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/IR/Module.h"
23 #include "llvm/MC/MCContext.h"
24 #include "llvm/MC/MCInstrDesc.h"
25 #include "llvm/MC/MCRegisterInfo.h"
26 #include "llvm/MC/MCSectionELF.h"
27 #include "llvm/MC/MCSubtargetInfo.h"
28 #include "llvm/MC/SubtargetFeature.h"
29 #include "llvm/Support/Casting.h"
30 #include "llvm/Support/ELF.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include <algorithm>
34 #include <cassert>
35 #include <cstdint>
36 #include <cstring>
37 #include <utility>
38 
39 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
40 
41 
42 #define GET_INSTRINFO_NAMED_OPS
43 #include "AMDGPUGenInstrInfo.inc"
44 #undef GET_INSTRINFO_NAMED_OPS
45 
46 namespace {
47 
48 /// \returns Bit mask for given bit \p Shift and bit \p Width.
49 unsigned getBitMask(unsigned Shift, unsigned Width) {
50   return ((1 << Width) - 1) << Shift;
51 }
52 
53 /// \brief Packs \p Src into \p Dst for given bit \p Shift and bit \p Width.
54 ///
55 /// \returns Packed \p Dst.
56 unsigned packBits(unsigned Src, unsigned Dst, unsigned Shift, unsigned Width) {
57   Dst &= ~(1 << Shift) & ~getBitMask(Shift, Width);
58   Dst |= (Src << Shift) & getBitMask(Shift, Width);
59   return Dst;
60 }
61 
62 /// \brief Unpacks bits from \p Src for given bit \p Shift and bit \p Width.
63 ///
64 /// \returns Unpacked bits.
65 unsigned unpackBits(unsigned Src, unsigned Shift, unsigned Width) {
66   return (Src & getBitMask(Shift, Width)) >> Shift;
67 }
68 
69 /// \returns Vmcnt bit shift (lower bits).
70 unsigned getVmcntBitShiftLo() { return 0; }
71 
72 /// \returns Vmcnt bit width (lower bits).
73 unsigned getVmcntBitWidthLo() { return 4; }
74 
75 /// \returns Expcnt bit shift.
76 unsigned getExpcntBitShift() { return 4; }
77 
78 /// \returns Expcnt bit width.
79 unsigned getExpcntBitWidth() { return 3; }
80 
81 /// \returns Lgkmcnt bit shift.
82 unsigned getLgkmcntBitShift() { return 8; }
83 
84 /// \returns Lgkmcnt bit width.
85 unsigned getLgkmcntBitWidth() { return 4; }
86 
87 /// \returns Vmcnt bit shift (higher bits).
88 unsigned getVmcntBitShiftHi() { return 14; }
89 
90 /// \returns Vmcnt bit width (higher bits).
91 unsigned getVmcntBitWidthHi() { return 2; }
92 
93 } // end namespace anonymous
94 
95 namespace llvm {
96 namespace AMDGPU {
97 
98 namespace IsaInfo {
99 
100 IsaVersion getIsaVersion(const FeatureBitset &Features) {
101   // CI.
102   if (Features.test(FeatureISAVersion7_0_0))
103     return {7, 0, 0};
104   if (Features.test(FeatureISAVersion7_0_1))
105     return {7, 0, 1};
106   if (Features.test(FeatureISAVersion7_0_2))
107     return {7, 0, 2};
108 
109   // VI.
110   if (Features.test(FeatureISAVersion8_0_0))
111     return {8, 0, 0};
112   if (Features.test(FeatureISAVersion8_0_1))
113     return {8, 0, 1};
114   if (Features.test(FeatureISAVersion8_0_2))
115     return {8, 0, 2};
116   if (Features.test(FeatureISAVersion8_0_3))
117     return {8, 0, 3};
118   if (Features.test(FeatureISAVersion8_0_4))
119     return {8, 0, 4};
120   if (Features.test(FeatureISAVersion8_1_0))
121     return {8, 1, 0};
122 
123   // GFX9.
124   if (Features.test(FeatureISAVersion9_0_0))
125     return {9, 0, 0};
126   if (Features.test(FeatureISAVersion9_0_1))
127     return {9, 0, 1};
128 
129   if (!Features.test(FeatureGCN) || Features.test(FeatureSouthernIslands))
130     return {0, 0, 0};
131   return {7, 0, 0};
132 }
133 
134 unsigned getWavefrontSize(const FeatureBitset &Features) {
135   if (Features.test(FeatureWavefrontSize16))
136     return 16;
137   if (Features.test(FeatureWavefrontSize32))
138     return 32;
139 
140   return 64;
141 }
142 
143 unsigned getLocalMemorySize(const FeatureBitset &Features) {
144   if (Features.test(FeatureLocalMemorySize32768))
145     return 32768;
146   if (Features.test(FeatureLocalMemorySize65536))
147     return 65536;
148 
149   return 0;
150 }
151 
152 unsigned getEUsPerCU(const FeatureBitset &Features) {
153   return 4;
154 }
155 
156 unsigned getMaxWorkGroupsPerCU(const FeatureBitset &Features,
157                                unsigned FlatWorkGroupSize) {
158   if (!Features.test(FeatureGCN))
159     return 8;
160   unsigned N = getWavesPerWorkGroup(Features, FlatWorkGroupSize);
161   if (N == 1)
162     return 40;
163   N = 40 / N;
164   return std::min(N, 16u);
165 }
166 
167 unsigned getMaxWavesPerCU(const FeatureBitset &Features) {
168   return getMaxWavesPerEU(Features) * getEUsPerCU(Features);
169 }
170 
171 unsigned getMaxWavesPerCU(const FeatureBitset &Features,
172                           unsigned FlatWorkGroupSize) {
173   return getWavesPerWorkGroup(Features, FlatWorkGroupSize);
174 }
175 
176 unsigned getMinWavesPerEU(const FeatureBitset &Features) {
177   return 1;
178 }
179 
180 unsigned getMaxWavesPerEU(const FeatureBitset &Features) {
181   if (!Features.test(FeatureGCN))
182     return 8;
183   // FIXME: Need to take scratch memory into account.
184   return 10;
185 }
186 
187 unsigned getMaxWavesPerEU(const FeatureBitset &Features,
188                           unsigned FlatWorkGroupSize) {
189   return alignTo(getMaxWavesPerCU(Features, FlatWorkGroupSize),
190                  getEUsPerCU(Features)) / getEUsPerCU(Features);
191 }
192 
193 unsigned getMinFlatWorkGroupSize(const FeatureBitset &Features) {
194   return 1;
195 }
196 
197 unsigned getMaxFlatWorkGroupSize(const FeatureBitset &Features) {
198   return 2048;
199 }
200 
201 unsigned getWavesPerWorkGroup(const FeatureBitset &Features,
202                               unsigned FlatWorkGroupSize) {
203   return alignTo(FlatWorkGroupSize, getWavefrontSize(Features)) /
204                  getWavefrontSize(Features);
205 }
206 
207 unsigned getSGPRAllocGranule(const FeatureBitset &Features) {
208   IsaVersion Version = getIsaVersion(Features);
209   if (Version.Major >= 8)
210     return 16;
211   return 8;
212 }
213 
214 unsigned getSGPREncodingGranule(const FeatureBitset &Features) {
215   return 8;
216 }
217 
218 unsigned getTotalNumSGPRs(const FeatureBitset &Features) {
219   IsaVersion Version = getIsaVersion(Features);
220   if (Version.Major >= 8)
221     return 800;
222   return 512;
223 }
224 
225 unsigned getAddressableNumSGPRs(const FeatureBitset &Features) {
226   if (Features.test(FeatureSGPRInitBug))
227     return FIXED_NUM_SGPRS_FOR_INIT_BUG;
228 
229   IsaVersion Version = getIsaVersion(Features);
230   if (Version.Major >= 8)
231     return 102;
232   return 104;
233 }
234 
235 unsigned getMinNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
236   assert(WavesPerEU != 0);
237 
238   if (WavesPerEU >= getMaxWavesPerEU(Features))
239     return 0;
240   unsigned MinNumSGPRs =
241       alignDown(getTotalNumSGPRs(Features) / (WavesPerEU + 1),
242                 getSGPRAllocGranule(Features)) + 1;
243   return std::min(MinNumSGPRs, getAddressableNumSGPRs(Features));
244 }
245 
246 unsigned getMaxNumSGPRs(const FeatureBitset &Features, unsigned WavesPerEU,
247                         bool Addressable) {
248   assert(WavesPerEU != 0);
249 
250   IsaVersion Version = getIsaVersion(Features);
251   unsigned MaxNumSGPRs = alignDown(getTotalNumSGPRs(Features) / WavesPerEU,
252                                    getSGPRAllocGranule(Features));
253   unsigned AddressableNumSGPRs = getAddressableNumSGPRs(Features);
254   if (Version.Major >= 8 && !Addressable)
255     AddressableNumSGPRs = 112;
256   return std::min(MaxNumSGPRs, AddressableNumSGPRs);
257 }
258 
259 unsigned getVGPRAllocGranule(const FeatureBitset &Features) {
260   return 4;
261 }
262 
263 unsigned getVGPREncodingGranule(const FeatureBitset &Features) {
264   return getVGPRAllocGranule(Features);
265 }
266 
267 unsigned getTotalNumVGPRs(const FeatureBitset &Features) {
268   return 256;
269 }
270 
271 unsigned getAddressableNumVGPRs(const FeatureBitset &Features) {
272   return getTotalNumVGPRs(Features);
273 }
274 
275 unsigned getMinNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
276   assert(WavesPerEU != 0);
277 
278   if (WavesPerEU >= getMaxWavesPerEU(Features))
279     return 0;
280   unsigned MinNumVGPRs =
281       alignDown(getTotalNumVGPRs(Features) / (WavesPerEU + 1),
282                 getVGPRAllocGranule(Features)) + 1;
283   return std::min(MinNumVGPRs, getAddressableNumVGPRs(Features));
284 }
285 
286 unsigned getMaxNumVGPRs(const FeatureBitset &Features, unsigned WavesPerEU) {
287   assert(WavesPerEU != 0);
288 
289   unsigned MaxNumVGPRs = alignDown(getTotalNumVGPRs(Features) / WavesPerEU,
290                                    getVGPRAllocGranule(Features));
291   unsigned AddressableNumVGPRs = getAddressableNumVGPRs(Features);
292   return std::min(MaxNumVGPRs, AddressableNumVGPRs);
293 }
294 
295 } // end namespace IsaInfo
296 
297 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header,
298                                const FeatureBitset &Features) {
299   IsaInfo::IsaVersion ISA = IsaInfo::getIsaVersion(Features);
300 
301   memset(&Header, 0, sizeof(Header));
302 
303   Header.amd_kernel_code_version_major = 1;
304   Header.amd_kernel_code_version_minor = 1;
305   Header.amd_machine_kind = 1; // AMD_MACHINE_KIND_AMDGPU
306   Header.amd_machine_version_major = ISA.Major;
307   Header.amd_machine_version_minor = ISA.Minor;
308   Header.amd_machine_version_stepping = ISA.Stepping;
309   Header.kernel_code_entry_byte_offset = sizeof(Header);
310   // wavefront_size is specified as a power of 2: 2^6 = 64 threads.
311   Header.wavefront_size = 6;
312 
313   // If the code object does not support indirect functions, then the value must
314   // be 0xffffffff.
315   Header.call_convention = -1;
316 
317   // These alignment values are specified in powers of two, so alignment =
318   // 2^n.  The minimum alignment is 2^4 = 16.
319   Header.kernarg_segment_alignment = 4;
320   Header.group_segment_alignment = 4;
321   Header.private_segment_alignment = 4;
322 }
323 
324 MCSection *getHSATextSection(MCContext &Ctx) {
325   return Ctx.getELFSection(".hsatext", ELF::SHT_PROGBITS,
326                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
327                            ELF::SHF_EXECINSTR |
328                            ELF::SHF_AMDGPU_HSA_AGENT |
329                            ELF::SHF_AMDGPU_HSA_CODE);
330 }
331 
332 MCSection *getHSADataGlobalAgentSection(MCContext &Ctx) {
333   return Ctx.getELFSection(".hsadata_global_agent", ELF::SHT_PROGBITS,
334                            ELF::SHF_ALLOC | ELF::SHF_WRITE |
335                            ELF::SHF_AMDGPU_HSA_GLOBAL |
336                            ELF::SHF_AMDGPU_HSA_AGENT);
337 }
338 
339 MCSection *getHSADataGlobalProgramSection(MCContext &Ctx) {
340   return  Ctx.getELFSection(".hsadata_global_program", ELF::SHT_PROGBITS,
341                             ELF::SHF_ALLOC | ELF::SHF_WRITE |
342                             ELF::SHF_AMDGPU_HSA_GLOBAL);
343 }
344 
345 MCSection *getHSARodataReadonlyAgentSection(MCContext &Ctx) {
346   return Ctx.getELFSection(".hsarodata_readonly_agent", ELF::SHT_PROGBITS,
347                            ELF::SHF_ALLOC | ELF::SHF_AMDGPU_HSA_READONLY |
348                            ELF::SHF_AMDGPU_HSA_AGENT);
349 }
350 
351 bool isGroupSegment(const GlobalValue *GV, AMDGPUAS AS) {
352   return GV->getType()->getAddressSpace() == AS.LOCAL_ADDRESS;
353 }
354 
355 bool isGlobalSegment(const GlobalValue *GV, AMDGPUAS AS) {
356   return GV->getType()->getAddressSpace() == AS.GLOBAL_ADDRESS;
357 }
358 
359 bool isReadOnlySegment(const GlobalValue *GV, AMDGPUAS AS) {
360   return GV->getType()->getAddressSpace() == AS.CONSTANT_ADDRESS;
361 }
362 
363 bool shouldEmitConstantsToTextSection(const Triple &TT) {
364   return TT.getOS() != Triple::AMDHSA;
365 }
366 
367 int getIntegerAttribute(const Function &F, StringRef Name, int Default) {
368   Attribute A = F.getFnAttribute(Name);
369   int Result = Default;
370 
371   if (A.isStringAttribute()) {
372     StringRef Str = A.getValueAsString();
373     if (Str.getAsInteger(0, Result)) {
374       LLVMContext &Ctx = F.getContext();
375       Ctx.emitError("can't parse integer attribute " + Name);
376     }
377   }
378 
379   return Result;
380 }
381 
382 std::pair<int, int> getIntegerPairAttribute(const Function &F,
383                                             StringRef Name,
384                                             std::pair<int, int> Default,
385                                             bool OnlyFirstRequired) {
386   Attribute A = F.getFnAttribute(Name);
387   if (!A.isStringAttribute())
388     return Default;
389 
390   LLVMContext &Ctx = F.getContext();
391   std::pair<int, int> Ints = Default;
392   std::pair<StringRef, StringRef> Strs = A.getValueAsString().split(',');
393   if (Strs.first.trim().getAsInteger(0, Ints.first)) {
394     Ctx.emitError("can't parse first integer attribute " + Name);
395     return Default;
396   }
397   if (Strs.second.trim().getAsInteger(0, Ints.second)) {
398     if (!OnlyFirstRequired || !Strs.second.trim().empty()) {
399       Ctx.emitError("can't parse second integer attribute " + Name);
400       return Default;
401     }
402   }
403 
404   return Ints;
405 }
406 
407 unsigned getVmcntBitMask(const IsaInfo::IsaVersion &Version) {
408   unsigned VmcntLo = (1 << getVmcntBitWidthLo()) - 1;
409   if (Version.Major < 9)
410     return VmcntLo;
411 
412   unsigned VmcntHi = ((1 << getVmcntBitWidthHi()) - 1) << getVmcntBitWidthLo();
413   return VmcntLo | VmcntHi;
414 }
415 
416 unsigned getExpcntBitMask(const IsaInfo::IsaVersion &Version) {
417   return (1 << getExpcntBitWidth()) - 1;
418 }
419 
420 unsigned getLgkmcntBitMask(const IsaInfo::IsaVersion &Version) {
421   return (1 << getLgkmcntBitWidth()) - 1;
422 }
423 
424 unsigned getWaitcntBitMask(const IsaInfo::IsaVersion &Version) {
425   unsigned VmcntLo = getBitMask(getVmcntBitShiftLo(), getVmcntBitWidthLo());
426   unsigned Expcnt = getBitMask(getExpcntBitShift(), getExpcntBitWidth());
427   unsigned Lgkmcnt = getBitMask(getLgkmcntBitShift(), getLgkmcntBitWidth());
428   unsigned Waitcnt = VmcntLo | Expcnt | Lgkmcnt;
429   if (Version.Major < 9)
430     return Waitcnt;
431 
432   unsigned VmcntHi = getBitMask(getVmcntBitShiftHi(), getVmcntBitWidthHi());
433   return Waitcnt | VmcntHi;
434 }
435 
436 unsigned decodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
437   unsigned VmcntLo =
438       unpackBits(Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo());
439   if (Version.Major < 9)
440     return VmcntLo;
441 
442   unsigned VmcntHi =
443       unpackBits(Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi());
444   VmcntHi <<= getVmcntBitWidthLo();
445   return VmcntLo | VmcntHi;
446 }
447 
448 unsigned decodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
449   return unpackBits(Waitcnt, getExpcntBitShift(), getExpcntBitWidth());
450 }
451 
452 unsigned decodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt) {
453   return unpackBits(Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth());
454 }
455 
456 void decodeWaitcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
457                    unsigned &Vmcnt, unsigned &Expcnt, unsigned &Lgkmcnt) {
458   Vmcnt = decodeVmcnt(Version, Waitcnt);
459   Expcnt = decodeExpcnt(Version, Waitcnt);
460   Lgkmcnt = decodeLgkmcnt(Version, Waitcnt);
461 }
462 
463 unsigned encodeVmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
464                      unsigned Vmcnt) {
465   Waitcnt =
466       packBits(Vmcnt, Waitcnt, getVmcntBitShiftLo(), getVmcntBitWidthLo());
467   if (Version.Major < 9)
468     return Waitcnt;
469 
470   Vmcnt >>= getVmcntBitWidthLo();
471   return packBits(Vmcnt, Waitcnt, getVmcntBitShiftHi(), getVmcntBitWidthHi());
472 }
473 
474 unsigned encodeExpcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
475                       unsigned Expcnt) {
476   return packBits(Expcnt, Waitcnt, getExpcntBitShift(), getExpcntBitWidth());
477 }
478 
479 unsigned encodeLgkmcnt(const IsaInfo::IsaVersion &Version, unsigned Waitcnt,
480                        unsigned Lgkmcnt) {
481   return packBits(Lgkmcnt, Waitcnt, getLgkmcntBitShift(), getLgkmcntBitWidth());
482 }
483 
484 unsigned encodeWaitcnt(const IsaInfo::IsaVersion &Version,
485                        unsigned Vmcnt, unsigned Expcnt, unsigned Lgkmcnt) {
486   unsigned Waitcnt = getWaitcntBitMask(Version);
487   Waitcnt = encodeVmcnt(Version, Waitcnt, Vmcnt);
488   Waitcnt = encodeExpcnt(Version, Waitcnt, Expcnt);
489   Waitcnt = encodeLgkmcnt(Version, Waitcnt, Lgkmcnt);
490   return Waitcnt;
491 }
492 
493 unsigned getInitialPSInputAddr(const Function &F) {
494   return getIntegerAttribute(F, "InitialPSInputAddr", 0);
495 }
496 
497 bool isShader(CallingConv::ID cc) {
498   switch(cc) {
499     case CallingConv::AMDGPU_VS:
500     case CallingConv::AMDGPU_GS:
501     case CallingConv::AMDGPU_PS:
502     case CallingConv::AMDGPU_CS:
503       return true;
504     default:
505       return false;
506   }
507 }
508 
509 bool isCompute(CallingConv::ID cc) {
510   return !isShader(cc) || cc == CallingConv::AMDGPU_CS;
511 }
512 
513 bool isSI(const MCSubtargetInfo &STI) {
514   return STI.getFeatureBits()[AMDGPU::FeatureSouthernIslands];
515 }
516 
517 bool isCI(const MCSubtargetInfo &STI) {
518   return STI.getFeatureBits()[AMDGPU::FeatureSeaIslands];
519 }
520 
521 bool isVI(const MCSubtargetInfo &STI) {
522   return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
523 }
524 
525 unsigned getMCReg(unsigned Reg, const MCSubtargetInfo &STI) {
526 
527   switch(Reg) {
528   default: break;
529   case AMDGPU::FLAT_SCR:
530     assert(!isSI(STI));
531     return isCI(STI) ? AMDGPU::FLAT_SCR_ci : AMDGPU::FLAT_SCR_vi;
532 
533   case AMDGPU::FLAT_SCR_LO:
534     assert(!isSI(STI));
535     return isCI(STI) ? AMDGPU::FLAT_SCR_LO_ci : AMDGPU::FLAT_SCR_LO_vi;
536 
537   case AMDGPU::FLAT_SCR_HI:
538     assert(!isSI(STI));
539     return isCI(STI) ? AMDGPU::FLAT_SCR_HI_ci : AMDGPU::FLAT_SCR_HI_vi;
540   }
541   return Reg;
542 }
543 
544 unsigned mc2PseudoReg(unsigned Reg) {
545   switch (Reg) {
546   case AMDGPU::FLAT_SCR_ci:
547   case AMDGPU::FLAT_SCR_vi:
548     return FLAT_SCR;
549 
550   case AMDGPU::FLAT_SCR_LO_ci:
551   case AMDGPU::FLAT_SCR_LO_vi:
552     return AMDGPU::FLAT_SCR_LO;
553 
554   case AMDGPU::FLAT_SCR_HI_ci:
555   case AMDGPU::FLAT_SCR_HI_vi:
556     return AMDGPU::FLAT_SCR_HI;
557 
558   default:
559     return Reg;
560   }
561 }
562 
563 bool isSISrcOperand(const MCInstrDesc &Desc, unsigned OpNo) {
564   assert(OpNo < Desc.NumOperands);
565   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
566   return OpType >= AMDGPU::OPERAND_SRC_FIRST &&
567          OpType <= AMDGPU::OPERAND_SRC_LAST;
568 }
569 
570 bool isSISrcFPOperand(const MCInstrDesc &Desc, unsigned OpNo) {
571   assert(OpNo < Desc.NumOperands);
572   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
573   switch (OpType) {
574   case AMDGPU::OPERAND_REG_IMM_FP32:
575   case AMDGPU::OPERAND_REG_IMM_FP64:
576   case AMDGPU::OPERAND_REG_IMM_FP16:
577   case AMDGPU::OPERAND_REG_INLINE_C_FP32:
578   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
579   case AMDGPU::OPERAND_REG_INLINE_C_FP16:
580   case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
581     return true;
582   default:
583     return false;
584   }
585 }
586 
587 bool isSISrcInlinableOperand(const MCInstrDesc &Desc, unsigned OpNo) {
588   assert(OpNo < Desc.NumOperands);
589   unsigned OpType = Desc.OpInfo[OpNo].OperandType;
590   return OpType >= AMDGPU::OPERAND_REG_INLINE_C_FIRST &&
591          OpType <= AMDGPU::OPERAND_REG_INLINE_C_LAST;
592 }
593 
594 // Avoid using MCRegisterClass::getSize, since that function will go away
595 // (move from MC* level to Target* level). Return size in bits.
596 unsigned getRegBitWidth(unsigned RCID) {
597   switch (RCID) {
598   case AMDGPU::SGPR_32RegClassID:
599   case AMDGPU::VGPR_32RegClassID:
600   case AMDGPU::VS_32RegClassID:
601   case AMDGPU::SReg_32RegClassID:
602   case AMDGPU::SReg_32_XM0RegClassID:
603     return 32;
604   case AMDGPU::SGPR_64RegClassID:
605   case AMDGPU::VS_64RegClassID:
606   case AMDGPU::SReg_64RegClassID:
607   case AMDGPU::VReg_64RegClassID:
608     return 64;
609   case AMDGPU::VReg_96RegClassID:
610     return 96;
611   case AMDGPU::SGPR_128RegClassID:
612   case AMDGPU::SReg_128RegClassID:
613   case AMDGPU::VReg_128RegClassID:
614     return 128;
615   case AMDGPU::SReg_256RegClassID:
616   case AMDGPU::VReg_256RegClassID:
617     return 256;
618   case AMDGPU::SReg_512RegClassID:
619   case AMDGPU::VReg_512RegClassID:
620     return 512;
621   default:
622     llvm_unreachable("Unexpected register class");
623   }
624 }
625 
626 unsigned getRegBitWidth(const MCRegisterClass &RC) {
627   return getRegBitWidth(RC.getID());
628 }
629 
630 unsigned getRegOperandSize(const MCRegisterInfo *MRI, const MCInstrDesc &Desc,
631                            unsigned OpNo) {
632   assert(OpNo < Desc.NumOperands);
633   unsigned RCID = Desc.OpInfo[OpNo].RegClass;
634   return getRegBitWidth(MRI->getRegClass(RCID)) / 8;
635 }
636 
637 bool isInlinableLiteral64(int64_t Literal, bool HasInv2Pi) {
638   if (Literal >= -16 && Literal <= 64)
639     return true;
640 
641   uint64_t Val = static_cast<uint64_t>(Literal);
642   return (Val == DoubleToBits(0.0)) ||
643          (Val == DoubleToBits(1.0)) ||
644          (Val == DoubleToBits(-1.0)) ||
645          (Val == DoubleToBits(0.5)) ||
646          (Val == DoubleToBits(-0.5)) ||
647          (Val == DoubleToBits(2.0)) ||
648          (Val == DoubleToBits(-2.0)) ||
649          (Val == DoubleToBits(4.0)) ||
650          (Val == DoubleToBits(-4.0)) ||
651          (Val == 0x3fc45f306dc9c882 && HasInv2Pi);
652 }
653 
654 bool isInlinableLiteral32(int32_t Literal, bool HasInv2Pi) {
655   if (Literal >= -16 && Literal <= 64)
656     return true;
657 
658   // The actual type of the operand does not seem to matter as long
659   // as the bits match one of the inline immediate values.  For example:
660   //
661   // -nan has the hexadecimal encoding of 0xfffffffe which is -2 in decimal,
662   // so it is a legal inline immediate.
663   //
664   // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
665   // floating-point, so it is a legal inline immediate.
666 
667   uint32_t Val = static_cast<uint32_t>(Literal);
668   return (Val == FloatToBits(0.0f)) ||
669          (Val == FloatToBits(1.0f)) ||
670          (Val == FloatToBits(-1.0f)) ||
671          (Val == FloatToBits(0.5f)) ||
672          (Val == FloatToBits(-0.5f)) ||
673          (Val == FloatToBits(2.0f)) ||
674          (Val == FloatToBits(-2.0f)) ||
675          (Val == FloatToBits(4.0f)) ||
676          (Val == FloatToBits(-4.0f)) ||
677          (Val == 0x3e22f983 && HasInv2Pi);
678 }
679 
680 bool isInlinableLiteral16(int16_t Literal, bool HasInv2Pi) {
681   if (!HasInv2Pi)
682     return false;
683 
684   if (Literal >= -16 && Literal <= 64)
685     return true;
686 
687   uint16_t Val = static_cast<uint16_t>(Literal);
688   return Val == 0x3C00 || // 1.0
689          Val == 0xBC00 || // -1.0
690          Val == 0x3800 || // 0.5
691          Val == 0xB800 || // -0.5
692          Val == 0x4000 || // 2.0
693          Val == 0xC000 || // -2.0
694          Val == 0x4400 || // 4.0
695          Val == 0xC400 || // -4.0
696          Val == 0x3118;   // 1/2pi
697 }
698 
699 bool isInlinableLiteralV216(int32_t Literal, bool HasInv2Pi) {
700   assert(HasInv2Pi);
701 
702   int16_t Lo16 = static_cast<int16_t>(Literal);
703   int16_t Hi16 = static_cast<int16_t>(Literal >> 16);
704   return Lo16 == Hi16 && isInlinableLiteral16(Lo16, HasInv2Pi);
705 }
706 
707 bool isUniformMMO(const MachineMemOperand *MMO) {
708   const Value *Ptr = MMO->getValue();
709   // UndefValue means this is a load of a kernel input.  These are uniform.
710   // Sometimes LDS instructions have constant pointers.
711   // If Ptr is null, then that means this mem operand contains a
712   // PseudoSourceValue like GOT.
713   if (!Ptr || isa<UndefValue>(Ptr) || isa<Argument>(Ptr) ||
714       isa<Constant>(Ptr) || isa<GlobalValue>(Ptr))
715     return true;
716 
717   const Instruction *I = dyn_cast<Instruction>(Ptr);
718   return I && I->getMetadata("amdgpu.uniform");
719 }
720 
721 int64_t getSMRDEncodedOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) {
722   if (isSI(ST) || isCI(ST))
723     return ByteOffset >> 2;
724 
725   return ByteOffset;
726 }
727 
728 bool isLegalSMRDImmOffset(const MCSubtargetInfo &ST, int64_t ByteOffset) {
729   int64_t EncodedOffset = getSMRDEncodedOffset(ST, ByteOffset);
730   return isSI(ST) || isCI(ST) ? isUInt<8>(EncodedOffset) :
731                                 isUInt<20>(EncodedOffset);
732 }
733 } // end namespace AMDGPU
734 
735 } // end namespace llvm
736 
737 const unsigned AMDGPUAS::MAX_COMMON_ADDRESS;
738 const unsigned AMDGPUAS::GLOBAL_ADDRESS;
739 const unsigned AMDGPUAS::LOCAL_ADDRESS;
740 const unsigned AMDGPUAS::PARAM_D_ADDRESS;
741 const unsigned AMDGPUAS::PARAM_I_ADDRESS;
742 const unsigned AMDGPUAS::CONSTANT_BUFFER_0;
743 const unsigned AMDGPUAS::CONSTANT_BUFFER_1;
744 const unsigned AMDGPUAS::CONSTANT_BUFFER_2;
745 const unsigned AMDGPUAS::CONSTANT_BUFFER_3;
746 const unsigned AMDGPUAS::CONSTANT_BUFFER_4;
747 const unsigned AMDGPUAS::CONSTANT_BUFFER_5;
748 const unsigned AMDGPUAS::CONSTANT_BUFFER_6;
749 const unsigned AMDGPUAS::CONSTANT_BUFFER_7;
750 const unsigned AMDGPUAS::CONSTANT_BUFFER_8;
751 const unsigned AMDGPUAS::CONSTANT_BUFFER_9;
752 const unsigned AMDGPUAS::CONSTANT_BUFFER_10;
753 const unsigned AMDGPUAS::CONSTANT_BUFFER_11;
754 const unsigned AMDGPUAS::CONSTANT_BUFFER_12;
755 const unsigned AMDGPUAS::CONSTANT_BUFFER_13;
756 const unsigned AMDGPUAS::CONSTANT_BUFFER_14;
757 const unsigned AMDGPUAS::CONSTANT_BUFFER_15;
758 const unsigned AMDGPUAS::UNKNOWN_ADDRESS_SPACE;
759 
760 namespace llvm {
761 namespace AMDGPU {
762 
763 AMDGPUAS getAMDGPUAS(Triple T) {
764   auto Env = T.getEnvironmentName();
765   AMDGPUAS AS;
766   if (Env == "amdgiz" || Env == "amdgizcl") {
767     AS.FLAT_ADDRESS     = 0;
768     AS.PRIVATE_ADDRESS  = 5;
769     AS.REGION_ADDRESS   = 4;
770   }
771   else {
772     AS.FLAT_ADDRESS     = 4;
773     AS.PRIVATE_ADDRESS  = 0;
774     AS.REGION_ADDRESS   = 5;
775    }
776   return AS;
777 }
778 
779 AMDGPUAS getAMDGPUAS(const TargetMachine &M) {
780   return getAMDGPUAS(M.getTargetTriple());
781 }
782 
783 AMDGPUAS getAMDGPUAS(const Module &M) {
784   return getAMDGPUAS(Triple(M.getTargetTriple()));
785 }
786 } // namespace AMDGPU
787 } // namespace llvm
788