1 //===-- AMDGPUDisassembler.cpp - Disassembler for AMDGPU ISA --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 //===----------------------------------------------------------------------===//
14 /// This file contains definition for AMDGPU ISA disassembler
16 //===----------------------------------------------------------------------===//
18 // ToDo: What to do with instruction suffixes (v_mov_b32 vs v_mov_b32_e32)?
20 #include "AMDGPUDisassembler.h"
22 #include "AMDGPURegisterInfo.h"
23 #include "Utils/AMDGPUBaseInfo.h"
25 #include "llvm/MC/MCContext.h"
26 #include "llvm/MC/MCFixedLenDisassembler.h"
27 #include "llvm/MC/MCInst.h"
28 #include "llvm/MC/MCInstrDesc.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/Support/Endian.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/TargetRegistry.h"
37 #define DEBUG_TYPE "amdgpu-disassembler"
39 typedef llvm::MCDisassembler::DecodeStatus DecodeStatus;
42 inline static MCDisassembler::DecodeStatus
43 addOperand(MCInst &Inst, const MCOperand& Opnd) {
44 Inst.addOperand(Opnd);
45 return Opnd.isValid() ?
46 MCDisassembler::Success :
47 MCDisassembler::SoftFail;
50 #define DECODE_OPERAND2(RegClass, DecName) \
51 static DecodeStatus Decode##RegClass##RegisterClass(MCInst &Inst, \
54 const void *Decoder) { \
55 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); \
56 return addOperand(Inst, DAsm->decodeOperand_##DecName(Imm)); \
59 #define DECODE_OPERAND(RegClass) DECODE_OPERAND2(RegClass, RegClass)
61 DECODE_OPERAND(VGPR_32)
65 DECODE_OPERAND(VReg_64)
66 DECODE_OPERAND(VReg_96)
67 DECODE_OPERAND(VReg_128)
69 DECODE_OPERAND(SGPR_32)
70 DECODE_OPERAND(SReg_32)
71 DECODE_OPERAND(SReg_64)
72 DECODE_OPERAND(SReg_128)
73 DECODE_OPERAND(SReg_256)
74 DECODE_OPERAND(SReg_512)
76 #define GET_SUBTARGETINFO_ENUM
77 #include "AMDGPUGenSubtargetInfo.inc"
78 #undef GET_SUBTARGETINFO_ENUM
80 #include "AMDGPUGenDisassemblerTables.inc"
82 //===----------------------------------------------------------------------===//
84 //===----------------------------------------------------------------------===//
86 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
87 assert(Bytes.size() >= sizeof(T));
88 const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data());
89 Bytes = Bytes.slice(sizeof(T));
93 DecodeStatus AMDGPUDisassembler::tryDecodeInst(const uint8_t* Table,
96 uint64_t Address) const {
97 assert(MI.getOpcode() == 0);
98 assert(MI.getNumOperands() == 0);
100 const auto SavedBytes = Bytes;
101 if (decodeInstruction(Table, TmpInst, Inst, Address, this, STI)) {
103 return MCDisassembler::Success;
106 return MCDisassembler::Fail;
109 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
110 ArrayRef<uint8_t> Bytes_,
113 raw_ostream &CS) const {
116 // ToDo: AMDGPUDisassembler supports only VI ISA.
117 assert(AMDGPU::isVI(STI) && "Can disassemble only VI ISA.");
119 const unsigned MaxInstBytesNum = (std::min)((size_t)8, Bytes_.size());
120 Bytes = Bytes_.slice(0, MaxInstBytesNum);
122 DecodeStatus Res = MCDisassembler::Fail;
124 // ToDo: better to switch encoding length using some bit predicate
125 // but it is unknown yet, so try all we can
127 // Try to decode DPP first to solve conflict with VOP1 and VOP2 encodings
128 if (Bytes.size() >= 8) {
129 const uint64_t QW = eatBytes<uint64_t>(Bytes);
130 Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address);
134 // Reinitialize Bytes as DPP64 could have eaten too much
135 Bytes = Bytes_.slice(0, MaxInstBytesNum);
137 // Try decode 32-bit instruction
138 if (Bytes.size() < 4) break;
139 const uint32_t DW = eatBytes<uint32_t>(Bytes);
140 Res = tryDecodeInst(DecoderTableVI32, MI, DW, Address);
143 Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address);
146 if (Bytes.size() < 4) break;
147 const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW;
148 Res = tryDecodeInst(DecoderTableVI64, MI, QW, Address);
151 Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address);
154 Size = Res ? (MaxInstBytesNum - Bytes.size()) : 0;
158 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
159 return getContext().getRegisterInfo()->
160 getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
164 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
165 const Twine& ErrMsg) const {
166 *CommentStream << "Error: " + ErrMsg;
168 // ToDo: add support for error operands to MCInst.h
169 // return MCOperand::createError(V);
174 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
175 return MCOperand::createReg(RegId);
179 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
180 unsigned Val) const {
181 const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
182 if (Val >= RegCl.getNumRegs())
183 return errOperand(Val, Twine(getRegClassName(RegClassID)) +
184 ": unknown register " + Twine(Val));
185 return createRegOperand(RegCl.getRegister(Val));
189 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
190 unsigned Val) const {
191 // ToDo: SI/CI have 104 SGPRs, VI - 102
192 // Valery: here we accepting as much as we can, let assembler sort it out
194 switch (SRegClassID) {
195 case AMDGPU::SGPR_32RegClassID:
196 case AMDGPU::SReg_32RegClassID: break;
197 case AMDGPU::SGPR_64RegClassID:
198 case AMDGPU::SReg_64RegClassID: shift = 1; break;
199 case AMDGPU::SReg_128RegClassID:
200 // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
202 case AMDGPU::SReg_256RegClassID:
203 // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
205 case AMDGPU::SReg_512RegClassID: shift = 2; break;
206 // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
208 default: assert(false); break;
210 if (Val % (1 << shift))
211 *CommentStream << "Warning: " << getRegClassName(SRegClassID)
212 << ": scalar reg isn't aligned " << Val;
213 return createRegOperand(SRegClassID, Val >> shift);
216 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const {
217 return decodeSrcOp(OP32, Val);
220 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const {
221 return decodeSrcOp(OP64, Val);
224 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const {
225 return createRegOperand(AMDGPU::VGPR_32RegClassID, Val);
228 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const {
229 return createRegOperand(AMDGPU::VReg_64RegClassID, Val);
232 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const {
233 return createRegOperand(AMDGPU::VReg_96RegClassID, Val);
236 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const {
237 return createRegOperand(AMDGPU::VReg_128RegClassID, Val);
240 MCOperand AMDGPUDisassembler::decodeOperand_SGPR_32(unsigned Val) const {
241 return createSRegOperand(AMDGPU::SGPR_32RegClassID, Val);
244 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const {
245 // table-gen generated disassembler doesn't care about operand types
246 // leaving only registry class so SSrc_32 operand turns into SReg_32
247 // and therefore we accept immediates and literals here as well
248 return decodeSrcOp(OP32, Val);
251 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const {
252 // see decodeOperand_SReg_32 comment
253 return decodeSrcOp(OP64, Val);
256 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const {
257 return createSRegOperand(AMDGPU::SReg_128RegClassID, Val);
260 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const {
261 return createSRegOperand(AMDGPU::SReg_256RegClassID, Val);
264 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const {
265 return createSRegOperand(AMDGPU::SReg_512RegClassID, Val);
269 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
270 // For now all literal constants are supposed to be unsigned integer
271 // ToDo: deal with signed/unsigned 64-bit integer constants
272 // ToDo: deal with float/double constants
273 if (Bytes.size() < 4)
274 return errOperand(0, "cannot read literal, inst bytes left " +
275 Twine(Bytes.size()));
276 return MCOperand::createImm(eatBytes<uint32_t>(Bytes));
279 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
280 assert(Imm >= 128 && Imm <= 208);
281 return MCOperand::createImm((Imm <= 192) ? (Imm - 128) : (192 - Imm));
284 MCOperand AMDGPUDisassembler::decodeFPImmed(bool Is32, unsigned Imm) {
285 assert(Imm >= 240 && Imm <= 248);
286 // ToDo: case 248: 1/(2*PI) - is allowed only on VI
287 // ToDo: AMDGPUInstPrinter does not support 1/(2*PI). It consider 1/(2*PI) as
291 case 240: V = 0.5f; break;
292 case 241: V = -0.5f; break;
293 case 242: V = 1.0f; break;
294 case 243: V = -1.0f; break;
295 case 244: V = 2.0f; break;
296 case 245: V = -2.0f; break;
297 case 246: V = 4.0f; break;
298 case 247: V = -4.0f; break;
299 case 248: return MCOperand::createImm(Is32 ? // 1/(2*PI)
304 return MCOperand::createImm(Is32? FloatToBits(V) : DoubleToBits(V));
307 MCOperand AMDGPUDisassembler::decodeSrcOp(bool Is32, unsigned Val) const {
308 using namespace AMDGPU;
309 assert(Val < 512); // enum9
312 return createRegOperand(Is32 ? VGPR_32RegClassID : VReg_64RegClassID,
315 return createSRegOperand(Is32 ? SGPR_32RegClassID : SGPR_64RegClassID,
318 if (Val >= 128 && Val <= 208)
319 return decodeIntImmed(Val);
321 if (Val >= 240 && Val <= 248)
322 return decodeFPImmed(Is32, Val);
325 return decodeLiteralConstant();
327 return Is32 ? decodeSpecialReg32(Val) : decodeSpecialReg64(Val);
330 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
331 using namespace AMDGPU;
333 case 102: return createRegOperand(getMCReg(FLAT_SCR_LO, STI));
334 case 103: return createRegOperand(getMCReg(FLAT_SCR_HI, STI));
335 // ToDo: no support for xnack_mask_lo/_hi register
338 case 106: return createRegOperand(VCC_LO);
339 case 107: return createRegOperand(VCC_HI);
340 // ToDo: no support for tba_lo/_hi register
343 // ToDo: no support for tma_lo/_hi register
346 // ToDo: no support for ttmp[0:11] register
359 case 124: return createRegOperand(M0);
360 case 126: return createRegOperand(EXEC_LO);
361 case 127: return createRegOperand(EXEC_HI);
362 // ToDo: no support for vccz register
364 // ToDo: no support for execz register
366 case 253: return createRegOperand(SCC);
369 return errOperand(Val, "unknown operand encoding " + Twine(Val));
372 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
373 using namespace AMDGPU;
375 case 102: return createRegOperand(getMCReg(FLAT_SCR, STI));
376 case 106: return createRegOperand(VCC);
377 case 126: return createRegOperand(EXEC);
380 return errOperand(Val, "unknown operand encoding " + Twine(Val));
383 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
384 const MCSubtargetInfo &STI,
386 return new AMDGPUDisassembler(STI, Ctx);
389 extern "C" void LLVMInitializeAMDGPUDisassembler() {
390 TargetRegistry::RegisterMCDisassembler(TheGCNTarget, createAMDGPUDisassembler);