xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 //===- AMDGPUDisassembler.cpp - Disassembler for AMDGPU ISA ---------------===//
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 //===----------------------------------------------------------------------===//
10 //
11 /// \file
12 ///
13 /// This file contains definition for AMDGPU ISA disassembler
14 //
15 //===----------------------------------------------------------------------===//
16 
17 // ToDo: What to do with instruction suffixes (v_mov_b32 vs v_mov_b32_e32)?
18 
19 #include "Disassembler/AMDGPUDisassembler.h"
20 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
21 #include "SIDefines.h"
22 #include "SIRegisterInfo.h"
23 #include "TargetInfo/AMDGPUTargetInfo.h"
24 #include "Utils/AMDGPUBaseInfo.h"
25 #include "llvm-c/DisassemblerTypes.h"
26 #include "llvm/BinaryFormat/ELF.h"
27 #include "llvm/MC/MCAsmInfo.h"
28 #include "llvm/MC/MCContext.h"
29 #include "llvm/MC/MCDecoderOps.h"
30 #include "llvm/MC/MCExpr.h"
31 #include "llvm/MC/MCInstrDesc.h"
32 #include "llvm/MC/MCRegisterInfo.h"
33 #include "llvm/MC/MCSubtargetInfo.h"
34 #include "llvm/MC/TargetRegistry.h"
35 #include "llvm/Support/AMDHSAKernelDescriptor.h"
36 
37 using namespace llvm;
38 
39 #define DEBUG_TYPE "amdgpu-disassembler"
40 
41 #define SGPR_MAX                                                               \
42   (isGFX10Plus() ? AMDGPU::EncValues::SGPR_MAX_GFX10                           \
43                  : AMDGPU::EncValues::SGPR_MAX_SI)
44 
45 using DecodeStatus = llvm::MCDisassembler::DecodeStatus;
46 
47 AMDGPUDisassembler::AMDGPUDisassembler(const MCSubtargetInfo &STI,
48                                        MCContext &Ctx,
49                                        MCInstrInfo const *MCII) :
50   MCDisassembler(STI, Ctx), MCII(MCII), MRI(*Ctx.getRegisterInfo()),
51   TargetMaxInstBytes(Ctx.getAsmInfo()->getMaxInstLength(&STI)) {
52 
53   // ToDo: AMDGPUDisassembler supports only VI ISA.
54   if (!STI.getFeatureBits()[AMDGPU::FeatureGCN3Encoding] && !isGFX10Plus())
55     report_fatal_error("Disassembly not yet supported for subtarget");
56 }
57 
58 inline static MCDisassembler::DecodeStatus
59 addOperand(MCInst &Inst, const MCOperand& Opnd) {
60   Inst.addOperand(Opnd);
61   return Opnd.isValid() ?
62     MCDisassembler::Success :
63     MCDisassembler::Fail;
64 }
65 
66 static int insertNamedMCOperand(MCInst &MI, const MCOperand &Op,
67                                 uint16_t NameIdx) {
68   int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), NameIdx);
69   if (OpIdx != -1) {
70     auto I = MI.begin();
71     std::advance(I, OpIdx);
72     MI.insert(I, Op);
73   }
74   return OpIdx;
75 }
76 
77 static DecodeStatus decodeSoppBrTarget(MCInst &Inst, unsigned Imm,
78                                        uint64_t Addr,
79                                        const MCDisassembler *Decoder) {
80   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
81 
82   // Our branches take a simm16, but we need two extra bits to account for the
83   // factor of 4.
84   APInt SignedOffset(18, Imm * 4, true);
85   int64_t Offset = (SignedOffset.sext(64) + 4 + Addr).getSExtValue();
86 
87   if (DAsm->tryAddingSymbolicOperand(Inst, Offset, Addr, true, 2, 2, 0))
88     return MCDisassembler::Success;
89   return addOperand(Inst, MCOperand::createImm(Imm));
90 }
91 
92 static DecodeStatus decodeSMEMOffset(MCInst &Inst, unsigned Imm, uint64_t Addr,
93                                      const MCDisassembler *Decoder) {
94   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
95   int64_t Offset;
96   if (DAsm->isVI()) {         // VI supports 20-bit unsigned offsets.
97     Offset = Imm & 0xFFFFF;
98   } else {                    // GFX9+ supports 21-bit signed offsets.
99     Offset = SignExtend64<21>(Imm);
100   }
101   return addOperand(Inst, MCOperand::createImm(Offset));
102 }
103 
104 static DecodeStatus decodeBoolReg(MCInst &Inst, unsigned Val, uint64_t Addr,
105                                   const MCDisassembler *Decoder) {
106   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
107   return addOperand(Inst, DAsm->decodeBoolReg(Val));
108 }
109 
110 #define DECODE_OPERAND(StaticDecoderName, DecoderName)                         \
111   static DecodeStatus StaticDecoderName(MCInst &Inst, unsigned Imm,            \
112                                         uint64_t /*Addr*/,                     \
113                                         const MCDisassembler *Decoder) {       \
114     auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);              \
115     return addOperand(Inst, DAsm->DecoderName(Imm));                           \
116   }
117 
118 #define DECODE_OPERAND_REG(RegClass) \
119 DECODE_OPERAND(Decode##RegClass##RegisterClass, decodeOperand_##RegClass)
120 
121 DECODE_OPERAND_REG(VGPR_32)
122 DECODE_OPERAND_REG(VGPR_32_Lo128)
123 DECODE_OPERAND_REG(VRegOrLds_32)
124 DECODE_OPERAND_REG(VS_32)
125 DECODE_OPERAND_REG(VS_64)
126 DECODE_OPERAND_REG(VS_128)
127 
128 DECODE_OPERAND_REG(VReg_64)
129 DECODE_OPERAND_REG(VReg_96)
130 DECODE_OPERAND_REG(VReg_128)
131 DECODE_OPERAND_REG(VReg_256)
132 DECODE_OPERAND_REG(VReg_288)
133 DECODE_OPERAND_REG(VReg_352)
134 DECODE_OPERAND_REG(VReg_384)
135 DECODE_OPERAND_REG(VReg_512)
136 DECODE_OPERAND_REG(VReg_1024)
137 
138 DECODE_OPERAND_REG(SReg_32)
139 DECODE_OPERAND_REG(SReg_32_XM0_XEXEC)
140 DECODE_OPERAND_REG(SReg_32_XEXEC_HI)
141 DECODE_OPERAND_REG(SRegOrLds_32)
142 DECODE_OPERAND_REG(SReg_64)
143 DECODE_OPERAND_REG(SReg_64_XEXEC)
144 DECODE_OPERAND_REG(SReg_128)
145 DECODE_OPERAND_REG(SReg_256)
146 DECODE_OPERAND_REG(SReg_512)
147 
148 DECODE_OPERAND_REG(AGPR_32)
149 DECODE_OPERAND_REG(AReg_64)
150 DECODE_OPERAND_REG(AReg_128)
151 DECODE_OPERAND_REG(AReg_256)
152 DECODE_OPERAND_REG(AReg_512)
153 DECODE_OPERAND_REG(AReg_1024)
154 DECODE_OPERAND_REG(AV_32)
155 DECODE_OPERAND_REG(AV_64)
156 DECODE_OPERAND_REG(AV_128)
157 DECODE_OPERAND_REG(AVDst_128)
158 DECODE_OPERAND_REG(AVDst_512)
159 
160 static DecodeStatus decodeOperand_VSrc16(MCInst &Inst, unsigned Imm,
161                                          uint64_t Addr,
162                                          const MCDisassembler *Decoder) {
163   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
164   return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
165 }
166 
167 static DecodeStatus decodeOperand_VSrcV216(MCInst &Inst, unsigned Imm,
168                                            uint64_t Addr,
169                                            const MCDisassembler *Decoder) {
170   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
171   return addOperand(Inst, DAsm->decodeOperand_VSrcV216(Imm));
172 }
173 
174 static DecodeStatus decodeOperand_VSrcV232(MCInst &Inst, unsigned Imm,
175                                            uint64_t Addr,
176                                            const MCDisassembler *Decoder) {
177   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
178   return addOperand(Inst, DAsm->decodeOperand_VSrcV232(Imm));
179 }
180 
181 static DecodeStatus decodeOperand_VS_16(MCInst &Inst, unsigned Imm,
182                                         uint64_t Addr,
183                                         const MCDisassembler *Decoder) {
184   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
185   return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
186 }
187 
188 static DecodeStatus decodeOperand_VS_32(MCInst &Inst, unsigned Imm,
189                                         uint64_t Addr,
190                                         const MCDisassembler *Decoder) {
191   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
192   return addOperand(Inst, DAsm->decodeOperand_VS_32(Imm));
193 }
194 
195 static DecodeStatus decodeOperand_AReg_64(MCInst &Inst, unsigned Imm,
196                                           uint64_t Addr,
197                                           const MCDisassembler *Decoder) {
198   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
199   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm | 512));
200 }
201 
202 static DecodeStatus decodeOperand_AReg_128(MCInst &Inst, unsigned Imm,
203                                            uint64_t Addr,
204                                            const MCDisassembler *Decoder) {
205   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
206   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm | 512));
207 }
208 
209 static DecodeStatus decodeOperand_AReg_256(MCInst &Inst, unsigned Imm,
210                                            uint64_t Addr,
211                                            const MCDisassembler *Decoder) {
212   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
213   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm | 512));
214 }
215 
216 static DecodeStatus decodeOperand_AReg_512(MCInst &Inst, unsigned Imm,
217                                            uint64_t Addr,
218                                            const MCDisassembler *Decoder) {
219   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
220   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm | 512));
221 }
222 
223 static DecodeStatus decodeOperand_AReg_1024(MCInst &Inst, unsigned Imm,
224                                             uint64_t Addr,
225                                             const MCDisassembler *Decoder) {
226   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
227   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm | 512));
228 }
229 
230 static DecodeStatus decodeOperand_VReg_64(MCInst &Inst, unsigned Imm,
231                                           uint64_t Addr,
232                                           const MCDisassembler *Decoder) {
233   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
234   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm));
235 }
236 
237 static DecodeStatus decodeOperand_VReg_128(MCInst &Inst, unsigned Imm,
238                                            uint64_t Addr,
239                                            const MCDisassembler *Decoder) {
240   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
241   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm));
242 }
243 
244 static DecodeStatus decodeOperand_VReg_256(MCInst &Inst, unsigned Imm,
245                                            uint64_t Addr,
246                                            const MCDisassembler *Decoder) {
247   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
248   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm));
249 }
250 
251 static DecodeStatus decodeOperand_VReg_512(MCInst &Inst, unsigned Imm,
252                                            uint64_t Addr,
253                                            const MCDisassembler *Decoder) {
254   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
255   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm));
256 }
257 
258 static DecodeStatus decodeOperand_VReg_1024(MCInst &Inst, unsigned Imm,
259                                             uint64_t Addr,
260                                             const MCDisassembler *Decoder) {
261   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
262   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm));
263 }
264 
265 static DecodeStatus decodeOperand_f32kimm(MCInst &Inst, unsigned Imm,
266                                           uint64_t Addr,
267                                           const MCDisassembler *Decoder) {
268   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
269   return addOperand(Inst, DAsm->decodeMandatoryLiteralConstant(Imm));
270 }
271 
272 static DecodeStatus decodeOperand_f16kimm(MCInst &Inst, unsigned Imm,
273                                           uint64_t Addr,
274                                           const MCDisassembler *Decoder) {
275   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
276   return addOperand(Inst, DAsm->decodeMandatoryLiteralConstant(Imm));
277 }
278 
279 static DecodeStatus
280 decodeOperand_VS_16_Deferred(MCInst &Inst, unsigned Imm, uint64_t Addr,
281                              const MCDisassembler *Decoder) {
282   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
283   return addOperand(
284       Inst, DAsm->decodeSrcOp(llvm::AMDGPUDisassembler::OPW16, Imm, true));
285 }
286 
287 static DecodeStatus
288 decodeOperand_VS_32_Deferred(MCInst &Inst, unsigned Imm, uint64_t Addr,
289                              const MCDisassembler *Decoder) {
290   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
291   return addOperand(
292       Inst, DAsm->decodeSrcOp(llvm::AMDGPUDisassembler::OPW32, Imm, true));
293 }
294 
295 static DecodeStatus decodeOperandVOPDDstY(MCInst &Inst, unsigned Val,
296                                           uint64_t Addr, const void *Decoder) {
297   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
298   return addOperand(Inst, DAsm->decodeVOPDDstYOp(Inst, Val));
299 }
300 
301 static bool IsAGPROperand(const MCInst &Inst, int OpIdx,
302                           const MCRegisterInfo *MRI) {
303   if (OpIdx < 0)
304     return false;
305 
306   const MCOperand &Op = Inst.getOperand(OpIdx);
307   if (!Op.isReg())
308     return false;
309 
310   unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
311   auto Reg = Sub ? Sub : Op.getReg();
312   return Reg >= AMDGPU::AGPR0 && Reg <= AMDGPU::AGPR255;
313 }
314 
315 static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
316                                              AMDGPUDisassembler::OpWidthTy Opw,
317                                              const MCDisassembler *Decoder) {
318   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
319   if (!DAsm->isGFX90A()) {
320     Imm &= 511;
321   } else {
322     // If atomic has both vdata and vdst their register classes are tied.
323     // The bit is decoded along with the vdst, first operand. We need to
324     // change register class to AGPR if vdst was AGPR.
325     // If a DS instruction has both data0 and data1 their register classes
326     // are also tied.
327     unsigned Opc = Inst.getOpcode();
328     uint64_t TSFlags = DAsm->getMCII()->get(Opc).TSFlags;
329     uint16_t DataNameIdx = (TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0
330                                                         : AMDGPU::OpName::vdata;
331     const MCRegisterInfo *MRI = DAsm->getContext().getRegisterInfo();
332     int DataIdx = AMDGPU::getNamedOperandIdx(Opc, DataNameIdx);
333     if ((int)Inst.getNumOperands() == DataIdx) {
334       int DstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
335       if (IsAGPROperand(Inst, DstIdx, MRI))
336         Imm |= 512;
337     }
338 
339     if (TSFlags & SIInstrFlags::DS) {
340       int Data2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
341       if ((int)Inst.getNumOperands() == Data2Idx &&
342           IsAGPROperand(Inst, DataIdx, MRI))
343         Imm |= 512;
344     }
345   }
346   return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
347 }
348 
349 static DecodeStatus
350 DecodeAVLdSt_32RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
351                              const MCDisassembler *Decoder) {
352   return decodeOperand_AVLdSt_Any(Inst, Imm,
353                                   AMDGPUDisassembler::OPW32, Decoder);
354 }
355 
356 static DecodeStatus
357 DecodeAVLdSt_64RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
358                              const MCDisassembler *Decoder) {
359   return decodeOperand_AVLdSt_Any(Inst, Imm,
360                                   AMDGPUDisassembler::OPW64, Decoder);
361 }
362 
363 static DecodeStatus
364 DecodeAVLdSt_96RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
365                              const MCDisassembler *Decoder) {
366   return decodeOperand_AVLdSt_Any(Inst, Imm,
367                                   AMDGPUDisassembler::OPW96, Decoder);
368 }
369 
370 static DecodeStatus
371 DecodeAVLdSt_128RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
372                               const MCDisassembler *Decoder) {
373   return decodeOperand_AVLdSt_Any(Inst, Imm,
374                                   AMDGPUDisassembler::OPW128, Decoder);
375 }
376 
377 static DecodeStatus
378 DecodeAVLdSt_160RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
379                               const MCDisassembler *Decoder) {
380   return decodeOperand_AVLdSt_Any(Inst, Imm, AMDGPUDisassembler::OPW160,
381                                   Decoder);
382 }
383 
384 static DecodeStatus decodeOperand_SReg_32(MCInst &Inst, unsigned Imm,
385                                           uint64_t Addr,
386                                           const MCDisassembler *Decoder) {
387   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
388   return addOperand(Inst, DAsm->decodeOperand_SReg_32(Imm));
389 }
390 
391 #define DECODE_SDWA(DecName) \
392 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName)
393 
394 DECODE_SDWA(Src32)
395 DECODE_SDWA(Src16)
396 DECODE_SDWA(VopcDst)
397 
398 #include "AMDGPUGenDisassemblerTables.inc"
399 
400 //===----------------------------------------------------------------------===//
401 //
402 //===----------------------------------------------------------------------===//
403 
404 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
405   assert(Bytes.size() >= sizeof(T));
406   const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data());
407   Bytes = Bytes.slice(sizeof(T));
408   return Res;
409 }
410 
411 static inline DecoderUInt128 eat12Bytes(ArrayRef<uint8_t> &Bytes) {
412   assert(Bytes.size() >= 12);
413   uint64_t Lo = support::endian::read<uint64_t, support::endianness::little>(
414       Bytes.data());
415   Bytes = Bytes.slice(8);
416   uint64_t Hi = support::endian::read<uint32_t, support::endianness::little>(
417       Bytes.data());
418   Bytes = Bytes.slice(4);
419   return DecoderUInt128(Lo, Hi);
420 }
421 
422 // The disassembler is greedy, so we need to check FI operand value to
423 // not parse a dpp if the correct literal is not set. For dpp16 the
424 // autogenerated decoder checks the dpp literal
425 static bool isValidDPP8(const MCInst &MI) {
426   using namespace llvm::AMDGPU::DPP;
427   int FiIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::fi);
428   assert(FiIdx != -1);
429   if ((unsigned)FiIdx >= MI.getNumOperands())
430     return false;
431   unsigned Fi = MI.getOperand(FiIdx).getImm();
432   return Fi == DPP8_FI_0 || Fi == DPP8_FI_1;
433 }
434 
435 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
436                                                 ArrayRef<uint8_t> Bytes_,
437                                                 uint64_t Address,
438                                                 raw_ostream &CS) const {
439   CommentStream = &CS;
440   bool IsSDWA = false;
441 
442   unsigned MaxInstBytesNum = std::min((size_t)TargetMaxInstBytes, Bytes_.size());
443   Bytes = Bytes_.slice(0, MaxInstBytesNum);
444 
445   DecodeStatus Res = MCDisassembler::Fail;
446   do {
447     // ToDo: better to switch encoding length using some bit predicate
448     // but it is unknown yet, so try all we can
449 
450     // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2
451     // encodings
452     if (isGFX11Plus() && Bytes.size() >= 12 ) {
453       DecoderUInt128 DecW = eat12Bytes(Bytes);
454       Res = tryDecodeInst(DecoderTableDPP8GFX1196, MI, DecW,
455                                           Address);
456       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
457         break;
458       MI = MCInst(); // clear
459       Res = tryDecodeInst(DecoderTableDPPGFX1196, MI, DecW,
460                                           Address);
461       if (Res) {
462         if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3P)
463           convertVOP3PDPPInst(MI);
464         else if (AMDGPU::isVOPC64DPP(MI.getOpcode()))
465           convertVOPCDPPInst(MI); // Special VOP3 case
466         else {
467           assert(MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3);
468           convertVOP3DPPInst(MI); // Regular VOP3 case
469         }
470         break;
471       }
472       Res = tryDecodeInst(DecoderTableGFX1196, MI, DecW, Address);
473       if (Res)
474         break;
475     }
476     // Reinitialize Bytes
477     Bytes = Bytes_.slice(0, MaxInstBytesNum);
478 
479     if (Bytes.size() >= 8) {
480       const uint64_t QW = eatBytes<uint64_t>(Bytes);
481 
482       if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) {
483         Res = tryDecodeInst(DecoderTableGFX10_B64, MI, QW, Address);
484         if (Res) {
485           if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dpp8)
486               == -1)
487             break;
488           if (convertDPP8Inst(MI) == MCDisassembler::Success)
489             break;
490           MI = MCInst(); // clear
491         }
492       }
493 
494       Res = tryDecodeInst(DecoderTableDPP864, MI, QW, Address);
495       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
496         break;
497       MI = MCInst(); // clear
498 
499       Res = tryDecodeInst(DecoderTableDPP8GFX1164, MI, QW, Address);
500       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
501         break;
502       MI = MCInst(); // clear
503 
504       Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address);
505       if (Res) break;
506 
507       Res = tryDecodeInst(DecoderTableDPPGFX1164, MI, QW, Address);
508       if (Res) {
509         if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOPC)
510           convertVOPCDPPInst(MI);
511         break;
512       }
513 
514       Res = tryDecodeInst(DecoderTableSDWA64, MI, QW, Address);
515       if (Res) { IsSDWA = true;  break; }
516 
517       Res = tryDecodeInst(DecoderTableSDWA964, MI, QW, Address);
518       if (Res) { IsSDWA = true;  break; }
519 
520       Res = tryDecodeInst(DecoderTableSDWA1064, MI, QW, Address);
521       if (Res) { IsSDWA = true;  break; }
522 
523       if (STI.getFeatureBits()[AMDGPU::FeatureUnpackedD16VMem]) {
524         Res = tryDecodeInst(DecoderTableGFX80_UNPACKED64, MI, QW, Address);
525         if (Res)
526           break;
527       }
528 
529       // Some GFX9 subtargets repurposed the v_mad_mix_f32, v_mad_mixlo_f16 and
530       // v_mad_mixhi_f16 for FMA variants. Try to decode using this special
531       // table first so we print the correct name.
532       if (STI.getFeatureBits()[AMDGPU::FeatureFmaMixInsts]) {
533         Res = tryDecodeInst(DecoderTableGFX9_DL64, MI, QW, Address);
534         if (Res)
535           break;
536       }
537     }
538 
539     // Reinitialize Bytes as DPP64 could have eaten too much
540     Bytes = Bytes_.slice(0, MaxInstBytesNum);
541 
542     // Try decode 32-bit instruction
543     if (Bytes.size() < 4) break;
544     const uint32_t DW = eatBytes<uint32_t>(Bytes);
545     Res = tryDecodeInst(DecoderTableGFX832, MI, DW, Address);
546     if (Res) break;
547 
548     Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address);
549     if (Res) break;
550 
551     Res = tryDecodeInst(DecoderTableGFX932, MI, DW, Address);
552     if (Res) break;
553 
554     if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) {
555       Res = tryDecodeInst(DecoderTableGFX90A32, MI, DW, Address);
556       if (Res)
557         break;
558     }
559 
560     if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) {
561       Res = tryDecodeInst(DecoderTableGFX10_B32, MI, DW, Address);
562       if (Res) break;
563     }
564 
565     Res = tryDecodeInst(DecoderTableGFX1032, MI, DW, Address);
566     if (Res) break;
567 
568     Res = tryDecodeInst(DecoderTableGFX1132, MI, DW, Address);
569     if (Res) break;
570 
571     if (Bytes.size() < 4) break;
572     const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW;
573 
574     if (STI.getFeatureBits()[AMDGPU::FeatureGFX940Insts]) {
575       Res = tryDecodeInst(DecoderTableGFX94064, MI, QW, Address);
576       if (Res)
577         break;
578     }
579 
580     if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) {
581       Res = tryDecodeInst(DecoderTableGFX90A64, MI, QW, Address);
582       if (Res)
583         break;
584     }
585 
586     Res = tryDecodeInst(DecoderTableGFX864, MI, QW, Address);
587     if (Res) break;
588 
589     Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address);
590     if (Res) break;
591 
592     Res = tryDecodeInst(DecoderTableGFX964, MI, QW, Address);
593     if (Res) break;
594 
595     Res = tryDecodeInst(DecoderTableGFX1064, MI, QW, Address);
596     if (Res) break;
597 
598     Res = tryDecodeInst(DecoderTableGFX1164, MI, QW, Address);
599     if (Res)
600       break;
601 
602     Res = tryDecodeInst(DecoderTableWMMAGFX1164, MI, QW, Address);
603   } while (false);
604 
605   if (Res && AMDGPU::isMAC(MI.getOpcode())) {
606     // Insert dummy unused src2_modifiers.
607     insertNamedMCOperand(MI, MCOperand::createImm(0),
608                          AMDGPU::OpName::src2_modifiers);
609   }
610 
611   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
612           (SIInstrFlags::MUBUF | SIInstrFlags::FLAT | SIInstrFlags::SMRD))) {
613     int CPolPos = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
614                                              AMDGPU::OpName::cpol);
615     if (CPolPos != -1) {
616       unsigned CPol =
617           (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::IsAtomicRet) ?
618               AMDGPU::CPol::GLC : 0;
619       if (MI.getNumOperands() <= (unsigned)CPolPos) {
620         insertNamedMCOperand(MI, MCOperand::createImm(CPol),
621                              AMDGPU::OpName::cpol);
622       } else if (CPol) {
623         MI.getOperand(CPolPos).setImm(MI.getOperand(CPolPos).getImm() | CPol);
624       }
625     }
626   }
627 
628   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
629               (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF)) &&
630              (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts])) {
631     // GFX90A lost TFE, its place is occupied by ACC.
632     int TFEOpIdx =
633         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::tfe);
634     if (TFEOpIdx != -1) {
635       auto TFEIter = MI.begin();
636       std::advance(TFEIter, TFEOpIdx);
637       MI.insert(TFEIter, MCOperand::createImm(0));
638     }
639   }
640 
641   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
642               (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF))) {
643     int SWZOpIdx =
644         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::swz);
645     if (SWZOpIdx != -1) {
646       auto SWZIter = MI.begin();
647       std::advance(SWZIter, SWZOpIdx);
648       MI.insert(SWZIter, MCOperand::createImm(0));
649     }
650   }
651 
652   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::MIMG)) {
653     int VAddr0Idx =
654         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
655     int RsrcIdx =
656         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
657     unsigned NSAArgs = RsrcIdx - VAddr0Idx - 1;
658     if (VAddr0Idx >= 0 && NSAArgs > 0) {
659       unsigned NSAWords = (NSAArgs + 3) / 4;
660       if (Bytes.size() < 4 * NSAWords) {
661         Res = MCDisassembler::Fail;
662       } else {
663         for (unsigned i = 0; i < NSAArgs; ++i) {
664           const unsigned VAddrIdx = VAddr0Idx + 1 + i;
665           auto VAddrRCID =
666               MCII->get(MI.getOpcode()).operands()[VAddrIdx].RegClass;
667           MI.insert(MI.begin() + VAddrIdx,
668                     createRegOperand(VAddrRCID, Bytes[i]));
669         }
670         Bytes = Bytes.slice(4 * NSAWords);
671       }
672     }
673 
674     if (Res)
675       Res = convertMIMGInst(MI);
676   }
677 
678   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::EXP))
679     Res = convertEXPInst(MI);
680 
681   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VINTERP))
682     Res = convertVINTERPInst(MI);
683 
684   if (Res && IsSDWA)
685     Res = convertSDWAInst(MI);
686 
687   int VDstIn_Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
688                                               AMDGPU::OpName::vdst_in);
689   if (VDstIn_Idx != -1) {
690     int Tied = MCII->get(MI.getOpcode()).getOperandConstraint(VDstIn_Idx,
691                            MCOI::OperandConstraint::TIED_TO);
692     if (Tied != -1 && (MI.getNumOperands() <= (unsigned)VDstIn_Idx ||
693          !MI.getOperand(VDstIn_Idx).isReg() ||
694          MI.getOperand(VDstIn_Idx).getReg() != MI.getOperand(Tied).getReg())) {
695       if (MI.getNumOperands() > (unsigned)VDstIn_Idx)
696         MI.erase(&MI.getOperand(VDstIn_Idx));
697       insertNamedMCOperand(MI,
698         MCOperand::createReg(MI.getOperand(Tied).getReg()),
699         AMDGPU::OpName::vdst_in);
700     }
701   }
702 
703   int ImmLitIdx =
704       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::imm);
705   bool IsSOPK = MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::SOPK;
706   if (Res && ImmLitIdx != -1 && !IsSOPK)
707     Res = convertFMAanyK(MI, ImmLitIdx);
708 
709   // if the opcode was not recognized we'll assume a Size of 4 bytes
710   // (unless there are fewer bytes left)
711   Size = Res ? (MaxInstBytesNum - Bytes.size())
712              : std::min((size_t)4, Bytes_.size());
713   return Res;
714 }
715 
716 DecodeStatus AMDGPUDisassembler::convertEXPInst(MCInst &MI) const {
717   if (STI.getFeatureBits()[AMDGPU::FeatureGFX11]) {
718     // The MCInst still has these fields even though they are no longer encoded
719     // in the GFX11 instruction.
720     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vm);
721     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::compr);
722   }
723   return MCDisassembler::Success;
724 }
725 
726 DecodeStatus AMDGPUDisassembler::convertVINTERPInst(MCInst &MI) const {
727   if (MI.getOpcode() == AMDGPU::V_INTERP_P10_F16_F32_inreg_gfx11 ||
728       MI.getOpcode() == AMDGPU::V_INTERP_P10_RTZ_F16_F32_inreg_gfx11 ||
729       MI.getOpcode() == AMDGPU::V_INTERP_P2_F16_F32_inreg_gfx11 ||
730       MI.getOpcode() == AMDGPU::V_INTERP_P2_RTZ_F16_F32_inreg_gfx11) {
731     // The MCInst has this field that is not directly encoded in the
732     // instruction.
733     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::op_sel);
734   }
735   return MCDisassembler::Success;
736 }
737 
738 DecodeStatus AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const {
739   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
740       STI.getFeatureBits()[AMDGPU::FeatureGFX10]) {
741     if (AMDGPU::hasNamedOperand(MI.getOpcode(), AMDGPU::OpName::sdst))
742       // VOPC - insert clamp
743       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp);
744   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
745     int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst);
746     if (SDst != -1) {
747       // VOPC - insert VCC register as sdst
748       insertNamedMCOperand(MI, createRegOperand(AMDGPU::VCC),
749                            AMDGPU::OpName::sdst);
750     } else {
751       // VOP1/2 - insert omod if present in instruction
752       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod);
753     }
754   }
755   return MCDisassembler::Success;
756 }
757 
758 struct VOPModifiers {
759   unsigned OpSel = 0;
760   unsigned OpSelHi = 0;
761   unsigned NegLo = 0;
762   unsigned NegHi = 0;
763 };
764 
765 // Reconstruct values of VOP3/VOP3P operands such as op_sel.
766 // Note that these values do not affect disassembler output,
767 // so this is only necessary for consistency with src_modifiers.
768 static VOPModifiers collectVOPModifiers(const MCInst &MI,
769                                         bool IsVOP3P = false) {
770   VOPModifiers Modifiers;
771   unsigned Opc = MI.getOpcode();
772   const int ModOps[] = {AMDGPU::OpName::src0_modifiers,
773                         AMDGPU::OpName::src1_modifiers,
774                         AMDGPU::OpName::src2_modifiers};
775   for (int J = 0; J < 3; ++J) {
776     int OpIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]);
777     if (OpIdx == -1)
778       continue;
779 
780     unsigned Val = MI.getOperand(OpIdx).getImm();
781 
782     Modifiers.OpSel |= !!(Val & SISrcMods::OP_SEL_0) << J;
783     if (IsVOP3P) {
784       Modifiers.OpSelHi |= !!(Val & SISrcMods::OP_SEL_1) << J;
785       Modifiers.NegLo |= !!(Val & SISrcMods::NEG) << J;
786       Modifiers.NegHi |= !!(Val & SISrcMods::NEG_HI) << J;
787     } else if (J == 0) {
788       Modifiers.OpSel |= !!(Val & SISrcMods::DST_OP_SEL) << 3;
789     }
790   }
791 
792   return Modifiers;
793 }
794 
795 // MAC opcodes have special old and src2 operands.
796 // src2 is tied to dst, while old is not tied (but assumed to be).
797 bool AMDGPUDisassembler::isMacDPP(MCInst &MI) const {
798   constexpr int DST_IDX = 0;
799   auto Opcode = MI.getOpcode();
800   const auto &Desc = MCII->get(Opcode);
801   auto OldIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::old);
802 
803   if (OldIdx != -1 && Desc.getOperandConstraint(
804                           OldIdx, MCOI::OperandConstraint::TIED_TO) == -1) {
805     assert(AMDGPU::hasNamedOperand(Opcode, AMDGPU::OpName::src2));
806     assert(Desc.getOperandConstraint(
807                AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2),
808                MCOI::OperandConstraint::TIED_TO) == DST_IDX);
809     (void)DST_IDX;
810     return true;
811   }
812 
813   return false;
814 }
815 
816 // Create dummy old operand and insert dummy unused src2_modifiers
817 void AMDGPUDisassembler::convertMacDPPInst(MCInst &MI) const {
818   assert(MI.getNumOperands() + 1 < MCII->get(MI.getOpcode()).getNumOperands());
819   insertNamedMCOperand(MI, MCOperand::createReg(0), AMDGPU::OpName::old);
820   insertNamedMCOperand(MI, MCOperand::createImm(0),
821                        AMDGPU::OpName::src2_modifiers);
822 }
823 
824 // We must check FI == literal to reject not genuine dpp8 insts, and we must
825 // first add optional MI operands to check FI
826 DecodeStatus AMDGPUDisassembler::convertDPP8Inst(MCInst &MI) const {
827   unsigned Opc = MI.getOpcode();
828   if (MCII->get(Opc).TSFlags & SIInstrFlags::VOP3P) {
829     convertVOP3PDPPInst(MI);
830   } else if ((MCII->get(Opc).TSFlags & SIInstrFlags::VOPC) ||
831              AMDGPU::isVOPC64DPP(Opc)) {
832     convertVOPCDPPInst(MI);
833   } else {
834     if (isMacDPP(MI))
835       convertMacDPPInst(MI);
836 
837     unsigned DescNumOps = MCII->get(Opc).getNumOperands();
838     if (MI.getNumOperands() < DescNumOps &&
839         AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel)) {
840       auto Mods = collectVOPModifiers(MI);
841       insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
842                            AMDGPU::OpName::op_sel);
843     } else {
844       // Insert dummy unused src modifiers.
845       if (MI.getNumOperands() < DescNumOps &&
846           AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src0_modifiers))
847         insertNamedMCOperand(MI, MCOperand::createImm(0),
848                              AMDGPU::OpName::src0_modifiers);
849 
850       if (MI.getNumOperands() < DescNumOps &&
851           AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src1_modifiers))
852         insertNamedMCOperand(MI, MCOperand::createImm(0),
853                              AMDGPU::OpName::src1_modifiers);
854     }
855   }
856   return isValidDPP8(MI) ? MCDisassembler::Success : MCDisassembler::SoftFail;
857 }
858 
859 DecodeStatus AMDGPUDisassembler::convertVOP3DPPInst(MCInst &MI) const {
860   if (isMacDPP(MI))
861     convertMacDPPInst(MI);
862 
863   unsigned Opc = MI.getOpcode();
864   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
865   if (MI.getNumOperands() < DescNumOps &&
866       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel)) {
867     auto Mods = collectVOPModifiers(MI);
868     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
869                          AMDGPU::OpName::op_sel);
870   }
871   return MCDisassembler::Success;
872 }
873 
874 // Note that before gfx10, the MIMG encoding provided no information about
875 // VADDR size. Consequently, decoded instructions always show address as if it
876 // has 1 dword, which could be not really so.
877 DecodeStatus AMDGPUDisassembler::convertMIMGInst(MCInst &MI) const {
878 
879   int VDstIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
880                                            AMDGPU::OpName::vdst);
881 
882   int VDataIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
883                                             AMDGPU::OpName::vdata);
884   int VAddr0Idx =
885       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
886   int DMaskIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
887                                             AMDGPU::OpName::dmask);
888 
889   int TFEIdx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
890                                             AMDGPU::OpName::tfe);
891   int D16Idx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
892                                             AMDGPU::OpName::d16);
893 
894   const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode());
895   const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
896       AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
897 
898   assert(VDataIdx != -1);
899   if (BaseOpcode->BVH) {
900     // Add A16 operand for intersect_ray instructions
901     if (AMDGPU::hasNamedOperand(MI.getOpcode(), AMDGPU::OpName::a16))
902       addOperand(MI, MCOperand::createImm(1));
903     return MCDisassembler::Success;
904   }
905 
906   bool IsAtomic = (VDstIdx != -1);
907   bool IsGather4 = MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::Gather4;
908   bool IsNSA = false;
909   unsigned AddrSize = Info->VAddrDwords;
910 
911   if (isGFX10Plus()) {
912     unsigned DimIdx =
913         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dim);
914     int A16Idx =
915         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16);
916     const AMDGPU::MIMGDimInfo *Dim =
917         AMDGPU::getMIMGDimInfoByEncoding(MI.getOperand(DimIdx).getImm());
918     const bool IsA16 = (A16Idx != -1 && MI.getOperand(A16Idx).getImm());
919 
920     AddrSize =
921         AMDGPU::getAddrSizeMIMGOp(BaseOpcode, Dim, IsA16, AMDGPU::hasG16(STI));
922 
923     IsNSA = Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA ||
924             Info->MIMGEncoding == AMDGPU::MIMGEncGfx11NSA;
925     if (!IsNSA) {
926       if (AddrSize > 12)
927         AddrSize = 16;
928     } else {
929       if (AddrSize > Info->VAddrDwords) {
930         // The NSA encoding does not contain enough operands for the combination
931         // of base opcode / dimension. Should this be an error?
932         return MCDisassembler::Success;
933       }
934     }
935   }
936 
937   unsigned DMask = MI.getOperand(DMaskIdx).getImm() & 0xf;
938   unsigned DstSize = IsGather4 ? 4 : std::max(llvm::popcount(DMask), 1);
939 
940   bool D16 = D16Idx >= 0 && MI.getOperand(D16Idx).getImm();
941   if (D16 && AMDGPU::hasPackedD16(STI)) {
942     DstSize = (DstSize + 1) / 2;
943   }
944 
945   if (TFEIdx != -1 && MI.getOperand(TFEIdx).getImm())
946     DstSize += 1;
947 
948   if (DstSize == Info->VDataDwords && AddrSize == Info->VAddrDwords)
949     return MCDisassembler::Success;
950 
951   int NewOpcode =
952       AMDGPU::getMIMGOpcode(Info->BaseOpcode, Info->MIMGEncoding, DstSize, AddrSize);
953   if (NewOpcode == -1)
954     return MCDisassembler::Success;
955 
956   // Widen the register to the correct number of enabled channels.
957   unsigned NewVdata = AMDGPU::NoRegister;
958   if (DstSize != Info->VDataDwords) {
959     auto DataRCID = MCII->get(NewOpcode).operands()[VDataIdx].RegClass;
960 
961     // Get first subregister of VData
962     unsigned Vdata0 = MI.getOperand(VDataIdx).getReg();
963     unsigned VdataSub0 = MRI.getSubReg(Vdata0, AMDGPU::sub0);
964     Vdata0 = (VdataSub0 != 0)? VdataSub0 : Vdata0;
965 
966     NewVdata = MRI.getMatchingSuperReg(Vdata0, AMDGPU::sub0,
967                                        &MRI.getRegClass(DataRCID));
968     if (NewVdata == AMDGPU::NoRegister) {
969       // It's possible to encode this such that the low register + enabled
970       // components exceeds the register count.
971       return MCDisassembler::Success;
972     }
973   }
974 
975   // If not using NSA on GFX10+, widen address register to correct size.
976   unsigned NewVAddr0 = AMDGPU::NoRegister;
977   if (isGFX10Plus() && !IsNSA && AddrSize != Info->VAddrDwords) {
978     unsigned VAddr0 = MI.getOperand(VAddr0Idx).getReg();
979     unsigned VAddrSub0 = MRI.getSubReg(VAddr0, AMDGPU::sub0);
980     VAddr0 = (VAddrSub0 != 0) ? VAddrSub0 : VAddr0;
981 
982     auto AddrRCID = MCII->get(NewOpcode).operands()[VAddr0Idx].RegClass;
983     NewVAddr0 = MRI.getMatchingSuperReg(VAddr0, AMDGPU::sub0,
984                                         &MRI.getRegClass(AddrRCID));
985     if (NewVAddr0 == AMDGPU::NoRegister)
986       return MCDisassembler::Success;
987   }
988 
989   MI.setOpcode(NewOpcode);
990 
991   if (NewVdata != AMDGPU::NoRegister) {
992     MI.getOperand(VDataIdx) = MCOperand::createReg(NewVdata);
993 
994     if (IsAtomic) {
995       // Atomic operations have an additional operand (a copy of data)
996       MI.getOperand(VDstIdx) = MCOperand::createReg(NewVdata);
997     }
998   }
999 
1000   if (NewVAddr0 != AMDGPU::NoRegister) {
1001     MI.getOperand(VAddr0Idx) = MCOperand::createReg(NewVAddr0);
1002   } else if (IsNSA) {
1003     assert(AddrSize <= Info->VAddrDwords);
1004     MI.erase(MI.begin() + VAddr0Idx + AddrSize,
1005              MI.begin() + VAddr0Idx + Info->VAddrDwords);
1006   }
1007 
1008   return MCDisassembler::Success;
1009 }
1010 
1011 // Opsel and neg bits are used in src_modifiers and standalone operands. Autogen
1012 // decoder only adds to src_modifiers, so manually add the bits to the other
1013 // operands.
1014 DecodeStatus AMDGPUDisassembler::convertVOP3PDPPInst(MCInst &MI) const {
1015   unsigned Opc = MI.getOpcode();
1016   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
1017   auto Mods = collectVOPModifiers(MI, true);
1018 
1019   if (MI.getNumOperands() < DescNumOps &&
1020       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::vdst_in))
1021     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vdst_in);
1022 
1023   if (MI.getNumOperands() < DescNumOps &&
1024       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel))
1025     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
1026                          AMDGPU::OpName::op_sel);
1027   if (MI.getNumOperands() < DescNumOps &&
1028       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel_hi))
1029     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSelHi),
1030                          AMDGPU::OpName::op_sel_hi);
1031   if (MI.getNumOperands() < DescNumOps &&
1032       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::neg_lo))
1033     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegLo),
1034                          AMDGPU::OpName::neg_lo);
1035   if (MI.getNumOperands() < DescNumOps &&
1036       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::neg_hi))
1037     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegHi),
1038                          AMDGPU::OpName::neg_hi);
1039 
1040   return MCDisassembler::Success;
1041 }
1042 
1043 // Create dummy old operand and insert optional operands
1044 DecodeStatus AMDGPUDisassembler::convertVOPCDPPInst(MCInst &MI) const {
1045   unsigned Opc = MI.getOpcode();
1046   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
1047 
1048   if (MI.getNumOperands() < DescNumOps &&
1049       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::old))
1050     insertNamedMCOperand(MI, MCOperand::createReg(0), AMDGPU::OpName::old);
1051 
1052   if (MI.getNumOperands() < DescNumOps &&
1053       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src0_modifiers))
1054     insertNamedMCOperand(MI, MCOperand::createImm(0),
1055                          AMDGPU::OpName::src0_modifiers);
1056 
1057   if (MI.getNumOperands() < DescNumOps &&
1058       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src1_modifiers))
1059     insertNamedMCOperand(MI, MCOperand::createImm(0),
1060                          AMDGPU::OpName::src1_modifiers);
1061   return MCDisassembler::Success;
1062 }
1063 
1064 DecodeStatus AMDGPUDisassembler::convertFMAanyK(MCInst &MI,
1065                                                 int ImmLitIdx) const {
1066   assert(HasLiteral && "Should have decoded a literal");
1067   const MCInstrDesc &Desc = MCII->get(MI.getOpcode());
1068   unsigned DescNumOps = Desc.getNumOperands();
1069   insertNamedMCOperand(MI, MCOperand::createImm(Literal),
1070                        AMDGPU::OpName::immDeferred);
1071   assert(DescNumOps == MI.getNumOperands());
1072   for (unsigned I = 0; I < DescNumOps; ++I) {
1073     auto &Op = MI.getOperand(I);
1074     auto OpType = Desc.operands()[I].OperandType;
1075     bool IsDeferredOp = (OpType == AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED ||
1076                          OpType == AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED);
1077     if (Op.isImm() && Op.getImm() == AMDGPU::EncValues::LITERAL_CONST &&
1078         IsDeferredOp)
1079       Op.setImm(Literal);
1080   }
1081   return MCDisassembler::Success;
1082 }
1083 
1084 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
1085   return getContext().getRegisterInfo()->
1086     getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
1087 }
1088 
1089 inline
1090 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
1091                                          const Twine& ErrMsg) const {
1092   *CommentStream << "Error: " + ErrMsg;
1093 
1094   // ToDo: add support for error operands to MCInst.h
1095   // return MCOperand::createError(V);
1096   return MCOperand();
1097 }
1098 
1099 inline
1100 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
1101   return MCOperand::createReg(AMDGPU::getMCReg(RegId, STI));
1102 }
1103 
1104 inline
1105 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
1106                                                unsigned Val) const {
1107   const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
1108   if (Val >= RegCl.getNumRegs())
1109     return errOperand(Val, Twine(getRegClassName(RegClassID)) +
1110                            ": unknown register " + Twine(Val));
1111   return createRegOperand(RegCl.getRegister(Val));
1112 }
1113 
1114 inline
1115 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
1116                                                 unsigned Val) const {
1117   // ToDo: SI/CI have 104 SGPRs, VI - 102
1118   // Valery: here we accepting as much as we can, let assembler sort it out
1119   int shift = 0;
1120   switch (SRegClassID) {
1121   case AMDGPU::SGPR_32RegClassID:
1122   case AMDGPU::TTMP_32RegClassID:
1123     break;
1124   case AMDGPU::SGPR_64RegClassID:
1125   case AMDGPU::TTMP_64RegClassID:
1126     shift = 1;
1127     break;
1128   case AMDGPU::SGPR_128RegClassID:
1129   case AMDGPU::TTMP_128RegClassID:
1130   // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
1131   // this bundle?
1132   case AMDGPU::SGPR_256RegClassID:
1133   case AMDGPU::TTMP_256RegClassID:
1134     // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
1135   // this bundle?
1136   case AMDGPU::SGPR_288RegClassID:
1137   case AMDGPU::TTMP_288RegClassID:
1138   case AMDGPU::SGPR_320RegClassID:
1139   case AMDGPU::TTMP_320RegClassID:
1140   case AMDGPU::SGPR_352RegClassID:
1141   case AMDGPU::TTMP_352RegClassID:
1142   case AMDGPU::SGPR_384RegClassID:
1143   case AMDGPU::TTMP_384RegClassID:
1144   case AMDGPU::SGPR_512RegClassID:
1145   case AMDGPU::TTMP_512RegClassID:
1146     shift = 2;
1147     break;
1148   // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
1149   // this bundle?
1150   default:
1151     llvm_unreachable("unhandled register class");
1152   }
1153 
1154   if (Val % (1 << shift)) {
1155     *CommentStream << "Warning: " << getRegClassName(SRegClassID)
1156                    << ": scalar reg isn't aligned " << Val;
1157   }
1158 
1159   return createRegOperand(SRegClassID, Val >> shift);
1160 }
1161 
1162 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const {
1163   return decodeSrcOp(OPW32, Val);
1164 }
1165 
1166 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const {
1167   return decodeSrcOp(OPW64, Val);
1168 }
1169 
1170 MCOperand AMDGPUDisassembler::decodeOperand_VS_128(unsigned Val) const {
1171   return decodeSrcOp(OPW128, Val);
1172 }
1173 
1174 MCOperand AMDGPUDisassembler::decodeOperand_VSrc16(unsigned Val) const {
1175   return decodeSrcOp(OPW16, Val);
1176 }
1177 
1178 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV216(unsigned Val) const {
1179   return decodeSrcOp(OPWV216, Val);
1180 }
1181 
1182 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV232(unsigned Val) const {
1183   return decodeSrcOp(OPWV232, Val);
1184 }
1185 
1186 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32_Lo128(unsigned Val) const {
1187   return createRegOperand(AMDGPU::VGPR_32_Lo128RegClassID, Val);
1188 }
1189 
1190 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const {
1191   // Some instructions have operand restrictions beyond what the encoding
1192   // allows. Some ordinarily VSrc_32 operands are VGPR_32, so clear the extra
1193   // high bit.
1194   Val &= 255;
1195 
1196   return createRegOperand(AMDGPU::VGPR_32RegClassID, Val);
1197 }
1198 
1199 MCOperand AMDGPUDisassembler::decodeOperand_VRegOrLds_32(unsigned Val) const {
1200   return decodeSrcOp(OPW32, Val);
1201 }
1202 
1203 MCOperand AMDGPUDisassembler::decodeOperand_AGPR_32(unsigned Val) const {
1204   return createRegOperand(AMDGPU::AGPR_32RegClassID, Val & 255);
1205 }
1206 
1207 MCOperand AMDGPUDisassembler::decodeOperand_AReg_64(unsigned Val) const {
1208   return createRegOperand(AMDGPU::AReg_64RegClassID, Val & 255);
1209 }
1210 
1211 MCOperand AMDGPUDisassembler::decodeOperand_AReg_128(unsigned Val) const {
1212   return createRegOperand(AMDGPU::AReg_128RegClassID, Val & 255);
1213 }
1214 
1215 MCOperand AMDGPUDisassembler::decodeOperand_AReg_256(unsigned Val) const {
1216   return createRegOperand(AMDGPU::AReg_256RegClassID, Val & 255);
1217 }
1218 
1219 MCOperand AMDGPUDisassembler::decodeOperand_AReg_288(unsigned Val) const {
1220   return createRegOperand(AMDGPU::AReg_288RegClassID, Val & 255);
1221 }
1222 
1223 MCOperand AMDGPUDisassembler::decodeOperand_AReg_320(unsigned Val) const {
1224   return createRegOperand(AMDGPU::AReg_320RegClassID, Val & 255);
1225 }
1226 
1227 MCOperand AMDGPUDisassembler::decodeOperand_AReg_352(unsigned Val) const {
1228   return createRegOperand(AMDGPU::AReg_352RegClassID, Val & 255);
1229 }
1230 
1231 MCOperand AMDGPUDisassembler::decodeOperand_AReg_384(unsigned Val) const {
1232   return createRegOperand(AMDGPU::AReg_384RegClassID, Val & 255);
1233 }
1234 
1235 
1236 MCOperand AMDGPUDisassembler::decodeOperand_AReg_512(unsigned Val) const {
1237   return createRegOperand(AMDGPU::AReg_512RegClassID, Val & 255);
1238 }
1239 
1240 MCOperand AMDGPUDisassembler::decodeOperand_AReg_1024(unsigned Val) const {
1241   return createRegOperand(AMDGPU::AReg_1024RegClassID, Val & 255);
1242 }
1243 
1244 MCOperand AMDGPUDisassembler::decodeOperand_AV_32(unsigned Val) const {
1245   return decodeSrcOp(OPW32, Val);
1246 }
1247 
1248 MCOperand AMDGPUDisassembler::decodeOperand_AV_64(unsigned Val) const {
1249   return decodeSrcOp(OPW64, Val);
1250 }
1251 
1252 MCOperand AMDGPUDisassembler::decodeOperand_AV_128(unsigned Val) const {
1253   return decodeSrcOp(OPW128, Val);
1254 }
1255 
1256 MCOperand AMDGPUDisassembler::decodeOperand_AVDst_128(unsigned Val) const {
1257   using namespace AMDGPU::EncValues;
1258   assert((Val & IS_VGPR) == 0); // Val{8} is not encoded but assumed to be 1.
1259   return decodeSrcOp(OPW128, Val | IS_VGPR);
1260 }
1261 
1262 MCOperand AMDGPUDisassembler::decodeOperand_AVDst_512(unsigned Val) const {
1263   using namespace AMDGPU::EncValues;
1264   assert((Val & IS_VGPR) == 0); // Val{8} is not encoded but assumed to be 1.
1265   return decodeSrcOp(OPW512, Val | IS_VGPR);
1266 }
1267 
1268 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const {
1269   return createRegOperand(AMDGPU::VReg_64RegClassID, Val);
1270 }
1271 
1272 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const {
1273   return createRegOperand(AMDGPU::VReg_96RegClassID, Val);
1274 }
1275 
1276 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const {
1277   return createRegOperand(AMDGPU::VReg_128RegClassID, Val);
1278 }
1279 
1280 MCOperand AMDGPUDisassembler::decodeOperand_VReg_256(unsigned Val) const {
1281   return createRegOperand(AMDGPU::VReg_256RegClassID, Val);
1282 }
1283 
1284 MCOperand AMDGPUDisassembler::decodeOperand_VReg_288(unsigned Val) const {
1285   return createRegOperand(AMDGPU::VReg_288RegClassID, Val);
1286 }
1287 
1288 MCOperand AMDGPUDisassembler::decodeOperand_VReg_320(unsigned Val) const {
1289   return createRegOperand(AMDGPU::VReg_320RegClassID, Val);
1290 }
1291 
1292 MCOperand AMDGPUDisassembler::decodeOperand_VReg_352(unsigned Val) const {
1293   return createRegOperand(AMDGPU::VReg_352RegClassID, Val);
1294 }
1295 
1296 MCOperand AMDGPUDisassembler::decodeOperand_VReg_384(unsigned Val) const {
1297   return createRegOperand(AMDGPU::VReg_384RegClassID, Val);
1298 }
1299 
1300 MCOperand AMDGPUDisassembler::decodeOperand_VReg_512(unsigned Val) const {
1301   return createRegOperand(AMDGPU::VReg_512RegClassID, Val);
1302 }
1303 
1304 MCOperand AMDGPUDisassembler::decodeOperand_VReg_1024(unsigned Val) const {
1305   return createRegOperand(AMDGPU::VReg_1024RegClassID, Val);
1306 }
1307 
1308 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const {
1309   // table-gen generated disassembler doesn't care about operand types
1310   // leaving only registry class so SSrc_32 operand turns into SReg_32
1311   // and therefore we accept immediates and literals here as well
1312   return decodeSrcOp(OPW32, Val);
1313 }
1314 
1315 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XM0_XEXEC(
1316   unsigned Val) const {
1317   // SReg_32_XM0 is SReg_32 without M0 or EXEC_LO/EXEC_HI
1318   return decodeOperand_SReg_32(Val);
1319 }
1320 
1321 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XEXEC_HI(
1322   unsigned Val) const {
1323   // SReg_32_XM0 is SReg_32 without EXEC_HI
1324   return decodeOperand_SReg_32(Val);
1325 }
1326 
1327 MCOperand AMDGPUDisassembler::decodeOperand_SRegOrLds_32(unsigned Val) const {
1328   // table-gen generated disassembler doesn't care about operand types
1329   // leaving only registry class so SSrc_32 operand turns into SReg_32
1330   // and therefore we accept immediates and literals here as well
1331   return decodeSrcOp(OPW32, Val);
1332 }
1333 
1334 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const {
1335   return decodeSrcOp(OPW64, Val);
1336 }
1337 
1338 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64_XEXEC(unsigned Val) const {
1339   return decodeSrcOp(OPW64, Val);
1340 }
1341 
1342 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const {
1343   return decodeSrcOp(OPW128, Val);
1344 }
1345 
1346 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const {
1347   return decodeDstOp(OPW256, Val);
1348 }
1349 
1350 MCOperand AMDGPUDisassembler::decodeOperand_SReg_288(unsigned Val) const {
1351   return decodeDstOp(OPW288, Val);
1352 }
1353 
1354 MCOperand AMDGPUDisassembler::decodeOperand_SReg_320(unsigned Val) const {
1355   return decodeDstOp(OPW320, Val);
1356 }
1357 
1358 MCOperand AMDGPUDisassembler::decodeOperand_SReg_352(unsigned Val) const {
1359   return decodeDstOp(OPW352, Val);
1360 }
1361 
1362 MCOperand AMDGPUDisassembler::decodeOperand_SReg_384(unsigned Val) const {
1363   return decodeDstOp(OPW384, Val);
1364 }
1365 
1366 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const {
1367   return decodeDstOp(OPW512, Val);
1368 }
1369 
1370 // Decode Literals for insts which always have a literal in the encoding
1371 MCOperand
1372 AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
1373   if (HasLiteral) {
1374     assert(
1375         AMDGPU::hasVOPD(STI) &&
1376         "Should only decode multiple kimm with VOPD, check VSrc operand types");
1377     if (Literal != Val)
1378       return errOperand(Val, "More than one unique literal is illegal");
1379   }
1380   HasLiteral = true;
1381   Literal = Val;
1382   return MCOperand::createImm(Literal);
1383 }
1384 
1385 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
1386   // For now all literal constants are supposed to be unsigned integer
1387   // ToDo: deal with signed/unsigned 64-bit integer constants
1388   // ToDo: deal with float/double constants
1389   if (!HasLiteral) {
1390     if (Bytes.size() < 4) {
1391       return errOperand(0, "cannot read literal, inst bytes left " +
1392                         Twine(Bytes.size()));
1393     }
1394     HasLiteral = true;
1395     Literal = eatBytes<uint32_t>(Bytes);
1396   }
1397   return MCOperand::createImm(Literal);
1398 }
1399 
1400 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
1401   using namespace AMDGPU::EncValues;
1402 
1403   assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX);
1404   return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ?
1405     (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) :
1406     (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm)));
1407       // Cast prevents negative overflow.
1408 }
1409 
1410 static int64_t getInlineImmVal32(unsigned Imm) {
1411   switch (Imm) {
1412   case 240:
1413     return FloatToBits(0.5f);
1414   case 241:
1415     return FloatToBits(-0.5f);
1416   case 242:
1417     return FloatToBits(1.0f);
1418   case 243:
1419     return FloatToBits(-1.0f);
1420   case 244:
1421     return FloatToBits(2.0f);
1422   case 245:
1423     return FloatToBits(-2.0f);
1424   case 246:
1425     return FloatToBits(4.0f);
1426   case 247:
1427     return FloatToBits(-4.0f);
1428   case 248: // 1 / (2 * PI)
1429     return 0x3e22f983;
1430   default:
1431     llvm_unreachable("invalid fp inline imm");
1432   }
1433 }
1434 
1435 static int64_t getInlineImmVal64(unsigned Imm) {
1436   switch (Imm) {
1437   case 240:
1438     return DoubleToBits(0.5);
1439   case 241:
1440     return DoubleToBits(-0.5);
1441   case 242:
1442     return DoubleToBits(1.0);
1443   case 243:
1444     return DoubleToBits(-1.0);
1445   case 244:
1446     return DoubleToBits(2.0);
1447   case 245:
1448     return DoubleToBits(-2.0);
1449   case 246:
1450     return DoubleToBits(4.0);
1451   case 247:
1452     return DoubleToBits(-4.0);
1453   case 248: // 1 / (2 * PI)
1454     return 0x3fc45f306dc9c882;
1455   default:
1456     llvm_unreachable("invalid fp inline imm");
1457   }
1458 }
1459 
1460 static int64_t getInlineImmVal16(unsigned Imm) {
1461   switch (Imm) {
1462   case 240:
1463     return 0x3800;
1464   case 241:
1465     return 0xB800;
1466   case 242:
1467     return 0x3C00;
1468   case 243:
1469     return 0xBC00;
1470   case 244:
1471     return 0x4000;
1472   case 245:
1473     return 0xC000;
1474   case 246:
1475     return 0x4400;
1476   case 247:
1477     return 0xC400;
1478   case 248: // 1 / (2 * PI)
1479     return 0x3118;
1480   default:
1481     llvm_unreachable("invalid fp inline imm");
1482   }
1483 }
1484 
1485 MCOperand AMDGPUDisassembler::decodeFPImmed(OpWidthTy Width, unsigned Imm) {
1486   assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN
1487       && Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX);
1488 
1489   // ToDo: case 248: 1/(2*PI) - is allowed only on VI
1490   switch (Width) {
1491   case OPW32:
1492   case OPW128: // splat constants
1493   case OPW512:
1494   case OPW1024:
1495   case OPWV232:
1496     return MCOperand::createImm(getInlineImmVal32(Imm));
1497   case OPW64:
1498   case OPW256:
1499     return MCOperand::createImm(getInlineImmVal64(Imm));
1500   case OPW16:
1501   case OPWV216:
1502     return MCOperand::createImm(getInlineImmVal16(Imm));
1503   default:
1504     llvm_unreachable("implement me");
1505   }
1506 }
1507 
1508 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const {
1509   using namespace AMDGPU;
1510 
1511   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1512   switch (Width) {
1513   default: // fall
1514   case OPW32:
1515   case OPW16:
1516   case OPWV216:
1517     return VGPR_32RegClassID;
1518   case OPW64:
1519   case OPWV232: return VReg_64RegClassID;
1520   case OPW96: return VReg_96RegClassID;
1521   case OPW128: return VReg_128RegClassID;
1522   case OPW160: return VReg_160RegClassID;
1523   case OPW256: return VReg_256RegClassID;
1524   case OPW288: return VReg_288RegClassID;
1525   case OPW320: return VReg_320RegClassID;
1526   case OPW352: return VReg_352RegClassID;
1527   case OPW384: return VReg_384RegClassID;
1528   case OPW512: return VReg_512RegClassID;
1529   case OPW1024: return VReg_1024RegClassID;
1530   }
1531 }
1532 
1533 unsigned AMDGPUDisassembler::getAgprClassId(const OpWidthTy Width) const {
1534   using namespace AMDGPU;
1535 
1536   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1537   switch (Width) {
1538   default: // fall
1539   case OPW32:
1540   case OPW16:
1541   case OPWV216:
1542     return AGPR_32RegClassID;
1543   case OPW64:
1544   case OPWV232: return AReg_64RegClassID;
1545   case OPW96: return AReg_96RegClassID;
1546   case OPW128: return AReg_128RegClassID;
1547   case OPW160: return AReg_160RegClassID;
1548   case OPW256: return AReg_256RegClassID;
1549   case OPW288: return AReg_288RegClassID;
1550   case OPW320: return AReg_320RegClassID;
1551   case OPW352: return AReg_352RegClassID;
1552   case OPW384: return AReg_384RegClassID;
1553   case OPW512: return AReg_512RegClassID;
1554   case OPW1024: return AReg_1024RegClassID;
1555   }
1556 }
1557 
1558 
1559 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const {
1560   using namespace AMDGPU;
1561 
1562   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1563   switch (Width) {
1564   default: // fall
1565   case OPW32:
1566   case OPW16:
1567   case OPWV216:
1568     return SGPR_32RegClassID;
1569   case OPW64:
1570   case OPWV232: return SGPR_64RegClassID;
1571   case OPW96: return SGPR_96RegClassID;
1572   case OPW128: return SGPR_128RegClassID;
1573   case OPW160: return SGPR_160RegClassID;
1574   case OPW256: return SGPR_256RegClassID;
1575   case OPW288: return SGPR_288RegClassID;
1576   case OPW320: return SGPR_320RegClassID;
1577   case OPW352: return SGPR_352RegClassID;
1578   case OPW384: return SGPR_384RegClassID;
1579   case OPW512: return SGPR_512RegClassID;
1580   }
1581 }
1582 
1583 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const {
1584   using namespace AMDGPU;
1585 
1586   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1587   switch (Width) {
1588   default: // fall
1589   case OPW32:
1590   case OPW16:
1591   case OPWV216:
1592     return TTMP_32RegClassID;
1593   case OPW64:
1594   case OPWV232: return TTMP_64RegClassID;
1595   case OPW128: return TTMP_128RegClassID;
1596   case OPW256: return TTMP_256RegClassID;
1597   case OPW288: return TTMP_288RegClassID;
1598   case OPW320: return TTMP_320RegClassID;
1599   case OPW352: return TTMP_352RegClassID;
1600   case OPW384: return TTMP_384RegClassID;
1601   case OPW512: return TTMP_512RegClassID;
1602   }
1603 }
1604 
1605 int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
1606   using namespace AMDGPU::EncValues;
1607 
1608   unsigned TTmpMin = isGFX9Plus() ? TTMP_GFX9PLUS_MIN : TTMP_VI_MIN;
1609   unsigned TTmpMax = isGFX9Plus() ? TTMP_GFX9PLUS_MAX : TTMP_VI_MAX;
1610 
1611   return (TTmpMin <= Val && Val <= TTmpMax)? Val - TTmpMin : -1;
1612 }
1613 
1614 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
1615                                           bool MandatoryLiteral) const {
1616   using namespace AMDGPU::EncValues;
1617 
1618   assert(Val < 1024); // enum10
1619 
1620   bool IsAGPR = Val & 512;
1621   Val &= 511;
1622 
1623   if (VGPR_MIN <= Val && Val <= VGPR_MAX) {
1624     return createRegOperand(IsAGPR ? getAgprClassId(Width)
1625                                    : getVgprClassId(Width), Val - VGPR_MIN);
1626   }
1627   if (Val <= SGPR_MAX) {
1628     // "SGPR_MIN <= Val" is always true and causes compilation warning.
1629     static_assert(SGPR_MIN == 0);
1630     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
1631   }
1632 
1633   int TTmpIdx = getTTmpIdx(Val);
1634   if (TTmpIdx >= 0) {
1635     return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
1636   }
1637 
1638   if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX)
1639     return decodeIntImmed(Val);
1640 
1641   if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX)
1642     return decodeFPImmed(Width, Val);
1643 
1644   if (Val == LITERAL_CONST) {
1645     if (MandatoryLiteral)
1646       // Keep a sentinel value for deferred setting
1647       return MCOperand::createImm(LITERAL_CONST);
1648     else
1649       return decodeLiteralConstant();
1650   }
1651 
1652   switch (Width) {
1653   case OPW32:
1654   case OPW16:
1655   case OPWV216:
1656     return decodeSpecialReg32(Val);
1657   case OPW64:
1658   case OPWV232:
1659     return decodeSpecialReg64(Val);
1660   default:
1661     llvm_unreachable("unexpected immediate type");
1662   }
1663 }
1664 
1665 MCOperand AMDGPUDisassembler::decodeDstOp(const OpWidthTy Width, unsigned Val) const {
1666   using namespace AMDGPU::EncValues;
1667 
1668   assert(Val < 128);
1669   assert(Width == OPW256 || Width == OPW512);
1670 
1671   if (Val <= SGPR_MAX) {
1672     // "SGPR_MIN <= Val" is always true and causes compilation warning.
1673     static_assert(SGPR_MIN == 0);
1674     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
1675   }
1676 
1677   int TTmpIdx = getTTmpIdx(Val);
1678   if (TTmpIdx >= 0) {
1679     return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
1680   }
1681 
1682   llvm_unreachable("unknown dst register");
1683 }
1684 
1685 // Bit 0 of DstY isn't stored in the instruction, because it's always the
1686 // opposite of bit 0 of DstX.
1687 MCOperand AMDGPUDisassembler::decodeVOPDDstYOp(MCInst &Inst,
1688                                                unsigned Val) const {
1689   int VDstXInd =
1690       AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::vdstX);
1691   assert(VDstXInd != -1);
1692   assert(Inst.getOperand(VDstXInd).isReg());
1693   unsigned XDstReg = MRI.getEncodingValue(Inst.getOperand(VDstXInd).getReg());
1694   Val |= ~XDstReg & 1;
1695   auto Width = llvm::AMDGPUDisassembler::OPW32;
1696   return createRegOperand(getVgprClassId(Width), Val);
1697 }
1698 
1699 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
1700   using namespace AMDGPU;
1701 
1702   switch (Val) {
1703   // clang-format off
1704   case 102: return createRegOperand(FLAT_SCR_LO);
1705   case 103: return createRegOperand(FLAT_SCR_HI);
1706   case 104: return createRegOperand(XNACK_MASK_LO);
1707   case 105: return createRegOperand(XNACK_MASK_HI);
1708   case 106: return createRegOperand(VCC_LO);
1709   case 107: return createRegOperand(VCC_HI);
1710   case 108: return createRegOperand(TBA_LO);
1711   case 109: return createRegOperand(TBA_HI);
1712   case 110: return createRegOperand(TMA_LO);
1713   case 111: return createRegOperand(TMA_HI);
1714   case 124:
1715     return isGFX11Plus() ? createRegOperand(SGPR_NULL) : createRegOperand(M0);
1716   case 125:
1717     return isGFX11Plus() ? createRegOperand(M0) : createRegOperand(SGPR_NULL);
1718   case 126: return createRegOperand(EXEC_LO);
1719   case 127: return createRegOperand(EXEC_HI);
1720   case 235: return createRegOperand(SRC_SHARED_BASE_LO);
1721   case 236: return createRegOperand(SRC_SHARED_LIMIT_LO);
1722   case 237: return createRegOperand(SRC_PRIVATE_BASE_LO);
1723   case 238: return createRegOperand(SRC_PRIVATE_LIMIT_LO);
1724   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1725   case 251: return createRegOperand(SRC_VCCZ);
1726   case 252: return createRegOperand(SRC_EXECZ);
1727   case 253: return createRegOperand(SRC_SCC);
1728   case 254: return createRegOperand(LDS_DIRECT);
1729   default: break;
1730     // clang-format on
1731   }
1732   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1733 }
1734 
1735 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
1736   using namespace AMDGPU;
1737 
1738   switch (Val) {
1739   case 102: return createRegOperand(FLAT_SCR);
1740   case 104: return createRegOperand(XNACK_MASK);
1741   case 106: return createRegOperand(VCC);
1742   case 108: return createRegOperand(TBA);
1743   case 110: return createRegOperand(TMA);
1744   case 124:
1745     if (isGFX11Plus())
1746       return createRegOperand(SGPR_NULL);
1747     break;
1748   case 125:
1749     if (!isGFX11Plus())
1750       return createRegOperand(SGPR_NULL);
1751     break;
1752   case 126: return createRegOperand(EXEC);
1753   case 235: return createRegOperand(SRC_SHARED_BASE);
1754   case 236: return createRegOperand(SRC_SHARED_LIMIT);
1755   case 237: return createRegOperand(SRC_PRIVATE_BASE);
1756   case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
1757   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1758   case 251: return createRegOperand(SRC_VCCZ);
1759   case 252: return createRegOperand(SRC_EXECZ);
1760   case 253: return createRegOperand(SRC_SCC);
1761   default: break;
1762   }
1763   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1764 }
1765 
1766 MCOperand AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width,
1767                                             const unsigned Val) const {
1768   using namespace AMDGPU::SDWA;
1769   using namespace AMDGPU::EncValues;
1770 
1771   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
1772       STI.getFeatureBits()[AMDGPU::FeatureGFX10]) {
1773     // XXX: cast to int is needed to avoid stupid warning:
1774     // compare with unsigned is always true
1775     if (int(SDWA9EncValues::SRC_VGPR_MIN) <= int(Val) &&
1776         Val <= SDWA9EncValues::SRC_VGPR_MAX) {
1777       return createRegOperand(getVgprClassId(Width),
1778                               Val - SDWA9EncValues::SRC_VGPR_MIN);
1779     }
1780     if (SDWA9EncValues::SRC_SGPR_MIN <= Val &&
1781         Val <= (isGFX10Plus() ? SDWA9EncValues::SRC_SGPR_MAX_GFX10
1782                               : SDWA9EncValues::SRC_SGPR_MAX_SI)) {
1783       return createSRegOperand(getSgprClassId(Width),
1784                                Val - SDWA9EncValues::SRC_SGPR_MIN);
1785     }
1786     if (SDWA9EncValues::SRC_TTMP_MIN <= Val &&
1787         Val <= SDWA9EncValues::SRC_TTMP_MAX) {
1788       return createSRegOperand(getTtmpClassId(Width),
1789                                Val - SDWA9EncValues::SRC_TTMP_MIN);
1790     }
1791 
1792     const unsigned SVal = Val - SDWA9EncValues::SRC_SGPR_MIN;
1793 
1794     if (INLINE_INTEGER_C_MIN <= SVal && SVal <= INLINE_INTEGER_C_MAX)
1795       return decodeIntImmed(SVal);
1796 
1797     if (INLINE_FLOATING_C_MIN <= SVal && SVal <= INLINE_FLOATING_C_MAX)
1798       return decodeFPImmed(Width, SVal);
1799 
1800     return decodeSpecialReg32(SVal);
1801   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
1802     return createRegOperand(getVgprClassId(Width), Val);
1803   }
1804   llvm_unreachable("unsupported target");
1805 }
1806 
1807 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const {
1808   return decodeSDWASrc(OPW16, Val);
1809 }
1810 
1811 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const {
1812   return decodeSDWASrc(OPW32, Val);
1813 }
1814 
1815 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const {
1816   using namespace AMDGPU::SDWA;
1817 
1818   assert((STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
1819           STI.getFeatureBits()[AMDGPU::FeatureGFX10]) &&
1820          "SDWAVopcDst should be present only on GFX9+");
1821 
1822   bool IsWave64 = STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64];
1823 
1824   if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) {
1825     Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
1826 
1827     int TTmpIdx = getTTmpIdx(Val);
1828     if (TTmpIdx >= 0) {
1829       auto TTmpClsId = getTtmpClassId(IsWave64 ? OPW64 : OPW32);
1830       return createSRegOperand(TTmpClsId, TTmpIdx);
1831     } else if (Val > SGPR_MAX) {
1832       return IsWave64 ? decodeSpecialReg64(Val)
1833                       : decodeSpecialReg32(Val);
1834     } else {
1835       return createSRegOperand(getSgprClassId(IsWave64 ? OPW64 : OPW32), Val);
1836     }
1837   } else {
1838     return createRegOperand(IsWave64 ? AMDGPU::VCC : AMDGPU::VCC_LO);
1839   }
1840 }
1841 
1842 MCOperand AMDGPUDisassembler::decodeBoolReg(unsigned Val) const {
1843   return STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64] ?
1844     decodeOperand_SReg_64(Val) : decodeOperand_SReg_32(Val);
1845 }
1846 
1847 bool AMDGPUDisassembler::isVI() const {
1848   return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
1849 }
1850 
1851 bool AMDGPUDisassembler::isGFX9() const { return AMDGPU::isGFX9(STI); }
1852 
1853 bool AMDGPUDisassembler::isGFX90A() const {
1854   return STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts];
1855 }
1856 
1857 bool AMDGPUDisassembler::isGFX9Plus() const { return AMDGPU::isGFX9Plus(STI); }
1858 
1859 bool AMDGPUDisassembler::isGFX10() const { return AMDGPU::isGFX10(STI); }
1860 
1861 bool AMDGPUDisassembler::isGFX10Plus() const {
1862   return AMDGPU::isGFX10Plus(STI);
1863 }
1864 
1865 bool AMDGPUDisassembler::isGFX11() const {
1866   return STI.getFeatureBits()[AMDGPU::FeatureGFX11];
1867 }
1868 
1869 bool AMDGPUDisassembler::isGFX11Plus() const {
1870   return AMDGPU::isGFX11Plus(STI);
1871 }
1872 
1873 
1874 bool AMDGPUDisassembler::hasArchitectedFlatScratch() const {
1875   return STI.getFeatureBits()[AMDGPU::FeatureArchitectedFlatScratch];
1876 }
1877 
1878 //===----------------------------------------------------------------------===//
1879 // AMDGPU specific symbol handling
1880 //===----------------------------------------------------------------------===//
1881 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
1882   do {                                                                         \
1883     KdStream << Indent << DIRECTIVE " "                                        \
1884              << ((FourByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';           \
1885   } while (0)
1886 
1887 // NOLINTNEXTLINE(readability-identifier-naming)
1888 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC1(
1889     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
1890   using namespace amdhsa;
1891   StringRef Indent = "\t";
1892 
1893   // We cannot accurately backward compute #VGPRs used from
1894   // GRANULATED_WORKITEM_VGPR_COUNT. But we are concerned with getting the same
1895   // value of GRANULATED_WORKITEM_VGPR_COUNT in the reassembled binary. So we
1896   // simply calculate the inverse of what the assembler does.
1897 
1898   uint32_t GranulatedWorkitemVGPRCount =
1899       (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT) >>
1900       COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT_SHIFT;
1901 
1902   uint32_t NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) *
1903                           AMDGPU::IsaInfo::getVGPREncodingGranule(&STI);
1904 
1905   KdStream << Indent << ".amdhsa_next_free_vgpr " << NextFreeVGPR << '\n';
1906 
1907   // We cannot backward compute values used to calculate
1908   // GRANULATED_WAVEFRONT_SGPR_COUNT. Hence the original values for following
1909   // directives can't be computed:
1910   // .amdhsa_reserve_vcc
1911   // .amdhsa_reserve_flat_scratch
1912   // .amdhsa_reserve_xnack_mask
1913   // They take their respective default values if not specified in the assembly.
1914   //
1915   // GRANULATED_WAVEFRONT_SGPR_COUNT
1916   //    = f(NEXT_FREE_SGPR + VCC + FLAT_SCRATCH + XNACK_MASK)
1917   //
1918   // We compute the inverse as though all directives apart from NEXT_FREE_SGPR
1919   // are set to 0. So while disassembling we consider that:
1920   //
1921   // GRANULATED_WAVEFRONT_SGPR_COUNT
1922   //    = f(NEXT_FREE_SGPR + 0 + 0 + 0)
1923   //
1924   // The disassembler cannot recover the original values of those 3 directives.
1925 
1926   uint32_t GranulatedWavefrontSGPRCount =
1927       (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT) >>
1928       COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT_SHIFT;
1929 
1930   if (isGFX10Plus() && GranulatedWavefrontSGPRCount)
1931     return MCDisassembler::Fail;
1932 
1933   uint32_t NextFreeSGPR = (GranulatedWavefrontSGPRCount + 1) *
1934                           AMDGPU::IsaInfo::getSGPREncodingGranule(&STI);
1935 
1936   KdStream << Indent << ".amdhsa_reserve_vcc " << 0 << '\n';
1937   if (!hasArchitectedFlatScratch())
1938     KdStream << Indent << ".amdhsa_reserve_flat_scratch " << 0 << '\n';
1939   KdStream << Indent << ".amdhsa_reserve_xnack_mask " << 0 << '\n';
1940   KdStream << Indent << ".amdhsa_next_free_sgpr " << NextFreeSGPR << "\n";
1941 
1942   if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIORITY)
1943     return MCDisassembler::Fail;
1944 
1945   PRINT_DIRECTIVE(".amdhsa_float_round_mode_32",
1946                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
1947   PRINT_DIRECTIVE(".amdhsa_float_round_mode_16_64",
1948                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
1949   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_32",
1950                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
1951   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_16_64",
1952                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
1953 
1954   if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIV)
1955     return MCDisassembler::Fail;
1956 
1957   PRINT_DIRECTIVE(".amdhsa_dx10_clamp", COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP);
1958 
1959   if (FourByteBuffer & COMPUTE_PGM_RSRC1_DEBUG_MODE)
1960     return MCDisassembler::Fail;
1961 
1962   PRINT_DIRECTIVE(".amdhsa_ieee_mode", COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE);
1963 
1964   if (FourByteBuffer & COMPUTE_PGM_RSRC1_BULKY)
1965     return MCDisassembler::Fail;
1966 
1967   if (FourByteBuffer & COMPUTE_PGM_RSRC1_CDBG_USER)
1968     return MCDisassembler::Fail;
1969 
1970   PRINT_DIRECTIVE(".amdhsa_fp16_overflow", COMPUTE_PGM_RSRC1_FP16_OVFL);
1971 
1972   if (FourByteBuffer & COMPUTE_PGM_RSRC1_RESERVED0)
1973     return MCDisassembler::Fail;
1974 
1975   if (isGFX10Plus()) {
1976     PRINT_DIRECTIVE(".amdhsa_workgroup_processor_mode",
1977                     COMPUTE_PGM_RSRC1_WGP_MODE);
1978     PRINT_DIRECTIVE(".amdhsa_memory_ordered", COMPUTE_PGM_RSRC1_MEM_ORDERED);
1979     PRINT_DIRECTIVE(".amdhsa_forward_progress", COMPUTE_PGM_RSRC1_FWD_PROGRESS);
1980   }
1981   return MCDisassembler::Success;
1982 }
1983 
1984 // NOLINTNEXTLINE(readability-identifier-naming)
1985 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC2(
1986     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
1987   using namespace amdhsa;
1988   StringRef Indent = "\t";
1989   if (hasArchitectedFlatScratch())
1990     PRINT_DIRECTIVE(".amdhsa_enable_private_segment",
1991                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
1992   else
1993     PRINT_DIRECTIVE(".amdhsa_system_sgpr_private_segment_wavefront_offset",
1994                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
1995   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_x",
1996                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
1997   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_y",
1998                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
1999   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_z",
2000                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
2001   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_info",
2002                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
2003   PRINT_DIRECTIVE(".amdhsa_system_vgpr_workitem_id",
2004                   COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
2005 
2006   if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_ADDRESS_WATCH)
2007     return MCDisassembler::Fail;
2008 
2009   if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_MEMORY)
2010     return MCDisassembler::Fail;
2011 
2012   if (FourByteBuffer & COMPUTE_PGM_RSRC2_GRANULATED_LDS_SIZE)
2013     return MCDisassembler::Fail;
2014 
2015   PRINT_DIRECTIVE(
2016       ".amdhsa_exception_fp_ieee_invalid_op",
2017       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
2018   PRINT_DIRECTIVE(".amdhsa_exception_fp_denorm_src",
2019                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
2020   PRINT_DIRECTIVE(
2021       ".amdhsa_exception_fp_ieee_div_zero",
2022       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
2023   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_overflow",
2024                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
2025   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_underflow",
2026                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
2027   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_inexact",
2028                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
2029   PRINT_DIRECTIVE(".amdhsa_exception_int_div_zero",
2030                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
2031 
2032   if (FourByteBuffer & COMPUTE_PGM_RSRC2_RESERVED0)
2033     return MCDisassembler::Fail;
2034 
2035   return MCDisassembler::Success;
2036 }
2037 
2038 #undef PRINT_DIRECTIVE
2039 
2040 MCDisassembler::DecodeStatus
2041 AMDGPUDisassembler::decodeKernelDescriptorDirective(
2042     DataExtractor::Cursor &Cursor, ArrayRef<uint8_t> Bytes,
2043     raw_string_ostream &KdStream) const {
2044 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
2045   do {                                                                         \
2046     KdStream << Indent << DIRECTIVE " "                                        \
2047              << ((TwoByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';            \
2048   } while (0)
2049 
2050   uint16_t TwoByteBuffer = 0;
2051   uint32_t FourByteBuffer = 0;
2052 
2053   StringRef ReservedBytes;
2054   StringRef Indent = "\t";
2055 
2056   assert(Bytes.size() == 64);
2057   DataExtractor DE(Bytes, /*IsLittleEndian=*/true, /*AddressSize=*/8);
2058 
2059   switch (Cursor.tell()) {
2060   case amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET:
2061     FourByteBuffer = DE.getU32(Cursor);
2062     KdStream << Indent << ".amdhsa_group_segment_fixed_size " << FourByteBuffer
2063              << '\n';
2064     return MCDisassembler::Success;
2065 
2066   case amdhsa::PRIVATE_SEGMENT_FIXED_SIZE_OFFSET:
2067     FourByteBuffer = DE.getU32(Cursor);
2068     KdStream << Indent << ".amdhsa_private_segment_fixed_size "
2069              << FourByteBuffer << '\n';
2070     return MCDisassembler::Success;
2071 
2072   case amdhsa::KERNARG_SIZE_OFFSET:
2073     FourByteBuffer = DE.getU32(Cursor);
2074     KdStream << Indent << ".amdhsa_kernarg_size "
2075              << FourByteBuffer << '\n';
2076     return MCDisassembler::Success;
2077 
2078   case amdhsa::RESERVED0_OFFSET:
2079     // 4 reserved bytes, must be 0.
2080     ReservedBytes = DE.getBytes(Cursor, 4);
2081     for (int I = 0; I < 4; ++I) {
2082       if (ReservedBytes[I] != 0) {
2083         return MCDisassembler::Fail;
2084       }
2085     }
2086     return MCDisassembler::Success;
2087 
2088   case amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET:
2089     // KERNEL_CODE_ENTRY_BYTE_OFFSET
2090     // So far no directive controls this for Code Object V3, so simply skip for
2091     // disassembly.
2092     DE.skip(Cursor, 8);
2093     return MCDisassembler::Success;
2094 
2095   case amdhsa::RESERVED1_OFFSET:
2096     // 20 reserved bytes, must be 0.
2097     ReservedBytes = DE.getBytes(Cursor, 20);
2098     for (int I = 0; I < 20; ++I) {
2099       if (ReservedBytes[I] != 0) {
2100         return MCDisassembler::Fail;
2101       }
2102     }
2103     return MCDisassembler::Success;
2104 
2105   case amdhsa::COMPUTE_PGM_RSRC3_OFFSET:
2106     // COMPUTE_PGM_RSRC3
2107     //  - Only set for GFX10, GFX6-9 have this to be 0.
2108     //  - Currently no directives directly control this.
2109     FourByteBuffer = DE.getU32(Cursor);
2110     if (!isGFX10Plus() && FourByteBuffer) {
2111       return MCDisassembler::Fail;
2112     }
2113     return MCDisassembler::Success;
2114 
2115   case amdhsa::COMPUTE_PGM_RSRC1_OFFSET:
2116     FourByteBuffer = DE.getU32(Cursor);
2117     if (decodeCOMPUTE_PGM_RSRC1(FourByteBuffer, KdStream) ==
2118         MCDisassembler::Fail) {
2119       return MCDisassembler::Fail;
2120     }
2121     return MCDisassembler::Success;
2122 
2123   case amdhsa::COMPUTE_PGM_RSRC2_OFFSET:
2124     FourByteBuffer = DE.getU32(Cursor);
2125     if (decodeCOMPUTE_PGM_RSRC2(FourByteBuffer, KdStream) ==
2126         MCDisassembler::Fail) {
2127       return MCDisassembler::Fail;
2128     }
2129     return MCDisassembler::Success;
2130 
2131   case amdhsa::KERNEL_CODE_PROPERTIES_OFFSET:
2132     using namespace amdhsa;
2133     TwoByteBuffer = DE.getU16(Cursor);
2134 
2135     if (!hasArchitectedFlatScratch())
2136       PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_buffer",
2137                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
2138     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_ptr",
2139                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
2140     PRINT_DIRECTIVE(".amdhsa_user_sgpr_queue_ptr",
2141                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
2142     PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_segment_ptr",
2143                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
2144     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_id",
2145                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
2146     if (!hasArchitectedFlatScratch())
2147       PRINT_DIRECTIVE(".amdhsa_user_sgpr_flat_scratch_init",
2148                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
2149     PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size",
2150                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
2151 
2152     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0)
2153       return MCDisassembler::Fail;
2154 
2155     // Reserved for GFX9
2156     if (isGFX9() &&
2157         (TwoByteBuffer & KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32)) {
2158       return MCDisassembler::Fail;
2159     } else if (isGFX10Plus()) {
2160       PRINT_DIRECTIVE(".amdhsa_wavefront_size32",
2161                       KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
2162     }
2163 
2164     if (AMDGPU::getAmdhsaCodeObjectVersion() >= 5)
2165       PRINT_DIRECTIVE(".amdhsa_uses_dynamic_stack",
2166                       KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK);
2167 
2168     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED1)
2169       return MCDisassembler::Fail;
2170 
2171     return MCDisassembler::Success;
2172 
2173   case amdhsa::RESERVED2_OFFSET:
2174     // 6 bytes from here are reserved, must be 0.
2175     ReservedBytes = DE.getBytes(Cursor, 6);
2176     for (int I = 0; I < 6; ++I) {
2177       if (ReservedBytes[I] != 0)
2178         return MCDisassembler::Fail;
2179     }
2180     return MCDisassembler::Success;
2181 
2182   default:
2183     llvm_unreachable("Unhandled index. Case statements cover everything.");
2184     return MCDisassembler::Fail;
2185   }
2186 #undef PRINT_DIRECTIVE
2187 }
2188 
2189 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeKernelDescriptor(
2190     StringRef KdName, ArrayRef<uint8_t> Bytes, uint64_t KdAddress) const {
2191   // CP microcode requires the kernel descriptor to be 64 aligned.
2192   if (Bytes.size() != 64 || KdAddress % 64 != 0)
2193     return MCDisassembler::Fail;
2194 
2195   std::string Kd;
2196   raw_string_ostream KdStream(Kd);
2197   KdStream << ".amdhsa_kernel " << KdName << '\n';
2198 
2199   DataExtractor::Cursor C(0);
2200   while (C && C.tell() < Bytes.size()) {
2201     MCDisassembler::DecodeStatus Status =
2202         decodeKernelDescriptorDirective(C, Bytes, KdStream);
2203 
2204     cantFail(C.takeError());
2205 
2206     if (Status == MCDisassembler::Fail)
2207       return MCDisassembler::Fail;
2208   }
2209   KdStream << ".end_amdhsa_kernel\n";
2210   outs() << KdStream.str();
2211   return MCDisassembler::Success;
2212 }
2213 
2214 std::optional<MCDisassembler::DecodeStatus>
2215 AMDGPUDisassembler::onSymbolStart(SymbolInfoTy &Symbol, uint64_t &Size,
2216                                   ArrayRef<uint8_t> Bytes, uint64_t Address,
2217                                   raw_ostream &CStream) const {
2218   // Right now only kernel descriptor needs to be handled.
2219   // We ignore all other symbols for target specific handling.
2220   // TODO:
2221   // Fix the spurious symbol issue for AMDGPU kernels. Exists for both Code
2222   // Object V2 and V3 when symbols are marked protected.
2223 
2224   // amd_kernel_code_t for Code Object V2.
2225   if (Symbol.Type == ELF::STT_AMDGPU_HSA_KERNEL) {
2226     Size = 256;
2227     return MCDisassembler::Fail;
2228   }
2229 
2230   // Code Object V3 kernel descriptors.
2231   StringRef Name = Symbol.Name;
2232   if (Symbol.Type == ELF::STT_OBJECT && Name.endswith(StringRef(".kd"))) {
2233     Size = 64; // Size = 64 regardless of success or failure.
2234     return decodeKernelDescriptor(Name.drop_back(3), Bytes, Address);
2235   }
2236   return std::nullopt;
2237 }
2238 
2239 //===----------------------------------------------------------------------===//
2240 // AMDGPUSymbolizer
2241 //===----------------------------------------------------------------------===//
2242 
2243 // Try to find symbol name for specified label
2244 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(
2245     MCInst &Inst, raw_ostream & /*cStream*/, int64_t Value,
2246     uint64_t /*Address*/, bool IsBranch, uint64_t /*Offset*/,
2247     uint64_t /*OpSize*/, uint64_t /*InstSize*/) {
2248 
2249   if (!IsBranch) {
2250     return false;
2251   }
2252 
2253   auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
2254   if (!Symbols)
2255     return false;
2256 
2257   auto Result = llvm::find_if(*Symbols, [Value](const SymbolInfoTy &Val) {
2258     return Val.Addr == static_cast<uint64_t>(Value) &&
2259            Val.Type == ELF::STT_NOTYPE;
2260   });
2261   if (Result != Symbols->end()) {
2262     auto *Sym = Ctx.getOrCreateSymbol(Result->Name);
2263     const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
2264     Inst.addOperand(MCOperand::createExpr(Add));
2265     return true;
2266   }
2267   // Add to list of referenced addresses, so caller can synthesize a label.
2268   ReferencedAddresses.push_back(static_cast<uint64_t>(Value));
2269   return false;
2270 }
2271 
2272 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
2273                                                        int64_t Value,
2274                                                        uint64_t Address) {
2275   llvm_unreachable("unimplemented");
2276 }
2277 
2278 //===----------------------------------------------------------------------===//
2279 // Initialization
2280 //===----------------------------------------------------------------------===//
2281 
2282 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
2283                               LLVMOpInfoCallback /*GetOpInfo*/,
2284                               LLVMSymbolLookupCallback /*SymbolLookUp*/,
2285                               void *DisInfo,
2286                               MCContext *Ctx,
2287                               std::unique_ptr<MCRelocationInfo> &&RelInfo) {
2288   return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
2289 }
2290 
2291 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
2292                                                 const MCSubtargetInfo &STI,
2293                                                 MCContext &Ctx) {
2294   return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo());
2295 }
2296 
2297 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUDisassembler() {
2298   TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
2299                                          createAMDGPUDisassembler);
2300   TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
2301                                        createAMDGPUSymbolizer);
2302 }
2303