xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp (revision 47ef2a131091508e049ab10cad7f91a3c1342cd9)
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/AMDGPUAsmUtils.h"
25 #include "Utils/AMDGPUBaseInfo.h"
26 #include "llvm-c/DisassemblerTypes.h"
27 #include "llvm/BinaryFormat/ELF.h"
28 #include "llvm/MC/MCAsmInfo.h"
29 #include "llvm/MC/MCContext.h"
30 #include "llvm/MC/MCDecoderOps.h"
31 #include "llvm/MC/MCExpr.h"
32 #include "llvm/MC/MCInstrDesc.h"
33 #include "llvm/MC/MCRegisterInfo.h"
34 #include "llvm/MC/MCSubtargetInfo.h"
35 #include "llvm/MC/TargetRegistry.h"
36 #include "llvm/Support/AMDHSAKernelDescriptor.h"
37 
38 using namespace llvm;
39 
40 #define DEBUG_TYPE "amdgpu-disassembler"
41 
42 #define SGPR_MAX                                                               \
43   (isGFX10Plus() ? AMDGPU::EncValues::SGPR_MAX_GFX10                           \
44                  : AMDGPU::EncValues::SGPR_MAX_SI)
45 
46 using DecodeStatus = llvm::MCDisassembler::DecodeStatus;
47 
48 static const MCSubtargetInfo &addDefaultWaveSize(const MCSubtargetInfo &STI,
49                                                  MCContext &Ctx) {
50   if (!STI.hasFeature(AMDGPU::FeatureWavefrontSize64) &&
51       !STI.hasFeature(AMDGPU::FeatureWavefrontSize32)) {
52     MCSubtargetInfo &STICopy = Ctx.getSubtargetCopy(STI);
53     // If there is no default wave size it must be a generation before gfx10,
54     // these have FeatureWavefrontSize64 in their definition already. For gfx10+
55     // set wave32 as a default.
56     STICopy.ToggleFeature(AMDGPU::FeatureWavefrontSize32);
57     return STICopy;
58   }
59 
60   return STI;
61 }
62 
63 AMDGPUDisassembler::AMDGPUDisassembler(const MCSubtargetInfo &STI,
64                                        MCContext &Ctx, MCInstrInfo const *MCII)
65     : MCDisassembler(addDefaultWaveSize(STI, Ctx), Ctx), MCII(MCII),
66       MRI(*Ctx.getRegisterInfo()), MAI(*Ctx.getAsmInfo()),
67       TargetMaxInstBytes(MAI.getMaxInstLength(&STI)),
68       CodeObjectVersion(AMDGPU::getDefaultAMDHSACodeObjectVersion()) {
69   // ToDo: AMDGPUDisassembler supports only VI ISA.
70   if (!STI.hasFeature(AMDGPU::FeatureGCN3Encoding) && !isGFX10Plus())
71     report_fatal_error("Disassembly not yet supported for subtarget");
72 
73   for (auto [Symbol, Code] : AMDGPU::UCVersion::getGFXVersions())
74     createConstantSymbolExpr(Symbol, Code);
75 
76   UCVersionW64Expr = createConstantSymbolExpr("UC_VERSION_W64_BIT", 0x2000);
77   UCVersionW32Expr = createConstantSymbolExpr("UC_VERSION_W32_BIT", 0x4000);
78   UCVersionMDPExpr = createConstantSymbolExpr("UC_VERSION_MDP_BIT", 0x8000);
79 }
80 
81 void AMDGPUDisassembler::setABIVersion(unsigned Version) {
82   CodeObjectVersion = AMDGPU::getAMDHSACodeObjectVersion(Version);
83 }
84 
85 inline static MCDisassembler::DecodeStatus
86 addOperand(MCInst &Inst, const MCOperand& Opnd) {
87   Inst.addOperand(Opnd);
88   return Opnd.isValid() ?
89     MCDisassembler::Success :
90     MCDisassembler::Fail;
91 }
92 
93 static int insertNamedMCOperand(MCInst &MI, const MCOperand &Op,
94                                 uint16_t NameIdx) {
95   int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), NameIdx);
96   if (OpIdx != -1) {
97     auto I = MI.begin();
98     std::advance(I, OpIdx);
99     MI.insert(I, Op);
100   }
101   return OpIdx;
102 }
103 
104 static DecodeStatus decodeSOPPBrTarget(MCInst &Inst, unsigned Imm,
105                                        uint64_t Addr,
106                                        const MCDisassembler *Decoder) {
107   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
108 
109   // Our branches take a simm16, but we need two extra bits to account for the
110   // factor of 4.
111   APInt SignedOffset(18, Imm * 4, true);
112   int64_t Offset = (SignedOffset.sext(64) + 4 + Addr).getSExtValue();
113 
114   if (DAsm->tryAddingSymbolicOperand(Inst, Offset, Addr, true, 2, 2, 0))
115     return MCDisassembler::Success;
116   return addOperand(Inst, MCOperand::createImm(Imm));
117 }
118 
119 static DecodeStatus decodeSMEMOffset(MCInst &Inst, unsigned Imm, uint64_t Addr,
120                                      const MCDisassembler *Decoder) {
121   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
122   int64_t Offset;
123   if (DAsm->isGFX12Plus()) { // GFX12 supports 24-bit signed offsets.
124     Offset = SignExtend64<24>(Imm);
125   } else if (DAsm->isVI()) { // VI supports 20-bit unsigned offsets.
126     Offset = Imm & 0xFFFFF;
127   } else { // GFX9+ supports 21-bit signed offsets.
128     Offset = SignExtend64<21>(Imm);
129   }
130   return addOperand(Inst, MCOperand::createImm(Offset));
131 }
132 
133 static DecodeStatus decodeBoolReg(MCInst &Inst, unsigned Val, uint64_t Addr,
134                                   const MCDisassembler *Decoder) {
135   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
136   return addOperand(Inst, DAsm->decodeBoolReg(Val));
137 }
138 
139 static DecodeStatus decodeSplitBarrier(MCInst &Inst, unsigned Val,
140                                        uint64_t Addr,
141                                        const MCDisassembler *Decoder) {
142   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
143   return addOperand(Inst, DAsm->decodeSplitBarrier(Val));
144 }
145 
146 static DecodeStatus decodeDpp8FI(MCInst &Inst, unsigned Val, uint64_t Addr,
147                                  const MCDisassembler *Decoder) {
148   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
149   return addOperand(Inst, DAsm->decodeDpp8FI(Val));
150 }
151 
152 #define DECODE_OPERAND(StaticDecoderName, DecoderName)                         \
153   static DecodeStatus StaticDecoderName(MCInst &Inst, unsigned Imm,            \
154                                         uint64_t /*Addr*/,                     \
155                                         const MCDisassembler *Decoder) {       \
156     auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);              \
157     return addOperand(Inst, DAsm->DecoderName(Imm));                           \
158   }
159 
160 // Decoder for registers, decode directly using RegClassID. Imm(8-bit) is
161 // number of register. Used by VGPR only and AGPR only operands.
162 #define DECODE_OPERAND_REG_8(RegClass)                                         \
163   static DecodeStatus Decode##RegClass##RegisterClass(                         \
164       MCInst &Inst, unsigned Imm, uint64_t /*Addr*/,                           \
165       const MCDisassembler *Decoder) {                                         \
166     assert(Imm < (1 << 8) && "8-bit encoding");                                \
167     auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);              \
168     return addOperand(                                                         \
169         Inst, DAsm->createRegOperand(AMDGPU::RegClass##RegClassID, Imm));      \
170   }
171 
172 #define DECODE_SrcOp(Name, EncSize, OpWidth, EncImm, MandatoryLiteral,         \
173                      ImmWidth)                                                 \
174   static DecodeStatus Name(MCInst &Inst, unsigned Imm, uint64_t /*Addr*/,      \
175                            const MCDisassembler *Decoder) {                    \
176     assert(Imm < (1 << EncSize) && #EncSize "-bit encoding");                  \
177     auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);              \
178     return addOperand(Inst,                                                    \
179                       DAsm->decodeSrcOp(AMDGPUDisassembler::OpWidth, EncImm,   \
180                                         MandatoryLiteral, ImmWidth));          \
181   }
182 
183 static DecodeStatus decodeSrcOp(MCInst &Inst, unsigned EncSize,
184                                 AMDGPUDisassembler::OpWidthTy OpWidth,
185                                 unsigned Imm, unsigned EncImm,
186                                 bool MandatoryLiteral, unsigned ImmWidth,
187                                 AMDGPU::OperandSemantics Sema,
188                                 const MCDisassembler *Decoder) {
189   assert(Imm < (1U << EncSize) && "Operand doesn't fit encoding!");
190   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
191   return addOperand(Inst, DAsm->decodeSrcOp(OpWidth, EncImm, MandatoryLiteral,
192                                             ImmWidth, Sema));
193 }
194 
195 // Decoder for registers. Imm(7-bit) is number of register, uses decodeSrcOp to
196 // get register class. Used by SGPR only operands.
197 #define DECODE_OPERAND_REG_7(RegClass, OpWidth)                                \
198   DECODE_SrcOp(Decode##RegClass##RegisterClass, 7, OpWidth, Imm, false, 0)
199 
200 // Decoder for registers. Imm(10-bit): Imm{7-0} is number of register,
201 // Imm{9} is acc(agpr or vgpr) Imm{8} should be 0 (see VOP3Pe_SMFMAC).
202 // Set Imm{8} to 1 (IS_VGPR) to decode using 'enum10' from decodeSrcOp.
203 // Used by AV_ register classes (AGPR or VGPR only register operands).
204 template <AMDGPUDisassembler::OpWidthTy OpWidth>
205 static DecodeStatus decodeAV10(MCInst &Inst, unsigned Imm, uint64_t /* Addr */,
206                                const MCDisassembler *Decoder) {
207   return decodeSrcOp(Inst, 10, OpWidth, Imm, Imm | AMDGPU::EncValues::IS_VGPR,
208                      false, 0, AMDGPU::OperandSemantics::INT, Decoder);
209 }
210 
211 // Decoder for Src(9-bit encoding) registers only.
212 template <AMDGPUDisassembler::OpWidthTy OpWidth>
213 static DecodeStatus decodeSrcReg9(MCInst &Inst, unsigned Imm,
214                                   uint64_t /* Addr */,
215                                   const MCDisassembler *Decoder) {
216   return decodeSrcOp(Inst, 9, OpWidth, Imm, Imm, false, 0,
217                      AMDGPU::OperandSemantics::INT, Decoder);
218 }
219 
220 // Decoder for Src(9-bit encoding) AGPR, register number encoded in 9bits, set
221 // Imm{9} to 1 (set acc) and decode using 'enum10' from decodeSrcOp, registers
222 // only.
223 template <AMDGPUDisassembler::OpWidthTy OpWidth>
224 static DecodeStatus decodeSrcA9(MCInst &Inst, unsigned Imm, uint64_t /* Addr */,
225                                 const MCDisassembler *Decoder) {
226   return decodeSrcOp(Inst, 9, OpWidth, Imm, Imm | 512, false, 0,
227                      AMDGPU::OperandSemantics::INT, Decoder);
228 }
229 
230 // Decoder for 'enum10' from decodeSrcOp, Imm{0-8} is 9-bit Src encoding
231 // Imm{9} is acc, registers only.
232 template <AMDGPUDisassembler::OpWidthTy OpWidth>
233 static DecodeStatus decodeSrcAV10(MCInst &Inst, unsigned Imm,
234                                   uint64_t /* Addr */,
235                                   const MCDisassembler *Decoder) {
236   return decodeSrcOp(Inst, 10, OpWidth, Imm, Imm, false, 0,
237                      AMDGPU::OperandSemantics::INT, Decoder);
238 }
239 
240 // Decoder for RegisterOperands using 9-bit Src encoding. Operand can be
241 // register from RegClass or immediate. Registers that don't belong to RegClass
242 // will be decoded and InstPrinter will report warning. Immediate will be
243 // decoded into constant of size ImmWidth, should match width of immediate used
244 // by OperandType (important for floating point types).
245 template <AMDGPUDisassembler::OpWidthTy OpWidth, unsigned ImmWidth,
246           unsigned OperandSemantics>
247 static DecodeStatus decodeSrcRegOrImm9(MCInst &Inst, unsigned Imm,
248                                        uint64_t /* Addr */,
249                                        const MCDisassembler *Decoder) {
250   return decodeSrcOp(Inst, 9, OpWidth, Imm, Imm, false, ImmWidth,
251                      (AMDGPU::OperandSemantics)OperandSemantics, Decoder);
252 }
253 
254 // Decoder for Src(9-bit encoding) AGPR or immediate. Set Imm{9} to 1 (set acc)
255 // and decode using 'enum10' from decodeSrcOp.
256 template <AMDGPUDisassembler::OpWidthTy OpWidth, unsigned ImmWidth,
257           unsigned OperandSemantics>
258 static DecodeStatus decodeSrcRegOrImmA9(MCInst &Inst, unsigned Imm,
259                                         uint64_t /* Addr */,
260                                         const MCDisassembler *Decoder) {
261   return decodeSrcOp(Inst, 9, OpWidth, Imm, Imm | 512, false, ImmWidth,
262                      (AMDGPU::OperandSemantics)OperandSemantics, Decoder);
263 }
264 
265 template <AMDGPUDisassembler::OpWidthTy OpWidth, unsigned ImmWidth,
266           unsigned OperandSemantics>
267 static DecodeStatus decodeSrcRegOrImmDeferred9(MCInst &Inst, unsigned Imm,
268                                                uint64_t /* Addr */,
269                                                const MCDisassembler *Decoder) {
270   return decodeSrcOp(Inst, 9, OpWidth, Imm, Imm, true, ImmWidth,
271                      (AMDGPU::OperandSemantics)OperandSemantics, Decoder);
272 }
273 
274 // Default decoders generated by tablegen: 'Decode<RegClass>RegisterClass'
275 // when RegisterClass is used as an operand. Most often used for destination
276 // operands.
277 
278 DECODE_OPERAND_REG_8(VGPR_32)
279 DECODE_OPERAND_REG_8(VGPR_32_Lo128)
280 DECODE_OPERAND_REG_8(VReg_64)
281 DECODE_OPERAND_REG_8(VReg_96)
282 DECODE_OPERAND_REG_8(VReg_128)
283 DECODE_OPERAND_REG_8(VReg_256)
284 DECODE_OPERAND_REG_8(VReg_288)
285 DECODE_OPERAND_REG_8(VReg_352)
286 DECODE_OPERAND_REG_8(VReg_384)
287 DECODE_OPERAND_REG_8(VReg_512)
288 DECODE_OPERAND_REG_8(VReg_1024)
289 
290 DECODE_OPERAND_REG_7(SReg_32, OPW32)
291 DECODE_OPERAND_REG_7(SReg_32_XEXEC, OPW32)
292 DECODE_OPERAND_REG_7(SReg_32_XM0_XEXEC, OPW32)
293 DECODE_OPERAND_REG_7(SReg_32_XEXEC_HI, OPW32)
294 DECODE_OPERAND_REG_7(SReg_64, OPW64)
295 DECODE_OPERAND_REG_7(SReg_64_XEXEC, OPW64)
296 DECODE_OPERAND_REG_7(SReg_96, OPW96)
297 DECODE_OPERAND_REG_7(SReg_128, OPW128)
298 DECODE_OPERAND_REG_7(SReg_256, OPW256)
299 DECODE_OPERAND_REG_7(SReg_512, OPW512)
300 
301 DECODE_OPERAND_REG_8(AGPR_32)
302 DECODE_OPERAND_REG_8(AReg_64)
303 DECODE_OPERAND_REG_8(AReg_128)
304 DECODE_OPERAND_REG_8(AReg_256)
305 DECODE_OPERAND_REG_8(AReg_512)
306 DECODE_OPERAND_REG_8(AReg_1024)
307 
308 static DecodeStatus DecodeVGPR_16RegisterClass(MCInst &Inst, unsigned Imm,
309                                                uint64_t /*Addr*/,
310                                                const MCDisassembler *Decoder) {
311   assert(isUInt<10>(Imm) && "10-bit encoding expected");
312   assert((Imm & (1 << 8)) == 0 && "Imm{8} should not be used");
313 
314   bool IsHi = Imm & (1 << 9);
315   unsigned RegIdx = Imm & 0xff;
316   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
317   return addOperand(Inst, DAsm->createVGPR16Operand(RegIdx, IsHi));
318 }
319 
320 static DecodeStatus
321 DecodeVGPR_16_Lo128RegisterClass(MCInst &Inst, unsigned Imm, uint64_t /*Addr*/,
322                                  const MCDisassembler *Decoder) {
323   assert(isUInt<8>(Imm) && "8-bit encoding expected");
324 
325   bool IsHi = Imm & (1 << 7);
326   unsigned RegIdx = Imm & 0x7f;
327   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
328   return addOperand(Inst, DAsm->createVGPR16Operand(RegIdx, IsHi));
329 }
330 
331 static DecodeStatus decodeOperand_VSrcT16_Lo128(MCInst &Inst, unsigned Imm,
332                                                 uint64_t /*Addr*/,
333                                                 const MCDisassembler *Decoder) {
334   assert(isUInt<9>(Imm) && "9-bit encoding expected");
335 
336   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
337   bool IsVGPR = Imm & (1 << 8);
338   if (IsVGPR) {
339     bool IsHi = Imm & (1 << 7);
340     unsigned RegIdx = Imm & 0x7f;
341     return addOperand(Inst, DAsm->createVGPR16Operand(RegIdx, IsHi));
342   }
343   return addOperand(Inst, DAsm->decodeNonVGPRSrcOp(AMDGPUDisassembler::OPW16,
344                                                    Imm & 0xFF, false, 16));
345 }
346 
347 static DecodeStatus decodeOperand_VSrcT16(MCInst &Inst, unsigned Imm,
348                                           uint64_t /*Addr*/,
349                                           const MCDisassembler *Decoder) {
350   assert(isUInt<10>(Imm) && "10-bit encoding expected");
351 
352   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
353   bool IsVGPR = Imm & (1 << 8);
354   if (IsVGPR) {
355     bool IsHi = Imm & (1 << 9);
356     unsigned RegIdx = Imm & 0xff;
357     return addOperand(Inst, DAsm->createVGPR16Operand(RegIdx, IsHi));
358   }
359   return addOperand(Inst, DAsm->decodeNonVGPRSrcOp(AMDGPUDisassembler::OPW16,
360                                                    Imm & 0xFF, false, 16));
361 }
362 
363 static DecodeStatus decodeOperand_KImmFP(MCInst &Inst, unsigned Imm,
364                                          uint64_t Addr,
365                                          const MCDisassembler *Decoder) {
366   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
367   return addOperand(Inst, DAsm->decodeMandatoryLiteralConstant(Imm));
368 }
369 
370 static DecodeStatus decodeOperandVOPDDstY(MCInst &Inst, unsigned Val,
371                                           uint64_t Addr, const void *Decoder) {
372   const auto *DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
373   return addOperand(Inst, DAsm->decodeVOPDDstYOp(Inst, Val));
374 }
375 
376 static bool IsAGPROperand(const MCInst &Inst, int OpIdx,
377                           const MCRegisterInfo *MRI) {
378   if (OpIdx < 0)
379     return false;
380 
381   const MCOperand &Op = Inst.getOperand(OpIdx);
382   if (!Op.isReg())
383     return false;
384 
385   unsigned Sub = MRI->getSubReg(Op.getReg(), AMDGPU::sub0);
386   auto Reg = Sub ? Sub : Op.getReg();
387   return Reg >= AMDGPU::AGPR0 && Reg <= AMDGPU::AGPR255;
388 }
389 
390 static DecodeStatus decodeAVLdSt(MCInst &Inst, unsigned Imm,
391                                  AMDGPUDisassembler::OpWidthTy Opw,
392                                  const MCDisassembler *Decoder) {
393   auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
394   if (!DAsm->isGFX90A()) {
395     Imm &= 511;
396   } else {
397     // If atomic has both vdata and vdst their register classes are tied.
398     // The bit is decoded along with the vdst, first operand. We need to
399     // change register class to AGPR if vdst was AGPR.
400     // If a DS instruction has both data0 and data1 their register classes
401     // are also tied.
402     unsigned Opc = Inst.getOpcode();
403     uint64_t TSFlags = DAsm->getMCII()->get(Opc).TSFlags;
404     uint16_t DataNameIdx = (TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0
405                                                         : AMDGPU::OpName::vdata;
406     const MCRegisterInfo *MRI = DAsm->getContext().getRegisterInfo();
407     int DataIdx = AMDGPU::getNamedOperandIdx(Opc, DataNameIdx);
408     if ((int)Inst.getNumOperands() == DataIdx) {
409       int DstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
410       if (IsAGPROperand(Inst, DstIdx, MRI))
411         Imm |= 512;
412     }
413 
414     if (TSFlags & SIInstrFlags::DS) {
415       int Data2Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
416       if ((int)Inst.getNumOperands() == Data2Idx &&
417           IsAGPROperand(Inst, DataIdx, MRI))
418         Imm |= 512;
419     }
420   }
421   return addOperand(Inst, DAsm->decodeSrcOp(Opw, Imm | 256));
422 }
423 
424 template <AMDGPUDisassembler::OpWidthTy Opw>
425 static DecodeStatus decodeAVLdSt(MCInst &Inst, unsigned Imm,
426                                  uint64_t /* Addr */,
427                                  const MCDisassembler *Decoder) {
428   return decodeAVLdSt(Inst, Imm, Opw, Decoder);
429 }
430 
431 static DecodeStatus decodeOperand_VSrc_f64(MCInst &Inst, unsigned Imm,
432                                            uint64_t Addr,
433                                            const MCDisassembler *Decoder) {
434   assert(Imm < (1 << 9) && "9-bit encoding");
435   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
436   return addOperand(Inst,
437                     DAsm->decodeSrcOp(AMDGPUDisassembler::OPW64, Imm, false, 64,
438                                       AMDGPU::OperandSemantics::FP64));
439 }
440 
441 #define DECODE_SDWA(DecName) \
442 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName)
443 
444 DECODE_SDWA(Src32)
445 DECODE_SDWA(Src16)
446 DECODE_SDWA(VopcDst)
447 
448 static DecodeStatus decodeVersionImm(MCInst &Inst, unsigned Imm,
449                                      uint64_t /* Addr */,
450                                      const MCDisassembler *Decoder) {
451   auto DAsm = static_cast<const AMDGPUDisassembler *>(Decoder);
452   return addOperand(Inst, DAsm->decodeVersionImm(Imm));
453 }
454 
455 #include "AMDGPUGenDisassemblerTables.inc"
456 
457 //===----------------------------------------------------------------------===//
458 //
459 //===----------------------------------------------------------------------===//
460 
461 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
462   assert(Bytes.size() >= sizeof(T));
463   const auto Res =
464       support::endian::read<T, llvm::endianness::little>(Bytes.data());
465   Bytes = Bytes.slice(sizeof(T));
466   return Res;
467 }
468 
469 static inline DecoderUInt128 eat12Bytes(ArrayRef<uint8_t> &Bytes) {
470   assert(Bytes.size() >= 12);
471   uint64_t Lo =
472       support::endian::read<uint64_t, llvm::endianness::little>(Bytes.data());
473   Bytes = Bytes.slice(8);
474   uint64_t Hi =
475       support::endian::read<uint32_t, llvm::endianness::little>(Bytes.data());
476   Bytes = Bytes.slice(4);
477   return DecoderUInt128(Lo, Hi);
478 }
479 
480 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
481                                                 ArrayRef<uint8_t> Bytes_,
482                                                 uint64_t Address,
483                                                 raw_ostream &CS) const {
484   unsigned MaxInstBytesNum = std::min((size_t)TargetMaxInstBytes, Bytes_.size());
485   Bytes = Bytes_.slice(0, MaxInstBytesNum);
486 
487   // In case the opcode is not recognized we'll assume a Size of 4 bytes (unless
488   // there are fewer bytes left). This will be overridden on success.
489   Size = std::min((size_t)4, Bytes_.size());
490 
491   do {
492     // ToDo: better to switch encoding length using some bit predicate
493     // but it is unknown yet, so try all we can
494 
495     // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2
496     // encodings
497     if (isGFX11Plus() && Bytes.size() >= 12 ) {
498       DecoderUInt128 DecW = eat12Bytes(Bytes);
499 
500       if (isGFX11() &&
501           tryDecodeInst(DecoderTableGFX1196, DecoderTableGFX11_FAKE1696, MI,
502                         DecW, Address, CS))
503         break;
504 
505       if (isGFX12() &&
506           tryDecodeInst(DecoderTableGFX1296, DecoderTableGFX12_FAKE1696, MI,
507                         DecW, Address, CS))
508         break;
509 
510       if (isGFX12() &&
511           tryDecodeInst(DecoderTableGFX12W6496, MI, DecW, Address, CS))
512         break;
513 
514       // Reinitialize Bytes
515       Bytes = Bytes_.slice(0, MaxInstBytesNum);
516     }
517 
518     if (Bytes.size() >= 8) {
519       const uint64_t QW = eatBytes<uint64_t>(Bytes);
520 
521       if (STI.hasFeature(AMDGPU::FeatureGFX10_BEncoding) &&
522           tryDecodeInst(DecoderTableGFX10_B64, MI, QW, Address, CS))
523         break;
524 
525       if (STI.hasFeature(AMDGPU::FeatureUnpackedD16VMem) &&
526           tryDecodeInst(DecoderTableGFX80_UNPACKED64, MI, QW, Address, CS))
527         break;
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.hasFeature(AMDGPU::FeatureFmaMixInsts) &&
533           tryDecodeInst(DecoderTableGFX9_DL64, MI, QW, Address, CS))
534         break;
535 
536       if (STI.hasFeature(AMDGPU::FeatureGFX940Insts) &&
537           tryDecodeInst(DecoderTableGFX94064, MI, QW, Address, CS))
538         break;
539 
540       if (STI.hasFeature(AMDGPU::FeatureGFX90AInsts) &&
541           tryDecodeInst(DecoderTableGFX90A64, MI, QW, Address, CS))
542         break;
543 
544       if ((isVI() || isGFX9()) &&
545           tryDecodeInst(DecoderTableGFX864, MI, QW, Address, CS))
546         break;
547 
548       if (isGFX9() && tryDecodeInst(DecoderTableGFX964, MI, QW, Address, CS))
549         break;
550 
551       if (isGFX10() && tryDecodeInst(DecoderTableGFX1064, MI, QW, Address, CS))
552         break;
553 
554       if (isGFX12() &&
555           tryDecodeInst(DecoderTableGFX1264, DecoderTableGFX12_FAKE1664, MI, QW,
556                         Address, CS))
557         break;
558 
559       if (isGFX11() &&
560           tryDecodeInst(DecoderTableGFX1164, DecoderTableGFX11_FAKE1664, MI, QW,
561                         Address, CS))
562         break;
563 
564       if (isGFX11() &&
565           tryDecodeInst(DecoderTableGFX11W6464, MI, QW, Address, CS))
566         break;
567 
568       if (isGFX12() &&
569           tryDecodeInst(DecoderTableGFX12W6464, MI, QW, Address, CS))
570         break;
571 
572       // Reinitialize Bytes
573       Bytes = Bytes_.slice(0, MaxInstBytesNum);
574     }
575 
576     // Try decode 32-bit instruction
577     if (Bytes.size() >= 4) {
578       const uint32_t DW = eatBytes<uint32_t>(Bytes);
579 
580       if ((isVI() || isGFX9()) &&
581           tryDecodeInst(DecoderTableGFX832, MI, DW, Address, CS))
582         break;
583 
584       if (tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address, CS))
585         break;
586 
587       if (isGFX9() && tryDecodeInst(DecoderTableGFX932, MI, DW, Address, CS))
588         break;
589 
590       if (STI.hasFeature(AMDGPU::FeatureGFX90AInsts) &&
591           tryDecodeInst(DecoderTableGFX90A32, MI, DW, Address, CS))
592         break;
593 
594       if (STI.hasFeature(AMDGPU::FeatureGFX10_BEncoding) &&
595           tryDecodeInst(DecoderTableGFX10_B32, MI, DW, Address, CS))
596         break;
597 
598       if (isGFX10() && tryDecodeInst(DecoderTableGFX1032, MI, DW, Address, CS))
599         break;
600 
601       if (isGFX11() &&
602           tryDecodeInst(DecoderTableGFX1132, DecoderTableGFX11_FAKE1632, MI, DW,
603                         Address, CS))
604         break;
605 
606       if (isGFX12() &&
607           tryDecodeInst(DecoderTableGFX1232, DecoderTableGFX12_FAKE1632, MI, DW,
608                         Address, CS))
609         break;
610     }
611 
612     return MCDisassembler::Fail;
613   } while (false);
614 
615   if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::DPP) {
616     if (isMacDPP(MI))
617       convertMacDPPInst(MI);
618 
619     if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3P)
620       convertVOP3PDPPInst(MI);
621     else if ((MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOPC) ||
622              AMDGPU::isVOPC64DPP(MI.getOpcode()))
623       convertVOPCDPPInst(MI); // Special VOP3 case
624     else if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dpp8) !=
625              -1)
626       convertDPP8Inst(MI);
627     else if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VOP3)
628       convertVOP3DPPInst(MI); // Regular VOP3 case
629   }
630 
631   if (AMDGPU::isMAC(MI.getOpcode())) {
632     // Insert dummy unused src2_modifiers.
633     insertNamedMCOperand(MI, MCOperand::createImm(0),
634                          AMDGPU::OpName::src2_modifiers);
635   }
636 
637   if (MI.getOpcode() == AMDGPU::V_CVT_SR_BF8_F32_e64_dpp ||
638       MI.getOpcode() == AMDGPU::V_CVT_SR_FP8_F32_e64_dpp) {
639     // Insert dummy unused src2_modifiers.
640     insertNamedMCOperand(MI, MCOperand::createImm(0),
641                          AMDGPU::OpName::src2_modifiers);
642   }
643 
644   if ((MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::DS) &&
645       !AMDGPU::hasGDS(STI)) {
646     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::gds);
647   }
648 
649   if (MCII->get(MI.getOpcode()).TSFlags &
650       (SIInstrFlags::MUBUF | SIInstrFlags::FLAT | SIInstrFlags::SMRD)) {
651     int CPolPos = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
652                                              AMDGPU::OpName::cpol);
653     if (CPolPos != -1) {
654       unsigned CPol =
655           (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::IsAtomicRet) ?
656               AMDGPU::CPol::GLC : 0;
657       if (MI.getNumOperands() <= (unsigned)CPolPos) {
658         insertNamedMCOperand(MI, MCOperand::createImm(CPol),
659                              AMDGPU::OpName::cpol);
660       } else if (CPol) {
661         MI.getOperand(CPolPos).setImm(MI.getOperand(CPolPos).getImm() | CPol);
662       }
663     }
664   }
665 
666   if ((MCII->get(MI.getOpcode()).TSFlags &
667        (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF)) &&
668       (STI.hasFeature(AMDGPU::FeatureGFX90AInsts))) {
669     // GFX90A lost TFE, its place is occupied by ACC.
670     int TFEOpIdx =
671         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::tfe);
672     if (TFEOpIdx != -1) {
673       auto TFEIter = MI.begin();
674       std::advance(TFEIter, TFEOpIdx);
675       MI.insert(TFEIter, MCOperand::createImm(0));
676     }
677   }
678 
679   if (MCII->get(MI.getOpcode()).TSFlags &
680       (SIInstrFlags::MTBUF | SIInstrFlags::MUBUF)) {
681     int SWZOpIdx =
682         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::swz);
683     if (SWZOpIdx != -1) {
684       auto SWZIter = MI.begin();
685       std::advance(SWZIter, SWZOpIdx);
686       MI.insert(SWZIter, MCOperand::createImm(0));
687     }
688   }
689 
690   if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::MIMG) {
691     int VAddr0Idx =
692         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
693     int RsrcIdx =
694         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
695     unsigned NSAArgs = RsrcIdx - VAddr0Idx - 1;
696     if (VAddr0Idx >= 0 && NSAArgs > 0) {
697       unsigned NSAWords = (NSAArgs + 3) / 4;
698       if (Bytes.size() < 4 * NSAWords)
699         return MCDisassembler::Fail;
700       for (unsigned i = 0; i < NSAArgs; ++i) {
701         const unsigned VAddrIdx = VAddr0Idx + 1 + i;
702         auto VAddrRCID =
703             MCII->get(MI.getOpcode()).operands()[VAddrIdx].RegClass;
704         MI.insert(MI.begin() + VAddrIdx, createRegOperand(VAddrRCID, Bytes[i]));
705       }
706       Bytes = Bytes.slice(4 * NSAWords);
707     }
708 
709     convertMIMGInst(MI);
710   }
711 
712   if (MCII->get(MI.getOpcode()).TSFlags &
713       (SIInstrFlags::VIMAGE | SIInstrFlags::VSAMPLE))
714     convertMIMGInst(MI);
715 
716   if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::EXP)
717     convertEXPInst(MI);
718 
719   if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::VINTERP)
720     convertVINTERPInst(MI);
721 
722   if (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::SDWA)
723     convertSDWAInst(MI);
724 
725   int VDstIn_Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
726                                               AMDGPU::OpName::vdst_in);
727   if (VDstIn_Idx != -1) {
728     int Tied = MCII->get(MI.getOpcode()).getOperandConstraint(VDstIn_Idx,
729                            MCOI::OperandConstraint::TIED_TO);
730     if (Tied != -1 && (MI.getNumOperands() <= (unsigned)VDstIn_Idx ||
731          !MI.getOperand(VDstIn_Idx).isReg() ||
732          MI.getOperand(VDstIn_Idx).getReg() != MI.getOperand(Tied).getReg())) {
733       if (MI.getNumOperands() > (unsigned)VDstIn_Idx)
734         MI.erase(&MI.getOperand(VDstIn_Idx));
735       insertNamedMCOperand(MI,
736         MCOperand::createReg(MI.getOperand(Tied).getReg()),
737         AMDGPU::OpName::vdst_in);
738     }
739   }
740 
741   int ImmLitIdx =
742       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::imm);
743   bool IsSOPK = MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::SOPK;
744   if (ImmLitIdx != -1 && !IsSOPK)
745     convertFMAanyK(MI, ImmLitIdx);
746 
747   Size = MaxInstBytesNum - Bytes.size();
748   return MCDisassembler::Success;
749 }
750 
751 void AMDGPUDisassembler::convertEXPInst(MCInst &MI) const {
752   if (STI.hasFeature(AMDGPU::FeatureGFX11Insts)) {
753     // The MCInst still has these fields even though they are no longer encoded
754     // in the GFX11 instruction.
755     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vm);
756     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::compr);
757   }
758 }
759 
760 void AMDGPUDisassembler::convertVINTERPInst(MCInst &MI) const {
761   if (MI.getOpcode() == AMDGPU::V_INTERP_P10_F16_F32_inreg_gfx11 ||
762       MI.getOpcode() == AMDGPU::V_INTERP_P10_F16_F32_inreg_gfx12 ||
763       MI.getOpcode() == AMDGPU::V_INTERP_P10_RTZ_F16_F32_inreg_gfx11 ||
764       MI.getOpcode() == AMDGPU::V_INTERP_P10_RTZ_F16_F32_inreg_gfx12 ||
765       MI.getOpcode() == AMDGPU::V_INTERP_P2_F16_F32_inreg_gfx11 ||
766       MI.getOpcode() == AMDGPU::V_INTERP_P2_F16_F32_inreg_gfx12 ||
767       MI.getOpcode() == AMDGPU::V_INTERP_P2_RTZ_F16_F32_inreg_gfx11 ||
768       MI.getOpcode() == AMDGPU::V_INTERP_P2_RTZ_F16_F32_inreg_gfx12) {
769     // The MCInst has this field that is not directly encoded in the
770     // instruction.
771     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::op_sel);
772   }
773 }
774 
775 void AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const {
776   if (STI.hasFeature(AMDGPU::FeatureGFX9) ||
777       STI.hasFeature(AMDGPU::FeatureGFX10)) {
778     if (AMDGPU::hasNamedOperand(MI.getOpcode(), AMDGPU::OpName::sdst))
779       // VOPC - insert clamp
780       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp);
781   } else if (STI.hasFeature(AMDGPU::FeatureVolcanicIslands)) {
782     int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst);
783     if (SDst != -1) {
784       // VOPC - insert VCC register as sdst
785       insertNamedMCOperand(MI, createRegOperand(AMDGPU::VCC),
786                            AMDGPU::OpName::sdst);
787     } else {
788       // VOP1/2 - insert omod if present in instruction
789       insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod);
790     }
791   }
792 }
793 
794 struct VOPModifiers {
795   unsigned OpSel = 0;
796   unsigned OpSelHi = 0;
797   unsigned NegLo = 0;
798   unsigned NegHi = 0;
799 };
800 
801 // Reconstruct values of VOP3/VOP3P operands such as op_sel.
802 // Note that these values do not affect disassembler output,
803 // so this is only necessary for consistency with src_modifiers.
804 static VOPModifiers collectVOPModifiers(const MCInst &MI,
805                                         bool IsVOP3P = false) {
806   VOPModifiers Modifiers;
807   unsigned Opc = MI.getOpcode();
808   const int ModOps[] = {AMDGPU::OpName::src0_modifiers,
809                         AMDGPU::OpName::src1_modifiers,
810                         AMDGPU::OpName::src2_modifiers};
811   for (int J = 0; J < 3; ++J) {
812     int OpIdx = AMDGPU::getNamedOperandIdx(Opc, ModOps[J]);
813     if (OpIdx == -1)
814       continue;
815 
816     unsigned Val = MI.getOperand(OpIdx).getImm();
817 
818     Modifiers.OpSel |= !!(Val & SISrcMods::OP_SEL_0) << J;
819     if (IsVOP3P) {
820       Modifiers.OpSelHi |= !!(Val & SISrcMods::OP_SEL_1) << J;
821       Modifiers.NegLo |= !!(Val & SISrcMods::NEG) << J;
822       Modifiers.NegHi |= !!(Val & SISrcMods::NEG_HI) << J;
823     } else if (J == 0) {
824       Modifiers.OpSel |= !!(Val & SISrcMods::DST_OP_SEL) << 3;
825     }
826   }
827 
828   return Modifiers;
829 }
830 
831 // Instructions decode the op_sel/suffix bits into the src_modifier
832 // operands. Copy those bits into the src operands for true16 VGPRs.
833 void AMDGPUDisassembler::convertTrue16OpSel(MCInst &MI) const {
834   const unsigned Opc = MI.getOpcode();
835   const MCRegisterClass &ConversionRC =
836       MRI.getRegClass(AMDGPU::VGPR_16RegClassID);
837   constexpr std::array<std::tuple<int, int, unsigned>, 4> OpAndOpMods = {
838       {{AMDGPU::OpName::src0, AMDGPU::OpName::src0_modifiers,
839         SISrcMods::OP_SEL_0},
840        {AMDGPU::OpName::src1, AMDGPU::OpName::src1_modifiers,
841         SISrcMods::OP_SEL_0},
842        {AMDGPU::OpName::src2, AMDGPU::OpName::src2_modifiers,
843         SISrcMods::OP_SEL_0},
844        {AMDGPU::OpName::vdst, AMDGPU::OpName::src0_modifiers,
845         SISrcMods::DST_OP_SEL}}};
846   for (const auto &[OpName, OpModsName, OpSelMask] : OpAndOpMods) {
847     int OpIdx = AMDGPU::getNamedOperandIdx(Opc, OpName);
848     int OpModsIdx = AMDGPU::getNamedOperandIdx(Opc, OpModsName);
849     if (OpIdx == -1 || OpModsIdx == -1)
850       continue;
851     MCOperand &Op = MI.getOperand(OpIdx);
852     if (!Op.isReg())
853       continue;
854     if (!ConversionRC.contains(Op.getReg()))
855       continue;
856     unsigned OpEnc = MRI.getEncodingValue(Op.getReg());
857     const MCOperand &OpMods = MI.getOperand(OpModsIdx);
858     unsigned ModVal = OpMods.getImm();
859     if (ModVal & OpSelMask) { // isHi
860       unsigned RegIdx = OpEnc & AMDGPU::HWEncoding::REG_IDX_MASK;
861       Op.setReg(ConversionRC.getRegister(RegIdx * 2 + 1));
862     }
863   }
864 }
865 
866 // MAC opcodes have special old and src2 operands.
867 // src2 is tied to dst, while old is not tied (but assumed to be).
868 bool AMDGPUDisassembler::isMacDPP(MCInst &MI) const {
869   constexpr int DST_IDX = 0;
870   auto Opcode = MI.getOpcode();
871   const auto &Desc = MCII->get(Opcode);
872   auto OldIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::old);
873 
874   if (OldIdx != -1 && Desc.getOperandConstraint(
875                           OldIdx, MCOI::OperandConstraint::TIED_TO) == -1) {
876     assert(AMDGPU::hasNamedOperand(Opcode, AMDGPU::OpName::src2));
877     assert(Desc.getOperandConstraint(
878                AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2),
879                MCOI::OperandConstraint::TIED_TO) == DST_IDX);
880     (void)DST_IDX;
881     return true;
882   }
883 
884   return false;
885 }
886 
887 // Create dummy old operand and insert dummy unused src2_modifiers
888 void AMDGPUDisassembler::convertMacDPPInst(MCInst &MI) const {
889   assert(MI.getNumOperands() + 1 < MCII->get(MI.getOpcode()).getNumOperands());
890   insertNamedMCOperand(MI, MCOperand::createReg(0), AMDGPU::OpName::old);
891   insertNamedMCOperand(MI, MCOperand::createImm(0),
892                        AMDGPU::OpName::src2_modifiers);
893 }
894 
895 void AMDGPUDisassembler::convertDPP8Inst(MCInst &MI) const {
896   unsigned Opc = MI.getOpcode();
897 
898   int VDstInIdx =
899       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdst_in);
900   if (VDstInIdx != -1)
901     insertNamedMCOperand(MI, MI.getOperand(0), AMDGPU::OpName::vdst_in);
902 
903   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
904   if (MI.getNumOperands() < DescNumOps &&
905       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel)) {
906     convertTrue16OpSel(MI);
907     auto Mods = collectVOPModifiers(MI);
908     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
909                          AMDGPU::OpName::op_sel);
910   } else {
911     // Insert dummy unused src modifiers.
912     if (MI.getNumOperands() < DescNumOps &&
913         AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src0_modifiers))
914       insertNamedMCOperand(MI, MCOperand::createImm(0),
915                            AMDGPU::OpName::src0_modifiers);
916 
917     if (MI.getNumOperands() < DescNumOps &&
918         AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src1_modifiers))
919       insertNamedMCOperand(MI, MCOperand::createImm(0),
920                            AMDGPU::OpName::src1_modifiers);
921   }
922 }
923 
924 void AMDGPUDisassembler::convertVOP3DPPInst(MCInst &MI) const {
925   convertTrue16OpSel(MI);
926 
927   int VDstInIdx =
928       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdst_in);
929   if (VDstInIdx != -1)
930     insertNamedMCOperand(MI, MI.getOperand(0), AMDGPU::OpName::vdst_in);
931 
932   unsigned Opc = MI.getOpcode();
933   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
934   if (MI.getNumOperands() < DescNumOps &&
935       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel)) {
936     auto Mods = collectVOPModifiers(MI);
937     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
938                          AMDGPU::OpName::op_sel);
939   }
940 }
941 
942 // Note that before gfx10, the MIMG encoding provided no information about
943 // VADDR size. Consequently, decoded instructions always show address as if it
944 // has 1 dword, which could be not really so.
945 void AMDGPUDisassembler::convertMIMGInst(MCInst &MI) const {
946   auto TSFlags = MCII->get(MI.getOpcode()).TSFlags;
947 
948   int VDstIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
949                                            AMDGPU::OpName::vdst);
950 
951   int VDataIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
952                                             AMDGPU::OpName::vdata);
953   int VAddr0Idx =
954       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vaddr0);
955   int RsrcOpName = (TSFlags & SIInstrFlags::MIMG) ? AMDGPU::OpName::srsrc
956                                                   : AMDGPU::OpName::rsrc;
957   int RsrcIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), RsrcOpName);
958   int DMaskIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
959                                             AMDGPU::OpName::dmask);
960 
961   int TFEIdx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
962                                             AMDGPU::OpName::tfe);
963   int D16Idx   = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
964                                             AMDGPU::OpName::d16);
965 
966   const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode());
967   const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
968       AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
969 
970   assert(VDataIdx != -1);
971   if (BaseOpcode->BVH) {
972     // Add A16 operand for intersect_ray instructions
973     addOperand(MI, MCOperand::createImm(BaseOpcode->A16));
974     return;
975   }
976 
977   bool IsAtomic = (VDstIdx != -1);
978   bool IsGather4 = TSFlags & SIInstrFlags::Gather4;
979   bool IsVSample = TSFlags & SIInstrFlags::VSAMPLE;
980   bool IsNSA = false;
981   bool IsPartialNSA = false;
982   unsigned AddrSize = Info->VAddrDwords;
983 
984   if (isGFX10Plus()) {
985     unsigned DimIdx =
986         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dim);
987     int A16Idx =
988         AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::a16);
989     const AMDGPU::MIMGDimInfo *Dim =
990         AMDGPU::getMIMGDimInfoByEncoding(MI.getOperand(DimIdx).getImm());
991     const bool IsA16 = (A16Idx != -1 && MI.getOperand(A16Idx).getImm());
992 
993     AddrSize =
994         AMDGPU::getAddrSizeMIMGOp(BaseOpcode, Dim, IsA16, AMDGPU::hasG16(STI));
995 
996     // VSAMPLE insts that do not use vaddr3 behave the same as NSA forms.
997     // VIMAGE insts other than BVH never use vaddr4.
998     IsNSA = Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA ||
999             Info->MIMGEncoding == AMDGPU::MIMGEncGfx11NSA ||
1000             Info->MIMGEncoding == AMDGPU::MIMGEncGfx12;
1001     if (!IsNSA) {
1002       if (!IsVSample && AddrSize > 12)
1003         AddrSize = 16;
1004     } else {
1005       if (AddrSize > Info->VAddrDwords) {
1006         if (!STI.hasFeature(AMDGPU::FeaturePartialNSAEncoding)) {
1007           // The NSA encoding does not contain enough operands for the
1008           // combination of base opcode / dimension. Should this be an error?
1009           return;
1010         }
1011         IsPartialNSA = true;
1012       }
1013     }
1014   }
1015 
1016   unsigned DMask = MI.getOperand(DMaskIdx).getImm() & 0xf;
1017   unsigned DstSize = IsGather4 ? 4 : std::max(llvm::popcount(DMask), 1);
1018 
1019   bool D16 = D16Idx >= 0 && MI.getOperand(D16Idx).getImm();
1020   if (D16 && AMDGPU::hasPackedD16(STI)) {
1021     DstSize = (DstSize + 1) / 2;
1022   }
1023 
1024   if (TFEIdx != -1 && MI.getOperand(TFEIdx).getImm())
1025     DstSize += 1;
1026 
1027   if (DstSize == Info->VDataDwords && AddrSize == Info->VAddrDwords)
1028     return;
1029 
1030   int NewOpcode =
1031       AMDGPU::getMIMGOpcode(Info->BaseOpcode, Info->MIMGEncoding, DstSize, AddrSize);
1032   if (NewOpcode == -1)
1033     return;
1034 
1035   // Widen the register to the correct number of enabled channels.
1036   unsigned NewVdata = AMDGPU::NoRegister;
1037   if (DstSize != Info->VDataDwords) {
1038     auto DataRCID = MCII->get(NewOpcode).operands()[VDataIdx].RegClass;
1039 
1040     // Get first subregister of VData
1041     unsigned Vdata0 = MI.getOperand(VDataIdx).getReg();
1042     unsigned VdataSub0 = MRI.getSubReg(Vdata0, AMDGPU::sub0);
1043     Vdata0 = (VdataSub0 != 0)? VdataSub0 : Vdata0;
1044 
1045     NewVdata = MRI.getMatchingSuperReg(Vdata0, AMDGPU::sub0,
1046                                        &MRI.getRegClass(DataRCID));
1047     if (NewVdata == AMDGPU::NoRegister) {
1048       // It's possible to encode this such that the low register + enabled
1049       // components exceeds the register count.
1050       return;
1051     }
1052   }
1053 
1054   // If not using NSA on GFX10+, widen vaddr0 address register to correct size.
1055   // If using partial NSA on GFX11+ widen last address register.
1056   int VAddrSAIdx = IsPartialNSA ? (RsrcIdx - 1) : VAddr0Idx;
1057   unsigned NewVAddrSA = AMDGPU::NoRegister;
1058   if (STI.hasFeature(AMDGPU::FeatureNSAEncoding) && (!IsNSA || IsPartialNSA) &&
1059       AddrSize != Info->VAddrDwords) {
1060     unsigned VAddrSA = MI.getOperand(VAddrSAIdx).getReg();
1061     unsigned VAddrSubSA = MRI.getSubReg(VAddrSA, AMDGPU::sub0);
1062     VAddrSA = VAddrSubSA ? VAddrSubSA : VAddrSA;
1063 
1064     auto AddrRCID = MCII->get(NewOpcode).operands()[VAddrSAIdx].RegClass;
1065     NewVAddrSA = MRI.getMatchingSuperReg(VAddrSA, AMDGPU::sub0,
1066                                         &MRI.getRegClass(AddrRCID));
1067     if (!NewVAddrSA)
1068       return;
1069   }
1070 
1071   MI.setOpcode(NewOpcode);
1072 
1073   if (NewVdata != AMDGPU::NoRegister) {
1074     MI.getOperand(VDataIdx) = MCOperand::createReg(NewVdata);
1075 
1076     if (IsAtomic) {
1077       // Atomic operations have an additional operand (a copy of data)
1078       MI.getOperand(VDstIdx) = MCOperand::createReg(NewVdata);
1079     }
1080   }
1081 
1082   if (NewVAddrSA) {
1083     MI.getOperand(VAddrSAIdx) = MCOperand::createReg(NewVAddrSA);
1084   } else if (IsNSA) {
1085     assert(AddrSize <= Info->VAddrDwords);
1086     MI.erase(MI.begin() + VAddr0Idx + AddrSize,
1087              MI.begin() + VAddr0Idx + Info->VAddrDwords);
1088   }
1089 }
1090 
1091 // Opsel and neg bits are used in src_modifiers and standalone operands. Autogen
1092 // decoder only adds to src_modifiers, so manually add the bits to the other
1093 // operands.
1094 void AMDGPUDisassembler::convertVOP3PDPPInst(MCInst &MI) const {
1095   unsigned Opc = MI.getOpcode();
1096   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
1097   auto Mods = collectVOPModifiers(MI, true);
1098 
1099   if (MI.getNumOperands() < DescNumOps &&
1100       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::vdst_in))
1101     insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::vdst_in);
1102 
1103   if (MI.getNumOperands() < DescNumOps &&
1104       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel))
1105     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSel),
1106                          AMDGPU::OpName::op_sel);
1107   if (MI.getNumOperands() < DescNumOps &&
1108       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::op_sel_hi))
1109     insertNamedMCOperand(MI, MCOperand::createImm(Mods.OpSelHi),
1110                          AMDGPU::OpName::op_sel_hi);
1111   if (MI.getNumOperands() < DescNumOps &&
1112       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::neg_lo))
1113     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegLo),
1114                          AMDGPU::OpName::neg_lo);
1115   if (MI.getNumOperands() < DescNumOps &&
1116       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::neg_hi))
1117     insertNamedMCOperand(MI, MCOperand::createImm(Mods.NegHi),
1118                          AMDGPU::OpName::neg_hi);
1119 }
1120 
1121 // Create dummy old operand and insert optional operands
1122 void AMDGPUDisassembler::convertVOPCDPPInst(MCInst &MI) const {
1123   unsigned Opc = MI.getOpcode();
1124   unsigned DescNumOps = MCII->get(Opc).getNumOperands();
1125 
1126   if (MI.getNumOperands() < DescNumOps &&
1127       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::old))
1128     insertNamedMCOperand(MI, MCOperand::createReg(0), AMDGPU::OpName::old);
1129 
1130   if (MI.getNumOperands() < DescNumOps &&
1131       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src0_modifiers))
1132     insertNamedMCOperand(MI, MCOperand::createImm(0),
1133                          AMDGPU::OpName::src0_modifiers);
1134 
1135   if (MI.getNumOperands() < DescNumOps &&
1136       AMDGPU::hasNamedOperand(Opc, AMDGPU::OpName::src1_modifiers))
1137     insertNamedMCOperand(MI, MCOperand::createImm(0),
1138                          AMDGPU::OpName::src1_modifiers);
1139 }
1140 
1141 void AMDGPUDisassembler::convertFMAanyK(MCInst &MI, int ImmLitIdx) const {
1142   assert(HasLiteral && "Should have decoded a literal");
1143   const MCInstrDesc &Desc = MCII->get(MI.getOpcode());
1144   unsigned DescNumOps = Desc.getNumOperands();
1145   insertNamedMCOperand(MI, MCOperand::createImm(Literal),
1146                        AMDGPU::OpName::immDeferred);
1147   assert(DescNumOps == MI.getNumOperands());
1148   for (unsigned I = 0; I < DescNumOps; ++I) {
1149     auto &Op = MI.getOperand(I);
1150     auto OpType = Desc.operands()[I].OperandType;
1151     bool IsDeferredOp = (OpType == AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED ||
1152                          OpType == AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED);
1153     if (Op.isImm() && Op.getImm() == AMDGPU::EncValues::LITERAL_CONST &&
1154         IsDeferredOp)
1155       Op.setImm(Literal);
1156   }
1157 }
1158 
1159 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
1160   return getContext().getRegisterInfo()->
1161     getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
1162 }
1163 
1164 inline
1165 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
1166                                          const Twine& ErrMsg) const {
1167   *CommentStream << "Error: " + ErrMsg;
1168 
1169   // ToDo: add support for error operands to MCInst.h
1170   // return MCOperand::createError(V);
1171   return MCOperand();
1172 }
1173 
1174 inline
1175 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
1176   return MCOperand::createReg(AMDGPU::getMCReg(RegId, STI));
1177 }
1178 
1179 inline
1180 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
1181                                                unsigned Val) const {
1182   const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
1183   if (Val >= RegCl.getNumRegs())
1184     return errOperand(Val, Twine(getRegClassName(RegClassID)) +
1185                            ": unknown register " + Twine(Val));
1186   return createRegOperand(RegCl.getRegister(Val));
1187 }
1188 
1189 inline
1190 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
1191                                                 unsigned Val) const {
1192   // ToDo: SI/CI have 104 SGPRs, VI - 102
1193   // Valery: here we accepting as much as we can, let assembler sort it out
1194   int shift = 0;
1195   switch (SRegClassID) {
1196   case AMDGPU::SGPR_32RegClassID:
1197   case AMDGPU::TTMP_32RegClassID:
1198     break;
1199   case AMDGPU::SGPR_64RegClassID:
1200   case AMDGPU::TTMP_64RegClassID:
1201     shift = 1;
1202     break;
1203   case AMDGPU::SGPR_96RegClassID:
1204   case AMDGPU::TTMP_96RegClassID:
1205   case AMDGPU::SGPR_128RegClassID:
1206   case AMDGPU::TTMP_128RegClassID:
1207   // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
1208   // this bundle?
1209   case AMDGPU::SGPR_256RegClassID:
1210   case AMDGPU::TTMP_256RegClassID:
1211     // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
1212   // this bundle?
1213   case AMDGPU::SGPR_288RegClassID:
1214   case AMDGPU::TTMP_288RegClassID:
1215   case AMDGPU::SGPR_320RegClassID:
1216   case AMDGPU::TTMP_320RegClassID:
1217   case AMDGPU::SGPR_352RegClassID:
1218   case AMDGPU::TTMP_352RegClassID:
1219   case AMDGPU::SGPR_384RegClassID:
1220   case AMDGPU::TTMP_384RegClassID:
1221   case AMDGPU::SGPR_512RegClassID:
1222   case AMDGPU::TTMP_512RegClassID:
1223     shift = 2;
1224     break;
1225   // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
1226   // this bundle?
1227   default:
1228     llvm_unreachable("unhandled register class");
1229   }
1230 
1231   if (Val % (1 << shift)) {
1232     *CommentStream << "Warning: " << getRegClassName(SRegClassID)
1233                    << ": scalar reg isn't aligned " << Val;
1234   }
1235 
1236   return createRegOperand(SRegClassID, Val >> shift);
1237 }
1238 
1239 MCOperand AMDGPUDisassembler::createVGPR16Operand(unsigned RegIdx,
1240                                                   bool IsHi) const {
1241   unsigned RegIdxInVGPR16 = RegIdx * 2 + (IsHi ? 1 : 0);
1242   return createRegOperand(AMDGPU::VGPR_16RegClassID, RegIdxInVGPR16);
1243 }
1244 
1245 // Decode Literals for insts which always have a literal in the encoding
1246 MCOperand
1247 AMDGPUDisassembler::decodeMandatoryLiteralConstant(unsigned Val) const {
1248   if (HasLiteral) {
1249     assert(
1250         AMDGPU::hasVOPD(STI) &&
1251         "Should only decode multiple kimm with VOPD, check VSrc operand types");
1252     if (Literal != Val)
1253       return errOperand(Val, "More than one unique literal is illegal");
1254   }
1255   HasLiteral = true;
1256   Literal = Val;
1257   return MCOperand::createImm(Literal);
1258 }
1259 
1260 MCOperand AMDGPUDisassembler::decodeLiteralConstant(bool ExtendFP64) const {
1261   // For now all literal constants are supposed to be unsigned integer
1262   // ToDo: deal with signed/unsigned 64-bit integer constants
1263   // ToDo: deal with float/double constants
1264   if (!HasLiteral) {
1265     if (Bytes.size() < 4) {
1266       return errOperand(0, "cannot read literal, inst bytes left " +
1267                         Twine(Bytes.size()));
1268     }
1269     HasLiteral = true;
1270     Literal = Literal64 = eatBytes<uint32_t>(Bytes);
1271     if (ExtendFP64)
1272       Literal64 <<= 32;
1273   }
1274   return MCOperand::createImm(ExtendFP64 ? Literal64 : Literal);
1275 }
1276 
1277 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
1278   using namespace AMDGPU::EncValues;
1279 
1280   assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX);
1281   return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ?
1282     (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) :
1283     (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm)));
1284       // Cast prevents negative overflow.
1285 }
1286 
1287 static int64_t getInlineImmVal32(unsigned Imm) {
1288   switch (Imm) {
1289   case 240:
1290     return llvm::bit_cast<uint32_t>(0.5f);
1291   case 241:
1292     return llvm::bit_cast<uint32_t>(-0.5f);
1293   case 242:
1294     return llvm::bit_cast<uint32_t>(1.0f);
1295   case 243:
1296     return llvm::bit_cast<uint32_t>(-1.0f);
1297   case 244:
1298     return llvm::bit_cast<uint32_t>(2.0f);
1299   case 245:
1300     return llvm::bit_cast<uint32_t>(-2.0f);
1301   case 246:
1302     return llvm::bit_cast<uint32_t>(4.0f);
1303   case 247:
1304     return llvm::bit_cast<uint32_t>(-4.0f);
1305   case 248: // 1 / (2 * PI)
1306     return 0x3e22f983;
1307   default:
1308     llvm_unreachable("invalid fp inline imm");
1309   }
1310 }
1311 
1312 static int64_t getInlineImmVal64(unsigned Imm) {
1313   switch (Imm) {
1314   case 240:
1315     return llvm::bit_cast<uint64_t>(0.5);
1316   case 241:
1317     return llvm::bit_cast<uint64_t>(-0.5);
1318   case 242:
1319     return llvm::bit_cast<uint64_t>(1.0);
1320   case 243:
1321     return llvm::bit_cast<uint64_t>(-1.0);
1322   case 244:
1323     return llvm::bit_cast<uint64_t>(2.0);
1324   case 245:
1325     return llvm::bit_cast<uint64_t>(-2.0);
1326   case 246:
1327     return llvm::bit_cast<uint64_t>(4.0);
1328   case 247:
1329     return llvm::bit_cast<uint64_t>(-4.0);
1330   case 248: // 1 / (2 * PI)
1331     return 0x3fc45f306dc9c882;
1332   default:
1333     llvm_unreachable("invalid fp inline imm");
1334   }
1335 }
1336 
1337 static int64_t getInlineImmValF16(unsigned Imm) {
1338   switch (Imm) {
1339   case 240:
1340     return 0x3800;
1341   case 241:
1342     return 0xB800;
1343   case 242:
1344     return 0x3C00;
1345   case 243:
1346     return 0xBC00;
1347   case 244:
1348     return 0x4000;
1349   case 245:
1350     return 0xC000;
1351   case 246:
1352     return 0x4400;
1353   case 247:
1354     return 0xC400;
1355   case 248: // 1 / (2 * PI)
1356     return 0x3118;
1357   default:
1358     llvm_unreachable("invalid fp inline imm");
1359   }
1360 }
1361 
1362 static int64_t getInlineImmValBF16(unsigned Imm) {
1363   switch (Imm) {
1364   case 240:
1365     return 0x3F00;
1366   case 241:
1367     return 0xBF00;
1368   case 242:
1369     return 0x3F80;
1370   case 243:
1371     return 0xBF80;
1372   case 244:
1373     return 0x4000;
1374   case 245:
1375     return 0xC000;
1376   case 246:
1377     return 0x4080;
1378   case 247:
1379     return 0xC080;
1380   case 248: // 1 / (2 * PI)
1381     return 0x3E22;
1382   default:
1383     llvm_unreachable("invalid fp inline imm");
1384   }
1385 }
1386 
1387 static int64_t getInlineImmVal16(unsigned Imm, AMDGPU::OperandSemantics Sema) {
1388   return (Sema == AMDGPU::OperandSemantics::BF16) ? getInlineImmValBF16(Imm)
1389                                                   : getInlineImmValF16(Imm);
1390 }
1391 
1392 MCOperand AMDGPUDisassembler::decodeFPImmed(unsigned ImmWidth, unsigned Imm,
1393                                             AMDGPU::OperandSemantics Sema) {
1394   assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN &&
1395          Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX);
1396 
1397   // ToDo: case 248: 1/(2*PI) - is allowed only on VI
1398   // ImmWidth 0 is a default case where operand should not allow immediates.
1399   // Imm value is still decoded into 32 bit immediate operand, inst printer will
1400   // use it to print verbose error message.
1401   switch (ImmWidth) {
1402   case 0:
1403   case 32:
1404     return MCOperand::createImm(getInlineImmVal32(Imm));
1405   case 64:
1406     return MCOperand::createImm(getInlineImmVal64(Imm));
1407   case 16:
1408     return MCOperand::createImm(getInlineImmVal16(Imm, Sema));
1409   default:
1410     llvm_unreachable("implement me");
1411   }
1412 }
1413 
1414 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const {
1415   using namespace AMDGPU;
1416 
1417   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1418   switch (Width) {
1419   default: // fall
1420   case OPW32:
1421   case OPW16:
1422   case OPWV216:
1423     return VGPR_32RegClassID;
1424   case OPW64:
1425   case OPWV232: return VReg_64RegClassID;
1426   case OPW96: return VReg_96RegClassID;
1427   case OPW128: return VReg_128RegClassID;
1428   case OPW160: return VReg_160RegClassID;
1429   case OPW256: return VReg_256RegClassID;
1430   case OPW288: return VReg_288RegClassID;
1431   case OPW320: return VReg_320RegClassID;
1432   case OPW352: return VReg_352RegClassID;
1433   case OPW384: return VReg_384RegClassID;
1434   case OPW512: return VReg_512RegClassID;
1435   case OPW1024: return VReg_1024RegClassID;
1436   }
1437 }
1438 
1439 unsigned AMDGPUDisassembler::getAgprClassId(const OpWidthTy Width) const {
1440   using namespace AMDGPU;
1441 
1442   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1443   switch (Width) {
1444   default: // fall
1445   case OPW32:
1446   case OPW16:
1447   case OPWV216:
1448     return AGPR_32RegClassID;
1449   case OPW64:
1450   case OPWV232: return AReg_64RegClassID;
1451   case OPW96: return AReg_96RegClassID;
1452   case OPW128: return AReg_128RegClassID;
1453   case OPW160: return AReg_160RegClassID;
1454   case OPW256: return AReg_256RegClassID;
1455   case OPW288: return AReg_288RegClassID;
1456   case OPW320: return AReg_320RegClassID;
1457   case OPW352: return AReg_352RegClassID;
1458   case OPW384: return AReg_384RegClassID;
1459   case OPW512: return AReg_512RegClassID;
1460   case OPW1024: return AReg_1024RegClassID;
1461   }
1462 }
1463 
1464 
1465 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const {
1466   using namespace AMDGPU;
1467 
1468   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1469   switch (Width) {
1470   default: // fall
1471   case OPW32:
1472   case OPW16:
1473   case OPWV216:
1474     return SGPR_32RegClassID;
1475   case OPW64:
1476   case OPWV232: return SGPR_64RegClassID;
1477   case OPW96: return SGPR_96RegClassID;
1478   case OPW128: return SGPR_128RegClassID;
1479   case OPW160: return SGPR_160RegClassID;
1480   case OPW256: return SGPR_256RegClassID;
1481   case OPW288: return SGPR_288RegClassID;
1482   case OPW320: return SGPR_320RegClassID;
1483   case OPW352: return SGPR_352RegClassID;
1484   case OPW384: return SGPR_384RegClassID;
1485   case OPW512: return SGPR_512RegClassID;
1486   }
1487 }
1488 
1489 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const {
1490   using namespace AMDGPU;
1491 
1492   assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
1493   switch (Width) {
1494   default: // fall
1495   case OPW32:
1496   case OPW16:
1497   case OPWV216:
1498     return TTMP_32RegClassID;
1499   case OPW64:
1500   case OPWV232: return TTMP_64RegClassID;
1501   case OPW128: return TTMP_128RegClassID;
1502   case OPW256: return TTMP_256RegClassID;
1503   case OPW288: return TTMP_288RegClassID;
1504   case OPW320: return TTMP_320RegClassID;
1505   case OPW352: return TTMP_352RegClassID;
1506   case OPW384: return TTMP_384RegClassID;
1507   case OPW512: return TTMP_512RegClassID;
1508   }
1509 }
1510 
1511 int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
1512   using namespace AMDGPU::EncValues;
1513 
1514   unsigned TTmpMin = isGFX9Plus() ? TTMP_GFX9PLUS_MIN : TTMP_VI_MIN;
1515   unsigned TTmpMax = isGFX9Plus() ? TTMP_GFX9PLUS_MAX : TTMP_VI_MAX;
1516 
1517   return (TTmpMin <= Val && Val <= TTmpMax)? Val - TTmpMin : -1;
1518 }
1519 
1520 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val,
1521                                           bool MandatoryLiteral,
1522                                           unsigned ImmWidth,
1523                                           AMDGPU::OperandSemantics Sema) const {
1524   using namespace AMDGPU::EncValues;
1525 
1526   assert(Val < 1024); // enum10
1527 
1528   bool IsAGPR = Val & 512;
1529   Val &= 511;
1530 
1531   if (VGPR_MIN <= Val && Val <= VGPR_MAX) {
1532     return createRegOperand(IsAGPR ? getAgprClassId(Width)
1533                                    : getVgprClassId(Width), Val - VGPR_MIN);
1534   }
1535   return decodeNonVGPRSrcOp(Width, Val & 0xFF, MandatoryLiteral, ImmWidth,
1536                             Sema);
1537 }
1538 
1539 MCOperand
1540 AMDGPUDisassembler::decodeNonVGPRSrcOp(const OpWidthTy Width, unsigned Val,
1541                                        bool MandatoryLiteral, unsigned ImmWidth,
1542                                        AMDGPU::OperandSemantics Sema) const {
1543   // Cases when Val{8} is 1 (vgpr, agpr or true 16 vgpr) should have been
1544   // decoded earlier.
1545   assert(Val < (1 << 8) && "9-bit Src encoding when Val{8} is 0");
1546   using namespace AMDGPU::EncValues;
1547 
1548   if (Val <= SGPR_MAX) {
1549     // "SGPR_MIN <= Val" is always true and causes compilation warning.
1550     static_assert(SGPR_MIN == 0);
1551     return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
1552   }
1553 
1554   int TTmpIdx = getTTmpIdx(Val);
1555   if (TTmpIdx >= 0) {
1556     return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
1557   }
1558 
1559   if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX)
1560     return decodeIntImmed(Val);
1561 
1562   if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX)
1563     return decodeFPImmed(ImmWidth, Val, Sema);
1564 
1565   if (Val == LITERAL_CONST) {
1566     if (MandatoryLiteral)
1567       // Keep a sentinel value for deferred setting
1568       return MCOperand::createImm(LITERAL_CONST);
1569     return decodeLiteralConstant(Sema == AMDGPU::OperandSemantics::FP64);
1570   }
1571 
1572   switch (Width) {
1573   case OPW32:
1574   case OPW16:
1575   case OPWV216:
1576     return decodeSpecialReg32(Val);
1577   case OPW64:
1578   case OPWV232:
1579     return decodeSpecialReg64(Val);
1580   default:
1581     llvm_unreachable("unexpected immediate type");
1582   }
1583 }
1584 
1585 // Bit 0 of DstY isn't stored in the instruction, because it's always the
1586 // opposite of bit 0 of DstX.
1587 MCOperand AMDGPUDisassembler::decodeVOPDDstYOp(MCInst &Inst,
1588                                                unsigned Val) const {
1589   int VDstXInd =
1590       AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::vdstX);
1591   assert(VDstXInd != -1);
1592   assert(Inst.getOperand(VDstXInd).isReg());
1593   unsigned XDstReg = MRI.getEncodingValue(Inst.getOperand(VDstXInd).getReg());
1594   Val |= ~XDstReg & 1;
1595   auto Width = llvm::AMDGPUDisassembler::OPW32;
1596   return createRegOperand(getVgprClassId(Width), Val);
1597 }
1598 
1599 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
1600   using namespace AMDGPU;
1601 
1602   switch (Val) {
1603   // clang-format off
1604   case 102: return createRegOperand(FLAT_SCR_LO);
1605   case 103: return createRegOperand(FLAT_SCR_HI);
1606   case 104: return createRegOperand(XNACK_MASK_LO);
1607   case 105: return createRegOperand(XNACK_MASK_HI);
1608   case 106: return createRegOperand(VCC_LO);
1609   case 107: return createRegOperand(VCC_HI);
1610   case 108: return createRegOperand(TBA_LO);
1611   case 109: return createRegOperand(TBA_HI);
1612   case 110: return createRegOperand(TMA_LO);
1613   case 111: return createRegOperand(TMA_HI);
1614   case 124:
1615     return isGFX11Plus() ? createRegOperand(SGPR_NULL) : createRegOperand(M0);
1616   case 125:
1617     return isGFX11Plus() ? createRegOperand(M0) : createRegOperand(SGPR_NULL);
1618   case 126: return createRegOperand(EXEC_LO);
1619   case 127: return createRegOperand(EXEC_HI);
1620   case 235: return createRegOperand(SRC_SHARED_BASE_LO);
1621   case 236: return createRegOperand(SRC_SHARED_LIMIT_LO);
1622   case 237: return createRegOperand(SRC_PRIVATE_BASE_LO);
1623   case 238: return createRegOperand(SRC_PRIVATE_LIMIT_LO);
1624   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1625   case 251: return createRegOperand(SRC_VCCZ);
1626   case 252: return createRegOperand(SRC_EXECZ);
1627   case 253: return createRegOperand(SRC_SCC);
1628   case 254: return createRegOperand(LDS_DIRECT);
1629   default: break;
1630     // clang-format on
1631   }
1632   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1633 }
1634 
1635 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
1636   using namespace AMDGPU;
1637 
1638   switch (Val) {
1639   case 102: return createRegOperand(FLAT_SCR);
1640   case 104: return createRegOperand(XNACK_MASK);
1641   case 106: return createRegOperand(VCC);
1642   case 108: return createRegOperand(TBA);
1643   case 110: return createRegOperand(TMA);
1644   case 124:
1645     if (isGFX11Plus())
1646       return createRegOperand(SGPR_NULL);
1647     break;
1648   case 125:
1649     if (!isGFX11Plus())
1650       return createRegOperand(SGPR_NULL);
1651     break;
1652   case 126: return createRegOperand(EXEC);
1653   case 235: return createRegOperand(SRC_SHARED_BASE);
1654   case 236: return createRegOperand(SRC_SHARED_LIMIT);
1655   case 237: return createRegOperand(SRC_PRIVATE_BASE);
1656   case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
1657   case 239: return createRegOperand(SRC_POPS_EXITING_WAVE_ID);
1658   case 251: return createRegOperand(SRC_VCCZ);
1659   case 252: return createRegOperand(SRC_EXECZ);
1660   case 253: return createRegOperand(SRC_SCC);
1661   default: break;
1662   }
1663   return errOperand(Val, "unknown operand encoding " + Twine(Val));
1664 }
1665 
1666 MCOperand
1667 AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width, const unsigned Val,
1668                                   unsigned ImmWidth,
1669                                   AMDGPU::OperandSemantics Sema) const {
1670   using namespace AMDGPU::SDWA;
1671   using namespace AMDGPU::EncValues;
1672 
1673   if (STI.hasFeature(AMDGPU::FeatureGFX9) ||
1674       STI.hasFeature(AMDGPU::FeatureGFX10)) {
1675     // XXX: cast to int is needed to avoid stupid warning:
1676     // compare with unsigned is always true
1677     if (int(SDWA9EncValues::SRC_VGPR_MIN) <= int(Val) &&
1678         Val <= SDWA9EncValues::SRC_VGPR_MAX) {
1679       return createRegOperand(getVgprClassId(Width),
1680                               Val - SDWA9EncValues::SRC_VGPR_MIN);
1681     }
1682     if (SDWA9EncValues::SRC_SGPR_MIN <= Val &&
1683         Val <= (isGFX10Plus() ? SDWA9EncValues::SRC_SGPR_MAX_GFX10
1684                               : SDWA9EncValues::SRC_SGPR_MAX_SI)) {
1685       return createSRegOperand(getSgprClassId(Width),
1686                                Val - SDWA9EncValues::SRC_SGPR_MIN);
1687     }
1688     if (SDWA9EncValues::SRC_TTMP_MIN <= Val &&
1689         Val <= SDWA9EncValues::SRC_TTMP_MAX) {
1690       return createSRegOperand(getTtmpClassId(Width),
1691                                Val - SDWA9EncValues::SRC_TTMP_MIN);
1692     }
1693 
1694     const unsigned SVal = Val - SDWA9EncValues::SRC_SGPR_MIN;
1695 
1696     if (INLINE_INTEGER_C_MIN <= SVal && SVal <= INLINE_INTEGER_C_MAX)
1697       return decodeIntImmed(SVal);
1698 
1699     if (INLINE_FLOATING_C_MIN <= SVal && SVal <= INLINE_FLOATING_C_MAX)
1700       return decodeFPImmed(ImmWidth, SVal, Sema);
1701 
1702     return decodeSpecialReg32(SVal);
1703   }
1704   if (STI.hasFeature(AMDGPU::FeatureVolcanicIslands))
1705     return createRegOperand(getVgprClassId(Width), Val);
1706   llvm_unreachable("unsupported target");
1707 }
1708 
1709 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const {
1710   return decodeSDWASrc(OPW16, Val, 16, AMDGPU::OperandSemantics::FP16);
1711 }
1712 
1713 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const {
1714   return decodeSDWASrc(OPW32, Val, 32, AMDGPU::OperandSemantics::FP32);
1715 }
1716 
1717 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const {
1718   using namespace AMDGPU::SDWA;
1719 
1720   assert((STI.hasFeature(AMDGPU::FeatureGFX9) ||
1721           STI.hasFeature(AMDGPU::FeatureGFX10)) &&
1722          "SDWAVopcDst should be present only on GFX9+");
1723 
1724   bool IsWave64 = STI.hasFeature(AMDGPU::FeatureWavefrontSize64);
1725 
1726   if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) {
1727     Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
1728 
1729     int TTmpIdx = getTTmpIdx(Val);
1730     if (TTmpIdx >= 0) {
1731       auto TTmpClsId = getTtmpClassId(IsWave64 ? OPW64 : OPW32);
1732       return createSRegOperand(TTmpClsId, TTmpIdx);
1733     }
1734     if (Val > SGPR_MAX) {
1735       return IsWave64 ? decodeSpecialReg64(Val) : decodeSpecialReg32(Val);
1736     }
1737     return createSRegOperand(getSgprClassId(IsWave64 ? OPW64 : OPW32), Val);
1738   }
1739   return createRegOperand(IsWave64 ? AMDGPU::VCC : AMDGPU::VCC_LO);
1740 }
1741 
1742 MCOperand AMDGPUDisassembler::decodeBoolReg(unsigned Val) const {
1743   return STI.hasFeature(AMDGPU::FeatureWavefrontSize64)
1744              ? decodeSrcOp(OPW64, Val)
1745              : decodeSrcOp(OPW32, Val);
1746 }
1747 
1748 MCOperand AMDGPUDisassembler::decodeSplitBarrier(unsigned Val) const {
1749   return decodeSrcOp(OPW32, Val);
1750 }
1751 
1752 MCOperand AMDGPUDisassembler::decodeDpp8FI(unsigned Val) const {
1753   if (Val != AMDGPU::DPP::DPP8_FI_0 && Val != AMDGPU::DPP::DPP8_FI_1)
1754     return MCOperand();
1755   return MCOperand::createImm(Val);
1756 }
1757 
1758 MCOperand AMDGPUDisassembler::decodeVersionImm(unsigned Imm) const {
1759   using VersionField = AMDGPU::EncodingField<7, 0>;
1760   using W64Bit = AMDGPU::EncodingBit<13>;
1761   using W32Bit = AMDGPU::EncodingBit<14>;
1762   using MDPBit = AMDGPU::EncodingBit<15>;
1763   using Encoding = AMDGPU::EncodingFields<VersionField, W64Bit, W32Bit, MDPBit>;
1764 
1765   auto [Version, W64, W32, MDP] = Encoding::decode(Imm);
1766 
1767   // Decode into a plain immediate if any unused bits are raised.
1768   if (Encoding::encode(Version, W64, W32, MDP) != Imm)
1769     return MCOperand::createImm(Imm);
1770 
1771   const auto &Versions = AMDGPU::UCVersion::getGFXVersions();
1772   auto I = find_if(Versions,
1773                    [Version = Version](const AMDGPU::UCVersion::GFXVersion &V) {
1774                      return V.Code == Version;
1775                    });
1776   MCContext &Ctx = getContext();
1777   const MCExpr *E;
1778   if (I == Versions.end())
1779     E = MCConstantExpr::create(Version, Ctx);
1780   else
1781     E = MCSymbolRefExpr::create(Ctx.getOrCreateSymbol(I->Symbol), Ctx);
1782 
1783   if (W64)
1784     E = MCBinaryExpr::createOr(E, UCVersionW64Expr, Ctx);
1785   if (W32)
1786     E = MCBinaryExpr::createOr(E, UCVersionW32Expr, Ctx);
1787   if (MDP)
1788     E = MCBinaryExpr::createOr(E, UCVersionMDPExpr, Ctx);
1789 
1790   return MCOperand::createExpr(E);
1791 }
1792 
1793 bool AMDGPUDisassembler::isVI() const {
1794   return STI.hasFeature(AMDGPU::FeatureVolcanicIslands);
1795 }
1796 
1797 bool AMDGPUDisassembler::isGFX9() const { return AMDGPU::isGFX9(STI); }
1798 
1799 bool AMDGPUDisassembler::isGFX90A() const {
1800   return STI.hasFeature(AMDGPU::FeatureGFX90AInsts);
1801 }
1802 
1803 bool AMDGPUDisassembler::isGFX9Plus() const { return AMDGPU::isGFX9Plus(STI); }
1804 
1805 bool AMDGPUDisassembler::isGFX10() const { return AMDGPU::isGFX10(STI); }
1806 
1807 bool AMDGPUDisassembler::isGFX10Plus() const {
1808   return AMDGPU::isGFX10Plus(STI);
1809 }
1810 
1811 bool AMDGPUDisassembler::isGFX11() const {
1812   return STI.hasFeature(AMDGPU::FeatureGFX11);
1813 }
1814 
1815 bool AMDGPUDisassembler::isGFX11Plus() const {
1816   return AMDGPU::isGFX11Plus(STI);
1817 }
1818 
1819 bool AMDGPUDisassembler::isGFX12() const {
1820   return STI.hasFeature(AMDGPU::FeatureGFX12);
1821 }
1822 
1823 bool AMDGPUDisassembler::isGFX12Plus() const {
1824   return AMDGPU::isGFX12Plus(STI);
1825 }
1826 
1827 bool AMDGPUDisassembler::hasArchitectedFlatScratch() const {
1828   return STI.hasFeature(AMDGPU::FeatureArchitectedFlatScratch);
1829 }
1830 
1831 bool AMDGPUDisassembler::hasKernargPreload() const {
1832   return AMDGPU::hasKernargPreload(STI);
1833 }
1834 
1835 //===----------------------------------------------------------------------===//
1836 // AMDGPU specific symbol handling
1837 //===----------------------------------------------------------------------===//
1838 
1839 /// Print a string describing the reserved bit range specified by Mask with
1840 /// offset BaseBytes for use in error comments. Mask is a single continuous
1841 /// range of 1s surrounded by zeros. The format here is meant to align with the
1842 /// tables that describe these bits in llvm.org/docs/AMDGPUUsage.html.
1843 static SmallString<32> getBitRangeFromMask(uint32_t Mask, unsigned BaseBytes) {
1844   SmallString<32> Result;
1845   raw_svector_ostream S(Result);
1846 
1847   int TrailingZeros = llvm::countr_zero(Mask);
1848   int PopCount = llvm::popcount(Mask);
1849 
1850   if (PopCount == 1) {
1851     S << "bit (" << (TrailingZeros + BaseBytes * CHAR_BIT) << ')';
1852   } else {
1853     S << "bits in range ("
1854       << (TrailingZeros + PopCount - 1 + BaseBytes * CHAR_BIT) << ':'
1855       << (TrailingZeros + BaseBytes * CHAR_BIT) << ')';
1856   }
1857 
1858   return Result;
1859 }
1860 
1861 #define GET_FIELD(MASK) (AMDHSA_BITS_GET(FourByteBuffer, MASK))
1862 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
1863   do {                                                                         \
1864     KdStream << Indent << DIRECTIVE " " << GET_FIELD(MASK) << '\n';            \
1865   } while (0)
1866 #define PRINT_PSEUDO_DIRECTIVE_COMMENT(DIRECTIVE, MASK)                        \
1867   do {                                                                         \
1868     KdStream << Indent << MAI.getCommentString() << ' ' << DIRECTIVE " "       \
1869              << GET_FIELD(MASK) << '\n';                                       \
1870   } while (0)
1871 
1872 #define CHECK_RESERVED_BITS_IMPL(MASK, DESC, MSG)                              \
1873   do {                                                                         \
1874     if (FourByteBuffer & (MASK)) {                                             \
1875       return createStringError(std::errc::invalid_argument,                    \
1876                                "kernel descriptor " DESC                       \
1877                                " reserved %s set" MSG,                         \
1878                                getBitRangeFromMask((MASK), 0).c_str());        \
1879     }                                                                          \
1880   } while (0)
1881 
1882 #define CHECK_RESERVED_BITS(MASK) CHECK_RESERVED_BITS_IMPL(MASK, #MASK, "")
1883 #define CHECK_RESERVED_BITS_MSG(MASK, MSG)                                     \
1884   CHECK_RESERVED_BITS_IMPL(MASK, #MASK, ", " MSG)
1885 #define CHECK_RESERVED_BITS_DESC(MASK, DESC)                                   \
1886   CHECK_RESERVED_BITS_IMPL(MASK, DESC, "")
1887 #define CHECK_RESERVED_BITS_DESC_MSG(MASK, DESC, MSG)                          \
1888   CHECK_RESERVED_BITS_IMPL(MASK, DESC, ", " MSG)
1889 
1890 // NOLINTNEXTLINE(readability-identifier-naming)
1891 Expected<bool> AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC1(
1892     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
1893   using namespace amdhsa;
1894   StringRef Indent = "\t";
1895 
1896   // We cannot accurately backward compute #VGPRs used from
1897   // GRANULATED_WORKITEM_VGPR_COUNT. But we are concerned with getting the same
1898   // value of GRANULATED_WORKITEM_VGPR_COUNT in the reassembled binary. So we
1899   // simply calculate the inverse of what the assembler does.
1900 
1901   uint32_t GranulatedWorkitemVGPRCount =
1902       GET_FIELD(COMPUTE_PGM_RSRC1_GRANULATED_WORKITEM_VGPR_COUNT);
1903 
1904   uint32_t NextFreeVGPR =
1905       (GranulatedWorkitemVGPRCount + 1) *
1906       AMDGPU::IsaInfo::getVGPREncodingGranule(&STI, EnableWavefrontSize32);
1907 
1908   KdStream << Indent << ".amdhsa_next_free_vgpr " << NextFreeVGPR << '\n';
1909 
1910   // We cannot backward compute values used to calculate
1911   // GRANULATED_WAVEFRONT_SGPR_COUNT. Hence the original values for following
1912   // directives can't be computed:
1913   // .amdhsa_reserve_vcc
1914   // .amdhsa_reserve_flat_scratch
1915   // .amdhsa_reserve_xnack_mask
1916   // They take their respective default values if not specified in the assembly.
1917   //
1918   // GRANULATED_WAVEFRONT_SGPR_COUNT
1919   //    = f(NEXT_FREE_SGPR + VCC + FLAT_SCRATCH + XNACK_MASK)
1920   //
1921   // We compute the inverse as though all directives apart from NEXT_FREE_SGPR
1922   // are set to 0. So while disassembling we consider that:
1923   //
1924   // GRANULATED_WAVEFRONT_SGPR_COUNT
1925   //    = f(NEXT_FREE_SGPR + 0 + 0 + 0)
1926   //
1927   // The disassembler cannot recover the original values of those 3 directives.
1928 
1929   uint32_t GranulatedWavefrontSGPRCount =
1930       GET_FIELD(COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT);
1931 
1932   if (isGFX10Plus())
1933     CHECK_RESERVED_BITS_MSG(COMPUTE_PGM_RSRC1_GRANULATED_WAVEFRONT_SGPR_COUNT,
1934                             "must be zero on gfx10+");
1935 
1936   uint32_t NextFreeSGPR = (GranulatedWavefrontSGPRCount + 1) *
1937                           AMDGPU::IsaInfo::getSGPREncodingGranule(&STI);
1938 
1939   KdStream << Indent << ".amdhsa_reserve_vcc " << 0 << '\n';
1940   if (!hasArchitectedFlatScratch())
1941     KdStream << Indent << ".amdhsa_reserve_flat_scratch " << 0 << '\n';
1942   KdStream << Indent << ".amdhsa_reserve_xnack_mask " << 0 << '\n';
1943   KdStream << Indent << ".amdhsa_next_free_sgpr " << NextFreeSGPR << "\n";
1944 
1945   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC1_PRIORITY);
1946 
1947   PRINT_DIRECTIVE(".amdhsa_float_round_mode_32",
1948                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_32);
1949   PRINT_DIRECTIVE(".amdhsa_float_round_mode_16_64",
1950                   COMPUTE_PGM_RSRC1_FLOAT_ROUND_MODE_16_64);
1951   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_32",
1952                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_32);
1953   PRINT_DIRECTIVE(".amdhsa_float_denorm_mode_16_64",
1954                   COMPUTE_PGM_RSRC1_FLOAT_DENORM_MODE_16_64);
1955 
1956   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC1_PRIV);
1957 
1958   if (!isGFX12Plus())
1959     PRINT_DIRECTIVE(".amdhsa_dx10_clamp",
1960                     COMPUTE_PGM_RSRC1_GFX6_GFX11_ENABLE_DX10_CLAMP);
1961 
1962   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC1_DEBUG_MODE);
1963 
1964   if (!isGFX12Plus())
1965     PRINT_DIRECTIVE(".amdhsa_ieee_mode",
1966                     COMPUTE_PGM_RSRC1_GFX6_GFX11_ENABLE_IEEE_MODE);
1967 
1968   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC1_BULKY);
1969   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC1_CDBG_USER);
1970 
1971   if (isGFX9Plus())
1972     PRINT_DIRECTIVE(".amdhsa_fp16_overflow", COMPUTE_PGM_RSRC1_GFX9_PLUS_FP16_OVFL);
1973 
1974   if (!isGFX9Plus())
1975     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC1_GFX6_GFX8_RESERVED0,
1976                                  "COMPUTE_PGM_RSRC1", "must be zero pre-gfx9");
1977 
1978   CHECK_RESERVED_BITS_DESC(COMPUTE_PGM_RSRC1_RESERVED1, "COMPUTE_PGM_RSRC1");
1979 
1980   if (!isGFX10Plus())
1981     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC1_GFX6_GFX9_RESERVED2,
1982                                  "COMPUTE_PGM_RSRC1", "must be zero pre-gfx10");
1983 
1984   if (isGFX10Plus()) {
1985     PRINT_DIRECTIVE(".amdhsa_workgroup_processor_mode",
1986                     COMPUTE_PGM_RSRC1_GFX10_PLUS_WGP_MODE);
1987     PRINT_DIRECTIVE(".amdhsa_memory_ordered", COMPUTE_PGM_RSRC1_GFX10_PLUS_MEM_ORDERED);
1988     PRINT_DIRECTIVE(".amdhsa_forward_progress", COMPUTE_PGM_RSRC1_GFX10_PLUS_FWD_PROGRESS);
1989   }
1990 
1991   if (isGFX12Plus())
1992     PRINT_DIRECTIVE(".amdhsa_round_robin_scheduling",
1993                     COMPUTE_PGM_RSRC1_GFX12_PLUS_ENABLE_WG_RR_EN);
1994 
1995   return true;
1996 }
1997 
1998 // NOLINTNEXTLINE(readability-identifier-naming)
1999 Expected<bool> AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC2(
2000     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
2001   using namespace amdhsa;
2002   StringRef Indent = "\t";
2003   if (hasArchitectedFlatScratch())
2004     PRINT_DIRECTIVE(".amdhsa_enable_private_segment",
2005                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
2006   else
2007     PRINT_DIRECTIVE(".amdhsa_system_sgpr_private_segment_wavefront_offset",
2008                     COMPUTE_PGM_RSRC2_ENABLE_PRIVATE_SEGMENT);
2009   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_x",
2010                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_X);
2011   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_y",
2012                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Y);
2013   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_id_z",
2014                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_ID_Z);
2015   PRINT_DIRECTIVE(".amdhsa_system_sgpr_workgroup_info",
2016                   COMPUTE_PGM_RSRC2_ENABLE_SGPR_WORKGROUP_INFO);
2017   PRINT_DIRECTIVE(".amdhsa_system_vgpr_workitem_id",
2018                   COMPUTE_PGM_RSRC2_ENABLE_VGPR_WORKITEM_ID);
2019 
2020   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_ADDRESS_WATCH);
2021   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_MEMORY);
2022   CHECK_RESERVED_BITS(COMPUTE_PGM_RSRC2_GRANULATED_LDS_SIZE);
2023 
2024   PRINT_DIRECTIVE(
2025       ".amdhsa_exception_fp_ieee_invalid_op",
2026       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INVALID_OPERATION);
2027   PRINT_DIRECTIVE(".amdhsa_exception_fp_denorm_src",
2028                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_FP_DENORMAL_SOURCE);
2029   PRINT_DIRECTIVE(
2030       ".amdhsa_exception_fp_ieee_div_zero",
2031       COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_DIVISION_BY_ZERO);
2032   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_overflow",
2033                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_OVERFLOW);
2034   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_underflow",
2035                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_UNDERFLOW);
2036   PRINT_DIRECTIVE(".amdhsa_exception_fp_ieee_inexact",
2037                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_IEEE_754_FP_INEXACT);
2038   PRINT_DIRECTIVE(".amdhsa_exception_int_div_zero",
2039                   COMPUTE_PGM_RSRC2_ENABLE_EXCEPTION_INT_DIVIDE_BY_ZERO);
2040 
2041   CHECK_RESERVED_BITS_DESC(COMPUTE_PGM_RSRC2_RESERVED0, "COMPUTE_PGM_RSRC2");
2042 
2043   return true;
2044 }
2045 
2046 // NOLINTNEXTLINE(readability-identifier-naming)
2047 Expected<bool> AMDGPUDisassembler::decodeCOMPUTE_PGM_RSRC3(
2048     uint32_t FourByteBuffer, raw_string_ostream &KdStream) const {
2049   using namespace amdhsa;
2050   StringRef Indent = "\t";
2051   if (isGFX90A()) {
2052     KdStream << Indent << ".amdhsa_accum_offset "
2053              << (GET_FIELD(COMPUTE_PGM_RSRC3_GFX90A_ACCUM_OFFSET) + 1) * 4
2054              << '\n';
2055 
2056     PRINT_DIRECTIVE(".amdhsa_tg_split", COMPUTE_PGM_RSRC3_GFX90A_TG_SPLIT);
2057 
2058     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX90A_RESERVED0,
2059                                  "COMPUTE_PGM_RSRC3", "must be zero on gfx90a");
2060     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX90A_RESERVED1,
2061                                  "COMPUTE_PGM_RSRC3", "must be zero on gfx90a");
2062   } else if (isGFX10Plus()) {
2063     // Bits [0-3].
2064     if (!isGFX12Plus()) {
2065       if (!EnableWavefrontSize32 || !*EnableWavefrontSize32) {
2066         PRINT_DIRECTIVE(".amdhsa_shared_vgpr_count",
2067                         COMPUTE_PGM_RSRC3_GFX10_GFX11_SHARED_VGPR_COUNT);
2068       } else {
2069         PRINT_PSEUDO_DIRECTIVE_COMMENT(
2070             "SHARED_VGPR_COUNT",
2071             COMPUTE_PGM_RSRC3_GFX10_GFX11_SHARED_VGPR_COUNT);
2072       }
2073     } else {
2074       CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX12_PLUS_RESERVED0,
2075                                    "COMPUTE_PGM_RSRC3",
2076                                    "must be zero on gfx12+");
2077     }
2078 
2079     // Bits [4-11].
2080     if (isGFX11()) {
2081       PRINT_PSEUDO_DIRECTIVE_COMMENT("INST_PREF_SIZE",
2082                                      COMPUTE_PGM_RSRC3_GFX11_INST_PREF_SIZE);
2083       PRINT_PSEUDO_DIRECTIVE_COMMENT("TRAP_ON_START",
2084                                      COMPUTE_PGM_RSRC3_GFX11_TRAP_ON_START);
2085       PRINT_PSEUDO_DIRECTIVE_COMMENT("TRAP_ON_END",
2086                                      COMPUTE_PGM_RSRC3_GFX11_TRAP_ON_END);
2087     } else if (isGFX12Plus()) {
2088       PRINT_PSEUDO_DIRECTIVE_COMMENT(
2089           "INST_PREF_SIZE", COMPUTE_PGM_RSRC3_GFX12_PLUS_INST_PREF_SIZE);
2090     } else {
2091       CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX10_RESERVED1,
2092                                    "COMPUTE_PGM_RSRC3",
2093                                    "must be zero on gfx10");
2094     }
2095 
2096     // Bits [12].
2097     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX10_PLUS_RESERVED2,
2098                                  "COMPUTE_PGM_RSRC3", "must be zero on gfx10+");
2099 
2100     // Bits [13].
2101     if (isGFX12Plus()) {
2102       PRINT_PSEUDO_DIRECTIVE_COMMENT("GLG_EN",
2103                                      COMPUTE_PGM_RSRC3_GFX12_PLUS_GLG_EN);
2104     } else {
2105       CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX10_GFX11_RESERVED3,
2106                                    "COMPUTE_PGM_RSRC3",
2107                                    "must be zero on gfx10 or gfx11");
2108     }
2109 
2110     // Bits [14-30].
2111     CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX10_PLUS_RESERVED4,
2112                                  "COMPUTE_PGM_RSRC3", "must be zero on gfx10+");
2113 
2114     // Bits [31].
2115     if (isGFX11Plus()) {
2116       PRINT_PSEUDO_DIRECTIVE_COMMENT("IMAGE_OP",
2117                                      COMPUTE_PGM_RSRC3_GFX11_PLUS_IMAGE_OP);
2118     } else {
2119       CHECK_RESERVED_BITS_DESC_MSG(COMPUTE_PGM_RSRC3_GFX10_RESERVED5,
2120                                    "COMPUTE_PGM_RSRC3",
2121                                    "must be zero on gfx10");
2122     }
2123   } else if (FourByteBuffer) {
2124     return createStringError(
2125         std::errc::invalid_argument,
2126         "kernel descriptor COMPUTE_PGM_RSRC3 must be all zero before gfx9");
2127   }
2128   return true;
2129 }
2130 #undef PRINT_PSEUDO_DIRECTIVE_COMMENT
2131 #undef PRINT_DIRECTIVE
2132 #undef GET_FIELD
2133 #undef CHECK_RESERVED_BITS_IMPL
2134 #undef CHECK_RESERVED_BITS
2135 #undef CHECK_RESERVED_BITS_MSG
2136 #undef CHECK_RESERVED_BITS_DESC
2137 #undef CHECK_RESERVED_BITS_DESC_MSG
2138 
2139 /// Create an error object to return from onSymbolStart for reserved kernel
2140 /// descriptor bits being set.
2141 static Error createReservedKDBitsError(uint32_t Mask, unsigned BaseBytes,
2142                                        const char *Msg = "") {
2143   return createStringError(
2144       std::errc::invalid_argument, "kernel descriptor reserved %s set%s%s",
2145       getBitRangeFromMask(Mask, BaseBytes).c_str(), *Msg ? ", " : "", Msg);
2146 }
2147 
2148 /// Create an error object to return from onSymbolStart for reserved kernel
2149 /// descriptor bytes being set.
2150 static Error createReservedKDBytesError(unsigned BaseInBytes,
2151                                         unsigned WidthInBytes) {
2152   // Create an error comment in the same format as the "Kernel Descriptor"
2153   // table here: https://llvm.org/docs/AMDGPUUsage.html#kernel-descriptor .
2154   return createStringError(
2155       std::errc::invalid_argument,
2156       "kernel descriptor reserved bits in range (%u:%u) set",
2157       (BaseInBytes + WidthInBytes) * CHAR_BIT - 1, BaseInBytes * CHAR_BIT);
2158 }
2159 
2160 Expected<bool> AMDGPUDisassembler::decodeKernelDescriptorDirective(
2161     DataExtractor::Cursor &Cursor, ArrayRef<uint8_t> Bytes,
2162     raw_string_ostream &KdStream) const {
2163 #define PRINT_DIRECTIVE(DIRECTIVE, MASK)                                       \
2164   do {                                                                         \
2165     KdStream << Indent << DIRECTIVE " "                                        \
2166              << ((TwoByteBuffer & MASK) >> (MASK##_SHIFT)) << '\n';            \
2167   } while (0)
2168 
2169   uint16_t TwoByteBuffer = 0;
2170   uint32_t FourByteBuffer = 0;
2171 
2172   StringRef ReservedBytes;
2173   StringRef Indent = "\t";
2174 
2175   assert(Bytes.size() == 64);
2176   DataExtractor DE(Bytes, /*IsLittleEndian=*/true, /*AddressSize=*/8);
2177 
2178   switch (Cursor.tell()) {
2179   case amdhsa::GROUP_SEGMENT_FIXED_SIZE_OFFSET:
2180     FourByteBuffer = DE.getU32(Cursor);
2181     KdStream << Indent << ".amdhsa_group_segment_fixed_size " << FourByteBuffer
2182              << '\n';
2183     return true;
2184 
2185   case amdhsa::PRIVATE_SEGMENT_FIXED_SIZE_OFFSET:
2186     FourByteBuffer = DE.getU32(Cursor);
2187     KdStream << Indent << ".amdhsa_private_segment_fixed_size "
2188              << FourByteBuffer << '\n';
2189     return true;
2190 
2191   case amdhsa::KERNARG_SIZE_OFFSET:
2192     FourByteBuffer = DE.getU32(Cursor);
2193     KdStream << Indent << ".amdhsa_kernarg_size "
2194              << FourByteBuffer << '\n';
2195     return true;
2196 
2197   case amdhsa::RESERVED0_OFFSET:
2198     // 4 reserved bytes, must be 0.
2199     ReservedBytes = DE.getBytes(Cursor, 4);
2200     for (int I = 0; I < 4; ++I) {
2201       if (ReservedBytes[I] != 0)
2202         return createReservedKDBytesError(amdhsa::RESERVED0_OFFSET, 4);
2203     }
2204     return true;
2205 
2206   case amdhsa::KERNEL_CODE_ENTRY_BYTE_OFFSET_OFFSET:
2207     // KERNEL_CODE_ENTRY_BYTE_OFFSET
2208     // So far no directive controls this for Code Object V3, so simply skip for
2209     // disassembly.
2210     DE.skip(Cursor, 8);
2211     return true;
2212 
2213   case amdhsa::RESERVED1_OFFSET:
2214     // 20 reserved bytes, must be 0.
2215     ReservedBytes = DE.getBytes(Cursor, 20);
2216     for (int I = 0; I < 20; ++I) {
2217       if (ReservedBytes[I] != 0)
2218         return createReservedKDBytesError(amdhsa::RESERVED1_OFFSET, 20);
2219     }
2220     return true;
2221 
2222   case amdhsa::COMPUTE_PGM_RSRC3_OFFSET:
2223     FourByteBuffer = DE.getU32(Cursor);
2224     return decodeCOMPUTE_PGM_RSRC3(FourByteBuffer, KdStream);
2225 
2226   case amdhsa::COMPUTE_PGM_RSRC1_OFFSET:
2227     FourByteBuffer = DE.getU32(Cursor);
2228     return decodeCOMPUTE_PGM_RSRC1(FourByteBuffer, KdStream);
2229 
2230   case amdhsa::COMPUTE_PGM_RSRC2_OFFSET:
2231     FourByteBuffer = DE.getU32(Cursor);
2232     return decodeCOMPUTE_PGM_RSRC2(FourByteBuffer, KdStream);
2233 
2234   case amdhsa::KERNEL_CODE_PROPERTIES_OFFSET:
2235     using namespace amdhsa;
2236     TwoByteBuffer = DE.getU16(Cursor);
2237 
2238     if (!hasArchitectedFlatScratch())
2239       PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_buffer",
2240                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_BUFFER);
2241     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_ptr",
2242                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_PTR);
2243     PRINT_DIRECTIVE(".amdhsa_user_sgpr_queue_ptr",
2244                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_QUEUE_PTR);
2245     PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_segment_ptr",
2246                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_KERNARG_SEGMENT_PTR);
2247     PRINT_DIRECTIVE(".amdhsa_user_sgpr_dispatch_id",
2248                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_DISPATCH_ID);
2249     if (!hasArchitectedFlatScratch())
2250       PRINT_DIRECTIVE(".amdhsa_user_sgpr_flat_scratch_init",
2251                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
2252     PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size",
2253                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
2254 
2255     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0)
2256       return createReservedKDBitsError(KERNEL_CODE_PROPERTY_RESERVED0,
2257                                        amdhsa::KERNEL_CODE_PROPERTIES_OFFSET);
2258 
2259     // Reserved for GFX9
2260     if (isGFX9() &&
2261         (TwoByteBuffer & KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32)) {
2262       return createReservedKDBitsError(
2263           KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32,
2264           amdhsa::KERNEL_CODE_PROPERTIES_OFFSET, "must be zero on gfx9");
2265     }
2266     if (isGFX10Plus()) {
2267       PRINT_DIRECTIVE(".amdhsa_wavefront_size32",
2268                       KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
2269     }
2270 
2271     if (CodeObjectVersion >= AMDGPU::AMDHSA_COV5)
2272       PRINT_DIRECTIVE(".amdhsa_uses_dynamic_stack",
2273                       KERNEL_CODE_PROPERTY_USES_DYNAMIC_STACK);
2274 
2275     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED1) {
2276       return createReservedKDBitsError(KERNEL_CODE_PROPERTY_RESERVED1,
2277                                        amdhsa::KERNEL_CODE_PROPERTIES_OFFSET);
2278     }
2279 
2280     return true;
2281 
2282   case amdhsa::KERNARG_PRELOAD_OFFSET:
2283     using namespace amdhsa;
2284     TwoByteBuffer = DE.getU16(Cursor);
2285     if (TwoByteBuffer & KERNARG_PRELOAD_SPEC_LENGTH) {
2286       PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_preload_length",
2287                       KERNARG_PRELOAD_SPEC_LENGTH);
2288     }
2289 
2290     if (TwoByteBuffer & KERNARG_PRELOAD_SPEC_OFFSET) {
2291       PRINT_DIRECTIVE(".amdhsa_user_sgpr_kernarg_preload_offset",
2292                       KERNARG_PRELOAD_SPEC_OFFSET);
2293     }
2294     return true;
2295 
2296   case amdhsa::RESERVED3_OFFSET:
2297     // 4 bytes from here are reserved, must be 0.
2298     ReservedBytes = DE.getBytes(Cursor, 4);
2299     for (int I = 0; I < 4; ++I) {
2300       if (ReservedBytes[I] != 0)
2301         return createReservedKDBytesError(amdhsa::RESERVED3_OFFSET, 4);
2302     }
2303     return true;
2304 
2305   default:
2306     llvm_unreachable("Unhandled index. Case statements cover everything.");
2307     return true;
2308   }
2309 #undef PRINT_DIRECTIVE
2310 }
2311 
2312 Expected<bool> AMDGPUDisassembler::decodeKernelDescriptor(
2313     StringRef KdName, ArrayRef<uint8_t> Bytes, uint64_t KdAddress) const {
2314 
2315   // CP microcode requires the kernel descriptor to be 64 aligned.
2316   if (Bytes.size() != 64 || KdAddress % 64 != 0)
2317     return createStringError(std::errc::invalid_argument,
2318                              "kernel descriptor must be 64-byte aligned");
2319 
2320   // FIXME: We can't actually decode "in order" as is done below, as e.g. GFX10
2321   // requires us to know the setting of .amdhsa_wavefront_size32 in order to
2322   // accurately produce .amdhsa_next_free_vgpr, and they appear in the wrong
2323   // order. Workaround this by first looking up .amdhsa_wavefront_size32 here
2324   // when required.
2325   if (isGFX10Plus()) {
2326     uint16_t KernelCodeProperties =
2327         support::endian::read16(&Bytes[amdhsa::KERNEL_CODE_PROPERTIES_OFFSET],
2328                                 llvm::endianness::little);
2329     EnableWavefrontSize32 =
2330         AMDHSA_BITS_GET(KernelCodeProperties,
2331                         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32);
2332   }
2333 
2334   std::string Kd;
2335   raw_string_ostream KdStream(Kd);
2336   KdStream << ".amdhsa_kernel " << KdName << '\n';
2337 
2338   DataExtractor::Cursor C(0);
2339   while (C && C.tell() < Bytes.size()) {
2340     Expected<bool> Res = decodeKernelDescriptorDirective(C, Bytes, KdStream);
2341 
2342     cantFail(C.takeError());
2343 
2344     if (!Res)
2345       return Res;
2346   }
2347   KdStream << ".end_amdhsa_kernel\n";
2348   outs() << KdStream.str();
2349   return true;
2350 }
2351 
2352 Expected<bool> AMDGPUDisassembler::onSymbolStart(SymbolInfoTy &Symbol,
2353                                                  uint64_t &Size,
2354                                                  ArrayRef<uint8_t> Bytes,
2355                                                  uint64_t Address) const {
2356   // Right now only kernel descriptor needs to be handled.
2357   // We ignore all other symbols for target specific handling.
2358   // TODO:
2359   // Fix the spurious symbol issue for AMDGPU kernels. Exists for both Code
2360   // Object V2 and V3 when symbols are marked protected.
2361 
2362   // amd_kernel_code_t for Code Object V2.
2363   if (Symbol.Type == ELF::STT_AMDGPU_HSA_KERNEL) {
2364     Size = 256;
2365     return createStringError(std::errc::invalid_argument,
2366                              "code object v2 is not supported");
2367   }
2368 
2369   // Code Object V3 kernel descriptors.
2370   StringRef Name = Symbol.Name;
2371   if (Symbol.Type == ELF::STT_OBJECT && Name.ends_with(StringRef(".kd"))) {
2372     Size = 64; // Size = 64 regardless of success or failure.
2373     return decodeKernelDescriptor(Name.drop_back(3), Bytes, Address);
2374   }
2375 
2376   return false;
2377 }
2378 
2379 const MCExpr *AMDGPUDisassembler::createConstantSymbolExpr(StringRef Id,
2380                                                            int64_t Val) {
2381   MCContext &Ctx = getContext();
2382   MCSymbol *Sym = Ctx.getOrCreateSymbol(Id);
2383   // Note: only set value to Val on a new symbol in case an dissassembler
2384   // has already been initialized in this context.
2385   if (!Sym->isVariable()) {
2386     Sym->setVariableValue(MCConstantExpr::create(Val, Ctx));
2387   } else {
2388     int64_t Res = ~Val;
2389     bool Valid = Sym->getVariableValue()->evaluateAsAbsolute(Res);
2390     if (!Valid || Res != Val)
2391       Ctx.reportWarning(SMLoc(), "unsupported redefinition of " + Id);
2392   }
2393   return MCSymbolRefExpr::create(Sym, Ctx);
2394 }
2395 
2396 //===----------------------------------------------------------------------===//
2397 // AMDGPUSymbolizer
2398 //===----------------------------------------------------------------------===//
2399 
2400 // Try to find symbol name for specified label
2401 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(
2402     MCInst &Inst, raw_ostream & /*cStream*/, int64_t Value,
2403     uint64_t /*Address*/, bool IsBranch, uint64_t /*Offset*/,
2404     uint64_t /*OpSize*/, uint64_t /*InstSize*/) {
2405 
2406   if (!IsBranch) {
2407     return false;
2408   }
2409 
2410   auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
2411   if (!Symbols)
2412     return false;
2413 
2414   auto Result = llvm::find_if(*Symbols, [Value](const SymbolInfoTy &Val) {
2415     return Val.Addr == static_cast<uint64_t>(Value) &&
2416            Val.Type == ELF::STT_NOTYPE;
2417   });
2418   if (Result != Symbols->end()) {
2419     auto *Sym = Ctx.getOrCreateSymbol(Result->Name);
2420     const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
2421     Inst.addOperand(MCOperand::createExpr(Add));
2422     return true;
2423   }
2424   // Add to list of referenced addresses, so caller can synthesize a label.
2425   ReferencedAddresses.push_back(static_cast<uint64_t>(Value));
2426   return false;
2427 }
2428 
2429 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
2430                                                        int64_t Value,
2431                                                        uint64_t Address) {
2432   llvm_unreachable("unimplemented");
2433 }
2434 
2435 //===----------------------------------------------------------------------===//
2436 // Initialization
2437 //===----------------------------------------------------------------------===//
2438 
2439 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
2440                               LLVMOpInfoCallback /*GetOpInfo*/,
2441                               LLVMSymbolLookupCallback /*SymbolLookUp*/,
2442                               void *DisInfo,
2443                               MCContext *Ctx,
2444                               std::unique_ptr<MCRelocationInfo> &&RelInfo) {
2445   return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
2446 }
2447 
2448 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
2449                                                 const MCSubtargetInfo &STI,
2450                                                 MCContext &Ctx) {
2451   return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo());
2452 }
2453 
2454 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUDisassembler() {
2455   TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
2456                                          createAMDGPUDisassembler);
2457   TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
2458                                        createAMDGPUSymbolizer);
2459 }
2460