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 "Disassembler/AMDGPUDisassembler.h"
22 #include "AMDGPURegisterInfo.h"
23 #include "SIDefines.h"
24 #include "Utils/AMDGPUBaseInfo.h"
25 #include "llvm-c/Disassembler.h"
26 #include "llvm/ADT/APInt.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/Twine.h"
29 #include "llvm/BinaryFormat/ELF.h"
30 #include "llvm/MC/MCContext.h"
31 #include "llvm/MC/MCDisassembler/MCDisassembler.h"
32 #include "llvm/MC/MCExpr.h"
33 #include "llvm/MC/MCFixedLenDisassembler.h"
34 #include "llvm/MC/MCInst.h"
35 #include "llvm/MC/MCSubtargetInfo.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Support/TargetRegistry.h"
40 #include "llvm/Support/raw_ostream.h"
51 #define DEBUG_TYPE "amdgpu-disassembler"
53 using DecodeStatus = llvm::MCDisassembler::DecodeStatus;
55 inline static MCDisassembler::DecodeStatus
56 addOperand(MCInst &Inst, const MCOperand& Opnd) {
57 Inst.addOperand(Opnd);
58 return Opnd.isValid() ?
59 MCDisassembler::Success :
60 MCDisassembler::SoftFail;
63 static int insertNamedMCOperand(MCInst &MI, const MCOperand &Op,
65 int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), NameIdx);
68 std::advance(I, OpIdx);
74 static DecodeStatus decodeSoppBrTarget(MCInst &Inst, unsigned Imm,
75 uint64_t Addr, const void *Decoder) {
76 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
78 APInt SignedOffset(18, Imm * 4, true);
79 int64_t Offset = (SignedOffset.sext(64) + 4 + Addr).getSExtValue();
81 if (DAsm->tryAddingSymbolicOperand(Inst, Offset, Addr, true, 2, 2))
82 return MCDisassembler::Success;
83 return addOperand(Inst, MCOperand::createImm(Imm));
86 #define DECODE_OPERAND(StaticDecoderName, DecoderName) \
87 static DecodeStatus StaticDecoderName(MCInst &Inst, \
90 const void *Decoder) { \
91 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder); \
92 return addOperand(Inst, DAsm->DecoderName(Imm)); \
95 #define DECODE_OPERAND_REG(RegClass) \
96 DECODE_OPERAND(Decode##RegClass##RegisterClass, decodeOperand_##RegClass)
98 DECODE_OPERAND_REG(VGPR_32)
99 DECODE_OPERAND_REG(VS_32)
100 DECODE_OPERAND_REG(VS_64)
101 DECODE_OPERAND_REG(VS_128)
103 DECODE_OPERAND_REG(VReg_64)
104 DECODE_OPERAND_REG(VReg_96)
105 DECODE_OPERAND_REG(VReg_128)
107 DECODE_OPERAND_REG(SReg_32)
108 DECODE_OPERAND_REG(SReg_32_XM0_XEXEC)
109 DECODE_OPERAND_REG(SReg_32_XEXEC_HI)
110 DECODE_OPERAND_REG(SReg_64)
111 DECODE_OPERAND_REG(SReg_64_XEXEC)
112 DECODE_OPERAND_REG(SReg_128)
113 DECODE_OPERAND_REG(SReg_256)
114 DECODE_OPERAND_REG(SReg_512)
116 static DecodeStatus decodeOperand_VSrc16(MCInst &Inst,
119 const void *Decoder) {
120 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
121 return addOperand(Inst, DAsm->decodeOperand_VSrc16(Imm));
124 static DecodeStatus decodeOperand_VSrcV216(MCInst &Inst,
127 const void *Decoder) {
128 auto DAsm = static_cast<const AMDGPUDisassembler*>(Decoder);
129 return addOperand(Inst, DAsm->decodeOperand_VSrcV216(Imm));
132 #define DECODE_SDWA(DecName) \
133 DECODE_OPERAND(decodeSDWA##DecName, decodeSDWA##DecName)
139 #include "AMDGPUGenDisassemblerTables.inc"
141 //===----------------------------------------------------------------------===//
143 //===----------------------------------------------------------------------===//
145 template <typename T> static inline T eatBytes(ArrayRef<uint8_t>& Bytes) {
146 assert(Bytes.size() >= sizeof(T));
147 const auto Res = support::endian::read<T, support::endianness::little>(Bytes.data());
148 Bytes = Bytes.slice(sizeof(T));
152 DecodeStatus AMDGPUDisassembler::tryDecodeInst(const uint8_t* Table,
155 uint64_t Address) const {
156 assert(MI.getOpcode() == 0);
157 assert(MI.getNumOperands() == 0);
160 const auto SavedBytes = Bytes;
161 if (decodeInstruction(Table, TmpInst, Inst, Address, this, STI)) {
163 return MCDisassembler::Success;
166 return MCDisassembler::Fail;
169 DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
170 ArrayRef<uint8_t> Bytes_,
173 raw_ostream &CS) const {
177 // ToDo: AMDGPUDisassembler supports only VI ISA.
178 if (!STI.getFeatureBits()[AMDGPU::FeatureGCN3Encoding])
179 report_fatal_error("Disassembly not yet supported for subtarget");
181 const unsigned MaxInstBytesNum = (std::min)((size_t)8, Bytes_.size());
182 Bytes = Bytes_.slice(0, MaxInstBytesNum);
184 DecodeStatus Res = MCDisassembler::Fail;
186 // ToDo: better to switch encoding length using some bit predicate
187 // but it is unknown yet, so try all we can
189 // Try to decode DPP and SDWA first to solve conflict with VOP1 and VOP2
191 if (Bytes.size() >= 8) {
192 const uint64_t QW = eatBytes<uint64_t>(Bytes);
193 Res = tryDecodeInst(DecoderTableDPP64, MI, QW, Address);
196 Res = tryDecodeInst(DecoderTableSDWA64, MI, QW, Address);
197 if (Res) { IsSDWA = true; break; }
199 Res = tryDecodeInst(DecoderTableSDWA964, MI, QW, Address);
200 if (Res) { IsSDWA = true; break; }
203 // Reinitialize Bytes as DPP64 could have eaten too much
204 Bytes = Bytes_.slice(0, MaxInstBytesNum);
206 // Try decode 32-bit instruction
207 if (Bytes.size() < 4) break;
208 const uint32_t DW = eatBytes<uint32_t>(Bytes);
209 Res = tryDecodeInst(DecoderTableVI32, MI, DW, Address);
212 Res = tryDecodeInst(DecoderTableAMDGPU32, MI, DW, Address);
215 Res = tryDecodeInst(DecoderTableGFX932, MI, DW, Address);
218 if (Bytes.size() < 4) break;
219 const uint64_t QW = ((uint64_t)eatBytes<uint32_t>(Bytes) << 32) | DW;
220 Res = tryDecodeInst(DecoderTableVI64, MI, QW, Address);
223 Res = tryDecodeInst(DecoderTableAMDGPU64, MI, QW, Address);
226 Res = tryDecodeInst(DecoderTableGFX964, MI, QW, Address);
229 if (Res && (MI.getOpcode() == AMDGPU::V_MAC_F32_e64_vi ||
230 MI.getOpcode() == AMDGPU::V_MAC_F32_e64_si ||
231 MI.getOpcode() == AMDGPU::V_MAC_F16_e64_vi)) {
232 // Insert dummy unused src2_modifiers.
233 insertNamedMCOperand(MI, MCOperand::createImm(0),
234 AMDGPU::OpName::src2_modifiers);
237 if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::MIMG)) {
238 Res = convertMIMGInst(MI);
242 Res = convertSDWAInst(MI);
244 Size = Res ? (MaxInstBytesNum - Bytes.size()) : 0;
248 DecodeStatus AMDGPUDisassembler::convertSDWAInst(MCInst &MI) const {
249 if (STI.getFeatureBits()[AMDGPU::FeatureGFX9]) {
250 if (AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst) != -1)
251 // VOPC - insert clamp
252 insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::clamp);
253 } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
254 int SDst = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sdst);
256 // VOPC - insert VCC register as sdst
257 insertNamedMCOperand(MI, createRegOperand(AMDGPU::VCC),
258 AMDGPU::OpName::sdst);
260 // VOP1/2 - insert omod if present in instruction
261 insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::omod);
264 return MCDisassembler::Success;
267 DecodeStatus AMDGPUDisassembler::convertMIMGInst(MCInst &MI) const {
268 int VDataIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
269 AMDGPU::OpName::vdata);
271 int DMaskIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
272 AMDGPU::OpName::dmask);
273 unsigned DMask = MI.getOperand(DMaskIdx).getImm() & 0xf;
275 return MCDisassembler::Success;
277 unsigned ChannelCount = countPopulation(DMask);
278 if (ChannelCount == 1)
279 return MCDisassembler::Success;
281 int NewOpcode = AMDGPU::getMaskedMIMGOp(*MCII, MI.getOpcode(), ChannelCount);
282 assert(NewOpcode != -1 && "could not find matching mimg channel instruction");
283 auto RCID = MCII->get(NewOpcode).OpInfo[VDataIdx].RegClass;
285 // Widen the register to the correct number of enabled channels.
286 unsigned Vdata0 = MI.getOperand(VDataIdx).getReg();
287 auto NewVdata = MRI.getMatchingSuperReg(Vdata0, AMDGPU::sub0,
288 &MRI.getRegClass(RCID));
289 if (NewVdata == AMDGPU::NoRegister) {
290 // It's possible to encode this such that the low register + enabled
291 // components exceeds the register count.
292 return MCDisassembler::Success;
295 MI.setOpcode(NewOpcode);
296 // vaddr will be always appear as a single VGPR. This will look different than
297 // how it is usually emitted because the number of register components is not
298 // in the instruction encoding.
299 MI.getOperand(VDataIdx) = MCOperand::createReg(NewVdata);
300 return MCDisassembler::Success;
303 const char* AMDGPUDisassembler::getRegClassName(unsigned RegClassID) const {
304 return getContext().getRegisterInfo()->
305 getRegClassName(&AMDGPUMCRegisterClasses[RegClassID]);
309 MCOperand AMDGPUDisassembler::errOperand(unsigned V,
310 const Twine& ErrMsg) const {
311 *CommentStream << "Error: " + ErrMsg;
313 // ToDo: add support for error operands to MCInst.h
314 // return MCOperand::createError(V);
319 MCOperand AMDGPUDisassembler::createRegOperand(unsigned int RegId) const {
320 return MCOperand::createReg(AMDGPU::getMCReg(RegId, STI));
324 MCOperand AMDGPUDisassembler::createRegOperand(unsigned RegClassID,
325 unsigned Val) const {
326 const auto& RegCl = AMDGPUMCRegisterClasses[RegClassID];
327 if (Val >= RegCl.getNumRegs())
328 return errOperand(Val, Twine(getRegClassName(RegClassID)) +
329 ": unknown register " + Twine(Val));
330 return createRegOperand(RegCl.getRegister(Val));
334 MCOperand AMDGPUDisassembler::createSRegOperand(unsigned SRegClassID,
335 unsigned Val) const {
336 // ToDo: SI/CI have 104 SGPRs, VI - 102
337 // Valery: here we accepting as much as we can, let assembler sort it out
339 switch (SRegClassID) {
340 case AMDGPU::SGPR_32RegClassID:
341 case AMDGPU::TTMP_32RegClassID:
343 case AMDGPU::SGPR_64RegClassID:
344 case AMDGPU::TTMP_64RegClassID:
347 case AMDGPU::SGPR_128RegClassID:
348 case AMDGPU::TTMP_128RegClassID:
349 // ToDo: unclear if s[100:104] is available on VI. Can we use VCC as SGPR in
351 case AMDGPU::SGPR_256RegClassID:
352 case AMDGPU::TTMP_256RegClassID:
353 // ToDo: unclear if s[96:104] is available on VI. Can we use VCC as SGPR in
355 case AMDGPU::SGPR_512RegClassID:
356 case AMDGPU::TTMP_512RegClassID:
359 // ToDo: unclear if s[88:104] is available on VI. Can we use VCC as SGPR in
362 llvm_unreachable("unhandled register class");
365 if (Val % (1 << shift)) {
366 *CommentStream << "Warning: " << getRegClassName(SRegClassID)
367 << ": scalar reg isn't aligned " << Val;
370 return createRegOperand(SRegClassID, Val >> shift);
373 MCOperand AMDGPUDisassembler::decodeOperand_VS_32(unsigned Val) const {
374 return decodeSrcOp(OPW32, Val);
377 MCOperand AMDGPUDisassembler::decodeOperand_VS_64(unsigned Val) const {
378 return decodeSrcOp(OPW64, Val);
381 MCOperand AMDGPUDisassembler::decodeOperand_VS_128(unsigned Val) const {
382 return decodeSrcOp(OPW128, Val);
385 MCOperand AMDGPUDisassembler::decodeOperand_VSrc16(unsigned Val) const {
386 return decodeSrcOp(OPW16, Val);
389 MCOperand AMDGPUDisassembler::decodeOperand_VSrcV216(unsigned Val) const {
390 return decodeSrcOp(OPWV216, Val);
393 MCOperand AMDGPUDisassembler::decodeOperand_VGPR_32(unsigned Val) const {
394 // Some instructions have operand restrictions beyond what the encoding
395 // allows. Some ordinarily VSrc_32 operands are VGPR_32, so clear the extra
399 return createRegOperand(AMDGPU::VGPR_32RegClassID, Val);
402 MCOperand AMDGPUDisassembler::decodeOperand_VReg_64(unsigned Val) const {
403 return createRegOperand(AMDGPU::VReg_64RegClassID, Val);
406 MCOperand AMDGPUDisassembler::decodeOperand_VReg_96(unsigned Val) const {
407 return createRegOperand(AMDGPU::VReg_96RegClassID, Val);
410 MCOperand AMDGPUDisassembler::decodeOperand_VReg_128(unsigned Val) const {
411 return createRegOperand(AMDGPU::VReg_128RegClassID, Val);
414 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32(unsigned Val) const {
415 // table-gen generated disassembler doesn't care about operand types
416 // leaving only registry class so SSrc_32 operand turns into SReg_32
417 // and therefore we accept immediates and literals here as well
418 return decodeSrcOp(OPW32, Val);
421 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XM0_XEXEC(
422 unsigned Val) const {
423 // SReg_32_XM0 is SReg_32 without M0 or EXEC_LO/EXEC_HI
424 return decodeOperand_SReg_32(Val);
427 MCOperand AMDGPUDisassembler::decodeOperand_SReg_32_XEXEC_HI(
428 unsigned Val) const {
429 // SReg_32_XM0 is SReg_32 without EXEC_HI
430 return decodeOperand_SReg_32(Val);
433 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64(unsigned Val) const {
434 return decodeSrcOp(OPW64, Val);
437 MCOperand AMDGPUDisassembler::decodeOperand_SReg_64_XEXEC(unsigned Val) const {
438 return decodeSrcOp(OPW64, Val);
441 MCOperand AMDGPUDisassembler::decodeOperand_SReg_128(unsigned Val) const {
442 return decodeSrcOp(OPW128, Val);
445 MCOperand AMDGPUDisassembler::decodeOperand_SReg_256(unsigned Val) const {
446 return decodeDstOp(OPW256, Val);
449 MCOperand AMDGPUDisassembler::decodeOperand_SReg_512(unsigned Val) const {
450 return decodeDstOp(OPW512, Val);
453 MCOperand AMDGPUDisassembler::decodeLiteralConstant() const {
454 // For now all literal constants are supposed to be unsigned integer
455 // ToDo: deal with signed/unsigned 64-bit integer constants
456 // ToDo: deal with float/double constants
458 if (Bytes.size() < 4) {
459 return errOperand(0, "cannot read literal, inst bytes left " +
460 Twine(Bytes.size()));
463 Literal = eatBytes<uint32_t>(Bytes);
465 return MCOperand::createImm(Literal);
468 MCOperand AMDGPUDisassembler::decodeIntImmed(unsigned Imm) {
469 using namespace AMDGPU::EncValues;
471 assert(Imm >= INLINE_INTEGER_C_MIN && Imm <= INLINE_INTEGER_C_MAX);
472 return MCOperand::createImm((Imm <= INLINE_INTEGER_C_POSITIVE_MAX) ?
473 (static_cast<int64_t>(Imm) - INLINE_INTEGER_C_MIN) :
474 (INLINE_INTEGER_C_POSITIVE_MAX - static_cast<int64_t>(Imm)));
475 // Cast prevents negative overflow.
478 static int64_t getInlineImmVal32(unsigned Imm) {
481 return FloatToBits(0.5f);
483 return FloatToBits(-0.5f);
485 return FloatToBits(1.0f);
487 return FloatToBits(-1.0f);
489 return FloatToBits(2.0f);
491 return FloatToBits(-2.0f);
493 return FloatToBits(4.0f);
495 return FloatToBits(-4.0f);
496 case 248: // 1 / (2 * PI)
499 llvm_unreachable("invalid fp inline imm");
503 static int64_t getInlineImmVal64(unsigned Imm) {
506 return DoubleToBits(0.5);
508 return DoubleToBits(-0.5);
510 return DoubleToBits(1.0);
512 return DoubleToBits(-1.0);
514 return DoubleToBits(2.0);
516 return DoubleToBits(-2.0);
518 return DoubleToBits(4.0);
520 return DoubleToBits(-4.0);
521 case 248: // 1 / (2 * PI)
522 return 0x3fc45f306dc9c882;
524 llvm_unreachable("invalid fp inline imm");
528 static int64_t getInlineImmVal16(unsigned Imm) {
546 case 248: // 1 / (2 * PI)
549 llvm_unreachable("invalid fp inline imm");
553 MCOperand AMDGPUDisassembler::decodeFPImmed(OpWidthTy Width, unsigned Imm) {
554 assert(Imm >= AMDGPU::EncValues::INLINE_FLOATING_C_MIN
555 && Imm <= AMDGPU::EncValues::INLINE_FLOATING_C_MAX);
557 // ToDo: case 248: 1/(2*PI) - is allowed only on VI
560 return MCOperand::createImm(getInlineImmVal32(Imm));
562 return MCOperand::createImm(getInlineImmVal64(Imm));
565 return MCOperand::createImm(getInlineImmVal16(Imm));
567 llvm_unreachable("implement me");
571 unsigned AMDGPUDisassembler::getVgprClassId(const OpWidthTy Width) const {
572 using namespace AMDGPU;
574 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
580 return VGPR_32RegClassID;
581 case OPW64: return VReg_64RegClassID;
582 case OPW128: return VReg_128RegClassID;
586 unsigned AMDGPUDisassembler::getSgprClassId(const OpWidthTy Width) const {
587 using namespace AMDGPU;
589 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
595 return SGPR_32RegClassID;
596 case OPW64: return SGPR_64RegClassID;
597 case OPW128: return SGPR_128RegClassID;
598 case OPW256: return SGPR_256RegClassID;
599 case OPW512: return SGPR_512RegClassID;
603 unsigned AMDGPUDisassembler::getTtmpClassId(const OpWidthTy Width) const {
604 using namespace AMDGPU;
606 assert(OPW_FIRST_ <= Width && Width < OPW_LAST_);
612 return TTMP_32RegClassID;
613 case OPW64: return TTMP_64RegClassID;
614 case OPW128: return TTMP_128RegClassID;
615 case OPW256: return TTMP_256RegClassID;
616 case OPW512: return TTMP_512RegClassID;
620 int AMDGPUDisassembler::getTTmpIdx(unsigned Val) const {
621 using namespace AMDGPU::EncValues;
623 unsigned TTmpMin = isGFX9() ? TTMP_GFX9_MIN : TTMP_VI_MIN;
624 unsigned TTmpMax = isGFX9() ? TTMP_GFX9_MAX : TTMP_VI_MAX;
626 return (TTmpMin <= Val && Val <= TTmpMax)? Val - TTmpMin : -1;
629 MCOperand AMDGPUDisassembler::decodeSrcOp(const OpWidthTy Width, unsigned Val) const {
630 using namespace AMDGPU::EncValues;
632 assert(Val < 512); // enum9
634 if (VGPR_MIN <= Val && Val <= VGPR_MAX) {
635 return createRegOperand(getVgprClassId(Width), Val - VGPR_MIN);
637 if (Val <= SGPR_MAX) {
638 assert(SGPR_MIN == 0); // "SGPR_MIN <= Val" is always true and causes compilation warning.
639 return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
642 int TTmpIdx = getTTmpIdx(Val);
644 return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
647 if (INLINE_INTEGER_C_MIN <= Val && Val <= INLINE_INTEGER_C_MAX)
648 return decodeIntImmed(Val);
650 if (INLINE_FLOATING_C_MIN <= Val && Val <= INLINE_FLOATING_C_MAX)
651 return decodeFPImmed(Width, Val);
653 if (Val == LITERAL_CONST)
654 return decodeLiteralConstant();
660 return decodeSpecialReg32(Val);
662 return decodeSpecialReg64(Val);
664 llvm_unreachable("unexpected immediate type");
668 MCOperand AMDGPUDisassembler::decodeDstOp(const OpWidthTy Width, unsigned Val) const {
669 using namespace AMDGPU::EncValues;
672 assert(Width == OPW256 || Width == OPW512);
674 if (Val <= SGPR_MAX) {
675 assert(SGPR_MIN == 0); // "SGPR_MIN <= Val" is always true and causes compilation warning.
676 return createSRegOperand(getSgprClassId(Width), Val - SGPR_MIN);
679 int TTmpIdx = getTTmpIdx(Val);
681 return createSRegOperand(getTtmpClassId(Width), TTmpIdx);
684 llvm_unreachable("unknown dst register");
687 MCOperand AMDGPUDisassembler::decodeSpecialReg32(unsigned Val) const {
688 using namespace AMDGPU;
691 case 102: return createRegOperand(FLAT_SCR_LO);
692 case 103: return createRegOperand(FLAT_SCR_HI);
693 // ToDo: no support for xnack_mask_lo/_hi register
696 case 106: return createRegOperand(VCC_LO);
697 case 107: return createRegOperand(VCC_HI);
698 case 108: assert(!isGFX9()); return createRegOperand(TBA_LO);
699 case 109: assert(!isGFX9()); return createRegOperand(TBA_HI);
700 case 110: assert(!isGFX9()); return createRegOperand(TMA_LO);
701 case 111: assert(!isGFX9()); return createRegOperand(TMA_HI);
702 case 124: return createRegOperand(M0);
703 case 126: return createRegOperand(EXEC_LO);
704 case 127: return createRegOperand(EXEC_HI);
705 case 235: return createRegOperand(SRC_SHARED_BASE);
706 case 236: return createRegOperand(SRC_SHARED_LIMIT);
707 case 237: return createRegOperand(SRC_PRIVATE_BASE);
708 case 238: return createRegOperand(SRC_PRIVATE_LIMIT);
709 // TODO: SRC_POPS_EXITING_WAVE_ID
710 // ToDo: no support for vccz register
712 // ToDo: no support for execz register
714 case 253: return createRegOperand(SCC);
717 return errOperand(Val, "unknown operand encoding " + Twine(Val));
720 MCOperand AMDGPUDisassembler::decodeSpecialReg64(unsigned Val) const {
721 using namespace AMDGPU;
724 case 102: return createRegOperand(FLAT_SCR);
725 case 106: return createRegOperand(VCC);
726 case 108: assert(!isGFX9()); return createRegOperand(TBA);
727 case 110: assert(!isGFX9()); return createRegOperand(TMA);
728 case 126: return createRegOperand(EXEC);
731 return errOperand(Val, "unknown operand encoding " + Twine(Val));
734 MCOperand AMDGPUDisassembler::decodeSDWASrc(const OpWidthTy Width,
735 unsigned Val) const {
736 using namespace AMDGPU::SDWA;
738 if (STI.getFeatureBits()[AMDGPU::FeatureGFX9]) {
739 // XXX: static_cast<int> is needed to avoid stupid warning:
740 // compare with unsigned is always true
741 if (SDWA9EncValues::SRC_VGPR_MIN <= static_cast<int>(Val) &&
742 Val <= SDWA9EncValues::SRC_VGPR_MAX) {
743 return createRegOperand(getVgprClassId(Width),
744 Val - SDWA9EncValues::SRC_VGPR_MIN);
746 if (SDWA9EncValues::SRC_SGPR_MIN <= Val &&
747 Val <= SDWA9EncValues::SRC_SGPR_MAX) {
748 return createSRegOperand(getSgprClassId(Width),
749 Val - SDWA9EncValues::SRC_SGPR_MIN);
751 if (SDWA9EncValues::SRC_TTMP_MIN <= Val &&
752 Val <= SDWA9EncValues::SRC_TTMP_MAX) {
753 return createSRegOperand(getTtmpClassId(Width),
754 Val - SDWA9EncValues::SRC_TTMP_MIN);
757 return decodeSpecialReg32(Val - SDWA9EncValues::SRC_SGPR_MIN);
758 } else if (STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands]) {
759 return createRegOperand(getVgprClassId(Width), Val);
761 llvm_unreachable("unsupported target");
764 MCOperand AMDGPUDisassembler::decodeSDWASrc16(unsigned Val) const {
765 return decodeSDWASrc(OPW16, Val);
768 MCOperand AMDGPUDisassembler::decodeSDWASrc32(unsigned Val) const {
769 return decodeSDWASrc(OPW32, Val);
772 MCOperand AMDGPUDisassembler::decodeSDWAVopcDst(unsigned Val) const {
773 using namespace AMDGPU::SDWA;
775 assert(STI.getFeatureBits()[AMDGPU::FeatureGFX9] &&
776 "SDWAVopcDst should be present only on GFX9");
777 if (Val & SDWA9EncValues::VOPC_DST_VCC_MASK) {
778 Val &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
780 int TTmpIdx = getTTmpIdx(Val);
782 return createSRegOperand(getTtmpClassId(OPW64), TTmpIdx);
783 } else if (Val > AMDGPU::EncValues::SGPR_MAX) {
784 return decodeSpecialReg64(Val);
786 return createSRegOperand(getSgprClassId(OPW64), Val);
789 return createRegOperand(AMDGPU::VCC);
793 bool AMDGPUDisassembler::isVI() const {
794 return STI.getFeatureBits()[AMDGPU::FeatureVolcanicIslands];
797 bool AMDGPUDisassembler::isGFX9() const {
798 return STI.getFeatureBits()[AMDGPU::FeatureGFX9];
801 //===----------------------------------------------------------------------===//
803 //===----------------------------------------------------------------------===//
805 // Try to find symbol name for specified label
806 bool AMDGPUSymbolizer::tryAddingSymbolicOperand(MCInst &Inst,
807 raw_ostream &/*cStream*/, int64_t Value,
808 uint64_t /*Address*/, bool IsBranch,
809 uint64_t /*Offset*/, uint64_t /*InstSize*/) {
810 using SymbolInfoTy = std::tuple<uint64_t, StringRef, uint8_t>;
811 using SectionSymbolsTy = std::vector<SymbolInfoTy>;
817 auto *Symbols = static_cast<SectionSymbolsTy *>(DisInfo);
818 auto Result = std::find_if(Symbols->begin(), Symbols->end(),
819 [Value](const SymbolInfoTy& Val) {
820 return std::get<0>(Val) == static_cast<uint64_t>(Value)
821 && std::get<2>(Val) == ELF::STT_NOTYPE;
823 if (Result != Symbols->end()) {
824 auto *Sym = Ctx.getOrCreateSymbol(std::get<1>(*Result));
825 const auto *Add = MCSymbolRefExpr::create(Sym, Ctx);
826 Inst.addOperand(MCOperand::createExpr(Add));
832 void AMDGPUSymbolizer::tryAddingPcLoadReferenceComment(raw_ostream &cStream,
835 llvm_unreachable("unimplemented");
838 //===----------------------------------------------------------------------===//
840 //===----------------------------------------------------------------------===//
842 static MCSymbolizer *createAMDGPUSymbolizer(const Triple &/*TT*/,
843 LLVMOpInfoCallback /*GetOpInfo*/,
844 LLVMSymbolLookupCallback /*SymbolLookUp*/,
847 std::unique_ptr<MCRelocationInfo> &&RelInfo) {
848 return new AMDGPUSymbolizer(*Ctx, std::move(RelInfo), DisInfo);
851 static MCDisassembler *createAMDGPUDisassembler(const Target &T,
852 const MCSubtargetInfo &STI,
854 return new AMDGPUDisassembler(STI, Ctx, T.createMCInstrInfo());
857 extern "C" void LLVMInitializeAMDGPUDisassembler() {
858 TargetRegistry::RegisterMCDisassembler(getTheGCNTarget(),
859 createAMDGPUDisassembler);
860 TargetRegistry::RegisterMCSymbolizer(getTheGCNTarget(),
861 createAMDGPUSymbolizer);