1 //==- SIMachineFunctionInfo.h - SIMachineFunctionInfo interface --*- C++ -*-==//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H
14 #define LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H
16 #include "AMDGPUArgumentUsageInfo.h"
17 #include "AMDGPUMachineFunction.h"
18 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
19 #include "SIInstrInfo.h"
20 #include "SIRegisterInfo.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/ADT/DenseMap.h"
23 #include "llvm/ADT/Optional.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/SparseBitVector.h"
27 #include "llvm/CodeGen/MIRYamlMapping.h"
28 #include "llvm/CodeGen/PseudoSourceValue.h"
29 #include "llvm/CodeGen/TargetInstrInfo.h"
30 #include "llvm/MC/MCRegisterInfo.h"
31 #include "llvm/Support/ErrorHandling.h"
39 class MachineFrameInfo;
40 class MachineFunction;
41 class TargetRegisterClass;
43 class AMDGPUPseudoSourceValue : public PseudoSourceValue {
45 enum AMDGPUPSVKind : unsigned {
46 PSVBuffer = PseudoSourceValue::TargetCustom,
52 AMDGPUPseudoSourceValue(unsigned Kind, const TargetInstrInfo &TII)
53 : PseudoSourceValue(Kind, TII) {}
56 bool isConstant(const MachineFrameInfo *) const override {
57 // This should probably be true for most images, but we will start by being
62 bool isAliased(const MachineFrameInfo *) const override {
66 bool mayAlias(const MachineFrameInfo *) const override {
71 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue {
73 explicit AMDGPUBufferPseudoSourceValue(const TargetInstrInfo &TII)
74 : AMDGPUPseudoSourceValue(PSVBuffer, TII) {}
76 static bool classof(const PseudoSourceValue *V) {
77 return V->kind() == PSVBuffer;
81 class AMDGPUImagePseudoSourceValue final : public AMDGPUPseudoSourceValue {
83 // TODO: Is the img rsrc useful?
84 explicit AMDGPUImagePseudoSourceValue(const TargetInstrInfo &TII)
85 : AMDGPUPseudoSourceValue(PSVImage, TII) {}
87 static bool classof(const PseudoSourceValue *V) {
88 return V->kind() == PSVImage;
92 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue {
94 explicit AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo &TII)
95 : AMDGPUPseudoSourceValue(GWSResource, TII) {}
97 static bool classof(const PseudoSourceValue *V) {
98 return V->kind() == GWSResource;
101 // These are inaccessible memory from IR.
102 bool isAliased(const MachineFrameInfo *) const override {
106 // These are inaccessible memory from IR.
107 bool mayAlias(const MachineFrameInfo *) const override {
111 void printCustom(raw_ostream &OS) const override {
121 StringValue RegisterName;
122 unsigned StackOffset;
124 Optional<unsigned> Mask;
126 // Default constructor, which creates a stack argument.
127 SIArgument() : IsRegister(false), StackOffset(0) {}
128 SIArgument(const SIArgument &Other) {
129 IsRegister = Other.IsRegister;
131 ::new ((void *)std::addressof(RegisterName))
132 StringValue(Other.RegisterName);
134 StackOffset = Other.StackOffset;
137 SIArgument &operator=(const SIArgument &Other) {
138 IsRegister = Other.IsRegister;
140 ::new ((void *)std::addressof(RegisterName))
141 StringValue(Other.RegisterName);
143 StackOffset = Other.StackOffset;
149 RegisterName.~StringValue();
152 // Helper to create a register or stack argument.
153 static inline SIArgument createArgument(bool IsReg) {
155 return SIArgument(IsReg);
160 // Construct a register argument.
161 SIArgument(bool) : IsRegister(true), RegisterName() {}
164 template <> struct MappingTraits<SIArgument> {
165 static void mapping(IO &YamlIO, SIArgument &A) {
166 if (YamlIO.outputting()) {
168 YamlIO.mapRequired("reg", A.RegisterName);
170 YamlIO.mapRequired("offset", A.StackOffset);
172 auto Keys = YamlIO.keys();
173 if (is_contained(Keys, "reg")) {
174 A = SIArgument::createArgument(true);
175 YamlIO.mapRequired("reg", A.RegisterName);
176 } else if (is_contained(Keys, "offset"))
177 YamlIO.mapRequired("offset", A.StackOffset);
179 YamlIO.setError("missing required key 'reg' or 'offset'");
181 YamlIO.mapOptional("mask", A.Mask);
183 static const bool flow = true;
186 struct SIArgumentInfo {
187 Optional<SIArgument> PrivateSegmentBuffer;
188 Optional<SIArgument> DispatchPtr;
189 Optional<SIArgument> QueuePtr;
190 Optional<SIArgument> KernargSegmentPtr;
191 Optional<SIArgument> DispatchID;
192 Optional<SIArgument> FlatScratchInit;
193 Optional<SIArgument> PrivateSegmentSize;
195 Optional<SIArgument> WorkGroupIDX;
196 Optional<SIArgument> WorkGroupIDY;
197 Optional<SIArgument> WorkGroupIDZ;
198 Optional<SIArgument> WorkGroupInfo;
199 Optional<SIArgument> PrivateSegmentWaveByteOffset;
201 Optional<SIArgument> ImplicitArgPtr;
202 Optional<SIArgument> ImplicitBufferPtr;
204 Optional<SIArgument> WorkItemIDX;
205 Optional<SIArgument> WorkItemIDY;
206 Optional<SIArgument> WorkItemIDZ;
209 template <> struct MappingTraits<SIArgumentInfo> {
210 static void mapping(IO &YamlIO, SIArgumentInfo &AI) {
211 YamlIO.mapOptional("privateSegmentBuffer", AI.PrivateSegmentBuffer);
212 YamlIO.mapOptional("dispatchPtr", AI.DispatchPtr);
213 YamlIO.mapOptional("queuePtr", AI.QueuePtr);
214 YamlIO.mapOptional("kernargSegmentPtr", AI.KernargSegmentPtr);
215 YamlIO.mapOptional("dispatchID", AI.DispatchID);
216 YamlIO.mapOptional("flatScratchInit", AI.FlatScratchInit);
217 YamlIO.mapOptional("privateSegmentSize", AI.PrivateSegmentSize);
219 YamlIO.mapOptional("workGroupIDX", AI.WorkGroupIDX);
220 YamlIO.mapOptional("workGroupIDY", AI.WorkGroupIDY);
221 YamlIO.mapOptional("workGroupIDZ", AI.WorkGroupIDZ);
222 YamlIO.mapOptional("workGroupInfo", AI.WorkGroupInfo);
223 YamlIO.mapOptional("privateSegmentWaveByteOffset",
224 AI.PrivateSegmentWaveByteOffset);
226 YamlIO.mapOptional("implicitArgPtr", AI.ImplicitArgPtr);
227 YamlIO.mapOptional("implicitBufferPtr", AI.ImplicitBufferPtr);
229 YamlIO.mapOptional("workItemIDX", AI.WorkItemIDX);
230 YamlIO.mapOptional("workItemIDY", AI.WorkItemIDY);
231 YamlIO.mapOptional("workItemIDZ", AI.WorkItemIDZ);
235 // Default to default mode for default calling convention.
238 bool DX10Clamp = true;
239 bool FP32InputDenormals = true;
240 bool FP32OutputDenormals = true;
241 bool FP64FP16InputDenormals = true;
242 bool FP64FP16OutputDenormals = true;
246 SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) {
248 DX10Clamp = Mode.DX10Clamp;
249 FP32InputDenormals = Mode.FP32InputDenormals;
250 FP32OutputDenormals = Mode.FP32OutputDenormals;
251 FP64FP16InputDenormals = Mode.FP64FP16InputDenormals;
252 FP64FP16OutputDenormals = Mode.FP64FP16OutputDenormals;
255 bool operator ==(const SIMode Other) const {
256 return IEEE == Other.IEEE &&
257 DX10Clamp == Other.DX10Clamp &&
258 FP32InputDenormals == Other.FP32InputDenormals &&
259 FP32OutputDenormals == Other.FP32OutputDenormals &&
260 FP64FP16InputDenormals == Other.FP64FP16InputDenormals &&
261 FP64FP16OutputDenormals == Other.FP64FP16OutputDenormals;
265 template <> struct MappingTraits<SIMode> {
266 static void mapping(IO &YamlIO, SIMode &Mode) {
267 YamlIO.mapOptional("ieee", Mode.IEEE, true);
268 YamlIO.mapOptional("dx10-clamp", Mode.DX10Clamp, true);
269 YamlIO.mapOptional("fp32-input-denormals", Mode.FP32InputDenormals, true);
270 YamlIO.mapOptional("fp32-output-denormals", Mode.FP32OutputDenormals, true);
271 YamlIO.mapOptional("fp64-fp16-input-denormals", Mode.FP64FP16InputDenormals, true);
272 YamlIO.mapOptional("fp64-fp16-output-denormals", Mode.FP64FP16OutputDenormals, true);
276 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo {
277 uint64_t ExplicitKernArgSize = 0;
278 unsigned MaxKernArgAlign = 0;
279 unsigned LDSSize = 0;
280 bool IsEntryFunction = false;
281 bool NoSignedZerosFPMath = false;
282 bool MemoryBound = false;
283 bool WaveLimiter = false;
284 uint32_t HighBitsOf32BitAddress = 0;
286 StringValue ScratchRSrcReg = "$private_rsrc_reg";
287 StringValue FrameOffsetReg = "$fp_reg";
288 StringValue StackPtrOffsetReg = "$sp_reg";
290 Optional<SIArgumentInfo> ArgInfo;
293 SIMachineFunctionInfo() = default;
294 SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &,
295 const TargetRegisterInfo &TRI);
297 void mappingImpl(yaml::IO &YamlIO) override;
298 ~SIMachineFunctionInfo() = default;
301 template <> struct MappingTraits<SIMachineFunctionInfo> {
302 static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) {
303 YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize,
305 YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u);
306 YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u);
307 YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false);
308 YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false);
309 YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false);
310 YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false);
311 YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg,
312 StringValue("$private_rsrc_reg"));
313 YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg,
314 StringValue("$fp_reg"));
315 YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg,
316 StringValue("$sp_reg"));
317 YamlIO.mapOptional("argumentInfo", MFI.ArgInfo);
318 YamlIO.mapOptional("mode", MFI.Mode, SIMode());
319 YamlIO.mapOptional("highBitsOf32BitAddress",
320 MFI.HighBitsOf32BitAddress, 0u);
324 } // end namespace yaml
326 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which
327 /// tells the hardware which interpolation parameters to load.
328 class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
329 friend class GCNTargetMachine;
331 Register TIDReg = AMDGPU::NoRegister;
333 // Registers that may be reserved for spilling purposes. These may be the same
334 // as the input registers.
335 Register ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG;
337 // This is the the unswizzled offset from the current dispatch's scratch wave
338 // base to the beginning of the current function's frame.
339 Register FrameOffsetReg = AMDGPU::FP_REG;
341 // This is an ABI register used in the non-entry calling convention to
342 // communicate the unswizzled offset from the current dispatch's scratch wave
343 // base to the beginning of the new function's frame.
344 Register StackPtrOffsetReg = AMDGPU::SP_REG;
346 AMDGPUFunctionArgInfo ArgInfo;
349 unsigned PSInputAddr = 0;
350 unsigned PSInputEnable = 0;
352 /// Number of bytes of arguments this function has on the stack. If the callee
353 /// is expected to restore the argument stack this should be a multiple of 16,
354 /// all usable during a tail call.
356 /// The alternative would forbid tail call optimisation in some cases: if we
357 /// want to transfer control from a function with 8-bytes of stack-argument
358 /// space to a function with 16-bytes then misalignment of this value would
359 /// make a stack adjustment necessary, which could not be undone by the
361 unsigned BytesInStackArgArea = 0;
363 bool ReturnsVoid = true;
365 // A pair of default/requested minimum/maximum flat work group sizes.
366 // Minimum - first, maximum - second.
367 std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0};
369 // A pair of default/requested minimum/maximum number of waves per execution
370 // unit. Minimum - first, maximum - second.
371 std::pair<unsigned, unsigned> WavesPerEU = {0, 0};
373 DenseMap<const Value *,
374 std::unique_ptr<const AMDGPUBufferPseudoSourceValue>> BufferPSVs;
375 DenseMap<const Value *,
376 std::unique_ptr<const AMDGPUImagePseudoSourceValue>> ImagePSVs;
377 std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV;
380 unsigned LDSWaveSpillSize = 0;
381 unsigned NumUserSGPRs = 0;
382 unsigned NumSystemSGPRs = 0;
384 bool HasSpilledSGPRs = false;
385 bool HasSpilledVGPRs = false;
386 bool HasNonSpillStackObjects = false;
387 bool IsStackRealigned = false;
389 unsigned NumSpilledSGPRs = 0;
390 unsigned NumSpilledVGPRs = 0;
392 // Feature bits required for inputs passed in user SGPRs.
393 bool PrivateSegmentBuffer : 1;
394 bool DispatchPtr : 1;
396 bool KernargSegmentPtr : 1;
398 bool FlatScratchInit : 1;
400 // Feature bits required for inputs passed in system SGPRs.
401 bool WorkGroupIDX : 1; // Always initialized.
402 bool WorkGroupIDY : 1;
403 bool WorkGroupIDZ : 1;
404 bool WorkGroupInfo : 1;
405 bool PrivateSegmentWaveByteOffset : 1;
407 bool WorkItemIDX : 1; // Always initialized.
408 bool WorkItemIDY : 1;
409 bool WorkItemIDZ : 1;
411 // Private memory buffer
412 // Compute directly in sgpr[0:1]
413 // Other shaders indirect 64-bits at sgpr[0:1]
414 bool ImplicitBufferPtr : 1;
416 // Pointer to where the ABI inserts special kernel arguments separate from the
417 // user arguments. This is an offset from the KernargSegmentPtr.
418 bool ImplicitArgPtr : 1;
420 // The hard-wired high half of the address of the global information table
421 // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since
422 // current hardware only allows a 16 bit value.
425 unsigned HighBitsOf32BitAddress;
428 // Current recorded maximum possible occupancy.
431 MCPhysReg getNextUserSGPR() const;
433 MCPhysReg getNextSystemSGPR() const;
440 SpilledReg() = default;
441 SpilledReg(Register R, int L) : VGPR (R), Lane (L) {}
443 bool hasLane() { return Lane != -1;}
444 bool hasReg() { return VGPR != 0;}
447 struct SGPRSpillVGPRCSR {
448 // VGPR used for SGPR spills
451 // If the VGPR is a CSR, the stack slot used to save/restore it in the
455 SGPRSpillVGPRCSR(Register V, Optional<int> F) : VGPR(V), FI(F) {}
458 struct VGPRSpillToAGPR {
459 SmallVector<MCPhysReg, 32> Lanes;
460 bool FullyAllocated = false;
463 SparseBitVector<> WWMReservedRegs;
465 void ReserveWWMRegister(Register Reg) { WWMReservedRegs.set(Reg); }
468 // Track VGPR + wave index for each subregister of the SGPR spilled to
470 DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills;
471 unsigned NumVGPRSpillLanes = 0;
472 SmallVector<SGPRSpillVGPRCSR, 2> SpillVGPRs;
474 DenseMap<int, VGPRSpillToAGPR> VGPRToAGPRSpills;
476 // AGPRs used for VGPR spills.
477 SmallVector<MCPhysReg, 32> SpillAGPR;
479 // VGPRs used for AGPR spills.
480 SmallVector<MCPhysReg, 32> SpillVGPR;
483 /// If this is set, an SGPR used for save/restore of the register used for the
485 Register SGPRForFPSaveRestoreCopy;
486 Optional<int> FramePointerSaveIndex;
488 /// If this is set, an SGPR used for save/restore of the register used for the
490 Register SGPRForBPSaveRestoreCopy;
491 Optional<int> BasePointerSaveIndex;
493 Register VGPRReservedForSGPRSpill;
494 bool isCalleeSavedReg(const MCPhysReg *CSRegs, MCPhysReg Reg);
497 SIMachineFunctionInfo(const MachineFunction &MF);
499 bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI);
501 ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const {
502 auto I = SGPRToVGPRSpills.find(FrameIndex);
503 return (I == SGPRToVGPRSpills.end()) ?
504 ArrayRef<SpilledReg>() : makeArrayRef(I->second);
507 ArrayRef<SGPRSpillVGPRCSR> getSGPRSpillVGPRs() const {
511 void setSGPRSpillVGPRs(Register NewVGPR, Optional<int> newFI, int Index) {
512 SpillVGPRs[Index].VGPR = NewVGPR;
513 SpillVGPRs[Index].FI = newFI;
514 VGPRReservedForSGPRSpill = NewVGPR;
517 bool removeVGPRForSGPRSpill(Register ReservedVGPR, MachineFunction &MF);
519 ArrayRef<MCPhysReg> getAGPRSpillVGPRs() const {
523 ArrayRef<MCPhysReg> getVGPRSpillAGPRs() const {
527 MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const {
528 auto I = VGPRToAGPRSpills.find(FrameIndex);
529 return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister
530 : I->second.Lanes[Lane];
533 bool haveFreeLanesForSGPRSpill(const MachineFunction &MF,
534 unsigned NumLane) const;
535 bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI);
536 bool reserveVGPRforSGPRSpills(MachineFunction &MF);
537 bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR);
538 void removeDeadFrameIndices(MachineFrameInfo &MFI);
540 bool hasCalculatedTID() const { return TIDReg != 0; };
541 Register getTIDReg() const { return TIDReg; };
542 void setTIDReg(Register Reg) { TIDReg = Reg; }
544 unsigned getBytesInStackArgArea() const {
545 return BytesInStackArgArea;
548 void setBytesInStackArgArea(unsigned Bytes) {
549 BytesInStackArgArea = Bytes;
553 Register addPrivateSegmentBuffer(const SIRegisterInfo &TRI);
554 Register addDispatchPtr(const SIRegisterInfo &TRI);
555 Register addQueuePtr(const SIRegisterInfo &TRI);
556 Register addKernargSegmentPtr(const SIRegisterInfo &TRI);
557 Register addDispatchID(const SIRegisterInfo &TRI);
558 Register addFlatScratchInit(const SIRegisterInfo &TRI);
559 Register addImplicitBufferPtr(const SIRegisterInfo &TRI);
562 Register addWorkGroupIDX() {
563 ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR());
565 return ArgInfo.WorkGroupIDX.getRegister();
568 Register addWorkGroupIDY() {
569 ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR());
571 return ArgInfo.WorkGroupIDY.getRegister();
574 Register addWorkGroupIDZ() {
575 ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR());
577 return ArgInfo.WorkGroupIDZ.getRegister();
580 Register addWorkGroupInfo() {
581 ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR());
583 return ArgInfo.WorkGroupInfo.getRegister();
586 // Add special VGPR inputs
587 void setWorkItemIDX(ArgDescriptor Arg) {
588 ArgInfo.WorkItemIDX = Arg;
591 void setWorkItemIDY(ArgDescriptor Arg) {
592 ArgInfo.WorkItemIDY = Arg;
595 void setWorkItemIDZ(ArgDescriptor Arg) {
596 ArgInfo.WorkItemIDZ = Arg;
599 Register addPrivateSegmentWaveByteOffset() {
600 ArgInfo.PrivateSegmentWaveByteOffset
601 = ArgDescriptor::createRegister(getNextSystemSGPR());
603 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
606 void setPrivateSegmentWaveByteOffset(Register Reg) {
607 ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg);
610 bool hasPrivateSegmentBuffer() const {
611 return PrivateSegmentBuffer;
614 bool hasDispatchPtr() const {
618 bool hasQueuePtr() const {
622 bool hasKernargSegmentPtr() const {
623 return KernargSegmentPtr;
626 bool hasDispatchID() const {
630 bool hasFlatScratchInit() const {
631 return FlatScratchInit;
634 bool hasWorkGroupIDX() const {
638 bool hasWorkGroupIDY() const {
642 bool hasWorkGroupIDZ() const {
646 bool hasWorkGroupInfo() const {
647 return WorkGroupInfo;
650 bool hasPrivateSegmentWaveByteOffset() const {
651 return PrivateSegmentWaveByteOffset;
654 bool hasWorkItemIDX() const {
658 bool hasWorkItemIDY() const {
662 bool hasWorkItemIDZ() const {
666 bool hasImplicitArgPtr() const {
667 return ImplicitArgPtr;
670 bool hasImplicitBufferPtr() const {
671 return ImplicitBufferPtr;
674 AMDGPUFunctionArgInfo &getArgInfo() {
678 const AMDGPUFunctionArgInfo &getArgInfo() const {
682 std::tuple<const ArgDescriptor *, const TargetRegisterClass *, LLT>
683 getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const {
684 return ArgInfo.getPreloadedValue(Value);
687 Register getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const {
688 auto Arg = std::get<0>(ArgInfo.getPreloadedValue(Value));
689 return Arg ? Arg->getRegister() : Register();
692 unsigned getGITPtrHigh() const {
696 Register getGITPtrLoReg(const MachineFunction &MF) const;
698 uint32_t get32BitAddressHighBits() const {
699 return HighBitsOf32BitAddress;
702 unsigned getGDSSize() const {
706 unsigned getNumUserSGPRs() const {
710 unsigned getNumPreloadedSGPRs() const {
711 return NumUserSGPRs + NumSystemSGPRs;
714 Register getPrivateSegmentWaveByteOffsetSystemSGPR() const {
715 return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
718 /// Returns the physical register reserved for use as the resource
719 /// descriptor for scratch accesses.
720 Register getScratchRSrcReg() const {
721 return ScratchRSrcReg;
724 void setScratchRSrcReg(Register Reg) {
725 assert(Reg != 0 && "Should never be unset");
726 ScratchRSrcReg = Reg;
729 Register getFrameOffsetReg() const {
730 return FrameOffsetReg;
733 void setFrameOffsetReg(Register Reg) {
734 assert(Reg != 0 && "Should never be unset");
735 FrameOffsetReg = Reg;
738 void setStackPtrOffsetReg(Register Reg) {
739 assert(Reg != 0 && "Should never be unset");
740 StackPtrOffsetReg = Reg;
743 // Note the unset value for this is AMDGPU::SP_REG rather than
744 // NoRegister. This is mostly a workaround for MIR tests where state that
745 // can't be directly computed from the function is not preserved in serialized
747 Register getStackPtrOffsetReg() const {
748 return StackPtrOffsetReg;
751 Register getQueuePtrUserSGPR() const {
752 return ArgInfo.QueuePtr.getRegister();
755 Register getImplicitBufferPtrUserSGPR() const {
756 return ArgInfo.ImplicitBufferPtr.getRegister();
759 bool hasSpilledSGPRs() const {
760 return HasSpilledSGPRs;
763 void setHasSpilledSGPRs(bool Spill = true) {
764 HasSpilledSGPRs = Spill;
767 bool hasSpilledVGPRs() const {
768 return HasSpilledVGPRs;
771 void setHasSpilledVGPRs(bool Spill = true) {
772 HasSpilledVGPRs = Spill;
775 bool hasNonSpillStackObjects() const {
776 return HasNonSpillStackObjects;
779 void setHasNonSpillStackObjects(bool StackObject = true) {
780 HasNonSpillStackObjects = StackObject;
783 bool isStackRealigned() const {
784 return IsStackRealigned;
787 void setIsStackRealigned(bool Realigned = true) {
788 IsStackRealigned = Realigned;
791 unsigned getNumSpilledSGPRs() const {
792 return NumSpilledSGPRs;
795 unsigned getNumSpilledVGPRs() const {
796 return NumSpilledVGPRs;
799 void addToSpilledSGPRs(unsigned num) {
800 NumSpilledSGPRs += num;
803 void addToSpilledVGPRs(unsigned num) {
804 NumSpilledVGPRs += num;
807 unsigned getPSInputAddr() const {
811 unsigned getPSInputEnable() const {
812 return PSInputEnable;
815 bool isPSInputAllocated(unsigned Index) const {
816 return PSInputAddr & (1 << Index);
819 void markPSInputAllocated(unsigned Index) {
820 PSInputAddr |= 1 << Index;
823 void markPSInputEnabled(unsigned Index) {
824 PSInputEnable |= 1 << Index;
827 bool returnsVoid() const {
831 void setIfReturnsVoid(bool Value) {
835 /// \returns A pair of default/requested minimum/maximum flat work group sizes
836 /// for this function.
837 std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const {
838 return FlatWorkGroupSizes;
841 /// \returns Default/requested minimum flat work group size for this function.
842 unsigned getMinFlatWorkGroupSize() const {
843 return FlatWorkGroupSizes.first;
846 /// \returns Default/requested maximum flat work group size for this function.
847 unsigned getMaxFlatWorkGroupSize() const {
848 return FlatWorkGroupSizes.second;
851 /// \returns A pair of default/requested minimum/maximum number of waves per
853 std::pair<unsigned, unsigned> getWavesPerEU() const {
857 /// \returns Default/requested minimum number of waves per execution unit.
858 unsigned getMinWavesPerEU() const {
859 return WavesPerEU.first;
862 /// \returns Default/requested maximum number of waves per execution unit.
863 unsigned getMaxWavesPerEU() const {
864 return WavesPerEU.second;
867 /// \returns SGPR used for \p Dim's work group ID.
868 Register getWorkGroupIDSGPR(unsigned Dim) const {
871 assert(hasWorkGroupIDX());
872 return ArgInfo.WorkGroupIDX.getRegister();
874 assert(hasWorkGroupIDY());
875 return ArgInfo.WorkGroupIDY.getRegister();
877 assert(hasWorkGroupIDZ());
878 return ArgInfo.WorkGroupIDZ.getRegister();
880 llvm_unreachable("unexpected dimension");
883 unsigned getLDSWaveSpillSize() const {
884 return LDSWaveSpillSize;
887 const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII,
888 const Value *BufferRsrc) {
890 auto PSV = BufferPSVs.try_emplace(
892 std::make_unique<AMDGPUBufferPseudoSourceValue>(TII));
893 return PSV.first->second.get();
896 const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII,
897 const Value *ImgRsrc) {
899 auto PSV = ImagePSVs.try_emplace(
901 std::make_unique<AMDGPUImagePseudoSourceValue>(TII));
902 return PSV.first->second.get();
905 const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) {
906 if (!GWSResourcePSV) {
908 std::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII);
911 return GWSResourcePSV.get();
914 unsigned getOccupancy() const {
918 unsigned getMinAllowedOccupancy() const {
919 if (!isMemoryBound() && !needsWaveLimiter())
921 return (Occupancy < 4) ? Occupancy : 4;
924 void limitOccupancy(const MachineFunction &MF);
926 void limitOccupancy(unsigned Limit) {
927 if (Occupancy > Limit)
931 void increaseOccupancy(const MachineFunction &MF, unsigned Limit) {
932 if (Occupancy < Limit)
938 } // end namespace llvm
940 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H