xref: /llvm-project/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp (revision c8fbf6ffeaeb4fc9a2b5fcd6c993d64bcbb52818)
1 //===- AMDGPUDisassembler.cpp - Disassembler for AMDGPU ISA ---------------===//
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 //===----------------------------------------------------------------------===//
11 //
12 /// \file
13 ///
14 /// This file contains definition for AMDGPU ISA disassembler
15 //
16 //===----------------------------------------------------------------------===//
17 
18 // ToDo: What to do with instruction suffixes (v_mov_b32 vs v_mov_b32_e32)?
19 
20 #include "Disassembler/AMDGPUDisassembler.h"
21 #include "AMDGPU.h"
22 #include "AMDGPURegisterInfo.h"
23 #include "SIDefines.h"
24 #include "Utils/AMDGPUBaseInfo.h"
25 #include "llvm-c/Disassembler.h"
26 #include "llvm/ADT/APInt.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/Twine.h"
29 #include "llvm/BinaryFormat/ELF.h"
30 #include "llvm/MC/MCContext.h"
31 #include "llvm/MC/MCDisassembler/MCDisassembler.h"
32 #include "llvm/MC/MCExpr.h"
33 #include "llvm/MC/MCFixedLenDisassembler.h"
34 #include "llvm/MC/MCInst.h"
35 #include "llvm/MC/MCSubtargetInfo.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/TargetRegistry.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include <algorithm>
42 #include <cassert>
43 #include <cstddef>
44 #include <cstdint>
45 #include <iterator>
46 #include <tuple>
47 #include <vector>
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "amdgpu-disassembler"
52 
53 using DecodeStatus = llvm::MCDisassembler::DecodeStatus;
54 
55 inline static MCDisassembler::DecodeStatus
56 addOperand(MCInst &Inst, const MCOperand& Opnd) {
57   Inst.addOperand(Opnd);
58   return Opnd.isValid() ?
59     MCDisassembler::Success :
60     MCDisassembler::SoftFail;
61 }
62 
63 static int insertNamedMCOperand(MCInst &MI, const MCOperand &Op,
64                                 uint16_t NameIdx) {
65   int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), NameIdx);
66   if (OpIdx != -1) {
67     auto I = MI.begin();
68     std::advance(I, OpIdx);
69     MI.insert(I, Op);
70   }
71   return OpIdx;
72 }
73 
74 static DecodeStatus decodeSoppBrTarget(MCInst &Inst, unsigned Imm,
75                                        uint64_t Addr, const void *Decoder) {
76   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
77 
78   APInt SignedOffset(18, Imm * 4, true);
79   int64_t Offset = (SignedOffset.sext(64) + 4 + Addr).getSExtValue();
80 
81   if (DAsm->tryAddingSymbolicOperand(Inst, Offset, Addr, true, 2, 2))
82     return MCDisassembler::Success;
83   return addOperand(Inst, MCOperand::createImm(Imm));
84 }
85 
86 #define DECODE_OPERAND(StaticDecoderName, DecoderName) \
87 static DecodeStatus StaticDecoderName(MCInst &Inst, \
88                                        unsigned Imm, \
89                                        uint64_t /*Addr*/, \
90                                        const void *Decoder) { \
91   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); \
92   return addOperand(Inst, DAsm->DecoderName(Imm)); \
93 }
94 
95 #define DECODE_OPERAND_REG(RegClass) \
96 DECODE_OPERAND(Decode##RegClass##RegisterClass, decodeOperand_##RegClass)
97 
98 DECODE_OPERAND_REG(VGPR_32)
99 DECODE_OPERAND_REG(VS_32)
100 DECODE_OPERAND_REG(VS_64)
101 DECODE_OPERAND_REG(VS_128)
102 
103 DECODE_OPERAND_REG(VReg_64)
104 DECODE_OPERAND_REG(VReg_96)
105 DECODE_OPERAND_REG(VReg_128)
106 
107 DECODE_OPERAND_REG(SReg_32)
108 DECODE_OPERAND_REG(SReg_32_XM0_XEXEC)
109 DECODE_OPERAND_REG(SReg_32_XEXEC_HI)
110 DECODE_OPERAND_REG(SReg_64)
111 DECODE_OPERAND_REG(SReg_64_XEXEC)
112 DECODE_OPERAND_REG(SReg_128)
113 DECODE_OPERAND_REG(SReg_256)
114 DECODE_OPERAND_REG(SReg_512)
115 
116 static DecodeStatus decodeOperand_VSrc16(MCInst &Inst,
117                                          unsigned Imm,
118                                          uint64_t Addr,
119                                          const void *Decoder) {
120   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
121   return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
122 }
123 
124 static DecodeStatus decodeOperand_VSrcV216(MCInst &Inst,
125                                          unsigned Imm,
126                                          uint64_t Addr,
127                                          const void *Decoder) {
128   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
129   return addOperand(Inst, DAsm->decodeOperand_VSrcV216(Imm));
130 }
131 
132 #define DECODE_SDWA(DecName) \
133 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName)
134 
135 DECODE_SDWA(Src32)
136 DECODE_SDWA(Src16)
137 DECODE_SDWA(VopcDst)
138 
139 #include "AMDGPUGenDisassemblerTables.inc"
140 
141 //===----------------------------------------------------------------------===//
142 //
143 //===----------------------------------------------------------------------===//
144 
145 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
146   assert(Bytes.size() >= sizeof(T));
147   const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data());
148   Bytes = Bytes.slice(sizeof(T));
149   return Res;
150 }
151 
152 DecodeStatus AMDGPUDisassembler::tryDecodeInst(const uint8_t* Table,
153                                                MCInst &MI,
154                                                uint64_t Inst,
155                                                uint64_t Address) const {
156   assert(MI.getOpcode() == 0);
157   assert(MI.getNumOperands() == 0);
158   MCInst TmpInst;
159   HasLiteral = false;
160   const auto SavedBytes = Bytes;
161   if (decodeInstruction(Table, TmpInst, Inst, Address, this, STI)) {
162     MI = TmpInst;
163     return MCDisassembler::Success;
164   }
165   Bytes = SavedBytes;
166   return MCDisassembler::Fail;
167 }
168 
169 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
170                                                 ArrayRef<uint8_t> Bytes_,
171                                                 uint64_t Address,
172                                                 raw_ostream &WS,
173                                                 raw_ostream &CS) const {
174   CommentStream = &CS;
175   bool IsSDWA = false;
176 
177   // ToDo: AMDGPUDisassembler supports only VI ISA.
178   if (!STI.getFeatureBits()[AMDGPU::FeatureGCN3Encoding])
179     report_fatal_error("Disassembly not yet supported for subtarget");
180 
181   const unsigned MaxInstBytesNum = (std::min)((size_t)8, Bytes_.size());
182   Bytes = Bytes_.slice(0, MaxInstBytesNum);
183 
184   DecodeStatus Res = MCDisassembler::Fail;
185   do {
186     // ToDo: better to switch encoding length using some bit predicate
187     // but it is unknown yet, so try all we can
188 
189     // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2
190     // encodings
191     if (Bytes.size() >= 8) {
192       const uint64_t QW = eatBytes<uint64_t>(Bytes);
193       Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address);
194       if (Res) break;
195 
196       Res = tryDecodeInst(DecoderTableSDWA64, MI, QW, Address);
197       if (Res) { IsSDWA = true;  break; }
198 
199       Res = tryDecodeInst(DecoderTableSDWA964, MI, QW, Address);
200       if (Res) { IsSDWA = true;  break; }
201     }
202 
203     // Reinitialize Bytes as DPP64 could have eaten too much
204     Bytes = Bytes_.slice(0, MaxInstBytesNum);
205 
206     // Try decode 32-bit instruction
207     if (Bytes.size() < 4) break;
208     const uint32_t DW = eatBytes<uint32_t>(Bytes);
209     Res = tryDecodeInst(DecoderTableVI32, MI, DW, Address);
210     if (Res) break;
211 
212     Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address);
213     if (Res) break;
214 
215     if (Bytes.size() < 4) break;
216     const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW;
217     Res = tryDecodeInst(DecoderTableVI64, MI, QW, Address);
218     if (Res) break;
219 
220     Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address);
221     if (Res) break;
222 
223     Res = tryDecodeInst(DecoderTableGFX964, MI, QW, Address);
224   } while (false);
225 
226   if (Res && (MI.getOpcode() == AMDGPU::V_MAC_F32_e64_vi ||
227               MI.getOpcode() == AMDGPU::V_MAC_F32_e64_si ||
228               MI.getOpcode() == AMDGPU::V_MAC_F16_e64_vi)) {
229     // Insert dummy unused src2_modifiers.
230     insertNamedMCOperand(MI, MCOperand::createImm(0),
231                          AMDGPU::OpName::src2_modifiers);
232   }
233 
234   if (Res && IsSDWA)
235     Res = convertSDWAInst(MI);
236 
237   Size = Res ? (MaxInstBytesNum - Bytes.size()) : 0;
238   return Res;
239 }
240 
241 DecodeStatus AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const {
242   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9]) {
243     if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst) != -1)
244       // VOPC - insert clamp
245       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp);
246   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
247     int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst);
248     if (SDst != -1) {
249       // VOPC - insert VCC register as sdst
250       insertNamedMCOperand(MI, MCOperand::createReg(AMDGPU::VCC),
251                            AMDGPU::OpName::sdst);
252     } else {
253       // VOP1/2 - insert omod if present in instruction
254       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod);
255     }
256   }
257   return MCDisassembler::Success;
258 }
259 
260 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
261   return getContext().getRegisterInfo()->
262     getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
263 }
264 
265 inline
266 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
267                                          const Twine& ErrMsg) const {
268   *CommentStream << "Error: " + ErrMsg;
269 
270   // ToDo: add support for error operands to MCInst.h
271   // return MCOperand::createError(V);
272   return MCOperand();
273 }
274 
275 inline
276 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
277   return MCOperand::createReg(RegId);
278 }
279 
280 inline
281 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
282                                                unsigned Val) const {
283   const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
284   if (Val >= RegCl.getNumRegs())
285     return errOperand(Val, Twine(getRegClassName(RegClassID)) +
286                            ": unknown register " + Twine(Val));
287   return createRegOperand(RegCl.getRegister(Val));
288 }
289 
290 inline
291 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
292                                                 unsigned Val) const {
293   // ToDo: SI/CI have 104 SGPRs, VI - 102
294   // Valery: here we accepting as much as we can, let assembler sort it out
295   int shift = 0;
296   switch (SRegClassID) {
297   case AMDGPU::SGPR_32RegClassID:
298   case AMDGPU::TTMP_32RegClassID:
299     break;
300   case AMDGPU::SGPR_64RegClassID:
301   case AMDGPU::TTMP_64RegClassID:
302     shift = 1;
303     break;
304   case AMDGPU::SGPR_128RegClassID:
305   case AMDGPU::TTMP_128RegClassID:
306   // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
307   // this bundle?
308   case AMDGPU::SReg_256RegClassID:
309   // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
310   // this bundle?
311   case AMDGPU::SReg_512RegClassID:
312     shift = 2;
313     break;
314   // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
315   // this bundle?
316   default:
317     llvm_unreachable("unhandled register class");
318   }
319 
320   if (Val % (1 << shift)) {
321     *CommentStream << "Warning: " << getRegClassName(SRegClassID)
322                    << ": scalar reg isn't aligned " << Val;
323   }
324 
325   return createRegOperand(SRegClassID, Val >> shift);
326 }
327 
328 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const {
329   return decodeSrcOp(OPW32, Val);
330 }
331 
332 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const {
333   return decodeSrcOp(OPW64, Val);
334 }
335 
336 MCOperand AMDGPUDisassembler::decodeOperand_VS_128(unsigned Val) const {
337   return decodeSrcOp(OPW128, Val);
338 }
339 
340 MCOperand AMDGPUDisassembler::decodeOperand_VSrc16(unsigned Val) const {
341   return decodeSrcOp(OPW16, Val);
342 }
343 
344 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV216(unsigned Val) const {
345   return decodeSrcOp(OPWV216, Val);
346 }
347 
348 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const {
349   // Some instructions have operand restrictions beyond what the encoding
350   // allows. Some ordinarily VSrc_32 operands are VGPR_32, so clear the extra
351   // high bit.
352   Val &= 255;
353 
354   return createRegOperand(AMDGPU::VGPR_32RegClassID, Val);
355 }
356 
357 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const {
358   return createRegOperand(AMDGPU::VReg_64RegClassID, Val);
359 }
360 
361 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const {
362   return createRegOperand(AMDGPU::VReg_96RegClassID, Val);
363 }
364 
365 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const {
366   return createRegOperand(AMDGPU::VReg_128RegClassID, Val);
367 }
368 
369 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const {
370   // table-gen generated disassembler doesn't care about operand types
371   // leaving only registry class so SSrc_32 operand turns into SReg_32
372   // and therefore we accept immediates and literals here as well
373   return decodeSrcOp(OPW32, Val);
374 }
375 
376 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XM0_XEXEC(
377   unsigned Val) const {
378   // SReg_32_XM0 is SReg_32 without M0 or EXEC_LO/EXEC_HI
379   return decodeOperand_SReg_32(Val);
380 }
381 
382 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XEXEC_HI(
383   unsigned Val) const {
384   // SReg_32_XM0 is SReg_32 without EXEC_HI
385   return decodeOperand_SReg_32(Val);
386 }
387 
388 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const {
389   return decodeSrcOp(OPW64, Val);
390 }
391 
392 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64_XEXEC(unsigned Val) const {
393   return decodeSrcOp(OPW64, Val);
394 }
395 
396 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const {
397   return decodeSrcOp(OPW128, Val);
398 }
399 
400 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const {
401   return createSRegOperand(AMDGPU::SReg_256RegClassID, Val);
402 }
403 
404 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const {
405   return createSRegOperand(AMDGPU::SReg_512RegClassID, Val);
406 }
407 
408 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
409   // For now all literal constants are supposed to be unsigned integer
410   // ToDo: deal with signed/unsigned 64-bit integer constants
411   // ToDo: deal with float/double constants
412   if (!HasLiteral) {
413     if (Bytes.size() < 4) {
414       return errOperand(0, "cannot read literal, inst bytes left " +
415                         Twine(Bytes.size()));
416     }
417     HasLiteral = true;
418     Literal = eatBytes<uint32_t>(Bytes);
419   }
420   return MCOperand::createImm(Literal);
421 }
422 
423 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
424   using namespace AMDGPU::EncValues;
425 
426   assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX);
427   return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ?
428     (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) :
429     (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm)));
430       // Cast prevents negative overflow.
431 }
432 
433 static int64_t getInlineImmVal32(unsigned Imm) {
434   switch (Imm) {
435   case 240:
436     return FloatToBits(0.5f);
437   case 241:
438     return FloatToBits(-0.5f);
439   case 242:
440     return FloatToBits(1.0f);
441   case 243:
442     return FloatToBits(-1.0f);
443   case 244:
444     return FloatToBits(2.0f);
445   case 245:
446     return FloatToBits(-2.0f);
447   case 246:
448     return FloatToBits(4.0f);
449   case 247:
450     return FloatToBits(-4.0f);
451   case 248: // 1 / (2 * PI)
452     return 0x3e22f983;
453   default:
454     llvm_unreachable("invalid fp inline imm");
455   }
456 }
457 
458 static int64_t getInlineImmVal64(unsigned Imm) {
459   switch (Imm) {
460   case 240:
461     return DoubleToBits(0.5);
462   case 241:
463     return DoubleToBits(-0.5);
464   case 242:
465     return DoubleToBits(1.0);
466   case 243:
467     return DoubleToBits(-1.0);
468   case 244:
469     return DoubleToBits(2.0);
470   case 245:
471     return DoubleToBits(-2.0);
472   case 246:
473     return DoubleToBits(4.0);
474   case 247:
475     return DoubleToBits(-4.0);
476   case 248: // 1 / (2 * PI)
477     return 0x3fc45f306dc9c882;
478   default:
479     llvm_unreachable("invalid fp inline imm");
480   }
481 }
482 
483 static int64_t getInlineImmVal16(unsigned Imm) {
484   switch (Imm) {
485   case 240:
486     return 0x3800;
487   case 241:
488     return 0xB800;
489   case 242:
490     return 0x3C00;
491   case 243:
492     return 0xBC00;
493   case 244:
494     return 0x4000;
495   case 245:
496     return 0xC000;
497   case 246:
498     return 0x4400;
499   case 247:
500     return 0xC400;
501   case 248: // 1 / (2 * PI)
502     return 0x3118;
503   default:
504     llvm_unreachable("invalid fp inline imm");
505   }
506 }
507 
508 MCOperand AMDGPUDisassembler::decodeFPImmed(OpWidthTy Width, unsigned Imm) {
509   assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN
510       && Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX);
511 
512   // ToDo: case 248: 1/(2*PI) - is allowed only on VI
513   switch (Width) {
514   case OPW32:
515     return MCOperand::createImm(getInlineImmVal32(Imm));
516   case OPW64:
517     return MCOperand::createImm(getInlineImmVal64(Imm));
518   case OPW16:
519   case OPWV216:
520     return MCOperand::createImm(getInlineImmVal16(Imm));
521   default:
522     llvm_unreachable("implement me");
523   }
524 }
525 
526 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const {
527   using namespace AMDGPU;
528 
529   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
530   switch (Width) {
531   default: // fall
532   case OPW32:
533   case OPW16:
534   case OPWV216:
535     return VGPR_32RegClassID;
536   case OPW64: return VReg_64RegClassID;
537   case OPW128: return VReg_128RegClassID;
538   }
539 }
540 
541 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const {
542   using namespace AMDGPU;
543 
544   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
545   switch (Width) {
546   default: // fall
547   case OPW32:
548   case OPW16:
549   case OPWV216:
550     return SGPR_32RegClassID;
551   case OPW64: return SGPR_64RegClassID;
552   case OPW128: return SGPR_128RegClassID;
553   }
554 }
555 
556 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const {
557   using namespace AMDGPU;
558 
559   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
560   switch (Width) {
561   default: // fall
562   case OPW32:
563   case OPW16:
564   case OPWV216:
565     return TTMP_32RegClassID;
566   case OPW64: return TTMP_64RegClassID;
567   case OPW128: return TTMP_128RegClassID;
568   }
569 }
570 
571 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val) const {
572   using namespace AMDGPU::EncValues;
573 
574   assert(Val < 512); // enum9
575 
576   if (VGPR_MIN <= Val && Val <= VGPR_MAX) {
577     return createRegOperand(getVgprClassId(Width), Val - VGPR_MIN);
578   }
579   if (Val <= SGPR_MAX) {
580     assert(SGPR_MIN == 0); // "SGPR_MIN <= Val" is always true and causes compilation warning.
581     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
582   }
583   if (TTMP_MIN <= Val && Val <= TTMP_MAX) {
584     return createSRegOperand(getTtmpClassId(Width), Val - TTMP_MIN);
585   }
586 
587   if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX)
588     return decodeIntImmed(Val);
589 
590   if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX)
591     return decodeFPImmed(Width, Val);
592 
593   if (Val == LITERAL_CONST)
594     return decodeLiteralConstant();
595 
596   switch (Width) {
597   case OPW32:
598   case OPW16:
599   case OPWV216:
600     return decodeSpecialReg32(Val);
601   case OPW64:
602     return decodeSpecialReg64(Val);
603   default:
604     llvm_unreachable("unexpected immediate type");
605   }
606 }
607 
608 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
609   using namespace AMDGPU;
610 
611   switch (Val) {
612   case 102: return createRegOperand(getMCReg(FLAT_SCR_LO, STI));
613   case 103: return createRegOperand(getMCReg(FLAT_SCR_HI, STI));
614     // ToDo: no support for xnack_mask_lo/_hi register
615   case 104:
616   case 105: break;
617   case 106: return createRegOperand(VCC_LO);
618   case 107: return createRegOperand(VCC_HI);
619   case 108: return createRegOperand(TBA_LO);
620   case 109: return createRegOperand(TBA_HI);
621   case 110: return createRegOperand(TMA_LO);
622   case 111: return createRegOperand(TMA_HI);
623   case 124: return createRegOperand(M0);
624   case 126: return createRegOperand(EXEC_LO);
625   case 127: return createRegOperand(EXEC_HI);
626   case 235: return createRegOperand(SRC_SHARED_BASE);
627   case 236: return createRegOperand(SRC_SHARED_LIMIT);
628   case 237: return createRegOperand(SRC_PRIVATE_BASE);
629   case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
630     // TODO: SRC_POPS_EXITING_WAVE_ID
631     // ToDo: no support for vccz register
632   case 251: break;
633     // ToDo: no support for execz register
634   case 252: break;
635   case 253: return createRegOperand(SCC);
636   default: break;
637   }
638   return errOperand(Val, "unknown operand encoding " + Twine(Val));
639 }
640 
641 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
642   using namespace AMDGPU;
643 
644   switch (Val) {
645   case 102: return createRegOperand(getMCReg(FLAT_SCR, STI));
646   case 106: return createRegOperand(VCC);
647   case 108: return createRegOperand(TBA);
648   case 110: return createRegOperand(TMA);
649   case 126: return createRegOperand(EXEC);
650   default: break;
651   }
652   return errOperand(Val, "unknown operand encoding " + Twine(Val));
653 }
654 
655 MCOperand AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width,
656                                             unsigned Val) const {
657   using namespace AMDGPU::SDWA;
658 
659   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9]) {
660     // XXX: static_cast<int> is needed to avoid stupid warning:
661     // compare with unsigned is always true
662     if (SDWA9EncValues::SRC_VGPR_MIN <= static_cast<int>(Val) &&
663         Val <= SDWA9EncValues::SRC_VGPR_MAX) {
664       return createRegOperand(getVgprClassId(Width),
665                               Val - SDWA9EncValues::SRC_VGPR_MIN);
666     }
667     if (SDWA9EncValues::SRC_SGPR_MIN <= Val &&
668         Val <= SDWA9EncValues::SRC_SGPR_MAX) {
669       return createSRegOperand(getSgprClassId(Width),
670                                Val - SDWA9EncValues::SRC_SGPR_MIN);
671     }
672 
673     return decodeSpecialReg32(Val - SDWA9EncValues::SRC_SGPR_MIN);
674   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
675     return createRegOperand(getVgprClassId(Width), Val);
676   }
677   llvm_unreachable("unsupported target");
678 }
679 
680 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const {
681   return decodeSDWASrc(OPW16, Val);
682 }
683 
684 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const {
685   return decodeSDWASrc(OPW32, Val);
686 }
687 
688 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const {
689   using namespace AMDGPU::SDWA;
690 
691   assert(STI.getFeatureBits()[AMDGPU::FeatureGFX9] &&
692          "SDWAVopcDst should be present only on GFX9");
693   if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) {
694     Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
695     if (Val > AMDGPU::EncValues::SGPR_MAX) {
696       return decodeSpecialReg64(Val);
697     } else {
698       return createSRegOperand(getSgprClassId(OPW64), Val);
699     }
700   } else {
701     return createRegOperand(AMDGPU::VCC);
702   }
703 }
704 
705 //===----------------------------------------------------------------------===//
706 // AMDGPUSymbolizer
707 //===----------------------------------------------------------------------===//
708 
709 // Try to find symbol name for specified label
710 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(MCInst &Inst,
711                                 raw_ostream &/*cStream*/, int64_t Value,
712                                 uint64_t /*Address*/, bool IsBranch,
713                                 uint64_t /*Offset*/, uint64_t /*InstSize*/) {
714   using SymbolInfoTy = std::tuple<uint64_t, StringRef, uint8_t>;
715   using SectionSymbolsTy = std::vector<SymbolInfoTy>;
716 
717   if (!IsBranch) {
718     return false;
719   }
720 
721   auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
722   auto Result = std::find_if(Symbols->begin(), Symbols->end(),
723                              [Value](const SymbolInfoTy& Val) {
724                                 return std::get<0>(Val) == static_cast<uint64_t>(Value)
725                                     && std::get<2>(Val) == ELF::STT_NOTYPE;
726                              });
727   if (Result != Symbols->end()) {
728     auto *Sym = Ctx.getOrCreateSymbol(std::get<1>(*Result));
729     const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
730     Inst.addOperand(MCOperand::createExpr(Add));
731     return true;
732   }
733   return false;
734 }
735 
736 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
737                                                        int64_t Value,
738                                                        uint64_t Address) {
739   llvm_unreachable("unimplemented");
740 }
741 
742 //===----------------------------------------------------------------------===//
743 // Initialization
744 //===----------------------------------------------------------------------===//
745 
746 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
747                               LLVMOpInfoCallback /*GetOpInfo*/,
748                               LLVMSymbolLookupCallback /*SymbolLookUp*/,
749                               void *DisInfo,
750                               MCContext *Ctx,
751                               std::unique_ptr<MCRelocationInfo> &&RelInfo) {
752   return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
753 }
754 
755 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
756                                                 const MCSubtargetInfo &STI,
757                                                 MCContext &Ctx) {
758   return new AMDGPUDisassembler(STI, Ctx);
759 }
760 
761 extern "C" void LLVMInitializeAMDGPUDisassembler() {
762   TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
763                                          createAMDGPUDisassembler);
764   TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
765                                        createAMDGPUSymbolizer);
766 }
767