1 //===- AMDGPUBaseInfo.cpp - AMDGPU Base encoding information --------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 
10 #include "AMDGPU.h"
11 #include "AMDGPUBaseInfo.h"
12 #include "SIDefines.h"
13 #include "llvm/ADT/StringRef.h"
14 #include "llvm/ADT/Triple.h"
15 #include "llvm/CodeGen/MachineMemOperand.h"
16 #include "llvm/IR/Attributes.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/GlobalValue.h"
20 #include "llvm/IR/Instruction.h"
21 #include "llvm/IR/LLVMContext.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCInstrDesc.h"
24 #include "llvm/MC/MCRegisterInfo.h"
25 #include "llvm/MC/MCSectionELF.h"
26 #include "llvm/MC/MCSubtargetInfo.h"
27 #include "llvm/MC/SubtargetFeature.h"
28 #include "llvm/Support/Casting.h"
29 #include "llvm/Support/ELF.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/MathExtras.h"
32 #include <algorithm>
33 #include <cassert>
34 #include <cstdint>
35 #include <cstring>
36 #include <utility>
37 
38 #define GET_SUBTARGETINFO_ENUM
39 #include "AMDGPUGenSubtargetInfo.inc"
40 #undef GET_SUBTARGETINFO_ENUM
41 
42 #define GET_REGINFO_ENUM
43 #include "AMDGPUGenRegisterInfo.inc"
44 #undef GET_REGINFO_ENUM
45 
46 #define GET_INSTRINFO_NAMED_OPS
47 #define GET_INSTRINFO_ENUM
48 #include "AMDGPUGenInstrInfo.inc"
49 #undef GET_INSTRINFO_NAMED_OPS
50 #undef GET_INSTRINFO_ENUM
51 
52 namespace {
53 
54 /// \returns Bit mask for given bit \p Shift and bit \p Width.
55 unsigned getBitMask(unsigned Shift, unsigned Width) {
56   return ((1 << Width) - 1) << Shift;
57 }
58 
59 /// \brief Packs \p Src into \p Dst for given bit \p Shift and bit \p Width.
60 ///
61 /// \returns Packed \p Dst.
62 unsigned packBits(unsigned Src, unsigned Dst, unsigned Shift, unsigned Width) {
63   Dst &= ~(1 << Shift) & ~getBitMask(Shift, Width);
64   Dst |= (Src << Shift) & getBitMask(Shift, Width);
65   return Dst;
66 }
67 
68 /// \brief Unpacks bits from \p Src for given bit \p Shift and bit \p Width.
69 ///
70 /// \returns Unpacked bits.
71 unsigned unpackBits(unsigned Src, unsigned Shift, unsigned Width) {
72   return (Src & getBitMask(Shift, Width)) >> Shift;
73 }
74 
75 /// \returns Vmcnt bit shift (lower bits).
76 unsigned getVmcntBitShiftLo() { return 0; }
77 
78 /// \returns Vmcnt bit width (lower bits).
79 unsigned getVmcntBitWidthLo() { return 4; }
80 
81 /// \returns Expcnt bit shift.
82 unsigned getExpcntBitShift() { return 4; }
83 
84 /// \returns Expcnt bit width.
85 unsigned getExpcntBitWidth() { return 3; }
86 
87 /// \returns Lgkmcnt bit shift.
88 unsigned getLgkmcntBitShift() { return 8; }
89 
90 /// \returns Lgkmcnt bit width.
91 unsigned getLgkmcntBitWidth() { return 4; }
92 
93 /// \returns Vmcnt bit shift (higher bits).
94 unsigned getVmcntBitShiftHi() { return 14; }
95 
96 /// \returns Vmcnt bit width (higher bits).
97 unsigned getVmcntBitWidthHi() { return 2; }
98 
99 } // end namespace anonymous
100 
101 namespace llvm {
102 namespace AMDGPU {
103 
104 namespace IsaInfo {
105 
106 IsaVersion getIsaVersion(const FeatureBitset &Features) {
107   // CI.
108   if (Features.test(FeatureISAVersion7_0_0))
109     return {7, 0, 0};
110   if (Features.test(FeatureISAVersion7_0_1))
111     return {7, 0, 1};
112   if (Features.test(FeatureISAVersion7_0_2))
113     return {7, 0, 2};
114 
115   // VI.
116   if (Features.test(FeatureISAVersion8_0_0))
117     return {8, 0, 0};
118   if (Features.test(FeatureISAVersion8_0_1))
119     return {8, 0, 1};
120   if (Features.test(FeatureISAVersion8_0_2))
121     return {8, 0, 2};
122   if (Features.test(FeatureISAVersion8_0_3))
123     return {8, 0, 3};
124   if (Features.test(FeatureISAVersion8_0_4))
125     return {8, 0, 4};
126   if (Features.test(FeatureISAVersion8_1_0))
127     return {8, 1, 0};
128 
129   // GFX9.
130   if (Features.test(FeatureISAVersion9_0_0))
131     return {9, 0, 0};
132   if (Features.test(FeatureISAVersion9_0_1))
133     return {9, 0, 1};
134 
135   if (!Features.test(FeatureGCN) || Features.test(FeatureSouthernIslands))
136     return {0, 0, 0};
137   return {7, 0, 0};
138 }
139 
140 unsigned getWavefrontSize(const FeatureBitset &Features) {
141   if (Features.test(FeatureWavefrontSize16))
142     return 16;
143   if (Features.test(FeatureWavefrontSize32))
144     return 32;
145 
146   return 64;
147 }
148 
149 unsigned getLocalMemorySize(const FeatureBitset &Features) {
150   if (Features.test(FeatureLocalMemorySize32768))
151     return 32768;
152   if (Features.test(FeatureLocalMemorySize65536))
153     return 65536;
154 
155   return 0;
156 }
157 
158 unsigned getEUsPerCU(const FeatureBitset &Features) {
159   return 4;
160 }
161 
162 unsigned getMaxWorkGroupsPerCU(const FeatureBitset &Features,
163                                unsigned FlatWorkGroupSize) {
164   if (!Features.test(FeatureGCN))
165     return 8;
166   unsigned N = getWavesPerWorkGroup(Features, FlatWorkGroupSize);
167   if (N == 1)
168     return 40;
169   N = 40 / N;
170   return std::min(N, 16u);
171 }
172 
173 unsigned getMaxWavesPerCU(const FeatureBitset &Features) {
174   return getMaxWavesPerEU(Features) * getEUsPerCU(Features);
175 }
176 
177 unsigned getMaxWavesPerCU(const FeatureBitset &Features,
178                           unsigned FlatWorkGroupSize) {
179   return getWavesPerWorkGroup(Features, FlatWorkGroupSize);
180 }
181 
182 unsigned getMinWavesPerEU(const FeatureBitset &Features) {
183   return 1;
184 }
185 
186 unsigned getMaxWavesPerEU(const FeatureBitset &Features) {
187   if (!Features.test(FeatureGCN))
188     return 8;
189   // FIXME: Need to take scratch memory into account.
190   return 10;
191 }
192 
193 unsigned getMaxWavesPerEU(const FeatureBitset &Features,
194                           unsigned FlatWorkGroupSize) {
195   return alignTo(getMaxWavesPerCU(Features, FlatWorkGroupSize),
196                  getEUsPerCU(Features)) / getEUsPerCU(Features);
197 }
198 
199 unsigned getMinFlatWorkGroupSize(const FeatureBitset &Features) {
200   return 1;
201 }
202 
203 unsigned getMaxFlatWorkGroupSize(const FeatureBitset &Features) {
204   return 2048;
205 }
206 
207 unsigned getWavesPerWorkGroup(const FeatureBitset &Features,
208                               unsigned FlatWorkGroupSize) {
209   return alignTo(FlatWorkGroupSize, getWavefrontSize(Features)) /
210                  getWavefrontSize(Features);
211 }
212 
213 unsigned getSGPRAllocGranule(const FeatureBitset &Features) {
214   IsaVersion Version = getIsaVersion(Features);
215   if (Version.Major >= 8)
216     return 16;
217   return 8;
218 }
219 
220 unsigned getSGPREncodingGranule(const FeatureBitset &Features) {
221   return 8;
222 }
223 
224 unsigned getTotalNumSGPRs(const FeatureBitset &Features) {
225   IsaVersion Version = getIsaVersion(Features);
226   if (Version.Major >= 8)
227     return 800;
228   return 512;
229 }
230 
231 unsigned getAddressableNumSGPRs(const FeatureBitset &Features) {
232   if (Features.test(FeatureSGPRInitBug))
233     return FIXED_NUM_SGPRS_FOR_INIT_BUG;
234 
235   IsaVersion Version = getIsaVersion(Features);
236   if (Version.Major >= 8)
237     return 102;
238   return 104;
239 }
240 
241 unsigned getMinNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
242   assert(WavesPerEU != 0);
243 
244   if (WavesPerEU >= getMaxWavesPerEU(Features))
245     return 0;
246   unsigned MinNumSGPRs =
247       alignDown(getTotalNumSGPRs(Features) / (WavesPerEU + 1),
248                 getSGPRAllocGranule(Features)) + 1;
249   return std::min(MinNumSGPRs, getAddressableNumSGPRs(Features));
250 }
251 
252 unsigned getMaxNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU,
253                         bool Addressable) {
254   assert(WavesPerEU != 0);
255 
256   IsaVersion Version = getIsaVersion(Features);
257   unsigned MaxNumSGPRs = alignDown(getTotalNumSGPRs(Features) / WavesPerEU,
258                                    getSGPRAllocGranule(Features));
259   unsigned AddressableNumSGPRs = getAddressableNumSGPRs(Features);
260   if (Version.Major >= 8 && !Addressable)
261     AddressableNumSGPRs = 112;
262   return std::min(MaxNumSGPRs, AddressableNumSGPRs);
263 }
264 
265 unsigned getVGPRAllocGranule(const FeatureBitset &Features) {
266   return 4;
267 }
268 
269 unsigned getVGPREncodingGranule(const FeatureBitset &Features) {
270   return getVGPRAllocGranule(Features);
271 }
272 
273 unsigned getTotalNumVGPRs(const FeatureBitset &Features) {
274   return 256;
275 }
276 
277 unsigned getAddressableNumVGPRs(const FeatureBitset &Features) {
278   return getTotalNumVGPRs(Features);
279 }
280 
281 unsigned getMinNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
282   assert(WavesPerEU != 0);
283 
284   if (WavesPerEU >= getMaxWavesPerEU(Features))
285     return 0;
286   unsigned MinNumVGPRs =
287       alignDown(getTotalNumVGPRs(Features) / (WavesPerEU + 1),
288                 getVGPRAllocGranule(Features)) + 1;
289   return std::min(MinNumVGPRs, getAddressableNumVGPRs(Features));
290 }
291 
292 unsigned getMaxNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
293   assert(WavesPerEU != 0);
294 
295   unsigned MaxNumVGPRs = alignDown(getTotalNumVGPRs(Features) / WavesPerEU,
296                                    getVGPRAllocGranule(Features));
297   unsigned AddressableNumVGPRs = getAddressableNumVGPRs(Features);
298   return std::min(MaxNumVGPRs, AddressableNumVGPRs);
299 }
300 
301 } // end namespace IsaInfo
302 
303 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header,
304                                const FeatureBitset &Features) {
305   IsaInfo::IsaVersion ISA = IsaInfo::getIsaVersion(Features);
306 
307   memset(&Header, 0, sizeof(Header));
308 
309   Header.amd_kernel_code_version_major = 1;
310   Header.amd_kernel_code_version_minor = 1;
311   Header.amd_machine_kind = 1; // AMD_MACHINE_KIND_AMDGPU
312   Header.amd_machine_version_major = ISA.Major;
313   Header.amd_machine_version_minor = ISA.Minor;
314   Header.amd_machine_version_stepping = ISA.Stepping;
315   Header.kernel_code_entry_byte_offset = sizeof(Header);
316   // wavefront_size is specified as a power of 2: 2^6 = 64 threads.
317   Header.wavefront_size = 6;
318 
319   // If the code object does not support indirect functions, then the value must
320   // be 0xffffffff.
321   Header.call_convention = -1;
322 
323   // These alignment values are specified in powers of two, so alignment =
324   // 2^n.  The minimum alignment is 2^4 = 16.
325   Header.kernarg_segment_alignment = 4;
326   Header.group_segment_alignment = 4;
327   Header.private_segment_alignment = 4;
328 }
329 
330 MCSection *getHSATextSection(MCContext &Ctx) {
331   return Ctx.getELFSection(".hsatext", ELF::SHT_PROGBITS,
332                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
333                            ELF::SHF_EXECINSTR |
334                            ELF::SHF_AMDGPU_HSA_AGENT |
335                            ELF::SHF_AMDGPU_HSA_CODE);
336 }
337 
338 MCSection *getHSADataGlobalAgentSection(MCContext &Ctx) {
339   return Ctx.getELFSection(".hsadata_global_agent", ELF::SHT_PROGBITS,
340                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
341                            ELF::SHF_AMDGPU_HSA_GLOBAL |
342                            ELF::SHF_AMDGPU_HSA_AGENT);
343 }
344 
345 MCSection *getHSADataGlobalProgramSection(MCContext &Ctx) {
346   return  Ctx.getELFSection(".hsadata_global_program", ELF::SHT_PROGBITS,
347                             ELF::SHF_ALLOC | ELF::SHF_WRITE |
348                             ELF::SHF_AMDGPU_HSA_GLOBAL);
349 }
350 
351 MCSection *getHSARodataReadonlyAgentSection(MCContext &Ctx) {
352   return Ctx.getELFSection(".hsarodata_readonly_agent", ELF::SHT_PROGBITS,
353                            ELF::SHF_ALLOC | ELF::SHF_AMDGPU_HSA_READONLY |
354                            ELF::SHF_AMDGPU_HSA_AGENT);
355 }
356 
357 bool isGroupSegment(const GlobalValue *GV) {
358   return GV->getType()->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS;
359 }
360 
361 bool isGlobalSegment(const GlobalValue *GV) {
362   return GV->getType()->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS;
363 }
364 
365 bool isReadOnlySegment(const GlobalValue *GV) {
366   return GV->getType()->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS;
367 }
368 
369 bool shouldEmitConstantsToTextSection(const Triple &TT) {
370   return TT.getOS() != Triple::AMDHSA;
371 }
372 
373 int getIntegerAttribute(const Function &F, StringRef Name, int Default) {
374   Attribute A = F.getFnAttribute(Name);
375   int Result = Default;
376 
377   if (A.isStringAttribute()) {
378     StringRef Str = A.getValueAsString();
379     if (Str.getAsInteger(0, Result)) {
380       LLVMContext &Ctx = F.getContext();
381       Ctx.emitError("can't parse integer attribute " + Name);
382     }
383   }
384 
385   return Result;
386 }
387 
388 std::pair<int, int> getIntegerPairAttribute(const Function &F,
389                                             StringRef Name,
390                                             std::pair<int, int> Default,
391                                             bool OnlyFirstRequired) {
392   Attribute A = F.getFnAttribute(Name);
393   if (!A.isStringAttribute())
394     return Default;
395 
396   LLVMContext &Ctx = F.getContext();
397   std::pair<int, int> Ints = Default;
398   std::pair<StringRef, StringRef> Strs = A.getValueAsString().split(',');
399   if (Strs.first.trim().getAsInteger(0, Ints.first)) {
400     Ctx.emitError("can't parse first integer attribute " + Name);
401     return Default;
402   }
403   if (Strs.second.trim().getAsInteger(0, Ints.second)) {
404     if (!OnlyFirstRequired || !Strs.second.trim().empty()) {
405       Ctx.emitError("can't parse second integer attribute " + Name);
406       return Default;
407     }
408   }
409 
410   return Ints;
411 }
412 
413 unsigned getVmcntBitMask(const IsaInfo::IsaVersion &Version) {
414   unsigned VmcntLo = (1 << getVmcntBitWidthLo()) - 1;
415   if (Version.Major < 9)
416     return VmcntLo;
417 
418   unsigned VmcntHi = ((1 << getVmcntBitWidthHi()) - 1) << getVmcntBitWidthLo();
419   return VmcntLo | VmcntHi;
420 }
421 
422 unsigned getExpcntBitMask(const IsaInfo::IsaVersion &Version) {
423   return (1 << getExpcntBitWidth()) - 1;
424 }
425 
426 unsigned getLgkmcntBitMask(const IsaInfo::IsaVersion &Version) {
427   return (1 << getLgkmcntBitWidth()) - 1;
428 }
429 
430 unsigned getWaitcntBitMask(const IsaInfo::IsaVersion &Version) {
431   unsigned VmcntLo = getBitMask(getVmcntBitShiftLo(), getVmcntBitWidthLo());
432   unsigned Expcnt = getBitMask(getExpcntBitShift(), getExpcntBitWidth());
433   unsigned Lgkmcnt = getBitMask(getLgkmcntBitShift(), getLgkmcntBitWidth());
434   unsigned Waitcnt = VmcntLo | Expcnt | Lgkmcnt;
435   if (Version.Major < 9)
436     return Waitcnt;
437 
438   unsigned VmcntHi = getBitMask(getVmcntBitShiftHi(), getVmcntBitWidthHi());
439   return Waitcnt | VmcntHi;
440 }
441 
442 unsigned decodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
443   unsigned VmcntLo =
444       unpackBits(Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo());
445   if (Version.Major < 9)
446     return VmcntLo;
447 
448   unsigned VmcntHi =
449       unpackBits(Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi());
450   VmcntHi <<= getVmcntBitWidthLo();
451   return VmcntLo | VmcntHi;
452 }
453 
454 unsigned decodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
455   return unpackBits(Waitcnt, getExpcntBitShift(), getExpcntBitWidth());
456 }
457 
458 unsigned decodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
459   return unpackBits(Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth());
460 }
461 
462 void decodeWaitcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
463                    unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt) {
464   Vmcnt = decodeVmcnt(Version, Waitcnt);
465   Expcnt = decodeExpcnt(Version, Waitcnt);
466   Lgkmcnt = decodeLgkmcnt(Version, Waitcnt);
467 }
468 
469 unsigned encodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
470                      unsigned Vmcnt) {
471   Waitcnt =
472       packBits(Vmcnt, Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo());
473   if (Version.Major < 9)
474     return Waitcnt;
475 
476   Vmcnt >>= getVmcntBitWidthLo();
477   return packBits(Vmcnt, Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi());
478 }
479 
480 unsigned encodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
481                       unsigned Expcnt) {
482   return packBits(Expcnt, Waitcnt, getExpcntBitShift(), getExpcntBitWidth());
483 }
484 
485 unsigned encodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
486                        unsigned Lgkmcnt) {
487   return packBits(Lgkmcnt, Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth());
488 }
489 
490 unsigned encodeWaitcnt(const IsaInfo::IsaVersion &Version,
491                        unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt) {
492   unsigned Waitcnt = getWaitcntBitMask(Version);
493   Waitcnt = encodeVmcnt(Version, Waitcnt, Vmcnt);
494   Waitcnt = encodeExpcnt(Version, Waitcnt, Expcnt);
495   Waitcnt = encodeLgkmcnt(Version, Waitcnt, Lgkmcnt);
496   return Waitcnt;
497 }
498 
499 unsigned getInitialPSInputAddr(const Function &F) {
500   return getIntegerAttribute(F, "InitialPSInputAddr", 0);
501 }
502 
503 bool isShader(CallingConv::ID cc) {
504   switch(cc) {
505     case CallingConv::AMDGPU_VS:
506     case CallingConv::AMDGPU_GS:
507     case CallingConv::AMDGPU_PS:
508     case CallingConv::AMDGPU_CS:
509       return true;
510     default:
511       return false;
512   }
513 }
514 
515 bool isCompute(CallingConv::ID cc) {
516   return !isShader(cc) || cc == CallingConv::AMDGPU_CS;
517 }
518 
519 bool isSI(const MCSubtargetInfo &STI) {
520   return STI.getFeatureBits()[AMDGPU::FeatureSouthernIslands];
521 }
522 
523 bool isCI(const MCSubtargetInfo &STI) {
524   return STI.getFeatureBits()[AMDGPU::FeatureSeaIslands];
525 }
526 
527 bool isVI(const MCSubtargetInfo &STI) {
528   return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
529 }
530 
531 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) {
532 
533   switch(Reg) {
534   default: break;
535   case AMDGPU::FLAT_SCR:
536     assert(!isSI(STI));
537     return isCI(STI) ? AMDGPU::FLAT_SCR_ci : AMDGPU::FLAT_SCR_vi;
538 
539   case AMDGPU::FLAT_SCR_LO:
540     assert(!isSI(STI));
541     return isCI(STI) ? AMDGPU::FLAT_SCR_LO_ci : AMDGPU::FLAT_SCR_LO_vi;
542 
543   case AMDGPU::FLAT_SCR_HI:
544     assert(!isSI(STI));
545     return isCI(STI) ? AMDGPU::FLAT_SCR_HI_ci : AMDGPU::FLAT_SCR_HI_vi;
546   }
547   return Reg;
548 }
549 
550 unsigned mc2PseudoReg(unsigned Reg) {
551   switch (Reg) {
552   case AMDGPU::FLAT_SCR_ci:
553   case AMDGPU::FLAT_SCR_vi:
554     return FLAT_SCR;
555 
556   case AMDGPU::FLAT_SCR_LO_ci:
557   case AMDGPU::FLAT_SCR_LO_vi:
558     return AMDGPU::FLAT_SCR_LO;
559 
560   case AMDGPU::FLAT_SCR_HI_ci:
561   case AMDGPU::FLAT_SCR_HI_vi:
562     return AMDGPU::FLAT_SCR_HI;
563 
564   default:
565     return Reg;
566   }
567 }
568 
569 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) {
570   assert(OpNo < Desc.NumOperands);
571   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
572   return OpType >= AMDGPU::OPERAND_SRC_FIRST &&
573          OpType <= AMDGPU::OPERAND_SRC_LAST;
574 }
575 
576 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo) {
577   assert(OpNo < Desc.NumOperands);
578   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
579   switch (OpType) {
580   case AMDGPU::OPERAND_REG_IMM_FP32:
581   case AMDGPU::OPERAND_REG_IMM_FP64:
582   case AMDGPU::OPERAND_REG_IMM_FP16:
583   case AMDGPU::OPERAND_REG_INLINE_C_FP32:
584   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
585   case AMDGPU::OPERAND_REG_INLINE_C_FP16:
586   case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
587     return true;
588   default:
589     return false;
590   }
591 }
592 
593 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo) {
594   assert(OpNo < Desc.NumOperands);
595   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
596   return OpType >= AMDGPU::OPERAND_REG_INLINE_C_FIRST &&
597          OpType <= AMDGPU::OPERAND_REG_INLINE_C_LAST;
598 }
599 
600 // Avoid using MCRegisterClass::getSize, since that function will go away
601 // (move from MC* level to Target* level). Return size in bits.
602 unsigned getRegBitWidth(unsigned RCID) {
603   switch (RCID) {
604   case AMDGPU::SGPR_32RegClassID:
605   case AMDGPU::VGPR_32RegClassID:
606   case AMDGPU::VS_32RegClassID:
607   case AMDGPU::SReg_32RegClassID:
608   case AMDGPU::SReg_32_XM0RegClassID:
609     return 32;
610   case AMDGPU::SGPR_64RegClassID:
611   case AMDGPU::VS_64RegClassID:
612   case AMDGPU::SReg_64RegClassID:
613   case AMDGPU::VReg_64RegClassID:
614     return 64;
615   case AMDGPU::VReg_96RegClassID:
616     return 96;
617   case AMDGPU::SGPR_128RegClassID:
618   case AMDGPU::SReg_128RegClassID:
619   case AMDGPU::VReg_128RegClassID:
620     return 128;
621   case AMDGPU::SReg_256RegClassID:
622   case AMDGPU::VReg_256RegClassID:
623     return 256;
624   case AMDGPU::SReg_512RegClassID:
625   case AMDGPU::VReg_512RegClassID:
626     return 512;
627   default:
628     llvm_unreachable("Unexpected register class");
629   }
630 }
631 
632 unsigned getRegBitWidth(const MCRegisterClass &RC) {
633   return getRegBitWidth(RC.getID());
634 }
635 
636 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc,
637                            unsigned OpNo) {
638   assert(OpNo < Desc.NumOperands);
639   unsigned RCID = Desc.OpInfo[OpNo].RegClass;
640   return getRegBitWidth(MRI->getRegClass(RCID)) / 8;
641 }
642 
643 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi) {
644   if (Literal >= -16 && Literal <= 64)
645     return true;
646 
647   uint64_t Val = static_cast<uint64_t>(Literal);
648   return (Val == DoubleToBits(0.0)) ||
649          (Val == DoubleToBits(1.0)) ||
650          (Val == DoubleToBits(-1.0)) ||
651          (Val == DoubleToBits(0.5)) ||
652          (Val == DoubleToBits(-0.5)) ||
653          (Val == DoubleToBits(2.0)) ||
654          (Val == DoubleToBits(-2.0)) ||
655          (Val == DoubleToBits(4.0)) ||
656          (Val == DoubleToBits(-4.0)) ||
657          (Val == 0x3fc45f306dc9c882 && HasInv2Pi);
658 }
659 
660 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi) {
661   if (Literal >= -16 && Literal <= 64)
662     return true;
663 
664   // The actual type of the operand does not seem to matter as long
665   // as the bits match one of the inline immediate values.  For example:
666   //
667   // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
668   // so it is a legal inline immediate.
669   //
670   // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
671   // floating-point, so it is a legal inline immediate.
672 
673   uint32_t Val = static_cast<uint32_t>(Literal);
674   return (Val == FloatToBits(0.0f)) ||
675          (Val == FloatToBits(1.0f)) ||
676          (Val == FloatToBits(-1.0f)) ||
677          (Val == FloatToBits(0.5f)) ||
678          (Val == FloatToBits(-0.5f)) ||
679          (Val == FloatToBits(2.0f)) ||
680          (Val == FloatToBits(-2.0f)) ||
681          (Val == FloatToBits(4.0f)) ||
682          (Val == FloatToBits(-4.0f)) ||
683          (Val == 0x3e22f983 && HasInv2Pi);
684 }
685 
686 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi) {
687   if (!HasInv2Pi)
688     return false;
689 
690   if (Literal >= -16 && Literal <= 64)
691     return true;
692 
693   uint16_t Val = static_cast<uint16_t>(Literal);
694   return Val == 0x3C00 || // 1.0
695          Val == 0xBC00 || // -1.0
696          Val == 0x3800 || // 0.5
697          Val == 0xB800 || // -0.5
698          Val == 0x4000 || // 2.0
699          Val == 0xC000 || // -2.0
700          Val == 0x4400 || // 4.0
701          Val == 0xC400 || // -4.0
702          Val == 0x3118;   // 1/2pi
703 }
704 
705 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi) {
706   assert(HasInv2Pi);
707 
708   int16_t Lo16 = static_cast<int16_t>(Literal);
709   int16_t Hi16 = static_cast<int16_t>(Literal >> 16);
710   return Lo16 == Hi16 && isInlinableLiteral16(Lo16, HasInv2Pi);
711 }
712 
713 bool isUniformMMO(const MachineMemOperand *MMO) {
714   const Value *Ptr = MMO->getValue();
715   // UndefValue means this is a load of a kernel input.  These are uniform.
716   // Sometimes LDS instructions have constant pointers.
717   // If Ptr is null, then that means this mem operand contains a
718   // PseudoSourceValue like GOT.
719   if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
720       isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
721     return true;
722 
723   const Instruction *I = dyn_cast<Instruction>(Ptr);
724   return I && I->getMetadata("amdgpu.uniform");
725 }
726 
727 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) {
728   if (isSI(ST) || isCI(ST))
729     return ByteOffset >> 2;
730 
731   return ByteOffset;
732 }
733 
734 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) {
735   int64_t EncodedOffset = getSMRDEncodedOffset(ST, ByteOffset);
736   return isSI(ST) || isCI(ST) ? isUInt<8>(EncodedOffset) :
737                                 isUInt<20>(EncodedOffset);
738 }
739 
740 } // end namespace AMDGPU
741 } // end namespace llvm
742