xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
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(VRegOrLds_32)
123 DECODE_OPERAND_REG(VS_32)
124 DECODE_OPERAND_REG(VS_64)
125 DECODE_OPERAND_REG(VS_128)
126 
127 DECODE_OPERAND_REG(VReg_64)
128 DECODE_OPERAND_REG(VReg_96)
129 DECODE_OPERAND_REG(VReg_128)
130 DECODE_OPERAND_REG(VReg_256)
131 DECODE_OPERAND_REG(VReg_512)
132 DECODE_OPERAND_REG(VReg_1024)
133 
134 DECODE_OPERAND_REG(SReg_32)
135 DECODE_OPERAND_REG(SReg_32_XM0_XEXEC)
136 DECODE_OPERAND_REG(SReg_32_XEXEC_HI)
137 DECODE_OPERAND_REG(SRegOrLds_32)
138 DECODE_OPERAND_REG(SReg_64)
139 DECODE_OPERAND_REG(SReg_64_XEXEC)
140 DECODE_OPERAND_REG(SReg_128)
141 DECODE_OPERAND_REG(SReg_256)
142 DECODE_OPERAND_REG(SReg_512)
143 
144 DECODE_OPERAND_REG(AGPR_32)
145 DECODE_OPERAND_REG(AReg_64)
146 DECODE_OPERAND_REG(AReg_128)
147 DECODE_OPERAND_REG(AReg_256)
148 DECODE_OPERAND_REG(AReg_512)
149 DECODE_OPERAND_REG(AReg_1024)
150 DECODE_OPERAND_REG(AV_32)
151 DECODE_OPERAND_REG(AV_64)
152 DECODE_OPERAND_REG(AV_128)
153 DECODE_OPERAND_REG(AVDst_128)
154 DECODE_OPERAND_REG(AVDst_512)
155 
156 static DecodeStatus decodeOperand_VSrc16(MCInst &Inst, unsigned Imm,
157                                          uint64_t Addr,
158                                          const MCDisassembler *Decoder) {
159   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
160   return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
161 }
162 
163 static DecodeStatus decodeOperand_VSrcV216(MCInst &Inst, unsigned Imm,
164                                            uint64_t Addr,
165                                            const MCDisassembler *Decoder) {
166   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
167   return addOperand(Inst, DAsm->decodeOperand_VSrcV216(Imm));
168 }
169 
170 static DecodeStatus decodeOperand_VSrcV232(MCInst &Inst, unsigned Imm,
171                                            uint64_t Addr,
172                                            const MCDisassembler *Decoder) {
173   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
174   return addOperand(Inst, DAsm->decodeOperand_VSrcV232(Imm));
175 }
176 
177 static DecodeStatus decodeOperand_VS_16(MCInst &Inst, unsigned Imm,
178                                         uint64_t Addr,
179                                         const MCDisassembler *Decoder) {
180   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
181   return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
182 }
183 
184 static DecodeStatus decodeOperand_VS_32(MCInst &Inst, unsigned Imm,
185                                         uint64_t Addr,
186                                         const MCDisassembler *Decoder) {
187   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
188   return addOperand(Inst, DAsm->decodeOperand_VS_32(Imm));
189 }
190 
191 static DecodeStatus decodeOperand_AReg_64(MCInst &Inst, unsigned Imm,
192                                           uint64_t Addr,
193                                           const MCDisassembler *Decoder) {
194   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
195   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm | 512));
196 }
197 
198 static DecodeStatus decodeOperand_AReg_128(MCInst &Inst, unsigned Imm,
199                                            uint64_t Addr,
200                                            const MCDisassembler *Decoder) {
201   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
202   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm | 512));
203 }
204 
205 static DecodeStatus decodeOperand_AReg_256(MCInst &Inst, unsigned Imm,
206                                            uint64_t Addr,
207                                            const MCDisassembler *Decoder) {
208   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
209   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm | 512));
210 }
211 
212 static DecodeStatus decodeOperand_AReg_512(MCInst &Inst, unsigned Imm,
213                                            uint64_t Addr,
214                                            const MCDisassembler *Decoder) {
215   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
216   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm | 512));
217 }
218 
219 static DecodeStatus decodeOperand_AReg_1024(MCInst &Inst, unsigned Imm,
220                                             uint64_t Addr,
221                                             const MCDisassembler *Decoder) {
222   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
223   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm | 512));
224 }
225 
226 static DecodeStatus decodeOperand_VReg_64(MCInst &Inst, unsigned Imm,
227                                           uint64_t Addr,
228                                           const MCDisassembler *Decoder) {
229   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
230   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm));
231 }
232 
233 static DecodeStatus decodeOperand_VReg_128(MCInst &Inst, unsigned Imm,
234                                            uint64_t Addr,
235                                            const MCDisassembler *Decoder) {
236   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
237   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW128, Imm));
238 }
239 
240 static DecodeStatus decodeOperand_VReg_256(MCInst &Inst, unsigned Imm,
241                                            uint64_t Addr,
242                                            const MCDisassembler *Decoder) {
243   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
244   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW256, Imm));
245 }
246 
247 static DecodeStatus decodeOperand_VReg_512(MCInst &Inst, unsigned Imm,
248                                            uint64_t Addr,
249                                            const MCDisassembler *Decoder) {
250   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
251   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW512, Imm));
252 }
253 
254 static DecodeStatus decodeOperand_VReg_1024(MCInst &Inst, unsigned Imm,
255                                             uint64_t Addr,
256                                             const MCDisassembler *Decoder) {
257   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
258   return addOperand(Inst, DAsm->decodeSrcOp(AMDGPUDisassembler::OPW1024, Imm));
259 }
260 
261 static DecodeStatus decodeOperand_f32kimm(MCInst &Inst, unsigned Imm,
262                                           uint64_t Addr,
263                                           const MCDisassembler *Decoder) {
264   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
265   return addOperand(Inst, DAsm->decodeMandatoryLiteralConstant(Imm));
266 }
267 
268 static DecodeStatus decodeOperand_f16kimm(MCInst &Inst, unsigned Imm,
269                                           uint64_t Addr,
270                                           const MCDisassembler *Decoder) {
271   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
272   return addOperand(Inst, DAsm->decodeMandatoryLiteralConstant(Imm));
273 }
274 
275 static DecodeStatus
276 decodeOperand_VS_16_Deferred(MCInst &Inst, unsigned Imm, uint64_t Addr,
277                              const MCDisassembler *Decoder) {
278   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
279   return addOperand(
280       Inst, DAsm->decodeSrcOp(llvm::AMDGPUDisassembler::OPW16, Imm, true));
281 }
282 
283 static DecodeStatus
284 decodeOperand_VS_32_Deferred(MCInst &Inst, unsigned Imm, uint64_t Addr,
285                              const MCDisassembler *Decoder) {
286   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
287   return addOperand(
288       Inst, DAsm->decodeSrcOp(llvm::AMDGPUDisassembler::OPW32, Imm, true));
289 }
290 
291 static DecodeStatus decodeOperandVOPDDstY(MCInst &Inst, unsigned Val,
292                                           uint64_t Addr, const void *Decoder) {
293   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
294   return addOperand(Inst, DAsm->decodeVOPDDstYOp(Inst, Val));
295 }
296 
297 static bool IsAGPROperand(const MCInst &Inst, int OpIdx,
298                           const MCRegisterInfo *MRI) {
299   if (OpIdx < 0)
300     return false;
301 
302   const MCOperand &Op = Inst.getOperand(OpIdx);
303   if (!Op.isReg())
304     return false;
305 
306   unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
307   auto Reg = Sub ? Sub : Op.getReg();
308   return Reg >= AMDGPU::AGPR0 && Reg <= AMDGPU::AGPR255;
309 }
310 
311 static DecodeStatus decodeOperand_AVLdSt_Any(MCInst &Inst, unsigned Imm,
312                                              AMDGPUDisassembler::OpWidthTy Opw,
313                                              const MCDisassembler *Decoder) {
314   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
315   if (!DAsm->isGFX90A()) {
316     Imm &= 511;
317   } else {
318     // If atomic has both vdata and vdst their register classes are tied.
319     // The bit is decoded along with the vdst, first operand. We need to
320     // change register class to AGPR if vdst was AGPR.
321     // If a DS instruction has both data0 and data1 their register classes
322     // are also tied.
323     unsigned Opc = Inst.getOpcode();
324     uint64_t TSFlags = DAsm->getMCII()->get(Opc).TSFlags;
325     uint16_t DataNameIdx = (TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0
326                                                         : AMDGPU::OpName::vdata;
327     const MCRegisterInfo *MRI = DAsm->getContext().getRegisterInfo();
328     int DataIdx = AMDGPU::getNamedOperandIdx(Opc, DataNameIdx);
329     if ((int)Inst.getNumOperands() == DataIdx) {
330       int DstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
331       if (IsAGPROperand(Inst, DstIdx, MRI))
332         Imm |= 512;
333     }
334 
335     if (TSFlags & SIInstrFlags::DS) {
336       int Data2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
337       if ((int)Inst.getNumOperands() == Data2Idx &&
338           IsAGPROperand(Inst, DataIdx, MRI))
339         Imm |= 512;
340     }
341   }
342   return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
343 }
344 
345 static DecodeStatus
346 DecodeAVLdSt_32RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
347                              const MCDisassembler *Decoder) {
348   return decodeOperand_AVLdSt_Any(Inst, Imm,
349                                   AMDGPUDisassembler::OPW32, Decoder);
350 }
351 
352 static DecodeStatus
353 DecodeAVLdSt_64RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
354                              const MCDisassembler *Decoder) {
355   return decodeOperand_AVLdSt_Any(Inst, Imm,
356                                   AMDGPUDisassembler::OPW64, Decoder);
357 }
358 
359 static DecodeStatus
360 DecodeAVLdSt_96RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
361                              const MCDisassembler *Decoder) {
362   return decodeOperand_AVLdSt_Any(Inst, Imm,
363                                   AMDGPUDisassembler::OPW96, Decoder);
364 }
365 
366 static DecodeStatus
367 DecodeAVLdSt_128RegisterClass(MCInst &Inst, unsigned Imm, uint64_t Addr,
368                               const MCDisassembler *Decoder) {
369   return decodeOperand_AVLdSt_Any(Inst, Imm,
370                                   AMDGPUDisassembler::OPW128, Decoder);
371 }
372 
373 static DecodeStatus decodeOperand_SReg_32(MCInst &Inst, unsigned Imm,
374                                           uint64_t Addr,
375                                           const MCDisassembler *Decoder) {
376   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
377   return addOperand(Inst, DAsm->decodeOperand_SReg_32(Imm));
378 }
379 
380 #define DECODE_SDWA(DecName) \
381 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName)
382 
383 DECODE_SDWA(Src32)
384 DECODE_SDWA(Src16)
385 DECODE_SDWA(VopcDst)
386 
387 #include "AMDGPUGenDisassemblerTables.inc"
388 
389 //===----------------------------------------------------------------------===//
390 //
391 //===----------------------------------------------------------------------===//
392 
393 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
394   assert(Bytes.size() >= sizeof(T));
395   const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data());
396   Bytes = Bytes.slice(sizeof(T));
397   return Res;
398 }
399 
400 static inline DecoderUInt128 eat12Bytes(ArrayRef<uint8_t> &Bytes) {
401   assert(Bytes.size() >= 12);
402   uint64_t Lo = support::endian::read<uint64_t, support::endianness::little>(
403       Bytes.data());
404   Bytes = Bytes.slice(8);
405   uint64_t Hi = support::endian::read<uint32_t, support::endianness::little>(
406       Bytes.data());
407   Bytes = Bytes.slice(4);
408   return DecoderUInt128(Lo, Hi);
409 }
410 
411 // The disassembler is greedy, so we need to check FI operand value to
412 // not parse a dpp if the correct literal is not set. For dpp16 the
413 // autogenerated decoder checks the dpp literal
414 static bool isValidDPP8(const MCInst &MI) {
415   using namespace llvm::AMDGPU::DPP;
416   int FiIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::fi);
417   assert(FiIdx != -1);
418   if ((unsigned)FiIdx >= MI.getNumOperands())
419     return false;
420   unsigned Fi = MI.getOperand(FiIdx).getImm();
421   return Fi == DPP8_FI_0 || Fi == DPP8_FI_1;
422 }
423 
424 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
425                                                 ArrayRef<uint8_t> Bytes_,
426                                                 uint64_t Address,
427                                                 raw_ostream &CS) const {
428   CommentStream = &CS;
429   bool IsSDWA = false;
430 
431   unsigned MaxInstBytesNum = std::min((size_t)TargetMaxInstBytes, Bytes_.size());
432   Bytes = Bytes_.slice(0, MaxInstBytesNum);
433 
434   DecodeStatus Res = MCDisassembler::Fail;
435   do {
436     // ToDo: better to switch encoding length using some bit predicate
437     // but it is unknown yet, so try all we can
438 
439     // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2
440     // encodings
441     if (isGFX11Plus() && Bytes.size() >= 12 ) {
442       DecoderUInt128 DecW = eat12Bytes(Bytes);
443       Res = tryDecodeInst(DecoderTableDPP8GFX1196, MI, DecW,
444                                           Address);
445       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
446         break;
447       MI = MCInst(); // clear
448       Res = tryDecodeInst(DecoderTableDPPGFX1196, MI, DecW,
449                                           Address);
450       if (Res) {
451         if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3P)
452           convertVOP3PDPPInst(MI);
453         else if (AMDGPU::isVOPC64DPP(MI.getOpcode()))
454           convertVOPCDPPInst(MI); // Special VOP3 case
455         else {
456           assert(MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3);
457           convertVOP3DPPInst(MI); // Regular VOP3 case
458         }
459         break;
460       }
461       Res = tryDecodeInst(DecoderTableGFX1196, MI, DecW, Address);
462       if (Res)
463         break;
464     }
465     // Reinitialize Bytes
466     Bytes = Bytes_.slice(0, MaxInstBytesNum);
467 
468     if (Bytes.size() >= 8) {
469       const uint64_t QW = eatBytes<uint64_t>(Bytes);
470 
471       if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) {
472         Res = tryDecodeInst(DecoderTableGFX10_B64, MI, QW, Address);
473         if (Res) {
474           if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dpp8)
475               == -1)
476             break;
477           if (convertDPP8Inst(MI) == MCDisassembler::Success)
478             break;
479           MI = MCInst(); // clear
480         }
481       }
482 
483       Res = tryDecodeInst(DecoderTableDPP864, MI, QW, Address);
484       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
485         break;
486       MI = MCInst(); // clear
487 
488       Res = tryDecodeInst(DecoderTableDPP8GFX1164, MI, QW, Address);
489       if (Res && convertDPP8Inst(MI) == MCDisassembler::Success)
490         break;
491       MI = MCInst(); // clear
492 
493       Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address);
494       if (Res) break;
495 
496       Res = tryDecodeInst(DecoderTableDPPGFX1164, MI, QW, Address);
497       if (Res) {
498         if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOPC)
499           convertVOPCDPPInst(MI);
500         break;
501       }
502 
503       Res = tryDecodeInst(DecoderTableSDWA64, MI, QW, Address);
504       if (Res) { IsSDWA = true;  break; }
505 
506       Res = tryDecodeInst(DecoderTableSDWA964, MI, QW, Address);
507       if (Res) { IsSDWA = true;  break; }
508 
509       Res = tryDecodeInst(DecoderTableSDWA1064, MI, QW, Address);
510       if (Res) { IsSDWA = true;  break; }
511 
512       if (STI.getFeatureBits()[AMDGPU::FeatureUnpackedD16VMem]) {
513         Res = tryDecodeInst(DecoderTableGFX80_UNPACKED64, MI, QW, Address);
514         if (Res)
515           break;
516       }
517 
518       // Some GFX9 subtargets repurposed the v_mad_mix_f32, v_mad_mixlo_f16 and
519       // v_mad_mixhi_f16 for FMA variants. Try to decode using this special
520       // table first so we print the correct name.
521       if (STI.getFeatureBits()[AMDGPU::FeatureFmaMixInsts]) {
522         Res = tryDecodeInst(DecoderTableGFX9_DL64, MI, QW, Address);
523         if (Res)
524           break;
525       }
526     }
527 
528     // Reinitialize Bytes as DPP64 could have eaten too much
529     Bytes = Bytes_.slice(0, MaxInstBytesNum);
530 
531     // Try decode 32-bit instruction
532     if (Bytes.size() < 4) break;
533     const uint32_t DW = eatBytes<uint32_t>(Bytes);
534     Res = tryDecodeInst(DecoderTableGFX832, MI, DW, Address);
535     if (Res) break;
536 
537     Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address);
538     if (Res) break;
539 
540     Res = tryDecodeInst(DecoderTableGFX932, MI, DW, Address);
541     if (Res) break;
542 
543     if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) {
544       Res = tryDecodeInst(DecoderTableGFX90A32, MI, DW, Address);
545       if (Res)
546         break;
547     }
548 
549     if (STI.getFeatureBits()[AMDGPU::FeatureGFX10_BEncoding]) {
550       Res = tryDecodeInst(DecoderTableGFX10_B32, MI, DW, Address);
551       if (Res) break;
552     }
553 
554     Res = tryDecodeInst(DecoderTableGFX1032, MI, DW, Address);
555     if (Res) break;
556 
557     Res = tryDecodeInst(DecoderTableGFX1132, MI, DW, Address);
558     if (Res) break;
559 
560     if (Bytes.size() < 4) break;
561     const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW;
562 
563     if (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts]) {
564       Res = tryDecodeInst(DecoderTableGFX90A64, MI, QW, Address);
565       if (Res)
566         break;
567     }
568 
569     Res = tryDecodeInst(DecoderTableGFX864, MI, QW, Address);
570     if (Res) break;
571 
572     Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address);
573     if (Res) break;
574 
575     Res = tryDecodeInst(DecoderTableGFX964, MI, QW, Address);
576     if (Res) break;
577 
578     Res = tryDecodeInst(DecoderTableGFX1064, MI, QW, Address);
579     if (Res) break;
580 
581     Res = tryDecodeInst(DecoderTableGFX1164, MI, QW, Address);
582     if (Res)
583       break;
584 
585     Res = tryDecodeInst(DecoderTableWMMAGFX1164, MI, QW, Address);
586   } while (false);
587 
588   if (Res && (MI.getOpcode() == AMDGPU::V_MAC_F32_e64_vi ||
589               MI.getOpcode() == AMDGPU::V_MAC_F32_e64_gfx6_gfx7 ||
590               MI.getOpcode() == AMDGPU::V_MAC_F32_e64_gfx10 ||
591               MI.getOpcode() == AMDGPU::V_MAC_LEGACY_F32_e64_gfx6_gfx7 ||
592               MI.getOpcode() == AMDGPU::V_MAC_LEGACY_F32_e64_gfx10 ||
593               MI.getOpcode() == AMDGPU::V_MAC_F16_e64_vi ||
594               MI.getOpcode() == AMDGPU::V_FMAC_F64_e64_gfx90a ||
595               MI.getOpcode() == AMDGPU::V_FMAC_F32_e64_vi ||
596               MI.getOpcode() == AMDGPU::V_FMAC_F32_e64_gfx10 ||
597               MI.getOpcode() == AMDGPU::V_FMAC_F32_e64_gfx11 ||
598               MI.getOpcode() == AMDGPU::V_FMAC_LEGACY_F32_e64_gfx10 ||
599               MI.getOpcode() == AMDGPU::V_FMAC_DX9_ZERO_F32_e64_gfx11 ||
600               MI.getOpcode() == AMDGPU::V_FMAC_F16_e64_gfx10 ||
601               MI.getOpcode() == AMDGPU::V_FMAC_F16_e64_gfx11)) {
602     // Insert dummy unused src2_modifiers.
603     insertNamedMCOperand(MI, MCOperand::createImm(0),
604                          AMDGPU::OpName::src2_modifiers);
605   }
606 
607   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
608           (SIInstrFlags::MUBUF | SIInstrFlags::FLAT | SIInstrFlags::SMRD))) {
609     int CPolPos = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
610                                              AMDGPU::OpName::cpol);
611     if (CPolPos != -1) {
612       unsigned CPol =
613           (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::IsAtomicRet) ?
614               AMDGPU::CPol::GLC : 0;
615       if (MI.getNumOperands() <= (unsigned)CPolPos) {
616         insertNamedMCOperand(MI, MCOperand::createImm(CPol),
617                              AMDGPU::OpName::cpol);
618       } else if (CPol) {
619         MI.getOperand(CPolPos).setImm(MI.getOperand(CPolPos).getImm() | CPol);
620       }
621     }
622   }
623 
624   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
625               (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF)) &&
626              (STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts])) {
627     // GFX90A lost TFE, its place is occupied by ACC.
628     int TFEOpIdx =
629         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::tfe);
630     if (TFEOpIdx != -1) {
631       auto TFEIter = MI.begin();
632       std::advance(TFEIter, TFEOpIdx);
633       MI.insert(TFEIter, MCOperand::createImm(0));
634     }
635   }
636 
637   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
638               (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF))) {
639     int SWZOpIdx =
640         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::swz);
641     if (SWZOpIdx != -1) {
642       auto SWZIter = MI.begin();
643       std::advance(SWZIter, SWZOpIdx);
644       MI.insert(SWZIter, MCOperand::createImm(0));
645     }
646   }
647 
648   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::MIMG)) {
649     int VAddr0Idx =
650         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
651     int RsrcIdx =
652         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
653     unsigned NSAArgs = RsrcIdx - VAddr0Idx - 1;
654     if (VAddr0Idx >= 0 && NSAArgs > 0) {
655       unsigned NSAWords = (NSAArgs + 3) / 4;
656       if (Bytes.size() < 4 * NSAWords) {
657         Res = MCDisassembler::Fail;
658       } else {
659         for (unsigned i = 0; i < NSAArgs; ++i) {
660           const unsigned VAddrIdx = VAddr0Idx + 1 + i;
661           auto VAddrRCID = MCII->get(MI.getOpcode()).OpInfo[VAddrIdx].RegClass;
662           MI.insert(MI.begin() + VAddrIdx,
663                     createRegOperand(VAddrRCID, Bytes[i]));
664         }
665         Bytes = Bytes.slice(4 * NSAWords);
666       }
667     }
668 
669     if (Res)
670       Res = convertMIMGInst(MI);
671   }
672 
673   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::EXP))
674     Res = convertEXPInst(MI);
675 
676   if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VINTERP))
677     Res = convertVINTERPInst(MI);
678 
679   if (Res && IsSDWA)
680     Res = convertSDWAInst(MI);
681 
682   int VDstIn_Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
683                                               AMDGPU::OpName::vdst_in);
684   if (VDstIn_Idx != -1) {
685     int Tied = MCII->get(MI.getOpcode()).getOperandConstraint(VDstIn_Idx,
686                            MCOI::OperandConstraint::TIED_TO);
687     if (Tied != -1 && (MI.getNumOperands() <= (unsigned)VDstIn_Idx ||
688          !MI.getOperand(VDstIn_Idx).isReg() ||
689          MI.getOperand(VDstIn_Idx).getReg() != MI.getOperand(Tied).getReg())) {
690       if (MI.getNumOperands() > (unsigned)VDstIn_Idx)
691         MI.erase(&MI.getOperand(VDstIn_Idx));
692       insertNamedMCOperand(MI,
693         MCOperand::createReg(MI.getOperand(Tied).getReg()),
694         AMDGPU::OpName::vdst_in);
695     }
696   }
697 
698   int ImmLitIdx =
699       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::imm);
700   if (Res && ImmLitIdx != -1)
701     Res = convertFMAanyK(MI, ImmLitIdx);
702 
703   // if the opcode was not recognized we'll assume a Size of 4 bytes
704   // (unless there are fewer bytes left)
705   Size = Res ? (MaxInstBytesNum - Bytes.size())
706              : std::min((size_t)4, Bytes_.size());
707   return Res;
708 }
709 
710 DecodeStatus AMDGPUDisassembler::convertEXPInst(MCInst &MI) const {
711   if (STI.getFeatureBits()[AMDGPU::FeatureGFX11]) {
712     // The MCInst still has these fields even though they are no longer encoded
713     // in the GFX11 instruction.
714     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vm);
715     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::compr);
716   }
717   return MCDisassembler::Success;
718 }
719 
720 DecodeStatus AMDGPUDisassembler::convertVINTERPInst(MCInst &MI) const {
721   if (MI.getOpcode() == AMDGPU::V_INTERP_P10_F16_F32_inreg_gfx11 ||
722       MI.getOpcode() == AMDGPU::V_INTERP_P10_RTZ_F16_F32_inreg_gfx11 ||
723       MI.getOpcode() == AMDGPU::V_INTERP_P2_F16_F32_inreg_gfx11 ||
724       MI.getOpcode() == AMDGPU::V_INTERP_P2_RTZ_F16_F32_inreg_gfx11) {
725     // The MCInst has this field that is not directly encoded in the
726     // instruction.
727     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::op_sel);
728   }
729   return MCDisassembler::Success;
730 }
731 
732 DecodeStatus AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const {
733   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
734       STI.getFeatureBits()[AMDGPU::FeatureGFX10]) {
735     if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst) != -1)
736       // VOPC - insert clamp
737       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp);
738   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
739     int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst);
740     if (SDst != -1) {
741       // VOPC - insert VCC register as sdst
742       insertNamedMCOperand(MI, createRegOperand(AMDGPU::VCC),
743                            AMDGPU::OpName::sdst);
744     } else {
745       // VOP1/2 - insert omod if present in instruction
746       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod);
747     }
748   }
749   return MCDisassembler::Success;
750 }
751 
752 struct VOPModifiers {
753   unsigned OpSel = 0;
754   unsigned OpSelHi = 0;
755   unsigned NegLo = 0;
756   unsigned NegHi = 0;
757 };
758 
759 // Reconstruct values of VOP3/VOP3P operands such as op_sel.
760 // Note that these values do not affect disassembler output,
761 // so this is only necessary for consistency with src_modifiers.
762 static VOPModifiers collectVOPModifiers(const MCInst &MI,
763                                         bool IsVOP3P = false) {
764   VOPModifiers Modifiers;
765   unsigned Opc = MI.getOpcode();
766   const int ModOps[] = {AMDGPU::OpName::src0_modifiers,
767                         AMDGPU::OpName::src1_modifiers,
768                         AMDGPU::OpName::src2_modifiers};
769   for (int J = 0; J < 3; ++J) {
770     int OpIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]);
771     if (OpIdx == -1)
772       continue;
773 
774     unsigned Val = MI.getOperand(OpIdx).getImm();
775 
776     Modifiers.OpSel |= !!(Val & SISrcMods::OP_SEL_0) << J;
777     if (IsVOP3P) {
778       Modifiers.OpSelHi |= !!(Val & SISrcMods::OP_SEL_1) << J;
779       Modifiers.NegLo |= !!(Val & SISrcMods::NEG) << J;
780       Modifiers.NegHi |= !!(Val & SISrcMods::NEG_HI) << J;
781     } else if (J == 0) {
782       Modifiers.OpSel |= !!(Val & SISrcMods::DST_OP_SEL) << 3;
783     }
784   }
785 
786   return Modifiers;
787 }
788 
789 // We must check FI == literal to reject not genuine dpp8 insts, and we must
790 // first add optional MI operands to check FI
791 DecodeStatus AMDGPUDisassembler::convertDPP8Inst(MCInst &MI) const {
792   unsigned Opc = MI.getOpcode();
793   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
794   if (MCII->get(Opc).TSFlags & SIInstrFlags::VOP3P) {
795     convertVOP3PDPPInst(MI);
796   } else if ((MCII->get(Opc).TSFlags & SIInstrFlags::VOPC) ||
797              AMDGPU::isVOPC64DPP(Opc)) {
798     convertVOPCDPPInst(MI);
799   } else if (MI.getNumOperands() < DescNumOps &&
800              AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel) != -1) {
801     auto Mods = collectVOPModifiers(MI);
802     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
803                          AMDGPU::OpName::op_sel);
804   } else {
805     // Insert dummy unused src modifiers.
806     if (MI.getNumOperands() < DescNumOps &&
807         AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers) != -1)
808       insertNamedMCOperand(MI, MCOperand::createImm(0),
809                            AMDGPU::OpName::src0_modifiers);
810 
811     if (MI.getNumOperands() < DescNumOps &&
812         AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers) != -1)
813       insertNamedMCOperand(MI, MCOperand::createImm(0),
814                            AMDGPU::OpName::src1_modifiers);
815   }
816   return isValidDPP8(MI) ? MCDisassembler::Success : MCDisassembler::SoftFail;
817 }
818 
819 DecodeStatus AMDGPUDisassembler::convertVOP3DPPInst(MCInst &MI) const {
820   unsigned Opc = MI.getOpcode();
821   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
822   if (MI.getNumOperands() < DescNumOps &&
823       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel) != -1) {
824     auto Mods = collectVOPModifiers(MI);
825     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
826                          AMDGPU::OpName::op_sel);
827   }
828   return MCDisassembler::Success;
829 }
830 
831 // Note that before gfx10, the MIMG encoding provided no information about
832 // VADDR size. Consequently, decoded instructions always show address as if it
833 // has 1 dword, which could be not really so.
834 DecodeStatus AMDGPUDisassembler::convertMIMGInst(MCInst &MI) const {
835 
836   int VDstIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
837                                            AMDGPU::OpName::vdst);
838 
839   int VDataIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
840                                             AMDGPU::OpName::vdata);
841   int VAddr0Idx =
842       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
843   int DMaskIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
844                                             AMDGPU::OpName::dmask);
845 
846   int TFEIdx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
847                                             AMDGPU::OpName::tfe);
848   int D16Idx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
849                                             AMDGPU::OpName::d16);
850 
851   const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode());
852   const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
853       AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
854 
855   assert(VDataIdx != -1);
856   if (BaseOpcode->BVH) {
857     // Add A16 operand for intersect_ray instructions
858     if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16) > -1) {
859       addOperand(MI, MCOperand::createImm(1));
860     }
861     return MCDisassembler::Success;
862   }
863 
864   bool IsAtomic = (VDstIdx != -1);
865   bool IsGather4 = MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::Gather4;
866   bool IsNSA = false;
867   unsigned AddrSize = Info->VAddrDwords;
868 
869   if (isGFX10Plus()) {
870     unsigned DimIdx =
871         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dim);
872     int A16Idx =
873         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16);
874     const AMDGPU::MIMGDimInfo *Dim =
875         AMDGPU::getMIMGDimInfoByEncoding(MI.getOperand(DimIdx).getImm());
876     const bool IsA16 = (A16Idx != -1 && MI.getOperand(A16Idx).getImm());
877 
878     AddrSize =
879         AMDGPU::getAddrSizeMIMGOp(BaseOpcode, Dim, IsA16, AMDGPU::hasG16(STI));
880 
881     IsNSA = Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA ||
882             Info->MIMGEncoding == AMDGPU::MIMGEncGfx11NSA;
883     if (!IsNSA) {
884       if (AddrSize > 8)
885         AddrSize = 16;
886     } else {
887       if (AddrSize > Info->VAddrDwords) {
888         // The NSA encoding does not contain enough operands for the combination
889         // of base opcode / dimension. Should this be an error?
890         return MCDisassembler::Success;
891       }
892     }
893   }
894 
895   unsigned DMask = MI.getOperand(DMaskIdx).getImm() & 0xf;
896   unsigned DstSize = IsGather4 ? 4 : std::max(countPopulation(DMask), 1u);
897 
898   bool D16 = D16Idx >= 0 && MI.getOperand(D16Idx).getImm();
899   if (D16 && AMDGPU::hasPackedD16(STI)) {
900     DstSize = (DstSize + 1) / 2;
901   }
902 
903   if (TFEIdx != -1 && MI.getOperand(TFEIdx).getImm())
904     DstSize += 1;
905 
906   if (DstSize == Info->VDataDwords && AddrSize == Info->VAddrDwords)
907     return MCDisassembler::Success;
908 
909   int NewOpcode =
910       AMDGPU::getMIMGOpcode(Info->BaseOpcode, Info->MIMGEncoding, DstSize, AddrSize);
911   if (NewOpcode == -1)
912     return MCDisassembler::Success;
913 
914   // Widen the register to the correct number of enabled channels.
915   unsigned NewVdata = AMDGPU::NoRegister;
916   if (DstSize != Info->VDataDwords) {
917     auto DataRCID = MCII->get(NewOpcode).OpInfo[VDataIdx].RegClass;
918 
919     // Get first subregister of VData
920     unsigned Vdata0 = MI.getOperand(VDataIdx).getReg();
921     unsigned VdataSub0 = MRI.getSubReg(Vdata0, AMDGPU::sub0);
922     Vdata0 = (VdataSub0 != 0)? VdataSub0 : Vdata0;
923 
924     NewVdata = MRI.getMatchingSuperReg(Vdata0, AMDGPU::sub0,
925                                        &MRI.getRegClass(DataRCID));
926     if (NewVdata == AMDGPU::NoRegister) {
927       // It's possible to encode this such that the low register + enabled
928       // components exceeds the register count.
929       return MCDisassembler::Success;
930     }
931   }
932 
933   // If not using NSA on GFX10+, widen address register to correct size.
934   unsigned NewVAddr0 = AMDGPU::NoRegister;
935   if (isGFX10Plus() && !IsNSA && AddrSize != Info->VAddrDwords) {
936     unsigned VAddr0 = MI.getOperand(VAddr0Idx).getReg();
937     unsigned VAddrSub0 = MRI.getSubReg(VAddr0, AMDGPU::sub0);
938     VAddr0 = (VAddrSub0 != 0) ? VAddrSub0 : VAddr0;
939 
940     auto AddrRCID = MCII->get(NewOpcode).OpInfo[VAddr0Idx].RegClass;
941     NewVAddr0 = MRI.getMatchingSuperReg(VAddr0, AMDGPU::sub0,
942                                         &MRI.getRegClass(AddrRCID));
943     if (NewVAddr0 == AMDGPU::NoRegister)
944       return MCDisassembler::Success;
945   }
946 
947   MI.setOpcode(NewOpcode);
948 
949   if (NewVdata != AMDGPU::NoRegister) {
950     MI.getOperand(VDataIdx) = MCOperand::createReg(NewVdata);
951 
952     if (IsAtomic) {
953       // Atomic operations have an additional operand (a copy of data)
954       MI.getOperand(VDstIdx) = MCOperand::createReg(NewVdata);
955     }
956   }
957 
958   if (NewVAddr0 != AMDGPU::NoRegister) {
959     MI.getOperand(VAddr0Idx) = MCOperand::createReg(NewVAddr0);
960   } else if (IsNSA) {
961     assert(AddrSize <= Info->VAddrDwords);
962     MI.erase(MI.begin() + VAddr0Idx + AddrSize,
963              MI.begin() + VAddr0Idx + Info->VAddrDwords);
964   }
965 
966   return MCDisassembler::Success;
967 }
968 
969 // Opsel and neg bits are used in src_modifiers and standalone operands. Autogen
970 // decoder only adds to src_modifiers, so manually add the bits to the other
971 // operands.
972 DecodeStatus AMDGPUDisassembler::convertVOP3PDPPInst(MCInst &MI) const {
973   unsigned Opc = MI.getOpcode();
974   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
975   auto Mods = collectVOPModifiers(MI, true);
976 
977   if (MI.getNumOperands() < DescNumOps &&
978       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst_in) != -1)
979     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vdst_in);
980 
981   if (MI.getNumOperands() < DescNumOps &&
982       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel) != -1)
983     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
984                          AMDGPU::OpName::op_sel);
985   if (MI.getNumOperands() < DescNumOps &&
986       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::op_sel_hi) != -1)
987     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSelHi),
988                          AMDGPU::OpName::op_sel_hi);
989   if (MI.getNumOperands() < DescNumOps &&
990       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::neg_lo) != -1)
991     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegLo),
992                          AMDGPU::OpName::neg_lo);
993   if (MI.getNumOperands() < DescNumOps &&
994       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::neg_hi) != -1)
995     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegHi),
996                          AMDGPU::OpName::neg_hi);
997 
998   return MCDisassembler::Success;
999 }
1000 
1001 // Create dummy old operand and insert optional operands
1002 DecodeStatus AMDGPUDisassembler::convertVOPCDPPInst(MCInst &MI) const {
1003   unsigned Opc = MI.getOpcode();
1004   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
1005 
1006   if (MI.getNumOperands() < DescNumOps &&
1007       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::old) != -1)
1008     insertNamedMCOperand(MI, MCOperand::createReg(0), AMDGPU::OpName::old);
1009 
1010   if (MI.getNumOperands() < DescNumOps &&
1011       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0_modifiers) != -1)
1012     insertNamedMCOperand(MI, MCOperand::createImm(0),
1013                          AMDGPU::OpName::src0_modifiers);
1014 
1015   if (MI.getNumOperands() < DescNumOps &&
1016       AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1_modifiers) != -1)
1017     insertNamedMCOperand(MI, MCOperand::createImm(0),
1018                          AMDGPU::OpName::src1_modifiers);
1019   return MCDisassembler::Success;
1020 }
1021 
1022 DecodeStatus AMDGPUDisassembler::convertFMAanyK(MCInst &MI,
1023                                                 int ImmLitIdx) const {
1024   assert(HasLiteral && "Should have decoded a literal");
1025   const MCInstrDesc &Desc = MCII->get(MI.getOpcode());
1026   unsigned DescNumOps = Desc.getNumOperands();
1027   insertNamedMCOperand(MI, MCOperand::createImm(Literal),
1028                        AMDGPU::OpName::immDeferred);
1029   assert(DescNumOps == MI.getNumOperands());
1030   for (unsigned I = 0; I < DescNumOps; ++I) {
1031     auto &Op = MI.getOperand(I);
1032     auto OpType = Desc.OpInfo[I].OperandType;
1033     bool IsDeferredOp = (OpType == AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED ||
1034                          OpType == AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED);
1035     if (Op.isImm() && Op.getImm() == AMDGPU::EncValues::LITERAL_CONST &&
1036         IsDeferredOp)
1037       Op.setImm(Literal);
1038   }
1039   return MCDisassembler::Success;
1040 }
1041 
1042 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
1043   return getContext().getRegisterInfo()->
1044     getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
1045 }
1046 
1047 inline
1048 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
1049                                          const Twine& ErrMsg) const {
1050   *CommentStream << "Error: " + ErrMsg;
1051 
1052   // ToDo: add support for error operands to MCInst.h
1053   // return MCOperand::createError(V);
1054   return MCOperand();
1055 }
1056 
1057 inline
1058 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
1059   return MCOperand::createReg(AMDGPU::getMCReg(RegId, STI));
1060 }
1061 
1062 inline
1063 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
1064                                                unsigned Val) const {
1065   const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
1066   if (Val >= RegCl.getNumRegs())
1067     return errOperand(Val, Twine(getRegClassName(RegClassID)) +
1068                            ": unknown register " + Twine(Val));
1069   return createRegOperand(RegCl.getRegister(Val));
1070 }
1071 
1072 inline
1073 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
1074                                                 unsigned Val) const {
1075   // ToDo: SI/CI have 104 SGPRs, VI - 102
1076   // Valery: here we accepting as much as we can, let assembler sort it out
1077   int shift = 0;
1078   switch (SRegClassID) {
1079   case AMDGPU::SGPR_32RegClassID:
1080   case AMDGPU::TTMP_32RegClassID:
1081     break;
1082   case AMDGPU::SGPR_64RegClassID:
1083   case AMDGPU::TTMP_64RegClassID:
1084     shift = 1;
1085     break;
1086   case AMDGPU::SGPR_128RegClassID:
1087   case AMDGPU::TTMP_128RegClassID:
1088   // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
1089   // this bundle?
1090   case AMDGPU::SGPR_256RegClassID:
1091   case AMDGPU::TTMP_256RegClassID:
1092     // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
1093   // this bundle?
1094   case AMDGPU::SGPR_512RegClassID:
1095   case AMDGPU::TTMP_512RegClassID:
1096     shift = 2;
1097     break;
1098   // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
1099   // this bundle?
1100   default:
1101     llvm_unreachable("unhandled register class");
1102   }
1103 
1104   if (Val % (1 << shift)) {
1105     *CommentStream << "Warning: " << getRegClassName(SRegClassID)
1106                    << ": scalar reg isn't aligned " << Val;
1107   }
1108 
1109   return createRegOperand(SRegClassID, Val >> shift);
1110 }
1111 
1112 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const {
1113   return decodeSrcOp(OPW32, Val);
1114 }
1115 
1116 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const {
1117   return decodeSrcOp(OPW64, Val);
1118 }
1119 
1120 MCOperand AMDGPUDisassembler::decodeOperand_VS_128(unsigned Val) const {
1121   return decodeSrcOp(OPW128, Val);
1122 }
1123 
1124 MCOperand AMDGPUDisassembler::decodeOperand_VSrc16(unsigned Val) const {
1125   return decodeSrcOp(OPW16, Val);
1126 }
1127 
1128 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV216(unsigned Val) const {
1129   return decodeSrcOp(OPWV216, Val);
1130 }
1131 
1132 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV232(unsigned Val) const {
1133   return decodeSrcOp(OPWV232, Val);
1134 }
1135 
1136 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const {
1137   // Some instructions have operand restrictions beyond what the encoding
1138   // allows. Some ordinarily VSrc_32 operands are VGPR_32, so clear the extra
1139   // high bit.
1140   Val &= 255;
1141 
1142   return createRegOperand(AMDGPU::VGPR_32RegClassID, Val);
1143 }
1144 
1145 MCOperand AMDGPUDisassembler::decodeOperand_VRegOrLds_32(unsigned Val) const {
1146   return decodeSrcOp(OPW32, Val);
1147 }
1148 
1149 MCOperand AMDGPUDisassembler::decodeOperand_AGPR_32(unsigned Val) const {
1150   return createRegOperand(AMDGPU::AGPR_32RegClassID, Val & 255);
1151 }
1152 
1153 MCOperand AMDGPUDisassembler::decodeOperand_AReg_64(unsigned Val) const {
1154   return createRegOperand(AMDGPU::AReg_64RegClassID, Val & 255);
1155 }
1156 
1157 MCOperand AMDGPUDisassembler::decodeOperand_AReg_128(unsigned Val) const {
1158   return createRegOperand(AMDGPU::AReg_128RegClassID, Val & 255);
1159 }
1160 
1161 MCOperand AMDGPUDisassembler::decodeOperand_AReg_256(unsigned Val) const {
1162   return createRegOperand(AMDGPU::AReg_256RegClassID, Val & 255);
1163 }
1164 
1165 MCOperand AMDGPUDisassembler::decodeOperand_AReg_512(unsigned Val) const {
1166   return createRegOperand(AMDGPU::AReg_512RegClassID, Val & 255);
1167 }
1168 
1169 MCOperand AMDGPUDisassembler::decodeOperand_AReg_1024(unsigned Val) const {
1170   return createRegOperand(AMDGPU::AReg_1024RegClassID, Val & 255);
1171 }
1172 
1173 MCOperand AMDGPUDisassembler::decodeOperand_AV_32(unsigned Val) const {
1174   return decodeSrcOp(OPW32, Val);
1175 }
1176 
1177 MCOperand AMDGPUDisassembler::decodeOperand_AV_64(unsigned Val) const {
1178   return decodeSrcOp(OPW64, Val);
1179 }
1180 
1181 MCOperand AMDGPUDisassembler::decodeOperand_AV_128(unsigned Val) const {
1182   return decodeSrcOp(OPW128, Val);
1183 }
1184 
1185 MCOperand AMDGPUDisassembler::decodeOperand_AVDst_128(unsigned Val) const {
1186   using namespace AMDGPU::EncValues;
1187   assert((Val & IS_VGPR) == 0); // Val{8} is not encoded but assumed to be 1.
1188   return decodeSrcOp(OPW128, Val | IS_VGPR);
1189 }
1190 
1191 MCOperand AMDGPUDisassembler::decodeOperand_AVDst_512(unsigned Val) const {
1192   using namespace AMDGPU::EncValues;
1193   assert((Val & IS_VGPR) == 0); // Val{8} is not encoded but assumed to be 1.
1194   return decodeSrcOp(OPW512, Val | IS_VGPR);
1195 }
1196 
1197 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const {
1198   return createRegOperand(AMDGPU::VReg_64RegClassID, Val);
1199 }
1200 
1201 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const {
1202   return createRegOperand(AMDGPU::VReg_96RegClassID, Val);
1203 }
1204 
1205 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const {
1206   return createRegOperand(AMDGPU::VReg_128RegClassID, Val);
1207 }
1208 
1209 MCOperand AMDGPUDisassembler::decodeOperand_VReg_256(unsigned Val) const {
1210   return createRegOperand(AMDGPU::VReg_256RegClassID, Val);
1211 }
1212 
1213 MCOperand AMDGPUDisassembler::decodeOperand_VReg_512(unsigned Val) const {
1214   return createRegOperand(AMDGPU::VReg_512RegClassID, Val);
1215 }
1216 
1217 MCOperand AMDGPUDisassembler::decodeOperand_VReg_1024(unsigned Val) const {
1218   return createRegOperand(AMDGPU::VReg_1024RegClassID, Val);
1219 }
1220 
1221 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const {
1222   // table-gen generated disassembler doesn't care about operand types
1223   // leaving only registry class so SSrc_32 operand turns into SReg_32
1224   // and therefore we accept immediates and literals here as well
1225   return decodeSrcOp(OPW32, Val);
1226 }
1227 
1228 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XM0_XEXEC(
1229   unsigned Val) const {
1230   // SReg_32_XM0 is SReg_32 without M0 or EXEC_LO/EXEC_HI
1231   return decodeOperand_SReg_32(Val);
1232 }
1233 
1234 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XEXEC_HI(
1235   unsigned Val) const {
1236   // SReg_32_XM0 is SReg_32 without EXEC_HI
1237   return decodeOperand_SReg_32(Val);
1238 }
1239 
1240 MCOperand AMDGPUDisassembler::decodeOperand_SRegOrLds_32(unsigned Val) const {
1241   // table-gen generated disassembler doesn't care about operand types
1242   // leaving only registry class so SSrc_32 operand turns into SReg_32
1243   // and therefore we accept immediates and literals here as well
1244   return decodeSrcOp(OPW32, Val);
1245 }
1246 
1247 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const {
1248   return decodeSrcOp(OPW64, Val);
1249 }
1250 
1251 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64_XEXEC(unsigned Val) const {
1252   return decodeSrcOp(OPW64, Val);
1253 }
1254 
1255 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const {
1256   return decodeSrcOp(OPW128, Val);
1257 }
1258 
1259 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const {
1260   return decodeDstOp(OPW256, Val);
1261 }
1262 
1263 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const {
1264   return decodeDstOp(OPW512, Val);
1265 }
1266 
1267 // Decode Literals for insts which always have a literal in the encoding
1268 MCOperand
1269 AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
1270   if (HasLiteral) {
1271     assert(
1272         AMDGPU::hasVOPD(STI) &&
1273         "Should only decode multiple kimm with VOPD, check VSrc operand types");
1274     if (Literal != Val)
1275       return errOperand(Val, "More than one unique literal is illegal");
1276   }
1277   HasLiteral = true;
1278   Literal = Val;
1279   return MCOperand::createImm(Literal);
1280 }
1281 
1282 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
1283   // For now all literal constants are supposed to be unsigned integer
1284   // ToDo: deal with signed/unsigned 64-bit integer constants
1285   // ToDo: deal with float/double constants
1286   if (!HasLiteral) {
1287     if (Bytes.size() < 4) {
1288       return errOperand(0, "cannot read literal, inst bytes left " +
1289                         Twine(Bytes.size()));
1290     }
1291     HasLiteral = true;
1292     Literal = eatBytes<uint32_t>(Bytes);
1293   }
1294   return MCOperand::createImm(Literal);
1295 }
1296 
1297 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
1298   using namespace AMDGPU::EncValues;
1299 
1300   assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX);
1301   return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ?
1302     (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) :
1303     (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm)));
1304       // Cast prevents negative overflow.
1305 }
1306 
1307 static int64_t getInlineImmVal32(unsigned Imm) {
1308   switch (Imm) {
1309   case 240:
1310     return FloatToBits(0.5f);
1311   case 241:
1312     return FloatToBits(-0.5f);
1313   case 242:
1314     return FloatToBits(1.0f);
1315   case 243:
1316     return FloatToBits(-1.0f);
1317   case 244:
1318     return FloatToBits(2.0f);
1319   case 245:
1320     return FloatToBits(-2.0f);
1321   case 246:
1322     return FloatToBits(4.0f);
1323   case 247:
1324     return FloatToBits(-4.0f);
1325   case 248: // 1 / (2 * PI)
1326     return 0x3e22f983;
1327   default:
1328     llvm_unreachable("invalid fp inline imm");
1329   }
1330 }
1331 
1332 static int64_t getInlineImmVal64(unsigned Imm) {
1333   switch (Imm) {
1334   case 240:
1335     return DoubleToBits(0.5);
1336   case 241:
1337     return DoubleToBits(-0.5);
1338   case 242:
1339     return DoubleToBits(1.0);
1340   case 243:
1341     return DoubleToBits(-1.0);
1342   case 244:
1343     return DoubleToBits(2.0);
1344   case 245:
1345     return DoubleToBits(-2.0);
1346   case 246:
1347     return DoubleToBits(4.0);
1348   case 247:
1349     return DoubleToBits(-4.0);
1350   case 248: // 1 / (2 * PI)
1351     return 0x3fc45f306dc9c882;
1352   default:
1353     llvm_unreachable("invalid fp inline imm");
1354   }
1355 }
1356 
1357 static int64_t getInlineImmVal16(unsigned Imm) {
1358   switch (Imm) {
1359   case 240:
1360     return 0x3800;
1361   case 241:
1362     return 0xB800;
1363   case 242:
1364     return 0x3C00;
1365   case 243:
1366     return 0xBC00;
1367   case 244:
1368     return 0x4000;
1369   case 245:
1370     return 0xC000;
1371   case 246:
1372     return 0x4400;
1373   case 247:
1374     return 0xC400;
1375   case 248: // 1 / (2 * PI)
1376     return 0x3118;
1377   default:
1378     llvm_unreachable("invalid fp inline imm");
1379   }
1380 }
1381 
1382 MCOperand AMDGPUDisassembler::decodeFPImmed(OpWidthTy Width, unsigned Imm) {
1383   assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN
1384       && Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX);
1385 
1386   // ToDo: case 248: 1/(2*PI) - is allowed only on VI
1387   switch (Width) {
1388   case OPW32:
1389   case OPW128: // splat constants
1390   case OPW512:
1391   case OPW1024:
1392   case OPWV232:
1393     return MCOperand::createImm(getInlineImmVal32(Imm));
1394   case OPW64:
1395   case OPW256:
1396     return MCOperand::createImm(getInlineImmVal64(Imm));
1397   case OPW16:
1398   case OPWV216:
1399     return MCOperand::createImm(getInlineImmVal16(Imm));
1400   default:
1401     llvm_unreachable("implement me");
1402   }
1403 }
1404 
1405 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const {
1406   using namespace AMDGPU;
1407 
1408   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1409   switch (Width) {
1410   default: // fall
1411   case OPW32:
1412   case OPW16:
1413   case OPWV216:
1414     return VGPR_32RegClassID;
1415   case OPW64:
1416   case OPWV232: return VReg_64RegClassID;
1417   case OPW96: return VReg_96RegClassID;
1418   case OPW128: return VReg_128RegClassID;
1419   case OPW160: return VReg_160RegClassID;
1420   case OPW256: return VReg_256RegClassID;
1421   case OPW512: return VReg_512RegClassID;
1422   case OPW1024: return VReg_1024RegClassID;
1423   }
1424 }
1425 
1426 unsigned AMDGPUDisassembler::getAgprClassId(const OpWidthTy Width) const {
1427   using namespace AMDGPU;
1428 
1429   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1430   switch (Width) {
1431   default: // fall
1432   case OPW32:
1433   case OPW16:
1434   case OPWV216:
1435     return AGPR_32RegClassID;
1436   case OPW64:
1437   case OPWV232: return AReg_64RegClassID;
1438   case OPW96: return AReg_96RegClassID;
1439   case OPW128: return AReg_128RegClassID;
1440   case OPW160: return AReg_160RegClassID;
1441   case OPW256: return AReg_256RegClassID;
1442   case OPW512: return AReg_512RegClassID;
1443   case OPW1024: return AReg_1024RegClassID;
1444   }
1445 }
1446 
1447 
1448 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const {
1449   using namespace AMDGPU;
1450 
1451   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1452   switch (Width) {
1453   default: // fall
1454   case OPW32:
1455   case OPW16:
1456   case OPWV216:
1457     return SGPR_32RegClassID;
1458   case OPW64:
1459   case OPWV232: return SGPR_64RegClassID;
1460   case OPW96: return SGPR_96RegClassID;
1461   case OPW128: return SGPR_128RegClassID;
1462   case OPW160: return SGPR_160RegClassID;
1463   case OPW256: return SGPR_256RegClassID;
1464   case OPW512: return SGPR_512RegClassID;
1465   }
1466 }
1467 
1468 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const {
1469   using namespace AMDGPU;
1470 
1471   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1472   switch (Width) {
1473   default: // fall
1474   case OPW32:
1475   case OPW16:
1476   case OPWV216:
1477     return TTMP_32RegClassID;
1478   case OPW64:
1479   case OPWV232: return TTMP_64RegClassID;
1480   case OPW128: return TTMP_128RegClassID;
1481   case OPW256: return TTMP_256RegClassID;
1482   case OPW512: return TTMP_512RegClassID;
1483   }
1484 }
1485 
1486 int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
1487   using namespace AMDGPU::EncValues;
1488 
1489   unsigned TTmpMin = isGFX9Plus() ? TTMP_GFX9PLUS_MIN : TTMP_VI_MIN;
1490   unsigned TTmpMax = isGFX9Plus() ? TTMP_GFX9PLUS_MAX : TTMP_VI_MAX;
1491 
1492   return (TTmpMin <= Val && Val <= TTmpMax)? Val - TTmpMin : -1;
1493 }
1494 
1495 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
1496                                           bool MandatoryLiteral) const {
1497   using namespace AMDGPU::EncValues;
1498 
1499   assert(Val < 1024); // enum10
1500 
1501   bool IsAGPR = Val & 512;
1502   Val &= 511;
1503 
1504   if (VGPR_MIN <= Val && Val <= VGPR_MAX) {
1505     return createRegOperand(IsAGPR ? getAgprClassId(Width)
1506                                    : getVgprClassId(Width), Val - VGPR_MIN);
1507   }
1508   if (Val <= SGPR_MAX) {
1509     // "SGPR_MIN <= Val" is always true and causes compilation warning.
1510     static_assert(SGPR_MIN == 0, "");
1511     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
1512   }
1513 
1514   int TTmpIdx = getTTmpIdx(Val);
1515   if (TTmpIdx >= 0) {
1516     return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
1517   }
1518 
1519   if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX)
1520     return decodeIntImmed(Val);
1521 
1522   if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX)
1523     return decodeFPImmed(Width, Val);
1524 
1525   if (Val == LITERAL_CONST) {
1526     if (MandatoryLiteral)
1527       // Keep a sentinel value for deferred setting
1528       return MCOperand::createImm(LITERAL_CONST);
1529     else
1530       return decodeLiteralConstant();
1531   }
1532 
1533   switch (Width) {
1534   case OPW32:
1535   case OPW16:
1536   case OPWV216:
1537     return decodeSpecialReg32(Val);
1538   case OPW64:
1539   case OPWV232:
1540     return decodeSpecialReg64(Val);
1541   default:
1542     llvm_unreachable("unexpected immediate type");
1543   }
1544 }
1545 
1546 MCOperand AMDGPUDisassembler::decodeDstOp(const OpWidthTy Width, unsigned Val) const {
1547   using namespace AMDGPU::EncValues;
1548 
1549   assert(Val < 128);
1550   assert(Width == OPW256 || Width == OPW512);
1551 
1552   if (Val <= SGPR_MAX) {
1553     // "SGPR_MIN <= Val" is always true and causes compilation warning.
1554     static_assert(SGPR_MIN == 0, "");
1555     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
1556   }
1557 
1558   int TTmpIdx = getTTmpIdx(Val);
1559   if (TTmpIdx >= 0) {
1560     return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
1561   }
1562 
1563   llvm_unreachable("unknown dst register");
1564 }
1565 
1566 // Bit 0 of DstY isn't stored in the instruction, because it's always the
1567 // opposite of bit 0 of DstX.
1568 MCOperand AMDGPUDisassembler::decodeVOPDDstYOp(MCInst &Inst,
1569                                                unsigned Val) const {
1570   int VDstXInd =
1571       AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::vdstX);
1572   assert(VDstXInd != -1);
1573   assert(Inst.getOperand(VDstXInd).isReg());
1574   unsigned XDstReg = MRI.getEncodingValue(Inst.getOperand(VDstXInd).getReg());
1575   Val |= ~XDstReg & 1;
1576   auto Width = llvm::AMDGPUDisassembler::OPW32;
1577   return createRegOperand(getVgprClassId(Width), Val);
1578 }
1579 
1580 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
1581   using namespace AMDGPU;
1582 
1583   switch (Val) {
1584   case 102: return createRegOperand(FLAT_SCR_LO);
1585   case 103: return createRegOperand(FLAT_SCR_HI);
1586   case 104: return createRegOperand(XNACK_MASK_LO);
1587   case 105: return createRegOperand(XNACK_MASK_HI);
1588   case 106: return createRegOperand(VCC_LO);
1589   case 107: return createRegOperand(VCC_HI);
1590   case 108: return createRegOperand(TBA_LO);
1591   case 109: return createRegOperand(TBA_HI);
1592   case 110: return createRegOperand(TMA_LO);
1593   case 111: return createRegOperand(TMA_HI);
1594   case 124:
1595     return isGFX11Plus() ? createRegOperand(SGPR_NULL) : createRegOperand(M0);
1596   case 125:
1597     return isGFX11Plus() ? createRegOperand(M0) : createRegOperand(SGPR_NULL);
1598   case 126: return createRegOperand(EXEC_LO);
1599   case 127: return createRegOperand(EXEC_HI);
1600   case 235: return createRegOperand(SRC_SHARED_BASE);
1601   case 236: return createRegOperand(SRC_SHARED_LIMIT);
1602   case 237: return createRegOperand(SRC_PRIVATE_BASE);
1603   case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
1604   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1605   case 251: return createRegOperand(SRC_VCCZ);
1606   case 252: return createRegOperand(SRC_EXECZ);
1607   case 253: return createRegOperand(SRC_SCC);
1608   case 254: return createRegOperand(LDS_DIRECT);
1609   default: break;
1610   }
1611   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1612 }
1613 
1614 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
1615   using namespace AMDGPU;
1616 
1617   switch (Val) {
1618   case 102: return createRegOperand(FLAT_SCR);
1619   case 104: return createRegOperand(XNACK_MASK);
1620   case 106: return createRegOperand(VCC);
1621   case 108: return createRegOperand(TBA);
1622   case 110: return createRegOperand(TMA);
1623   case 124:
1624     if (isGFX11Plus())
1625       return createRegOperand(SGPR_NULL);
1626     break;
1627   case 125:
1628     if (!isGFX11Plus())
1629       return createRegOperand(SGPR_NULL);
1630     break;
1631   case 126: return createRegOperand(EXEC);
1632   case 235: return createRegOperand(SRC_SHARED_BASE);
1633   case 236: return createRegOperand(SRC_SHARED_LIMIT);
1634   case 237: return createRegOperand(SRC_PRIVATE_BASE);
1635   case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
1636   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1637   case 251: return createRegOperand(SRC_VCCZ);
1638   case 252: return createRegOperand(SRC_EXECZ);
1639   case 253: return createRegOperand(SRC_SCC);
1640   default: break;
1641   }
1642   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1643 }
1644 
1645 MCOperand AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width,
1646                                             const unsigned Val) const {
1647   using namespace AMDGPU::SDWA;
1648   using namespace AMDGPU::EncValues;
1649 
1650   if (STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
1651       STI.getFeatureBits()[AMDGPU::FeatureGFX10]) {
1652     // XXX: cast to int is needed to avoid stupid warning:
1653     // compare with unsigned is always true
1654     if (int(SDWA9EncValues::SRC_VGPR_MIN) <= int(Val) &&
1655         Val <= SDWA9EncValues::SRC_VGPR_MAX) {
1656       return createRegOperand(getVgprClassId(Width),
1657                               Val - SDWA9EncValues::SRC_VGPR_MIN);
1658     }
1659     if (SDWA9EncValues::SRC_SGPR_MIN <= Val &&
1660         Val <= (isGFX10Plus() ? SDWA9EncValues::SRC_SGPR_MAX_GFX10
1661                               : SDWA9EncValues::SRC_SGPR_MAX_SI)) {
1662       return createSRegOperand(getSgprClassId(Width),
1663                                Val - SDWA9EncValues::SRC_SGPR_MIN);
1664     }
1665     if (SDWA9EncValues::SRC_TTMP_MIN <= Val &&
1666         Val <= SDWA9EncValues::SRC_TTMP_MAX) {
1667       return createSRegOperand(getTtmpClassId(Width),
1668                                Val - SDWA9EncValues::SRC_TTMP_MIN);
1669     }
1670 
1671     const unsigned SVal = Val - SDWA9EncValues::SRC_SGPR_MIN;
1672 
1673     if (INLINE_INTEGER_C_MIN <= SVal && SVal <= INLINE_INTEGER_C_MAX)
1674       return decodeIntImmed(SVal);
1675 
1676     if (INLINE_FLOATING_C_MIN <= SVal && SVal <= INLINE_FLOATING_C_MAX)
1677       return decodeFPImmed(Width, SVal);
1678 
1679     return decodeSpecialReg32(SVal);
1680   } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
1681     return createRegOperand(getVgprClassId(Width), Val);
1682   }
1683   llvm_unreachable("unsupported target");
1684 }
1685 
1686 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const {
1687   return decodeSDWASrc(OPW16, Val);
1688 }
1689 
1690 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const {
1691   return decodeSDWASrc(OPW32, Val);
1692 }
1693 
1694 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const {
1695   using namespace AMDGPU::SDWA;
1696 
1697   assert((STI.getFeatureBits()[AMDGPU::FeatureGFX9] ||
1698           STI.getFeatureBits()[AMDGPU::FeatureGFX10]) &&
1699          "SDWAVopcDst should be present only on GFX9+");
1700 
1701   bool IsWave64 = STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64];
1702 
1703   if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) {
1704     Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
1705 
1706     int TTmpIdx = getTTmpIdx(Val);
1707     if (TTmpIdx >= 0) {
1708       auto TTmpClsId = getTtmpClassId(IsWave64 ? OPW64 : OPW32);
1709       return createSRegOperand(TTmpClsId, TTmpIdx);
1710     } else if (Val > SGPR_MAX) {
1711       return IsWave64 ? decodeSpecialReg64(Val)
1712                       : decodeSpecialReg32(Val);
1713     } else {
1714       return createSRegOperand(getSgprClassId(IsWave64 ? OPW64 : OPW32), Val);
1715     }
1716   } else {
1717     return createRegOperand(IsWave64 ? AMDGPU::VCC : AMDGPU::VCC_LO);
1718   }
1719 }
1720 
1721 MCOperand AMDGPUDisassembler::decodeBoolReg(unsigned Val) const {
1722   return STI.getFeatureBits()[AMDGPU::FeatureWavefrontSize64] ?
1723     decodeOperand_SReg_64(Val) : decodeOperand_SReg_32(Val);
1724 }
1725 
1726 bool AMDGPUDisassembler::isVI() const {
1727   return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
1728 }
1729 
1730 bool AMDGPUDisassembler::isGFX9() const { return AMDGPU::isGFX9(STI); }
1731 
1732 bool AMDGPUDisassembler::isGFX90A() const {
1733   return STI.getFeatureBits()[AMDGPU::FeatureGFX90AInsts];
1734 }
1735 
1736 bool AMDGPUDisassembler::isGFX9Plus() const { return AMDGPU::isGFX9Plus(STI); }
1737 
1738 bool AMDGPUDisassembler::isGFX10() const { return AMDGPU::isGFX10(STI); }
1739 
1740 bool AMDGPUDisassembler::isGFX10Plus() const {
1741   return AMDGPU::isGFX10Plus(STI);
1742 }
1743 
1744 bool AMDGPUDisassembler::isGFX11() const {
1745   return STI.getFeatureBits()[AMDGPU::FeatureGFX11];
1746 }
1747 
1748 bool AMDGPUDisassembler::isGFX11Plus() const {
1749   return AMDGPU::isGFX11Plus(STI);
1750 }
1751 
1752 
1753 bool AMDGPUDisassembler::hasArchitectedFlatScratch() const {
1754   return STI.getFeatureBits()[AMDGPU::FeatureArchitectedFlatScratch];
1755 }
1756 
1757 //===----------------------------------------------------------------------===//
1758 // AMDGPU specific symbol handling
1759 //===----------------------------------------------------------------------===//
1760 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
1761   do {                                                                         \
1762     KdStream << Indent << DIRECTIVE " "                                        \
1763              << ((FourByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';           \
1764   } while (0)
1765 
1766 // NOLINTNEXTLINE(readability-identifier-naming)
1767 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC1(
1768     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
1769   using namespace amdhsa;
1770   StringRef Indent = "\t";
1771 
1772   // We cannot accurately backward compute #VGPRs used from
1773   // GRANULATED_WORKITEM_VGPR_COUNT. But we are concerned with getting the same
1774   // value of GRANULATED_WORKITEM_VGPR_COUNT in the reassembled binary. So we
1775   // simply calculate the inverse of what the assembler does.
1776 
1777   uint32_t GranulatedWorkitemVGPRCount =
1778       (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT) >>
1779       COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT_SHIFT;
1780 
1781   uint32_t NextFreeVGPR = (GranulatedWorkitemVGPRCount + 1) *
1782                           AMDGPU::IsaInfo::getVGPREncodingGranule(&STI);
1783 
1784   KdStream << Indent << ".amdhsa_next_free_vgpr " << NextFreeVGPR << '\n';
1785 
1786   // We cannot backward compute values used to calculate
1787   // GRANULATED_WAVEFRONT_SGPR_COUNT. Hence the original values for following
1788   // directives can't be computed:
1789   // .amdhsa_reserve_vcc
1790   // .amdhsa_reserve_flat_scratch
1791   // .amdhsa_reserve_xnack_mask
1792   // They take their respective default values if not specified in the assembly.
1793   //
1794   // GRANULATED_WAVEFRONT_SGPR_COUNT
1795   //    = f(NEXT_FREE_SGPR + VCC + FLAT_SCRATCH + XNACK_MASK)
1796   //
1797   // We compute the inverse as though all directives apart from NEXT_FREE_SGPR
1798   // are set to 0. So while disassembling we consider that:
1799   //
1800   // GRANULATED_WAVEFRONT_SGPR_COUNT
1801   //    = f(NEXT_FREE_SGPR + 0 + 0 + 0)
1802   //
1803   // The disassembler cannot recover the original values of those 3 directives.
1804 
1805   uint32_t GranulatedWavefrontSGPRCount =
1806       (FourByteBuffer & COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT) >>
1807       COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT_SHIFT;
1808 
1809   if (isGFX10Plus() && GranulatedWavefrontSGPRCount)
1810     return MCDisassembler::Fail;
1811 
1812   uint32_t NextFreeSGPR = (GranulatedWavefrontSGPRCount + 1) *
1813                           AMDGPU::IsaInfo::getSGPREncodingGranule(&STI);
1814 
1815   KdStream << Indent << ".amdhsa_reserve_vcc " << 0 << '\n';
1816   if (!hasArchitectedFlatScratch())
1817     KdStream << Indent << ".amdhsa_reserve_flat_scratch " << 0 << '\n';
1818   KdStream << Indent << ".amdhsa_reserve_xnack_mask " << 0 << '\n';
1819   KdStream << Indent << ".amdhsa_next_free_sgpr " << NextFreeSGPR << "\n";
1820 
1821   if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIORITY)
1822     return MCDisassembler::Fail;
1823 
1824   PRINT_DIRECTIVE(".amdhsa_float_round_mode_32",
1825                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
1826   PRINT_DIRECTIVE(".amdhsa_float_round_mode_16_64",
1827                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
1828   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_32",
1829                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
1830   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_16_64",
1831                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
1832 
1833   if (FourByteBuffer & COMPUTE_PGM_RSRC1_PRIV)
1834     return MCDisassembler::Fail;
1835 
1836   PRINT_DIRECTIVE(".amdhsa_dx10_clamp", COMPUTE_PGM_RSRC1_ENABLE_DX10_CLAMP);
1837 
1838   if (FourByteBuffer & COMPUTE_PGM_RSRC1_DEBUG_MODE)
1839     return MCDisassembler::Fail;
1840 
1841   PRINT_DIRECTIVE(".amdhsa_ieee_mode", COMPUTE_PGM_RSRC1_ENABLE_IEEE_MODE);
1842 
1843   if (FourByteBuffer & COMPUTE_PGM_RSRC1_BULKY)
1844     return MCDisassembler::Fail;
1845 
1846   if (FourByteBuffer & COMPUTE_PGM_RSRC1_CDBG_USER)
1847     return MCDisassembler::Fail;
1848 
1849   PRINT_DIRECTIVE(".amdhsa_fp16_overflow", COMPUTE_PGM_RSRC1_FP16_OVFL);
1850 
1851   if (FourByteBuffer & COMPUTE_PGM_RSRC1_RESERVED0)
1852     return MCDisassembler::Fail;
1853 
1854   if (isGFX10Plus()) {
1855     PRINT_DIRECTIVE(".amdhsa_workgroup_processor_mode",
1856                     COMPUTE_PGM_RSRC1_WGP_MODE);
1857     PRINT_DIRECTIVE(".amdhsa_memory_ordered", COMPUTE_PGM_RSRC1_MEM_ORDERED);
1858     PRINT_DIRECTIVE(".amdhsa_forward_progress", COMPUTE_PGM_RSRC1_FWD_PROGRESS);
1859   }
1860   return MCDisassembler::Success;
1861 }
1862 
1863 // NOLINTNEXTLINE(readability-identifier-naming)
1864 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC2(
1865     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
1866   using namespace amdhsa;
1867   StringRef Indent = "\t";
1868   if (hasArchitectedFlatScratch())
1869     PRINT_DIRECTIVE(".amdhsa_enable_private_segment",
1870                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
1871   else
1872     PRINT_DIRECTIVE(".amdhsa_system_sgpr_private_segment_wavefront_offset",
1873                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
1874   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_x",
1875                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
1876   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_y",
1877                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
1878   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_z",
1879                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
1880   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_info",
1881                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
1882   PRINT_DIRECTIVE(".amdhsa_system_vgpr_workitem_id",
1883                   COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
1884 
1885   if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_ADDRESS_WATCH)
1886     return MCDisassembler::Fail;
1887 
1888   if (FourByteBuffer & COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_MEMORY)
1889     return MCDisassembler::Fail;
1890 
1891   if (FourByteBuffer & COMPUTE_PGM_RSRC2_GRANULATED_LDS_SIZE)
1892     return MCDisassembler::Fail;
1893 
1894   PRINT_DIRECTIVE(
1895       ".amdhsa_exception_fp_ieee_invalid_op",
1896       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
1897   PRINT_DIRECTIVE(".amdhsa_exception_fp_denorm_src",
1898                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
1899   PRINT_DIRECTIVE(
1900       ".amdhsa_exception_fp_ieee_div_zero",
1901       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
1902   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_overflow",
1903                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
1904   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_underflow",
1905                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
1906   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_inexact",
1907                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
1908   PRINT_DIRECTIVE(".amdhsa_exception_int_div_zero",
1909                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
1910 
1911   if (FourByteBuffer & COMPUTE_PGM_RSRC2_RESERVED0)
1912     return MCDisassembler::Fail;
1913 
1914   return MCDisassembler::Success;
1915 }
1916 
1917 #undef PRINT_DIRECTIVE
1918 
1919 MCDisassembler::DecodeStatus
1920 AMDGPUDisassembler::decodeKernelDescriptorDirective(
1921     DataExtractor::Cursor &Cursor, ArrayRef<uint8_t> Bytes,
1922     raw_string_ostream &KdStream) const {
1923 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
1924   do {                                                                         \
1925     KdStream << Indent << DIRECTIVE " "                                        \
1926              << ((TwoByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';            \
1927   } while (0)
1928 
1929   uint16_t TwoByteBuffer = 0;
1930   uint32_t FourByteBuffer = 0;
1931 
1932   StringRef ReservedBytes;
1933   StringRef Indent = "\t";
1934 
1935   assert(Bytes.size() == 64);
1936   DataExtractor DE(Bytes, /*IsLittleEndian=*/true, /*AddressSize=*/8);
1937 
1938   switch (Cursor.tell()) {
1939   case amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET:
1940     FourByteBuffer = DE.getU32(Cursor);
1941     KdStream << Indent << ".amdhsa_group_segment_fixed_size " << FourByteBuffer
1942              << '\n';
1943     return MCDisassembler::Success;
1944 
1945   case amdhsa::PRIVATE_SEGMENT_FIXED_SIZE_OFFSET:
1946     FourByteBuffer = DE.getU32(Cursor);
1947     KdStream << Indent << ".amdhsa_private_segment_fixed_size "
1948              << FourByteBuffer << '\n';
1949     return MCDisassembler::Success;
1950 
1951   case amdhsa::KERNARG_SIZE_OFFSET:
1952     FourByteBuffer = DE.getU32(Cursor);
1953     KdStream << Indent << ".amdhsa_kernarg_size "
1954              << FourByteBuffer << '\n';
1955     return MCDisassembler::Success;
1956 
1957   case amdhsa::RESERVED0_OFFSET:
1958     // 4 reserved bytes, must be 0.
1959     ReservedBytes = DE.getBytes(Cursor, 4);
1960     for (int I = 0; I < 4; ++I) {
1961       if (ReservedBytes[I] != 0) {
1962         return MCDisassembler::Fail;
1963       }
1964     }
1965     return MCDisassembler::Success;
1966 
1967   case amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET:
1968     // KERNEL_CODE_ENTRY_BYTE_OFFSET
1969     // So far no directive controls this for Code Object V3, so simply skip for
1970     // disassembly.
1971     DE.skip(Cursor, 8);
1972     return MCDisassembler::Success;
1973 
1974   case amdhsa::RESERVED1_OFFSET:
1975     // 20 reserved bytes, must be 0.
1976     ReservedBytes = DE.getBytes(Cursor, 20);
1977     for (int I = 0; I < 20; ++I) {
1978       if (ReservedBytes[I] != 0) {
1979         return MCDisassembler::Fail;
1980       }
1981     }
1982     return MCDisassembler::Success;
1983 
1984   case amdhsa::COMPUTE_PGM_RSRC3_OFFSET:
1985     // COMPUTE_PGM_RSRC3
1986     //  - Only set for GFX10, GFX6-9 have this to be 0.
1987     //  - Currently no directives directly control this.
1988     FourByteBuffer = DE.getU32(Cursor);
1989     if (!isGFX10Plus() && FourByteBuffer) {
1990       return MCDisassembler::Fail;
1991     }
1992     return MCDisassembler::Success;
1993 
1994   case amdhsa::COMPUTE_PGM_RSRC1_OFFSET:
1995     FourByteBuffer = DE.getU32(Cursor);
1996     if (decodeCOMPUTE_PGM_RSRC1(FourByteBuffer, KdStream) ==
1997         MCDisassembler::Fail) {
1998       return MCDisassembler::Fail;
1999     }
2000     return MCDisassembler::Success;
2001 
2002   case amdhsa::COMPUTE_PGM_RSRC2_OFFSET:
2003     FourByteBuffer = DE.getU32(Cursor);
2004     if (decodeCOMPUTE_PGM_RSRC2(FourByteBuffer, KdStream) ==
2005         MCDisassembler::Fail) {
2006       return MCDisassembler::Fail;
2007     }
2008     return MCDisassembler::Success;
2009 
2010   case amdhsa::KERNEL_CODE_PROPERTIES_OFFSET:
2011     using namespace amdhsa;
2012     TwoByteBuffer = DE.getU16(Cursor);
2013 
2014     if (!hasArchitectedFlatScratch())
2015       PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_buffer",
2016                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
2017     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_ptr",
2018                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
2019     PRINT_DIRECTIVE(".amdhsa_user_sgpr_queue_ptr",
2020                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
2021     PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_segment_ptr",
2022                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
2023     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_id",
2024                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
2025     if (!hasArchitectedFlatScratch())
2026       PRINT_DIRECTIVE(".amdhsa_user_sgpr_flat_scratch_init",
2027                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
2028     PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size",
2029                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
2030 
2031     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0)
2032       return MCDisassembler::Fail;
2033 
2034     // Reserved for GFX9
2035     if (isGFX9() &&
2036         (TwoByteBuffer & KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32)) {
2037       return MCDisassembler::Fail;
2038     } else if (isGFX10Plus()) {
2039       PRINT_DIRECTIVE(".amdhsa_wavefront_size32",
2040                       KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
2041     }
2042 
2043     PRINT_DIRECTIVE(".amdhsa_uses_dynamic_stack",
2044                     KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK);
2045 
2046     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED1)
2047       return MCDisassembler::Fail;
2048 
2049     return MCDisassembler::Success;
2050 
2051   case amdhsa::RESERVED2_OFFSET:
2052     // 6 bytes from here are reserved, must be 0.
2053     ReservedBytes = DE.getBytes(Cursor, 6);
2054     for (int I = 0; I < 6; ++I) {
2055       if (ReservedBytes[I] != 0)
2056         return MCDisassembler::Fail;
2057     }
2058     return MCDisassembler::Success;
2059 
2060   default:
2061     llvm_unreachable("Unhandled index. Case statements cover everything.");
2062     return MCDisassembler::Fail;
2063   }
2064 #undef PRINT_DIRECTIVE
2065 }
2066 
2067 MCDisassembler::DecodeStatus AMDGPUDisassembler::decodeKernelDescriptor(
2068     StringRef KdName, ArrayRef<uint8_t> Bytes, uint64_t KdAddress) const {
2069   // CP microcode requires the kernel descriptor to be 64 aligned.
2070   if (Bytes.size() != 64 || KdAddress % 64 != 0)
2071     return MCDisassembler::Fail;
2072 
2073   std::string Kd;
2074   raw_string_ostream KdStream(Kd);
2075   KdStream << ".amdhsa_kernel " << KdName << '\n';
2076 
2077   DataExtractor::Cursor C(0);
2078   while (C && C.tell() < Bytes.size()) {
2079     MCDisassembler::DecodeStatus Status =
2080         decodeKernelDescriptorDirective(C, Bytes, KdStream);
2081 
2082     cantFail(C.takeError());
2083 
2084     if (Status == MCDisassembler::Fail)
2085       return MCDisassembler::Fail;
2086   }
2087   KdStream << ".end_amdhsa_kernel\n";
2088   outs() << KdStream.str();
2089   return MCDisassembler::Success;
2090 }
2091 
2092 Optional<MCDisassembler::DecodeStatus>
2093 AMDGPUDisassembler::onSymbolStart(SymbolInfoTy &Symbol, uint64_t &Size,
2094                                   ArrayRef<uint8_t> Bytes, uint64_t Address,
2095                                   raw_ostream &CStream) const {
2096   // Right now only kernel descriptor needs to be handled.
2097   // We ignore all other symbols for target specific handling.
2098   // TODO:
2099   // Fix the spurious symbol issue for AMDGPU kernels. Exists for both Code
2100   // Object V2 and V3 when symbols are marked protected.
2101 
2102   // amd_kernel_code_t for Code Object V2.
2103   if (Symbol.Type == ELF::STT_AMDGPU_HSA_KERNEL) {
2104     Size = 256;
2105     return MCDisassembler::Fail;
2106   }
2107 
2108   // Code Object V3 kernel descriptors.
2109   StringRef Name = Symbol.Name;
2110   if (Symbol.Type == ELF::STT_OBJECT && Name.endswith(StringRef(".kd"))) {
2111     Size = 64; // Size = 64 regardless of success or failure.
2112     return decodeKernelDescriptor(Name.drop_back(3), Bytes, Address);
2113   }
2114   return None;
2115 }
2116 
2117 //===----------------------------------------------------------------------===//
2118 // AMDGPUSymbolizer
2119 //===----------------------------------------------------------------------===//
2120 
2121 // Try to find symbol name for specified label
2122 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(
2123     MCInst &Inst, raw_ostream & /*cStream*/, int64_t Value,
2124     uint64_t /*Address*/, bool IsBranch, uint64_t /*Offset*/,
2125     uint64_t /*OpSize*/, uint64_t /*InstSize*/) {
2126 
2127   if (!IsBranch) {
2128     return false;
2129   }
2130 
2131   auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
2132   if (!Symbols)
2133     return false;
2134 
2135   auto Result = llvm::find_if(*Symbols, [Value](const SymbolInfoTy &Val) {
2136     return Val.Addr == static_cast<uint64_t>(Value) &&
2137            Val.Type == ELF::STT_NOTYPE;
2138   });
2139   if (Result != Symbols->end()) {
2140     auto *Sym = Ctx.getOrCreateSymbol(Result->Name);
2141     const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
2142     Inst.addOperand(MCOperand::createExpr(Add));
2143     return true;
2144   }
2145   // Add to list of referenced addresses, so caller can synthesize a label.
2146   ReferencedAddresses.push_back(static_cast<uint64_t>(Value));
2147   return false;
2148 }
2149 
2150 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
2151                                                        int64_t Value,
2152                                                        uint64_t Address) {
2153   llvm_unreachable("unimplemented");
2154 }
2155 
2156 //===----------------------------------------------------------------------===//
2157 // Initialization
2158 //===----------------------------------------------------------------------===//
2159 
2160 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
2161                               LLVMOpInfoCallback /*GetOpInfo*/,
2162                               LLVMSymbolLookupCallback /*SymbolLookUp*/,
2163                               void *DisInfo,
2164                               MCContext *Ctx,
2165                               std::unique_ptr<MCRelocationInfo> &&RelInfo) {
2166   return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
2167 }
2168 
2169 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
2170                                                 const MCSubtargetInfo &STI,
2171                                                 MCContext &Ctx) {
2172   return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo());
2173 }
2174 
2175 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUDisassembler() {
2176   TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
2177                                          createAMDGPUDisassembler);
2178   TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
2179                                        createAMDGPUSymbolizer);
2180 }
2181