegal types or vectors. The private vector loads are not
630 // scalarized, and the illegal scalars hit an assertion. This technique will not
631 // work well with large initializers, and this should eventually be
632 // removed. Initialized globals should be placed into a data section that the
633 // runtime will load into a buffer before the kernel is executed. Uses of the
634 // global need to be replaced with a pointer loaded from an implicit kernel
635 // argument into this buffer holding the copy of the data, which will remove the
636 // need for any of this.
637 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
638 const GlobalValue *GV,
639 const SDValue &InitPtr,
641 SelectionDAG &DAG) const {
642 const DataLayout *TD = getTargetMachine().getDataLayout();
644 Type *InitTy = Init->getType();
646 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
647 EVT VT = EVT::getEVT(InitTy);
648 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
649 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
650 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
651 TD->getPrefTypeAlignment(InitTy));
654 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
655 EVT VT = EVT::getEVT(CFP->getType());
656 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
657 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
658 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
659 TD->getPrefTypeAlignment(CFP->getType()));
662 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
663 const StructLayout *SL = TD->getStructLayout(ST);
665 EVT PtrVT = InitPtr.getValueType();
666 SmallVector<SDValue, 8> Chains;
668 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
669 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
670 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
672 Constant *Elt = Init->getAggregateElement(I);
673 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
676 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
679 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
680 EVT PtrVT = InitPtr.getValueType();
682 unsigned NumElements;
683 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
684 NumElements = AT->getNumElements();
685 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
686 NumElements = VT->getNumElements();
688 llvm_unreachable("Unexpected type");
690 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
691 SmallVector<SDValue, 8> Chains;
692 for (unsigned i = 0; i < NumElements; ++i) {
693 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
694 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
696 Constant *Elt = Init->getAggregateElement(i);
697 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
700 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
703 if (isa<UndefValue>(Init)) {
704 EVT VT = EVT::getEVT(InitTy);
705 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
706 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
707 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
708 TD->getPrefTypeAlignment(InitTy));
712 llvm_unreachable("Unhandled constant initializer");
715 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
717 SelectionDAG &DAG) const {
719 const DataLayout *TD = getTargetMachine().getDataLayout();
720 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
721 const GlobalValue *GV = G->getGlobal();
723 switch (G->getAddressSpace()) {
724 default: llvm_unreachable("Global Address lowering not implemented for this "
726 case AMDGPUAS::LOCAL_ADDRESS: {
727 // XXX: What does the value of G->getOffset() mean?
728 assert(G->getOffset() == 0 &&
729 "Do not know what to do with an non-zero offset");
732 if (MFI->LocalMemoryObjects.count(GV) == 0) {
733 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
734 Offset = MFI->LDSSize;
735 MFI->LocalMemoryObjects[GV] = Offset;
736 // XXX: Account for alignment?
737 MFI->LDSSize += Size;
739 Offset = MFI->LocalMemoryObjects[GV];
742 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
744 case AMDGPUAS::CONSTANT_ADDRESS: {
745 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
746 Type *EltType = GV->getType()->getElementType();
747 unsigned Size = TD->getTypeAllocSize(EltType);
748 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
750 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
751 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
753 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
754 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
756 const GlobalVariable *Var = cast<GlobalVariable>(GV);
757 if (!Var->hasInitializer()) {
758 // This has no use, but bugpoint will hit it.
759 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
762 const Constant *Init = Var->getInitializer();
763 SmallVector<SDNode*, 8> WorkList;
765 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
766 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
767 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
769 WorkList.push_back(*I);
771 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
772 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
773 E = WorkList.end(); I != E; ++I) {
774 SmallVector<SDValue, 8> Ops;
775 Ops.push_back(Chain);
776 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
777 Ops.push_back((*I)->getOperand(i));
779 DAG.UpdateNodeOperands(*I, Ops);
781 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
786 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
787 SelectionDAG &DAG) const {
788 SmallVector<SDValue, 8> Args;
789 SDValue A = Op.getOperand(0);
790 SDValue B = Op.getOperand(1);
792 DAG.ExtractVectorElements(A, Args);
793 DAG.ExtractVectorElements(B, Args);
795 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
798 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
799 SelectionDAG &DAG) const {
801 SmallVector<SDValue, 8> Args;
802 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
803 EVT VT = Op.getValueType();
804 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
805 VT.getVectorNumElements());
807 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
810 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
811 SelectionDAG &DAG) const {
813 MachineFunction &MF = DAG.getMachineFunction();
814 const AMDGPUFrameLowering *TFL =
815 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
817 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
819 unsigned FrameIndex = FIN->getIndex();
820 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
821 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
825 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
826 SelectionDAG &DAG) const {
827 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
829 EVT VT = Op.getValueType();
831 switch (IntrinsicID) {
833 case AMDGPUIntrinsic::AMDGPU_abs:
834 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
835 return LowerIntrinsicIABS(Op, DAG);
836 case AMDGPUIntrinsic::AMDGPU_lrp:
837 return LowerIntrinsicLRP(Op, DAG);
838 case AMDGPUIntrinsic::AMDGPU_fract:
839 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
840 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
842 case AMDGPUIntrinsic::AMDGPU_clamp:
843 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
844 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
845 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
847 case Intrinsic::AMDGPU_div_scale:
848 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
849 Op.getOperand(1), Op.getOperand(2));
851 case Intrinsic::AMDGPU_div_fmas:
852 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
853 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
855 case Intrinsic::AMDGPU_div_fixup:
856 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
857 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
859 case Intrinsic::AMDGPU_trig_preop:
860 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
861 Op.getOperand(1), Op.getOperand(2));
863 case Intrinsic::AMDGPU_rcp:
864 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
866 case Intrinsic::AMDGPU_rsq:
867 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
869 case AMDGPUIntrinsic::AMDGPU_imax:
870 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
872 case AMDGPUIntrinsic::AMDGPU_umax:
873 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
875 case AMDGPUIntrinsic::AMDGPU_imin:
876 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
878 case AMDGPUIntrinsic::AMDGPU_umin:
879 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
882 case AMDGPUIntrinsic::AMDGPU_umul24:
883 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
884 Op.getOperand(1), Op.getOperand(2));
886 case AMDGPUIntrinsic::AMDGPU_imul24:
887 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
888 Op.getOperand(1), Op.getOperand(2));
890 case AMDGPUIntrinsic::AMDGPU_umad24:
891 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
892 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
894 case AMDGPUIntrinsic::AMDGPU_imad24:
895 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
896 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
898 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
899 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
901 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
902 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
904 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
905 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
907 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
908 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
910 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
911 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
916 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
917 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
922 case AMDGPUIntrinsic::AMDGPU_bfi:
923 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
928 case AMDGPUIntrinsic::AMDGPU_bfm:
929 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
933 case AMDGPUIntrinsic::AMDGPU_brev:
934 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
936 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
937 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
939 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
940 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
944 ///IABS(a) = SMAX(sub(0, a), a)
945 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
946 SelectionDAG &DAG) const {
948 EVT VT = Op.getValueType();
949 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
952 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
955 /// Linear Interpolation
956 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
957 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
958 SelectionDAG &DAG) const {
960 EVT VT = Op.getValueType();
961 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
962 DAG.getConstantFP(1.0f, MVT::f32),
964 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
966 return DAG.getNode(ISD::FADD, DL, VT,
967 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
971 /// \brief Generate Min/Max node
972 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
973 SelectionDAG &DAG) const {
975 EVT VT = N->getValueType(0);
977 SDValue LHS = N->getOperand(0);
978 SDValue RHS = N->getOperand(1);
979 SDValue True = N->getOperand(2);
980 SDValue False = N->getOperand(3);
981 SDValue CC = N->getOperand(4);
983 if (VT != MVT::f32 ||
984 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
988 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1002 llvm_unreachable("Operation should already be optimised!");
1009 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
1010 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1018 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
1019 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1021 case ISD::SETCC_INVALID:
1022 llvm_unreachable("Invalid setcc condcode!");
1027 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
1028 SelectionDAG &DAG) const {
1029 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
1030 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
1031 EVT EltVT = Op.getValueType().getVectorElementType();
1032 EVT PtrVT = Load->getBasePtr().getValueType();
1033 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1034 SmallVector<SDValue, 8> Loads;
1037 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1038 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1039 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
1040 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1041 Load->getChain(), Ptr,
1042 MachinePointerInfo(Load->getMemOperand()->getValue()),
1043 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1044 Load->getAlignment()));
1046 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
1049 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1050 SelectionDAG &DAG) const {
1051 StoreSDNode *Store = cast<StoreSDNode>(Op);
1052 EVT MemVT = Store->getMemoryVT();
1053 unsigned MemBits = MemVT.getSizeInBits();
1055 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1056 // truncating store into an i32 store.
1057 // XXX: We could also handle optimize other vector bitwidths.
1058 if (!MemVT.isVector() || MemBits > 32) {
1063 SDValue Value = Store->getValue();
1064 EVT VT = Value.getValueType();
1065 EVT ElemVT = VT.getVectorElementType();
1066 SDValue Ptr = Store->getBasePtr();
1067 EVT MemEltVT = MemVT.getVectorElementType();
1068 unsigned MemEltBits = MemEltVT.getSizeInBits();
1069 unsigned MemNumElements = MemVT.getVectorNumElements();
1070 unsigned PackedSize = MemVT.getStoreSizeInBits();
1071 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1073 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1075 SDValue PackedValue;
1076 for (unsigned i = 0; i < MemNumElements; ++i) {
1077 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1078 DAG.getConstant(i, MVT::i32));
1079 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1080 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1082 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1083 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1088 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1092 if (PackedSize < 32) {
1093 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1094 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1095 Store->getMemOperand()->getPointerInfo(),
1097 Store->isNonTemporal(), Store->isVolatile(),
1098 Store->getAlignment());
1101 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1102 Store->getMemOperand()->getPointerInfo(),
1103 Store->isVolatile(), Store->isNonTemporal(),
1104 Store->getAlignment());
1107 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1108 SelectionDAG &DAG) const {
1109 StoreSDNode *Store = cast<StoreSDNode>(Op);
1110 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1111 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1112 EVT PtrVT = Store->getBasePtr().getValueType();
1113 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1116 SmallVector<SDValue, 8> Chains;
1118 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1119 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1120 Store->getValue(), DAG.getConstant(i, MVT::i32));
1121 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1122 Store->getBasePtr(),
1123 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1125 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1126 MachinePointerInfo(Store->getMemOperand()->getValue()),
1127 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1128 Store->getAlignment()));
1130 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1133 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1135 LoadSDNode *Load = cast<LoadSDNode>(Op);
1136 ISD::LoadExtType ExtType = Load->getExtensionType();
1137 EVT VT = Op.getValueType();
1138 EVT MemVT = Load->getMemoryVT();
1140 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1141 // We can do the extload to 32-bits, and then need to separately extend to
1144 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1148 Load->getMemOperand());
1149 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1152 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1153 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1154 // FIXME: Copied from PPC
1155 // First, load into 32 bits, then truncate to 1 bit.
1157 SDValue Chain = Load->getChain();
1158 SDValue BasePtr = Load->getBasePtr();
1159 MachineMemOperand *MMO = Load->getMemOperand();
1161 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1162 BasePtr, MVT::i8, MMO);
1163 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1166 // Lower loads constant address space global variable loads
1167 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1168 isa<GlobalVariable>(
1169 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1171 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1172 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1173 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1174 DAG.getConstant(2, MVT::i32));
1175 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1176 Load->getChain(), Ptr,
1177 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1180 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1181 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1185 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1186 DAG.getConstant(2, MVT::i32));
1187 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1188 Load->getChain(), Ptr,
1189 DAG.getTargetConstant(0, MVT::i32),
1191 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1193 DAG.getConstant(0x3, MVT::i32));
1194 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1195 DAG.getConstant(3, MVT::i32));
1197 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1199 EVT MemEltVT = MemVT.getScalarType();
1200 if (ExtType == ISD::SEXTLOAD) {
1201 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1202 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1205 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1208 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1210 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1211 if (Result.getNode()) {
1215 StoreSDNode *Store = cast<StoreSDNode>(Op);
1216 SDValue Chain = Store->getChain();
1217 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1218 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1219 Store->getValue().getValueType().isVector()) {
1220 return SplitVectorStore(Op, DAG);
1223 EVT MemVT = Store->getMemoryVT();
1224 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1225 MemVT.bitsLT(MVT::i32)) {
1227 if (Store->getMemoryVT() == MVT::i8) {
1229 } else if (Store->getMemoryVT() == MVT::i16) {
1232 SDValue BasePtr = Store->getBasePtr();
1233 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1234 DAG.getConstant(2, MVT::i32));
1235 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1236 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1238 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1239 DAG.getConstant(0x3, MVT::i32));
1241 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1242 DAG.getConstant(3, MVT::i32));
1244 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1247 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1249 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1250 MaskedValue, ShiftAmt);
1252 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1254 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1255 DAG.getConstant(0xffffffff, MVT::i32));
1256 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1258 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1259 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1260 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1265 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1267 EVT OVT = Op.getValueType();
1268 SDValue LHS = Op.getOperand(0);
1269 SDValue RHS = Op.getOperand(1);
1272 if (!OVT.isVector()) {
1275 } else if (OVT.getVectorNumElements() == 2) {
1278 } else if (OVT.getVectorNumElements() == 4) {
1282 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1283 // char|short jq = ia ^ ib;
1284 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1286 // jq = jq >> (bitsize - 2)
1287 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1290 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1293 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1295 // int ia = (int)LHS;
1296 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1298 // int ib, (int)RHS;
1299 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1301 // float fa = (float)ia;
1302 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1304 // float fb = (float)ib;
1305 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1307 // float fq = native_divide(fa, fb);
1308 SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
1311 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1313 // float fqneg = -fq;
1314 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1316 // float fr = mad(fqneg, fb, fa);
1317 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1318 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1320 // int iq = (int)fq;
1321 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1324 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1327 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1329 // int cv = fr >= fb;
1331 if (INTTY == MVT::i32) {
1332 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1334 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1336 // jq = (cv ? jq : 0);
1337 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1338 DAG.getConstant(0, OVT));
1340 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1341 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1345 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1347 EVT OVT = Op.getValueType();
1348 SDValue LHS = Op.getOperand(0);
1349 SDValue RHS = Op.getOperand(1);
1350 // The LowerSDIV32 function generates equivalent to the following IL.
1360 // ixor r10, r10, r11
1362 // ixor DST, r0, r10
1371 SDValue r10 = DAG.getSelectCC(DL,
1372 r0, DAG.getConstant(0, OVT),
1373 DAG.getConstant(-1, OVT),
1374 DAG.getConstant(0, OVT),
1378 SDValue r11 = DAG.getSelectCC(DL,
1379 r1, DAG.getConstant(0, OVT),
1380 DAG.getConstant(-1, OVT),
1381 DAG.getConstant(0, OVT),
1385 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1388 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1391 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1394 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1397 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1399 // ixor r10, r10, r11
1400 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1403 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1405 // ixor DST, r0, r10
1406 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1410 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1411 return SDValue(Op.getNode(), 0);
1414 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1415 EVT OVT = Op.getValueType().getScalarType();
1417 if (OVT == MVT::i64)
1418 return LowerSDIV64(Op, DAG);
1420 if (OVT.getScalarType() == MVT::i32)
1421 return LowerSDIV32(Op, DAG);
1423 if (OVT == MVT::i16 || OVT == MVT::i8) {
1424 // FIXME: We should be checking for the masked bits. This isn't reached
1425 // because i8 and i16 are not legal types.
1426 return LowerSDIV24(Op, DAG);
1429 return SDValue(Op.getNode(), 0);
1432 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1434 EVT OVT = Op.getValueType();
1435 SDValue LHS = Op.getOperand(0);
1436 SDValue RHS = Op.getOperand(1);
1437 // The LowerSREM32 function generates equivalent to the following IL.
1447 // umul r20, r20, r1
1450 // ixor DST, r0, r10
1459 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1462 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1465 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1468 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1471 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1474 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1477 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1479 // umul r20, r20, r1
1480 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1483 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1486 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1488 // ixor DST, r0, r10
1489 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1493 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1494 return SDValue(Op.getNode(), 0);
1497 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1498 EVT OVT = Op.getValueType();
1500 if (OVT.getScalarType() == MVT::i64)
1501 return LowerSREM64(Op, DAG);
1503 if (OVT.getScalarType() == MVT::i32)
1504 return LowerSREM32(Op, DAG);
1506 return SDValue(Op.getNode(), 0);
1509 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1510 SelectionDAG &DAG) const {
1512 EVT VT = Op.getValueType();
1514 SDValue Num = Op.getOperand(0);
1515 SDValue Den = Op.getOperand(1);
1517 // RCP = URECIP(Den) = 2^32 / Den + e
1518 // e is rounding error.
1519 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1521 // RCP_LO = umulo(RCP, Den) */
1522 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1524 // RCP_HI = mulhu (RCP, Den) */
1525 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1527 // NEG_RCP_LO = -RCP_LO
1528 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1531 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1532 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1535 // Calculate the rounding error from the URECIP instruction
1536 // E = mulhu(ABS_RCP_LO, RCP)
1537 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1539 // RCP_A_E = RCP + E
1540 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1542 // RCP_S_E = RCP - E
1543 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1545 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1546 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1549 // Quotient = mulhu(Tmp0, Num)
1550 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1552 // Num_S_Remainder = Quotient * Den
1553 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1555 // Remainder = Num - Num_S_Remainder
1556 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1558 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1559 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1560 DAG.getConstant(-1, VT),
1561 DAG.getConstant(0, VT),
1563 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1564 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1566 DAG.getConstant(-1, VT),
1567 DAG.getConstant(0, VT),
1569 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1570 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1573 // Calculate Division result:
1575 // Quotient_A_One = Quotient + 1
1576 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1577 DAG.getConstant(1, VT));
1579 // Quotient_S_One = Quotient - 1
1580 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1581 DAG.getConstant(1, VT));
1583 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1584 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1585 Quotient, Quotient_A_One, ISD::SETEQ);
1587 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1588 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1589 Quotient_S_One, Div, ISD::SETEQ);
1591 // Calculate Rem result:
1593 // Remainder_S_Den = Remainder - Den
1594 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1596 // Remainder_A_Den = Remainder + Den
1597 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1599 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1600 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1601 Remainder, Remainder_S_Den, ISD::SETEQ);
1603 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1604 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1605 Remainder_A_Den, Rem, ISD::SETEQ);
1610 return DAG.getMergeValues(Ops, DL);
1613 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1615 SDValue Src = Op.getOperand(0);
1617 // result = trunc(src)
1618 // if (src > 0.0 && src != result)
1621 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1623 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1624 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1626 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1628 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1629 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1630 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1632 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1633 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1636 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1638 SDValue Src = Op.getOperand(0);
1640 assert(Op.getValueType() == MVT::f64);
1642 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1643 const SDValue One = DAG.getConstant(1, MVT::i32);
1645 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1647 // Extract the upper half, since this is where we will find the sign and
1649 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1651 const unsigned FractBits = 52;
1652 const unsigned ExpBits = 11;
1654 // Extract the exponent.
1655 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1657 DAG.getConstant(FractBits - 32, MVT::i32),
1658 DAG.getConstant(ExpBits, MVT::i32));
1659 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1660 DAG.getConstant(1023, MVT::i32));
1662 // Extract the sign bit.
1663 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1664 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1666 // Extend back to to 64-bits.
1667 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1669 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1671 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1672 const SDValue FractMask
1673 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1675 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1676 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1677 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1679 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1681 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1683 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1684 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1686 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1687 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1689 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1692 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1694 SDValue Src = Op.getOperand(0);
1696 assert(Op.getValueType() == MVT::f64);
1698 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1699 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1700 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1702 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1703 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1705 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1707 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1708 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1710 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1711 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1713 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1716 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1717 // FNEARBYINT and FRINT are the same, except in their handling of FP
1718 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1719 // rint, so just treat them as equivalent.
1720 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1723 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1725 SDValue Src = Op.getOperand(0);
1727 // result = trunc(src);
1728 // if (src < 0.0 && src != result)
1731 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1733 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1734 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1736 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1738 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1739 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1740 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1742 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1743 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1746 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1747 SelectionDAG &DAG) const {
1748 SDValue S0 = Op.getOperand(0);
1750 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1753 // f32 uint_to_fp i64
1754 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1755 DAG.getConstant(0, MVT::i32));
1756 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1757 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1758 DAG.getConstant(1, MVT::i32));
1759 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1760 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1761 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1762 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1765 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1767 SelectionDAG &DAG) const {
1768 MVT VT = Op.getSimpleValueType();
1770 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1771 // Shift left by 'Shift' bits.
1772 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1773 // Signed shift Right by 'Shift' bits.
1774 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1777 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1778 SelectionDAG &DAG) const {
1779 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1780 MVT VT = Op.getSimpleValueType();
1781 MVT ScalarVT = VT.getScalarType();
1786 SDValue Src = Op.getOperand(0);
1789 // TODO: Don't scalarize on Evergreen?
1790 unsigned NElts = VT.getVectorNumElements();
1791 SmallVector<SDValue, 8> Args;
1792 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1794 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1795 for (unsigned I = 0; I < NElts; ++I)
1796 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1798 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1801 //===----------------------------------------------------------------------===//
1802 // Custom DAG optimizations
1803 //===----------------------------------------------------------------------===//
1805 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1806 APInt KnownZero, KnownOne;
1807 EVT VT = Op.getValueType();
1808 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1810 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1813 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1814 EVT VT = Op.getValueType();
1816 // In order for this to be a signed 24-bit value, bit 23, must
1818 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1819 // as unsigned 24-bit values.
1820 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1823 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1825 SelectionDAG &DAG = DCI.DAG;
1826 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1827 EVT VT = Op.getValueType();
1829 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1830 APInt KnownZero, KnownOne;
1831 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1832 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1833 DCI.CommitTargetLoweringOpt(TLO);
1836 template <typename IntTy>
1837 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1838 uint32_t Offset, uint32_t Width) {
1839 if (Width + Offset < 32) {
1840 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1841 return DAG.getConstant(Result, MVT::i32);
1844 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1847 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1848 DAGCombinerInfo &DCI) const {
1849 SelectionDAG &DAG = DCI.DAG;
1852 switch(N->getOpcode()) {
1855 EVT VT = N->getValueType(0);
1856 SDValue N0 = N->getOperand(0);
1857 SDValue N1 = N->getOperand(1);
1860 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1861 if (VT.isVector() || VT.getSizeInBits() > 32)
1864 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1865 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1866 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1867 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1868 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1869 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1870 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1871 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1876 // We need to use sext even for MUL_U24, because MUL_U24 is used
1877 // for signed multiply of 8 and 16-bit types.
1878 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1882 case AMDGPUISD::MUL_I24:
1883 case AMDGPUISD::MUL_U24: {
1884 SDValue N0 = N->getOperand(0);
1885 SDValue N1 = N->getOperand(1);
1886 simplifyI24(N0, DCI);
1887 simplifyI24(N1, DCI);
1890 case ISD::SELECT_CC: {
1891 return CombineMinMax(N, DAG);
1893 case AMDGPUISD::BFE_I32:
1894 case AMDGPUISD::BFE_U32: {
1895 assert(!N->getValueType(0).isVector() &&
1896 "Vector handling of BFE not implemented");
1897 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1901 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1903 return DAG.getConstant(0, MVT::i32);
1905 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1909 SDValue BitsFrom = N->getOperand(0);
1910 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1912 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1914 if (OffsetVal == 0) {
1915 // This is already sign / zero extended, so try to fold away extra BFEs.
1916 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1918 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1919 if (OpSignBits >= SignBits)
1922 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1924 // This is a sign_extend_inreg. Replace it to take advantage of existing
1925 // DAG Combines. If not eliminated, we will match back to BFE during
1928 // TODO: The sext_inreg of extended types ends, although we can could
1929 // handle them in a single BFE.
1930 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1931 DAG.getValueType(SmallVT));
1934 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1937 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1939 return constantFoldBFE<int32_t>(DAG,
1940 Val->getSExtValue(),
1945 return constantFoldBFE<uint32_t>(DAG,
1946 Val->getZExtValue(),
1951 APInt Demanded = APInt::getBitsSet(32,
1953 OffsetVal + WidthVal);
1955 if ((OffsetVal + WidthVal) >= 32) {
1956 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1957 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1958 BitsFrom, ShiftVal);
1961 APInt KnownZero, KnownOne;
1962 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1963 !DCI.isBeforeLegalizeOps());
1964 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1965 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1966 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1967 DCI.CommitTargetLoweringOpt(TLO);
1976 //===----------------------------------------------------------------------===//
1978 //===----------------------------------------------------------------------===//
1980 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1983 const SmallVectorImpl<ISD::InputArg> &Ins,
1984 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1986 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1987 if (Ins[i].ArgVT == Ins[i].VT) {
1988 OrigIns.push_back(Ins[i]);
1993 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1994 // Vector has been split into scalars.
1995 VT = Ins[i].ArgVT.getVectorElementType();
1996 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1997 Ins[i].ArgVT.getVectorElementType() !=
1998 Ins[i].VT.getVectorElementType()) {
1999 // Vector elements have been promoted
2002 // Vector has been spilt into smaller vectors.
2006 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2007 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2008 OrigIns.push_back(Arg);
2012 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2013 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2014 return CFP->isExactlyValue(1.0);
2016 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2017 return C->isAllOnesValue();
2022 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2023 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2024 return CFP->getValueAPF().isZero();
2026 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2027 return C->isNullValue();
2032 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2033 const TargetRegisterClass *RC,
2034 unsigned Reg, EVT VT) const {
2035 MachineFunction &MF = DAG.getMachineFunction();
2036 MachineRegisterInfo &MRI = MF.getRegInfo();
2037 unsigned VirtualRegister;
2038 if (!MRI.isLiveIn(Reg)) {
2039 VirtualRegister = MRI.createVirtualRegister(RC);
2040 MRI.addLiveIn(Reg, VirtualRegister);
2042 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2044 return DAG.getRegister(VirtualRegister, VT);
2047 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2049 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2051 default: return nullptr;
2053 NODE_NAME_CASE(CALL);
2054 NODE_NAME_CASE(UMUL);
2055 NODE_NAME_CASE(DIV_INF);
2056 NODE_NAME_CASE(RET_FLAG);
2057 NODE_NAME_CASE(BRANCH_COND);
2060 NODE_NAME_CASE(DWORDADDR)
2061 NODE_NAME_CASE(FRACT)
2062 NODE_NAME_CASE(CLAMP)
2063 NODE_NAME_CASE(FMAX)
2064 NODE_NAME_CASE(SMAX)
2065 NODE_NAME_CASE(UMAX)
2066 NODE_NAME_CASE(FMIN)
2067 NODE_NAME_CASE(SMIN)
2068 NODE_NAME_CASE(UMIN)
2069 NODE_NAME_CASE(URECIP)
2070 NODE_NAME_CASE(DIV_SCALE)
2071 NODE_NAME_CASE(DIV_FMAS)
2072 NODE_NAME_CASE(DIV_FIXUP)
2073 NODE_NAME_CASE(TRIG_PREOP)
2076 NODE_NAME_CASE(DOT4)
2077 NODE_NAME_CASE(BFE_U32)
2078 NODE_NAME_CASE(BFE_I32)
2081 NODE_NAME_CASE(BREV)
2082 NODE_NAME_CASE(MUL_U24)
2083 NODE_NAME_CASE(MUL_I24)
2084 NODE_NAME_CASE(MAD_U24)
2085 NODE_NAME_CASE(MAD_I24)
2086 NODE_NAME_CASE(EXPORT)
2087 NODE_NAME_CASE(CONST_ADDRESS)
2088 NODE_NAME_CASE(REGISTER_LOAD)
2089 NODE_NAME_CASE(REGISTER_STORE)
2090 NODE_NAME_CASE(LOAD_CONSTANT)
2091 NODE_NAME_CASE(LOAD_INPUT)
2092 NODE_NAME_CASE(SAMPLE)
2093 NODE_NAME_CASE(SAMPLEB)
2094 NODE_NAME_CASE(SAMPLED)
2095 NODE_NAME_CASE(SAMPLEL)
2096 NODE_NAME_CASE(CVT_F32_UBYTE0)
2097 NODE_NAME_CASE(CVT_F32_UBYTE1)
2098 NODE_NAME_CASE(CVT_F32_UBYTE2)
2099 NODE_NAME_CASE(CVT_F32_UBYTE3)
2100 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2101 NODE_NAME_CASE(STORE_MSKOR)
2102 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2106 static void computeKnownBitsForMinMax(const SDValue Op0,
2110 const SelectionDAG &DAG,
2112 APInt Op0Zero, Op0One;
2113 APInt Op1Zero, Op1One;
2114 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2115 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2117 KnownZero = Op0Zero & Op1Zero;
2118 KnownOne = Op0One & Op1One;
2121 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2125 const SelectionDAG &DAG,
2126 unsigned Depth) const {
2128 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2132 unsigned Opc = Op.getOpcode();
2137 case ISD::INTRINSIC_WO_CHAIN: {
2138 // FIXME: The intrinsic should just use the node.
2139 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2140 case AMDGPUIntrinsic::AMDGPU_imax:
2141 case AMDGPUIntrinsic::AMDGPU_umax:
2142 case AMDGPUIntrinsic::AMDGPU_imin:
2143 case AMDGPUIntrinsic::AMDGPU_umin:
2144 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2145 KnownZero, KnownOne, DAG, Depth);
2153 case AMDGPUISD::SMAX:
2154 case AMDGPUISD::UMAX:
2155 case AMDGPUISD::SMIN:
2156 case AMDGPUISD::UMIN:
2157 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2158 KnownZero, KnownOne, DAG, Depth);
2161 case AMDGPUISD::BFE_I32:
2162 case AMDGPUISD::BFE_U32: {
2163 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2167 unsigned BitWidth = 32;
2168 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2170 KnownZero = APInt::getAllOnesValue(BitWidth);
2171 KnownOne = APInt::getNullValue(BitWidth);
2175 // FIXME: This could do a lot more. If offset is 0, should be the same as
2176 // sign_extend_inreg implementation, but that involves duplicating it.
2177 if (Opc == AMDGPUISD::BFE_I32)
2178 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2180 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2187 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2189 const SelectionDAG &DAG,
2190 unsigned Depth) const {
2191 switch (Op.getOpcode()) {
2192 case AMDGPUISD::BFE_I32: {
2193 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2197 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2198 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2199 if (!Offset || !Offset->isNullValue())
2202 // TODO: Could probably figure something out with non-0 offsets.
2203 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2204 return std::max(SignBits, Op0SignBits);
2207 case AMDGPUISD::BFE_U32: {
2208 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2209 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;