xref: /llvm-project/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h (revision 58426a3707dcae151134b56688135da32f83e5bf) 
1 //==- SIMachineFunctionInfo.h - SIMachineFunctionInfo interface --*- C++ -*-==//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #ifndef LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H
14 #define LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H
15 
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"
32 #include <array>
33 #include <cassert>
34 #include <utility>
35 #include <vector>
36 
37 namespace llvm {
38 
39 class MachineFrameInfo;
40 class MachineFunction;
41 class TargetRegisterClass;
42 
43 class AMDGPUPseudoSourceValue : public PseudoSourceValue {
44 public:
45   enum AMDGPUPSVKind : unsigned {
46     PSVBuffer = PseudoSourceValue::TargetCustom,
47     PSVImage,
48     GWSResource
49   };
50 
51 protected:
52   AMDGPUPseudoSourceValue(unsigned Kind, const TargetInstrInfo &TII)
53       : PseudoSourceValue(Kind, TII) {}
54 
55 public:
56   bool isConstant(const MachineFrameInfo *) const override {
57     // This should probably be true for most images, but we will start by being
58     // conservative.
59     return false;
60   }
61 
62   bool isAliased(const MachineFrameInfo *) const override {
63     return true;
64   }
65 
66   bool mayAlias(const MachineFrameInfo *) const override {
67     return true;
68   }
69 };
70 
71 class AMDGPUBufferPseudoSourceValue final : public AMDGPUPseudoSourceValue {
72 public:
73   explicit AMDGPUBufferPseudoSourceValue(const TargetInstrInfo &TII)
74       : AMDGPUPseudoSourceValue(PSVBuffer, TII) {}
75 
76   static bool classof(const PseudoSourceValue *V) {
77     return V->kind() == PSVBuffer;
78   }
79 };
80 
81 class AMDGPUImagePseudoSourceValue final : public AMDGPUPseudoSourceValue {
82 public:
83   // TODO: Is the img rsrc useful?
84   explicit AMDGPUImagePseudoSourceValue(const TargetInstrInfo &TII)
85       : AMDGPUPseudoSourceValue(PSVImage, TII) {}
86 
87   static bool classof(const PseudoSourceValue *V) {
88     return V->kind() == PSVImage;
89   }
90 };
91 
92 class AMDGPUGWSResourcePseudoSourceValue final : public AMDGPUPseudoSourceValue {
93 public:
94   explicit AMDGPUGWSResourcePseudoSourceValue(const TargetInstrInfo &TII)
95       : AMDGPUPseudoSourceValue(GWSResource, TII) {}
96 
97   static bool classof(const PseudoSourceValue *V) {
98     return V->kind() == GWSResource;
99   }
100 
101   // These are inaccessible memory from IR.
102   bool isAliased(const MachineFrameInfo *) const override {
103     return false;
104   }
105 
106   // These are inaccessible memory from IR.
107   bool mayAlias(const MachineFrameInfo *) const override {
108     return false;
109   }
110 
111   void printCustom(raw_ostream &OS) const override {
112     OS << "GWSResource";
113   }
114 };
115 
116 namespace yaml {
117 
118 struct SIArgument {
119   bool IsRegister;
120   union {
121     StringValue RegisterName;
122     unsigned StackOffset;
123   };
124   Optional<unsigned> Mask;
125 
126   // Default constructor, which creates a stack argument.
127   SIArgument() : IsRegister(false), StackOffset(0) {}
128   SIArgument(const SIArgument &Other) {
129     IsRegister = Other.IsRegister;
130     if (IsRegister) {
131       ::new ((void *)std::addressof(RegisterName))
132           StringValue(Other.RegisterName);
133     } else
134       StackOffset = Other.StackOffset;
135     Mask = Other.Mask;
136   }
137   SIArgument &operator=(const SIArgument &Other) {
138     IsRegister = Other.IsRegister;
139     if (IsRegister) {
140       ::new ((void *)std::addressof(RegisterName))
141           StringValue(Other.RegisterName);
142     } else
143       StackOffset = Other.StackOffset;
144     Mask = Other.Mask;
145     return *this;
146   }
147   ~SIArgument() {
148     if (IsRegister)
149       RegisterName.~StringValue();
150   }
151 
152   // Helper to create a register or stack argument.
153   static inline SIArgument createArgument(bool IsReg) {
154     if (IsReg)
155       return SIArgument(IsReg);
156     return SIArgument();
157   }
158 
159 private:
160   // Construct a register argument.
161   SIArgument(bool) : IsRegister(true), RegisterName() {}
162 };
163 
164 template <> struct MappingTraits<SIArgument> {
165   static void mapping(IO &YamlIO, SIArgument &A) {
166     if (YamlIO.outputting()) {
167       if (A.IsRegister)
168         YamlIO.mapRequired("reg", A.RegisterName);
169       else
170         YamlIO.mapRequired("offset", A.StackOffset);
171     } else {
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);
178       else
179         YamlIO.setError("missing required key 'reg' or 'offset'");
180     }
181     YamlIO.mapOptional("mask", A.Mask);
182   }
183   static const bool flow = true;
184 };
185 
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;
194 
195   Optional<SIArgument> WorkGroupIDX;
196   Optional<SIArgument> WorkGroupIDY;
197   Optional<SIArgument> WorkGroupIDZ;
198   Optional<SIArgument> WorkGroupInfo;
199   Optional<SIArgument> PrivateSegmentWaveByteOffset;
200 
201   Optional<SIArgument> ImplicitArgPtr;
202   Optional<SIArgument> ImplicitBufferPtr;
203 
204   Optional<SIArgument> WorkItemIDX;
205   Optional<SIArgument> WorkItemIDY;
206   Optional<SIArgument> WorkItemIDZ;
207 };
208 
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);
218 
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);
225 
226     YamlIO.mapOptional("implicitArgPtr", AI.ImplicitArgPtr);
227     YamlIO.mapOptional("implicitBufferPtr", AI.ImplicitBufferPtr);
228 
229     YamlIO.mapOptional("workItemIDX", AI.WorkItemIDX);
230     YamlIO.mapOptional("workItemIDY", AI.WorkItemIDY);
231     YamlIO.mapOptional("workItemIDZ", AI.WorkItemIDZ);
232   }
233 };
234 
235 // Default to default mode for default calling convention.
236 struct SIMode {
237   bool IEEE = true;
238   bool DX10Clamp = true;
239 
240   SIMode() = default;
241 
242 
243   SIMode(const AMDGPU::SIModeRegisterDefaults &Mode) {
244     IEEE = Mode.IEEE;
245     DX10Clamp = Mode.DX10Clamp;
246   }
247 
248   bool operator ==(const SIMode Other) const {
249     return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp;
250   }
251 };
252 
253 template <> struct MappingTraits<SIMode> {
254   static void mapping(IO &YamlIO, SIMode &Mode) {
255     YamlIO.mapOptional("ieee", Mode.IEEE, true);
256     YamlIO.mapOptional("dx10-clamp", Mode.DX10Clamp, true);
257   }
258 };
259 
260 struct SIMachineFunctionInfo final : public yaml::MachineFunctionInfo {
261   uint64_t ExplicitKernArgSize = 0;
262   unsigned MaxKernArgAlign = 0;
263   unsigned LDSSize = 0;
264   bool IsEntryFunction = false;
265   bool NoSignedZerosFPMath = false;
266   bool MemoryBound = false;
267   bool WaveLimiter = false;
268 
269   StringValue ScratchRSrcReg = "$private_rsrc_reg";
270   StringValue ScratchWaveOffsetReg = "$scratch_wave_offset_reg";
271   StringValue FrameOffsetReg = "$fp_reg";
272   StringValue StackPtrOffsetReg = "$sp_reg";
273 
274   Optional<SIArgumentInfo> ArgInfo;
275   SIMode Mode;
276 
277   SIMachineFunctionInfo() = default;
278   SIMachineFunctionInfo(const llvm::SIMachineFunctionInfo &,
279                         const TargetRegisterInfo &TRI);
280 
281   void mappingImpl(yaml::IO &YamlIO) override;
282   ~SIMachineFunctionInfo() = default;
283 };
284 
285 template <> struct MappingTraits<SIMachineFunctionInfo> {
286   static void mapping(IO &YamlIO, SIMachineFunctionInfo &MFI) {
287     YamlIO.mapOptional("explicitKernArgSize", MFI.ExplicitKernArgSize,
288                        UINT64_C(0));
289     YamlIO.mapOptional("maxKernArgAlign", MFI.MaxKernArgAlign, 0u);
290     YamlIO.mapOptional("ldsSize", MFI.LDSSize, 0u);
291     YamlIO.mapOptional("isEntryFunction", MFI.IsEntryFunction, false);
292     YamlIO.mapOptional("noSignedZerosFPMath", MFI.NoSignedZerosFPMath, false);
293     YamlIO.mapOptional("memoryBound", MFI.MemoryBound, false);
294     YamlIO.mapOptional("waveLimiter", MFI.WaveLimiter, false);
295     YamlIO.mapOptional("scratchRSrcReg", MFI.ScratchRSrcReg,
296                        StringValue("$private_rsrc_reg"));
297     YamlIO.mapOptional("scratchWaveOffsetReg", MFI.ScratchWaveOffsetReg,
298                        StringValue("$scratch_wave_offset_reg"));
299     YamlIO.mapOptional("frameOffsetReg", MFI.FrameOffsetReg,
300                        StringValue("$fp_reg"));
301     YamlIO.mapOptional("stackPtrOffsetReg", MFI.StackPtrOffsetReg,
302                        StringValue("$sp_reg"));
303     YamlIO.mapOptional("argumentInfo", MFI.ArgInfo);
304     YamlIO.mapOptional("mode", MFI.Mode, SIMode());
305   }
306 };
307 
308 } // end namespace yaml
309 
310 /// This class keeps track of the SPI_SP_INPUT_ADDR config register, which
311 /// tells the hardware which interpolation parameters to load.
312 class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
313   friend class GCNTargetMachine;
314 
315   unsigned TIDReg = AMDGPU::NoRegister;
316 
317   // Registers that may be reserved for spilling purposes. These may be the same
318   // as the input registers.
319   unsigned ScratchRSrcReg = AMDGPU::PRIVATE_RSRC_REG;
320   unsigned ScratchWaveOffsetReg = AMDGPU::SCRATCH_WAVE_OFFSET_REG;
321 
322   // This is the current function's incremented size from the kernel's scratch
323   // wave offset register. For an entry function, this is exactly the same as
324   // the ScratchWaveOffsetReg.
325   unsigned FrameOffsetReg = AMDGPU::FP_REG;
326 
327   // Top of the stack SGPR offset derived from the ScratchWaveOffsetReg.
328   unsigned StackPtrOffsetReg = AMDGPU::SP_REG;
329 
330   AMDGPUFunctionArgInfo ArgInfo;
331 
332   // State of MODE register, assumed FP mode.
333   AMDGPU::SIModeRegisterDefaults Mode;
334 
335   // Graphics info.
336   unsigned PSInputAddr = 0;
337   unsigned PSInputEnable = 0;
338 
339   /// Number of bytes of arguments this function has on the stack. If the callee
340   /// is expected to restore the argument stack this should be a multiple of 16,
341   /// all usable during a tail call.
342   ///
343   /// The alternative would forbid tail call optimisation in some cases: if we
344   /// want to transfer control from a function with 8-bytes of stack-argument
345   /// space to a function with 16-bytes then misalignment of this value would
346   /// make a stack adjustment necessary, which could not be undone by the
347   /// callee.
348   unsigned BytesInStackArgArea = 0;
349 
350   bool ReturnsVoid = true;
351 
352   // A pair of default/requested minimum/maximum flat work group sizes.
353   // Minimum - first, maximum - second.
354   std::pair<unsigned, unsigned> FlatWorkGroupSizes = {0, 0};
355 
356   // A pair of default/requested minimum/maximum number of waves per execution
357   // unit. Minimum - first, maximum - second.
358   std::pair<unsigned, unsigned> WavesPerEU = {0, 0};
359 
360   DenseMap<const Value *,
361            std::unique_ptr<const AMDGPUBufferPseudoSourceValue>> BufferPSVs;
362   DenseMap<const Value *,
363            std::unique_ptr<const AMDGPUImagePseudoSourceValue>> ImagePSVs;
364   std::unique_ptr<const AMDGPUGWSResourcePseudoSourceValue> GWSResourcePSV;
365 
366 private:
367   unsigned LDSWaveSpillSize = 0;
368   unsigned NumUserSGPRs = 0;
369   unsigned NumSystemSGPRs = 0;
370 
371   bool HasSpilledSGPRs = false;
372   bool HasSpilledVGPRs = false;
373   bool HasNonSpillStackObjects = false;
374   bool IsStackRealigned = false;
375 
376   unsigned NumSpilledSGPRs = 0;
377   unsigned NumSpilledVGPRs = 0;
378 
379   // Feature bits required for inputs passed in user SGPRs.
380   bool PrivateSegmentBuffer : 1;
381   bool DispatchPtr : 1;
382   bool QueuePtr : 1;
383   bool KernargSegmentPtr : 1;
384   bool DispatchID : 1;
385   bool FlatScratchInit : 1;
386 
387   // Feature bits required for inputs passed in system SGPRs.
388   bool WorkGroupIDX : 1; // Always initialized.
389   bool WorkGroupIDY : 1;
390   bool WorkGroupIDZ : 1;
391   bool WorkGroupInfo : 1;
392   bool PrivateSegmentWaveByteOffset : 1;
393 
394   bool WorkItemIDX : 1; // Always initialized.
395   bool WorkItemIDY : 1;
396   bool WorkItemIDZ : 1;
397 
398   // Private memory buffer
399   // Compute directly in sgpr[0:1]
400   // Other shaders indirect 64-bits at sgpr[0:1]
401   bool ImplicitBufferPtr : 1;
402 
403   // Pointer to where the ABI inserts special kernel arguments separate from the
404   // user arguments. This is an offset from the KernargSegmentPtr.
405   bool ImplicitArgPtr : 1;
406 
407   // The hard-wired high half of the address of the global information table
408   // for AMDPAL OS type. 0xffffffff represents no hard-wired high half, since
409   // current hardware only allows a 16 bit value.
410   unsigned GITPtrHigh;
411 
412   unsigned HighBitsOf32BitAddress;
413   unsigned GDSSize;
414 
415   // Current recorded maximum possible occupancy.
416   unsigned Occupancy;
417 
418   MCPhysReg getNextUserSGPR() const;
419 
420   MCPhysReg getNextSystemSGPR() const;
421 
422 public:
423   struct SpilledReg {
424     unsigned VGPR = 0;
425     int Lane = -1;
426 
427     SpilledReg() = default;
428     SpilledReg(unsigned R, int L) : VGPR (R), Lane (L) {}
429 
430     bool hasLane() { return Lane != -1;}
431     bool hasReg() { return VGPR != 0;}
432   };
433 
434   struct SGPRSpillVGPRCSR {
435     // VGPR used for SGPR spills
436     unsigned VGPR;
437 
438     // If the VGPR is a CSR, the stack slot used to save/restore it in the
439     // prolog/epilog.
440     Optional<int> FI;
441 
442     SGPRSpillVGPRCSR(unsigned V, Optional<int> F) : VGPR(V), FI(F) {}
443   };
444 
445   SparseBitVector<> WWMReservedRegs;
446 
447   void ReserveWWMRegister(unsigned reg) { WWMReservedRegs.set(reg); }
448 
449 private:
450   // SGPR->VGPR spilling support.
451   using SpillRegMask = std::pair<unsigned, unsigned>;
452 
453   // Track VGPR + wave index for each subregister of the SGPR spilled to
454   // frameindex key.
455   DenseMap<int, std::vector<SpilledReg>> SGPRToVGPRSpills;
456   unsigned NumVGPRSpillLanes = 0;
457   SmallVector<SGPRSpillVGPRCSR, 2> SpillVGPRs;
458 
459 public: // FIXME
460   /// If this is set, an SGPR used for save/restore of the register used for the
461   /// frame pointer.
462   unsigned SGPRForFPSaveRestoreCopy = 0;
463   Optional<int> FramePointerSaveIndex;
464 
465 public:
466   SIMachineFunctionInfo(const MachineFunction &MF);
467 
468   bool initializeBaseYamlFields(const yaml::SIMachineFunctionInfo &YamlMFI);
469 
470   ArrayRef<SpilledReg> getSGPRToVGPRSpills(int FrameIndex) const {
471     auto I = SGPRToVGPRSpills.find(FrameIndex);
472     return (I == SGPRToVGPRSpills.end()) ?
473       ArrayRef<SpilledReg>() : makeArrayRef(I->second);
474   }
475 
476   ArrayRef<SGPRSpillVGPRCSR> getSGPRSpillVGPRs() const {
477     return SpillVGPRs;
478   }
479 
480   AMDGPU::SIModeRegisterDefaults getMode() const {
481     return Mode;
482   }
483 
484   bool haveFreeLanesForSGPRSpill(const MachineFunction &MF,
485                                  unsigned NumLane) const;
486   bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI);
487   void removeSGPRToVGPRFrameIndices(MachineFrameInfo &MFI);
488 
489   bool hasCalculatedTID() const { return TIDReg != 0; };
490   unsigned getTIDReg() const { return TIDReg; };
491   void setTIDReg(unsigned Reg) { TIDReg = Reg; }
492 
493   unsigned getBytesInStackArgArea() const {
494     return BytesInStackArgArea;
495   }
496 
497   void setBytesInStackArgArea(unsigned Bytes) {
498     BytesInStackArgArea = Bytes;
499   }
500 
501   // Add user SGPRs.
502   unsigned addPrivateSegmentBuffer(const SIRegisterInfo &TRI);
503   unsigned addDispatchPtr(const SIRegisterInfo &TRI);
504   unsigned addQueuePtr(const SIRegisterInfo &TRI);
505   unsigned addKernargSegmentPtr(const SIRegisterInfo &TRI);
506   unsigned addDispatchID(const SIRegisterInfo &TRI);
507   unsigned addFlatScratchInit(const SIRegisterInfo &TRI);
508   unsigned addImplicitBufferPtr(const SIRegisterInfo &TRI);
509 
510   // Add system SGPRs.
511   unsigned addWorkGroupIDX() {
512     ArgInfo.WorkGroupIDX = ArgDescriptor::createRegister(getNextSystemSGPR());
513     NumSystemSGPRs += 1;
514     return ArgInfo.WorkGroupIDX.getRegister();
515   }
516 
517   unsigned addWorkGroupIDY() {
518     ArgInfo.WorkGroupIDY = ArgDescriptor::createRegister(getNextSystemSGPR());
519     NumSystemSGPRs += 1;
520     return ArgInfo.WorkGroupIDY.getRegister();
521   }
522 
523   unsigned addWorkGroupIDZ() {
524     ArgInfo.WorkGroupIDZ = ArgDescriptor::createRegister(getNextSystemSGPR());
525     NumSystemSGPRs += 1;
526     return ArgInfo.WorkGroupIDZ.getRegister();
527   }
528 
529   unsigned addWorkGroupInfo() {
530     ArgInfo.WorkGroupInfo = ArgDescriptor::createRegister(getNextSystemSGPR());
531     NumSystemSGPRs += 1;
532     return ArgInfo.WorkGroupInfo.getRegister();
533   }
534 
535   // Add special VGPR inputs
536   void setWorkItemIDX(ArgDescriptor Arg) {
537     ArgInfo.WorkItemIDX = Arg;
538   }
539 
540   void setWorkItemIDY(ArgDescriptor Arg) {
541     ArgInfo.WorkItemIDY = Arg;
542   }
543 
544   void setWorkItemIDZ(ArgDescriptor Arg) {
545     ArgInfo.WorkItemIDZ = Arg;
546   }
547 
548   unsigned addPrivateSegmentWaveByteOffset() {
549     ArgInfo.PrivateSegmentWaveByteOffset
550       = ArgDescriptor::createRegister(getNextSystemSGPR());
551     NumSystemSGPRs += 1;
552     return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
553   }
554 
555   void setPrivateSegmentWaveByteOffset(unsigned Reg) {
556     ArgInfo.PrivateSegmentWaveByteOffset = ArgDescriptor::createRegister(Reg);
557   }
558 
559   bool hasPrivateSegmentBuffer() const {
560     return PrivateSegmentBuffer;
561   }
562 
563   bool hasDispatchPtr() const {
564     return DispatchPtr;
565   }
566 
567   bool hasQueuePtr() const {
568     return QueuePtr;
569   }
570 
571   bool hasKernargSegmentPtr() const {
572     return KernargSegmentPtr;
573   }
574 
575   bool hasDispatchID() const {
576     return DispatchID;
577   }
578 
579   bool hasFlatScratchInit() const {
580     return FlatScratchInit;
581   }
582 
583   bool hasWorkGroupIDX() const {
584     return WorkGroupIDX;
585   }
586 
587   bool hasWorkGroupIDY() const {
588     return WorkGroupIDY;
589   }
590 
591   bool hasWorkGroupIDZ() const {
592     return WorkGroupIDZ;
593   }
594 
595   bool hasWorkGroupInfo() const {
596     return WorkGroupInfo;
597   }
598 
599   bool hasPrivateSegmentWaveByteOffset() const {
600     return PrivateSegmentWaveByteOffset;
601   }
602 
603   bool hasWorkItemIDX() const {
604     return WorkItemIDX;
605   }
606 
607   bool hasWorkItemIDY() const {
608     return WorkItemIDY;
609   }
610 
611   bool hasWorkItemIDZ() const {
612     return WorkItemIDZ;
613   }
614 
615   bool hasImplicitArgPtr() const {
616     return ImplicitArgPtr;
617   }
618 
619   bool hasImplicitBufferPtr() const {
620     return ImplicitBufferPtr;
621   }
622 
623   AMDGPUFunctionArgInfo &getArgInfo() {
624     return ArgInfo;
625   }
626 
627   const AMDGPUFunctionArgInfo &getArgInfo() const {
628     return ArgInfo;
629   }
630 
631   std::pair<const ArgDescriptor *, const TargetRegisterClass *>
632   getPreloadedValue(AMDGPUFunctionArgInfo::PreloadedValue Value) const {
633     return ArgInfo.getPreloadedValue(Value);
634   }
635 
636   Register getPreloadedReg(AMDGPUFunctionArgInfo::PreloadedValue Value) const {
637     auto Arg = ArgInfo.getPreloadedValue(Value).first;
638     return Arg ? Arg->getRegister() : Register();
639   }
640 
641   unsigned getGITPtrHigh() const {
642     return GITPtrHigh;
643   }
644 
645   unsigned get32BitAddressHighBits() const {
646     return HighBitsOf32BitAddress;
647   }
648 
649   unsigned getGDSSize() const {
650     return GDSSize;
651   }
652 
653   unsigned getNumUserSGPRs() const {
654     return NumUserSGPRs;
655   }
656 
657   unsigned getNumPreloadedSGPRs() const {
658     return NumUserSGPRs + NumSystemSGPRs;
659   }
660 
661   unsigned getPrivateSegmentWaveByteOffsetSystemSGPR() const {
662     return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
663   }
664 
665   /// Returns the physical register reserved for use as the resource
666   /// descriptor for scratch accesses.
667   unsigned getScratchRSrcReg() const {
668     return ScratchRSrcReg;
669   }
670 
671   void setScratchRSrcReg(unsigned Reg) {
672     assert(Reg != 0 && "Should never be unset");
673     ScratchRSrcReg = Reg;
674   }
675 
676   unsigned getScratchWaveOffsetReg() const {
677     return ScratchWaveOffsetReg;
678   }
679 
680   unsigned getFrameOffsetReg() const {
681     return FrameOffsetReg;
682   }
683 
684   void setFrameOffsetReg(unsigned Reg) {
685     assert(Reg != 0 && "Should never be unset");
686     FrameOffsetReg = Reg;
687   }
688 
689   void setStackPtrOffsetReg(unsigned Reg) {
690     assert(Reg != 0 && "Should never be unset");
691     StackPtrOffsetReg = Reg;
692   }
693 
694   // Note the unset value for this is AMDGPU::SP_REG rather than
695   // NoRegister. This is mostly a workaround for MIR tests where state that
696   // can't be directly computed from the function is not preserved in serialized
697   // MIR.
698   unsigned getStackPtrOffsetReg() const {
699     return StackPtrOffsetReg;
700   }
701 
702   void setScratchWaveOffsetReg(unsigned Reg) {
703     assert(Reg != 0 && "Should never be unset");
704     ScratchWaveOffsetReg = Reg;
705   }
706 
707   unsigned getQueuePtrUserSGPR() const {
708     return ArgInfo.QueuePtr.getRegister();
709   }
710 
711   unsigned getImplicitBufferPtrUserSGPR() const {
712     return ArgInfo.ImplicitBufferPtr.getRegister();
713   }
714 
715   bool hasSpilledSGPRs() const {
716     return HasSpilledSGPRs;
717   }
718 
719   void setHasSpilledSGPRs(bool Spill = true) {
720     HasSpilledSGPRs = Spill;
721   }
722 
723   bool hasSpilledVGPRs() const {
724     return HasSpilledVGPRs;
725   }
726 
727   void setHasSpilledVGPRs(bool Spill = true) {
728     HasSpilledVGPRs = Spill;
729   }
730 
731   bool hasNonSpillStackObjects() const {
732     return HasNonSpillStackObjects;
733   }
734 
735   void setHasNonSpillStackObjects(bool StackObject = true) {
736     HasNonSpillStackObjects = StackObject;
737   }
738 
739   bool isStackRealigned() const {
740     return IsStackRealigned;
741   }
742 
743   void setIsStackRealigned(bool Realigned = true) {
744     IsStackRealigned = Realigned;
745   }
746 
747   unsigned getNumSpilledSGPRs() const {
748     return NumSpilledSGPRs;
749   }
750 
751   unsigned getNumSpilledVGPRs() const {
752     return NumSpilledVGPRs;
753   }
754 
755   void addToSpilledSGPRs(unsigned num) {
756     NumSpilledSGPRs += num;
757   }
758 
759   void addToSpilledVGPRs(unsigned num) {
760     NumSpilledVGPRs += num;
761   }
762 
763   unsigned getPSInputAddr() const {
764     return PSInputAddr;
765   }
766 
767   unsigned getPSInputEnable() const {
768     return PSInputEnable;
769   }
770 
771   bool isPSInputAllocated(unsigned Index) const {
772     return PSInputAddr & (1 << Index);
773   }
774 
775   void markPSInputAllocated(unsigned Index) {
776     PSInputAddr |= 1 << Index;
777   }
778 
779   void markPSInputEnabled(unsigned Index) {
780     PSInputEnable |= 1 << Index;
781   }
782 
783   bool returnsVoid() const {
784     return ReturnsVoid;
785   }
786 
787   void setIfReturnsVoid(bool Value) {
788     ReturnsVoid = Value;
789   }
790 
791   /// \returns A pair of default/requested minimum/maximum flat work group sizes
792   /// for this function.
793   std::pair<unsigned, unsigned> getFlatWorkGroupSizes() const {
794     return FlatWorkGroupSizes;
795   }
796 
797   /// \returns Default/requested minimum flat work group size for this function.
798   unsigned getMinFlatWorkGroupSize() const {
799     return FlatWorkGroupSizes.first;
800   }
801 
802   /// \returns Default/requested maximum flat work group size for this function.
803   unsigned getMaxFlatWorkGroupSize() const {
804     return FlatWorkGroupSizes.second;
805   }
806 
807   /// \returns A pair of default/requested minimum/maximum number of waves per
808   /// execution unit.
809   std::pair<unsigned, unsigned> getWavesPerEU() const {
810     return WavesPerEU;
811   }
812 
813   /// \returns Default/requested minimum number of waves per execution unit.
814   unsigned getMinWavesPerEU() const {
815     return WavesPerEU.first;
816   }
817 
818   /// \returns Default/requested maximum number of waves per execution unit.
819   unsigned getMaxWavesPerEU() const {
820     return WavesPerEU.second;
821   }
822 
823   /// \returns SGPR used for \p Dim's work group ID.
824   unsigned getWorkGroupIDSGPR(unsigned Dim) const {
825     switch (Dim) {
826     case 0:
827       assert(hasWorkGroupIDX());
828       return ArgInfo.WorkGroupIDX.getRegister();
829     case 1:
830       assert(hasWorkGroupIDY());
831       return ArgInfo.WorkGroupIDY.getRegister();
832     case 2:
833       assert(hasWorkGroupIDZ());
834       return ArgInfo.WorkGroupIDZ.getRegister();
835     }
836     llvm_unreachable("unexpected dimension");
837   }
838 
839   unsigned getLDSWaveSpillSize() const {
840     return LDSWaveSpillSize;
841   }
842 
843   const AMDGPUBufferPseudoSourceValue *getBufferPSV(const SIInstrInfo &TII,
844                                                     const Value *BufferRsrc) {
845     assert(BufferRsrc);
846     auto PSV = BufferPSVs.try_emplace(
847       BufferRsrc,
848       llvm::make_unique<AMDGPUBufferPseudoSourceValue>(TII));
849     return PSV.first->second.get();
850   }
851 
852   const AMDGPUImagePseudoSourceValue *getImagePSV(const SIInstrInfo &TII,
853                                                   const Value *ImgRsrc) {
854     assert(ImgRsrc);
855     auto PSV = ImagePSVs.try_emplace(
856       ImgRsrc,
857       llvm::make_unique<AMDGPUImagePseudoSourceValue>(TII));
858     return PSV.first->second.get();
859   }
860 
861   const AMDGPUGWSResourcePseudoSourceValue *getGWSPSV(const SIInstrInfo &TII) {
862     if (!GWSResourcePSV) {
863       GWSResourcePSV =
864           llvm::make_unique<AMDGPUGWSResourcePseudoSourceValue>(TII);
865     }
866 
867     return GWSResourcePSV.get();
868   }
869 
870   unsigned getOccupancy() const {
871     return Occupancy;
872   }
873 
874   unsigned getMinAllowedOccupancy() const {
875     if (!isMemoryBound() && !needsWaveLimiter())
876       return Occupancy;
877     return (Occupancy < 4) ? Occupancy : 4;
878   }
879 
880   void limitOccupancy(const MachineFunction &MF);
881 
882   void limitOccupancy(unsigned Limit) {
883     if (Occupancy > Limit)
884       Occupancy = Limit;
885   }
886 
887   void increaseOccupancy(const MachineFunction &MF, unsigned Limit) {
888     if (Occupancy < Limit)
889       Occupancy = Limit;
890     limitOccupancy(MF);
891   }
892 };
893 
894 } // end namespace llvm
895 
896 #endif // LLVM_LIB_TARGET_AMDGPU_SIMACHINEFUNCTIONINFO_H
897