1 //===-- AMDGPUInstPrinter.cpp - AMDGPU MC Inst -> ASM ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
9 //===----------------------------------------------------------------------===//
11 #include "AMDGPUInstPrinter.h"
12 #include "SIDefines.h"
13 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
14 #include "Utils/AMDGPUAsmUtils.h"
15 #include "Utils/AMDGPUBaseInfo.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCInst.h"
18 #include "llvm/MC/MCInstrDesc.h"
19 #include "llvm/MC/MCInstrInfo.h"
20 #include "llvm/MC/MCRegisterInfo.h"
21 #include "llvm/MC/MCSubtargetInfo.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/MathExtras.h"
24 #include "llvm/Support/raw_ostream.h"
28 using namespace llvm::AMDGPU;
30 void AMDGPUInstPrinter::printInst(const MCInst *MI, raw_ostream &OS,
31 StringRef Annot, const MCSubtargetInfo &STI) {
33 printInstruction(MI, STI, OS);
34 printAnnotation(OS, Annot);
37 void AMDGPUInstPrinter::printU4ImmOperand(const MCInst *MI, unsigned OpNo,
38 const MCSubtargetInfo &STI,
40 O << formatHex(MI->getOperand(OpNo).getImm() & 0xf);
43 void AMDGPUInstPrinter::printU8ImmOperand(const MCInst *MI, unsigned OpNo,
45 O << formatHex(MI->getOperand(OpNo).getImm() & 0xff);
48 void AMDGPUInstPrinter::printU16ImmOperand(const MCInst *MI, unsigned OpNo,
49 const MCSubtargetInfo &STI,
51 // It's possible to end up with a 32-bit literal used with a 16-bit operand
52 // with ignored high bits. Print as 32-bit anyway in that case.
53 int64_t Imm = MI->getOperand(OpNo).getImm();
54 if (isInt<16>(Imm) || isUInt<16>(Imm))
55 O << formatHex(static_cast<uint64_t>(Imm & 0xffff));
57 printU32ImmOperand(MI, OpNo, STI, O);
60 void AMDGPUInstPrinter::printU4ImmDecOperand(const MCInst *MI, unsigned OpNo,
62 O << formatDec(MI->getOperand(OpNo).getImm() & 0xf);
65 void AMDGPUInstPrinter::printU8ImmDecOperand(const MCInst *MI, unsigned OpNo,
67 O << formatDec(MI->getOperand(OpNo).getImm() & 0xff);
70 void AMDGPUInstPrinter::printU16ImmDecOperand(const MCInst *MI, unsigned OpNo,
72 O << formatDec(MI->getOperand(OpNo).getImm() & 0xffff);
75 void AMDGPUInstPrinter::printU32ImmOperand(const MCInst *MI, unsigned OpNo,
76 const MCSubtargetInfo &STI,
78 O << formatHex(MI->getOperand(OpNo).getImm() & 0xffffffff);
81 void AMDGPUInstPrinter::printNamedBit(const MCInst *MI, unsigned OpNo,
82 raw_ostream &O, StringRef BitName) {
83 if (MI->getOperand(OpNo).getImm()) {
88 void AMDGPUInstPrinter::printOffen(const MCInst *MI, unsigned OpNo,
90 printNamedBit(MI, OpNo, O, "offen");
93 void AMDGPUInstPrinter::printIdxen(const MCInst *MI, unsigned OpNo,
95 printNamedBit(MI, OpNo, O, "idxen");
98 void AMDGPUInstPrinter::printAddr64(const MCInst *MI, unsigned OpNo,
100 printNamedBit(MI, OpNo, O, "addr64");
103 void AMDGPUInstPrinter::printMBUFOffset(const MCInst *MI, unsigned OpNo,
105 if (MI->getOperand(OpNo).getImm()) {
107 printU16ImmDecOperand(MI, OpNo, O);
111 void AMDGPUInstPrinter::printOffset(const MCInst *MI, unsigned OpNo,
112 const MCSubtargetInfo &STI,
114 uint16_t Imm = MI->getOperand(OpNo).getImm();
116 O << ((OpNo == 0)? "offset:" : " offset:");
117 printU16ImmDecOperand(MI, OpNo, O);
121 void AMDGPUInstPrinter::printOffset0(const MCInst *MI, unsigned OpNo,
122 const MCSubtargetInfo &STI,
124 if (MI->getOperand(OpNo).getImm()) {
126 printU8ImmDecOperand(MI, OpNo, O);
130 void AMDGPUInstPrinter::printOffset1(const MCInst *MI, unsigned OpNo,
131 const MCSubtargetInfo &STI,
133 if (MI->getOperand(OpNo).getImm()) {
135 printU8ImmDecOperand(MI, OpNo, O);
139 void AMDGPUInstPrinter::printSMRDOffset8(const MCInst *MI, unsigned OpNo,
140 const MCSubtargetInfo &STI,
142 printU32ImmOperand(MI, OpNo, STI, O);
145 void AMDGPUInstPrinter::printSMRDOffset20(const MCInst *MI, unsigned OpNo,
146 const MCSubtargetInfo &STI,
148 printU32ImmOperand(MI, OpNo, STI, O);
151 void AMDGPUInstPrinter::printSMRDLiteralOffset(const MCInst *MI, unsigned OpNo,
152 const MCSubtargetInfo &STI,
154 printU32ImmOperand(MI, OpNo, STI, O);
157 void AMDGPUInstPrinter::printGDS(const MCInst *MI, unsigned OpNo,
158 const MCSubtargetInfo &STI, raw_ostream &O) {
159 printNamedBit(MI, OpNo, O, "gds");
162 void AMDGPUInstPrinter::printGLC(const MCInst *MI, unsigned OpNo,
163 const MCSubtargetInfo &STI, raw_ostream &O) {
164 printNamedBit(MI, OpNo, O, "glc");
167 void AMDGPUInstPrinter::printSLC(const MCInst *MI, unsigned OpNo,
168 const MCSubtargetInfo &STI, raw_ostream &O) {
169 printNamedBit(MI, OpNo, O, "slc");
172 void AMDGPUInstPrinter::printTFE(const MCInst *MI, unsigned OpNo,
173 const MCSubtargetInfo &STI, raw_ostream &O) {
174 printNamedBit(MI, OpNo, O, "tfe");
177 void AMDGPUInstPrinter::printDMask(const MCInst *MI, unsigned OpNo,
178 const MCSubtargetInfo &STI, raw_ostream &O) {
179 if (MI->getOperand(OpNo).getImm()) {
181 printU16ImmOperand(MI, OpNo, STI, O);
185 void AMDGPUInstPrinter::printUNorm(const MCInst *MI, unsigned OpNo,
186 const MCSubtargetInfo &STI, raw_ostream &O) {
187 printNamedBit(MI, OpNo, O, "unorm");
190 void AMDGPUInstPrinter::printDA(const MCInst *MI, unsigned OpNo,
191 const MCSubtargetInfo &STI, raw_ostream &O) {
192 printNamedBit(MI, OpNo, O, "da");
195 void AMDGPUInstPrinter::printR128(const MCInst *MI, unsigned OpNo,
196 const MCSubtargetInfo &STI, raw_ostream &O) {
197 printNamedBit(MI, OpNo, O, "r128");
200 void AMDGPUInstPrinter::printLWE(const MCInst *MI, unsigned OpNo,
201 const MCSubtargetInfo &STI, raw_ostream &O) {
202 printNamedBit(MI, OpNo, O, "lwe");
205 void AMDGPUInstPrinter::printExpCompr(const MCInst *MI, unsigned OpNo,
206 const MCSubtargetInfo &STI,
208 if (MI->getOperand(OpNo).getImm())
212 void AMDGPUInstPrinter::printExpVM(const MCInst *MI, unsigned OpNo,
213 const MCSubtargetInfo &STI,
215 if (MI->getOperand(OpNo).getImm())
219 void AMDGPUInstPrinter::printRegOperand(unsigned RegNo, raw_ostream &O,
220 const MCRegisterInfo &MRI) {
234 case AMDGPU::FLAT_SCR:
255 case AMDGPU::EXEC_LO:
258 case AMDGPU::EXEC_HI:
261 case AMDGPU::FLAT_SCR_LO:
262 O << "flat_scratch_lo";
264 case AMDGPU::FLAT_SCR_HI:
265 O << "flat_scratch_hi";
271 // The low 8 bits of the encoding value is the register index, for both VGPRs
273 unsigned RegIdx = MRI.getEncodingValue(RegNo) & ((1 << 8) - 1);
276 if (MRI.getRegClass(AMDGPU::VGPR_32RegClassID).contains(RegNo)) {
279 } else if (MRI.getRegClass(AMDGPU::SGPR_32RegClassID).contains(RegNo)) {
282 } else if (MRI.getRegClass(AMDGPU::VReg_64RegClassID).contains(RegNo)) {
285 } else if (MRI.getRegClass(AMDGPU::SGPR_64RegClassID).contains(RegNo)) {
288 } else if (MRI.getRegClass(AMDGPU::VReg_128RegClassID).contains(RegNo)) {
291 } else if (MRI.getRegClass(AMDGPU::SGPR_128RegClassID).contains(RegNo)) {
294 } else if (MRI.getRegClass(AMDGPU::VReg_96RegClassID).contains(RegNo)) {
297 } else if (MRI.getRegClass(AMDGPU::VReg_256RegClassID).contains(RegNo)) {
300 } else if (MRI.getRegClass(AMDGPU::SReg_256RegClassID).contains(RegNo)) {
303 } else if (MRI.getRegClass(AMDGPU::VReg_512RegClassID).contains(RegNo)) {
306 } else if (MRI.getRegClass(AMDGPU::SReg_512RegClassID).contains(RegNo)) {
309 } else if (MRI.getRegClass(AMDGPU::TTMP_64RegClassID).contains(RegNo)) {
312 // Trap temps start at offset 112. TODO: Get this from tablegen.
314 } else if (MRI.getRegClass(AMDGPU::TTMP_128RegClassID).contains(RegNo)) {
317 // Trap temps start at offset 112. TODO: Get this from tablegen.
320 O << getRegisterName(RegNo);
329 O << '[' << RegIdx << ':' << (RegIdx + NumRegs - 1) << ']';
332 void AMDGPUInstPrinter::printVOPDst(const MCInst *MI, unsigned OpNo,
333 const MCSubtargetInfo &STI, raw_ostream &O) {
334 if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3)
336 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::DPP)
338 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::SDWA)
343 printOperand(MI, OpNo, STI, O);
346 void AMDGPUInstPrinter::printImmediate16(uint32_t Imm,
347 const MCSubtargetInfo &STI,
349 int16_t SImm = static_cast<int16_t>(Imm);
350 if (SImm >= -16 && SImm <= 64) {
357 else if (Imm == 0xBC00)
359 else if (Imm == 0x3800)
361 else if (Imm == 0xB800)
363 else if (Imm == 0x4000)
365 else if (Imm == 0xC000)
367 else if (Imm == 0x4400)
369 else if (Imm == 0xC400)
371 else if (Imm == 0x3118) {
372 assert(STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]);
375 O << formatHex(static_cast<uint64_t>(Imm));
378 void AMDGPUInstPrinter::printImmediateV216(uint32_t Imm,
379 const MCSubtargetInfo &STI,
381 uint16_t Lo16 = static_cast<uint16_t>(Imm);
382 assert(Lo16 == static_cast<uint16_t>(Imm >> 16));
383 printImmediate16(Lo16, STI, O);
386 void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
387 const MCSubtargetInfo &STI,
389 int32_t SImm = static_cast<int32_t>(Imm);
390 if (SImm >= -16 && SImm <= 64) {
395 if (Imm == FloatToBits(0.0f))
397 else if (Imm == FloatToBits(1.0f))
399 else if (Imm == FloatToBits(-1.0f))
401 else if (Imm == FloatToBits(0.5f))
403 else if (Imm == FloatToBits(-0.5f))
405 else if (Imm == FloatToBits(2.0f))
407 else if (Imm == FloatToBits(-2.0f))
409 else if (Imm == FloatToBits(4.0f))
411 else if (Imm == FloatToBits(-4.0f))
413 else if (Imm == 0x3e22f983 &&
414 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
417 O << formatHex(static_cast<uint64_t>(Imm));
420 void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
421 const MCSubtargetInfo &STI,
423 int64_t SImm = static_cast<int64_t>(Imm);
424 if (SImm >= -16 && SImm <= 64) {
429 if (Imm == DoubleToBits(0.0))
431 else if (Imm == DoubleToBits(1.0))
433 else if (Imm == DoubleToBits(-1.0))
435 else if (Imm == DoubleToBits(0.5))
437 else if (Imm == DoubleToBits(-0.5))
439 else if (Imm == DoubleToBits(2.0))
441 else if (Imm == DoubleToBits(-2.0))
443 else if (Imm == DoubleToBits(4.0))
445 else if (Imm == DoubleToBits(-4.0))
447 else if (Imm == 0x3fc45f306dc9c882 &&
448 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
451 assert(isUInt<32>(Imm) || Imm == 0x3fc45f306dc9c882);
453 // In rare situations, we will have a 32-bit literal in a 64-bit
454 // operand. This is technically allowed for the encoding of s_mov_b64.
455 O << formatHex(static_cast<uint64_t>(Imm));
459 void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
460 const MCSubtargetInfo &STI,
462 if (OpNo >= MI->getNumOperands()) {
463 O << "/*Missing OP" << OpNo << "*/";
467 const MCOperand &Op = MI->getOperand(OpNo);
469 switch (Op.getReg()) {
470 // This is the default predicate state, so we don't need to print it.
471 case AMDGPU::PRED_SEL_OFF:
475 printRegOperand(Op.getReg(), O, MRI);
478 } else if (Op.isImm()) {
479 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
480 switch (Desc.OpInfo[OpNo].OperandType) {
481 case AMDGPU::OPERAND_REG_IMM_INT32:
482 case AMDGPU::OPERAND_REG_IMM_FP32:
483 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
484 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
485 case MCOI::OPERAND_IMMEDIATE:
486 printImmediate32(Op.getImm(), STI, O);
488 case AMDGPU::OPERAND_REG_IMM_INT64:
489 case AMDGPU::OPERAND_REG_IMM_FP64:
490 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
491 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
492 printImmediate64(Op.getImm(), STI, O);
494 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
495 case AMDGPU::OPERAND_REG_INLINE_C_FP16:
496 case AMDGPU::OPERAND_REG_IMM_INT16:
497 case AMDGPU::OPERAND_REG_IMM_FP16:
498 printImmediate16(Op.getImm(), STI, O);
500 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
501 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
502 printImmediateV216(Op.getImm(), STI, O);
504 case MCOI::OPERAND_UNKNOWN:
505 case MCOI::OPERAND_PCREL:
506 O << formatDec(Op.getImm());
508 case MCOI::OPERAND_REGISTER:
509 // FIXME: This should be removed and handled somewhere else. Seems to come
510 // from a disassembler bug.
511 O << "/*invalid immediate*/";
514 // We hit this for the immediate instruction bits that don't yet have a
516 llvm_unreachable("unexpected immediate operand type");
518 } else if (Op.isFPImm()) {
519 // We special case 0.0 because otherwise it will be printed as an integer.
520 if (Op.getFPImm() == 0.0)
523 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
524 int RCID = Desc.OpInfo[OpNo].RegClass;
525 unsigned RCBits = AMDGPU::getRegBitWidth(MRI.getRegClass(RCID));
527 printImmediate32(FloatToBits(Op.getFPImm()), STI, O);
528 else if (RCBits == 64)
529 printImmediate64(DoubleToBits(Op.getFPImm()), STI, O);
531 llvm_unreachable("Invalid register class size");
533 } else if (Op.isExpr()) {
534 const MCExpr *Exp = Op.getExpr();
541 void AMDGPUInstPrinter::printOperandAndFPInputMods(const MCInst *MI,
543 const MCSubtargetInfo &STI,
545 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
547 // Use 'neg(...)' instead of '-' to avoid ambiguity.
548 // This is important for integer literals because
549 // -1 is not the same value as neg(1).
550 bool NegMnemo = false;
552 if (InputModifiers & SISrcMods::NEG) {
553 if (OpNo + 1 < MI->getNumOperands() &&
554 (InputModifiers & SISrcMods::ABS) == 0) {
555 const MCOperand &Op = MI->getOperand(OpNo + 1);
556 NegMnemo = Op.isImm() || Op.isFPImm();
565 if (InputModifiers & SISrcMods::ABS)
567 printOperand(MI, OpNo + 1, STI, O);
568 if (InputModifiers & SISrcMods::ABS)
576 void AMDGPUInstPrinter::printOperandAndIntInputMods(const MCInst *MI,
578 const MCSubtargetInfo &STI,
580 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
581 if (InputModifiers & SISrcMods::SEXT)
583 printOperand(MI, OpNo + 1, STI, O);
584 if (InputModifiers & SISrcMods::SEXT)
588 void AMDGPUInstPrinter::printDPPCtrl(const MCInst *MI, unsigned OpNo,
589 const MCSubtargetInfo &STI,
591 unsigned Imm = MI->getOperand(OpNo).getImm();
594 O << formatDec(Imm & 0x3) << ',';
595 O << formatDec((Imm & 0xc) >> 2) << ',';
596 O << formatDec((Imm & 0x30) >> 4) << ',';
597 O << formatDec((Imm & 0xc0) >> 6) << ']';
598 } else if ((Imm >= 0x101) && (Imm <= 0x10f)) {
600 printU4ImmDecOperand(MI, OpNo, O);
601 } else if ((Imm >= 0x111) && (Imm <= 0x11f)) {
603 printU4ImmDecOperand(MI, OpNo, O);
604 } else if ((Imm >= 0x121) && (Imm <= 0x12f)) {
606 printU4ImmDecOperand(MI, OpNo, O);
607 } else if (Imm == 0x130) {
609 } else if (Imm == 0x134) {
611 } else if (Imm == 0x138) {
613 } else if (Imm == 0x13c) {
615 } else if (Imm == 0x140) {
617 } else if (Imm == 0x141) {
618 O << " row_half_mirror";
619 } else if (Imm == 0x142) {
620 O << " row_bcast:15";
621 } else if (Imm == 0x143) {
622 O << " row_bcast:31";
624 llvm_unreachable("Invalid dpp_ctrl value");
628 void AMDGPUInstPrinter::printRowMask(const MCInst *MI, unsigned OpNo,
629 const MCSubtargetInfo &STI,
632 printU4ImmOperand(MI, OpNo, STI, O);
635 void AMDGPUInstPrinter::printBankMask(const MCInst *MI, unsigned OpNo,
636 const MCSubtargetInfo &STI,
639 printU4ImmOperand(MI, OpNo, STI, O);
642 void AMDGPUInstPrinter::printBoundCtrl(const MCInst *MI, unsigned OpNo,
643 const MCSubtargetInfo &STI,
645 unsigned Imm = MI->getOperand(OpNo).getImm();
647 O << " bound_ctrl:0"; // XXX - this syntax is used in sp3
651 void AMDGPUInstPrinter::printSDWASel(const MCInst *MI, unsigned OpNo,
653 using namespace llvm::AMDGPU::SDWA;
655 unsigned Imm = MI->getOperand(OpNo).getImm();
657 case SdwaSel::BYTE_0: O << "BYTE_0"; break;
658 case SdwaSel::BYTE_1: O << "BYTE_1"; break;
659 case SdwaSel::BYTE_2: O << "BYTE_2"; break;
660 case SdwaSel::BYTE_3: O << "BYTE_3"; break;
661 case SdwaSel::WORD_0: O << "WORD_0"; break;
662 case SdwaSel::WORD_1: O << "WORD_1"; break;
663 case SdwaSel::DWORD: O << "DWORD"; break;
664 default: llvm_unreachable("Invalid SDWA data select operand");
668 void AMDGPUInstPrinter::printSDWADstSel(const MCInst *MI, unsigned OpNo,
669 const MCSubtargetInfo &STI,
672 printSDWASel(MI, OpNo, O);
675 void AMDGPUInstPrinter::printSDWASrc0Sel(const MCInst *MI, unsigned OpNo,
676 const MCSubtargetInfo &STI,
679 printSDWASel(MI, OpNo, O);
682 void AMDGPUInstPrinter::printSDWASrc1Sel(const MCInst *MI, unsigned OpNo,
683 const MCSubtargetInfo &STI,
686 printSDWASel(MI, OpNo, O);
689 void AMDGPUInstPrinter::printSDWADstUnused(const MCInst *MI, unsigned OpNo,
690 const MCSubtargetInfo &STI,
692 using namespace llvm::AMDGPU::SDWA;
695 unsigned Imm = MI->getOperand(OpNo).getImm();
697 case DstUnused::UNUSED_PAD: O << "UNUSED_PAD"; break;
698 case DstUnused::UNUSED_SEXT: O << "UNUSED_SEXT"; break;
699 case DstUnused::UNUSED_PRESERVE: O << "UNUSED_PRESERVE"; break;
700 default: llvm_unreachable("Invalid SDWA dest_unused operand");
704 template <unsigned N>
705 void AMDGPUInstPrinter::printExpSrcN(const MCInst *MI, unsigned OpNo,
706 const MCSubtargetInfo &STI,
708 unsigned Opc = MI->getOpcode();
709 int EnIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::en);
710 unsigned En = MI->getOperand(EnIdx).getImm();
712 int ComprIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::compr);
714 // If compr is set, print as src0, src0, src1, src1
715 if (MI->getOperand(ComprIdx).getImm()) {
716 if (N == 1 || N == 2)
723 printRegOperand(MI->getOperand(OpNo).getReg(), O, MRI);
728 void AMDGPUInstPrinter::printExpSrc0(const MCInst *MI, unsigned OpNo,
729 const MCSubtargetInfo &STI,
731 printExpSrcN<0>(MI, OpNo, STI, O);
734 void AMDGPUInstPrinter::printExpSrc1(const MCInst *MI, unsigned OpNo,
735 const MCSubtargetInfo &STI,
737 printExpSrcN<1>(MI, OpNo, STI, O);
740 void AMDGPUInstPrinter::printExpSrc2(const MCInst *MI, unsigned OpNo,
741 const MCSubtargetInfo &STI,
743 printExpSrcN<2>(MI, OpNo, STI, O);
746 void AMDGPUInstPrinter::printExpSrc3(const MCInst *MI, unsigned OpNo,
747 const MCSubtargetInfo &STI,
749 printExpSrcN<3>(MI, OpNo, STI, O);
752 void AMDGPUInstPrinter::printExpTgt(const MCInst *MI, unsigned OpNo,
753 const MCSubtargetInfo &STI,
755 // This is really a 6 bit field.
756 uint32_t Tgt = MI->getOperand(OpNo).getImm() & ((1 << 6) - 1);
764 else if (Tgt >= 12 && Tgt <= 15)
765 O << " pos" << Tgt - 12;
766 else if (Tgt >= 32 && Tgt <= 63)
767 O << " param" << Tgt - 32;
769 // Reserved values 10, 11
770 O << " invalid_target_" << Tgt;
774 static bool allOpsDefaultValue(const int* Ops, int NumOps, int Mod) {
775 int DefaultValue = (Mod == SISrcMods::OP_SEL_1);
777 for (int I = 0; I < NumOps; ++I) {
778 if (!!(Ops[I] & Mod) != DefaultValue)
785 static void printPackedModifier(const MCInst *MI, StringRef Name, unsigned Mod,
787 unsigned Opc = MI->getOpcode();
791 for (int OpName : { AMDGPU::OpName::src0_modifiers,
792 AMDGPU::OpName::src1_modifiers,
793 AMDGPU::OpName::src2_modifiers }) {
794 int Idx = AMDGPU::getNamedOperandIdx(Opc, OpName);
798 Ops[NumOps++] = MI->getOperand(Idx).getImm();
801 if (allOpsDefaultValue(Ops, NumOps, Mod))
805 for (int I = 0; I < NumOps; ++I) {
809 O << !!(Ops[I] & Mod);
815 void AMDGPUInstPrinter::printOpSel(const MCInst *MI, unsigned,
816 const MCSubtargetInfo &STI,
818 printPackedModifier(MI, " op_sel:[", SISrcMods::OP_SEL_0, O);
821 void AMDGPUInstPrinter::printOpSelHi(const MCInst *MI, unsigned OpNo,
822 const MCSubtargetInfo &STI,
824 printPackedModifier(MI, " op_sel_hi:[", SISrcMods::OP_SEL_1, O);
827 void AMDGPUInstPrinter::printNegLo(const MCInst *MI, unsigned OpNo,
828 const MCSubtargetInfo &STI,
830 printPackedModifier(MI, " neg_lo:[", SISrcMods::NEG, O);
833 void AMDGPUInstPrinter::printNegHi(const MCInst *MI, unsigned OpNo,
834 const MCSubtargetInfo &STI,
836 printPackedModifier(MI, " neg_hi:[", SISrcMods::NEG_HI, O);
839 void AMDGPUInstPrinter::printInterpSlot(const MCInst *MI, unsigned OpNum,
840 const MCSubtargetInfo &STI,
842 unsigned Imm = MI->getOperand(OpNum).getImm();
854 O << "invalid_param_" << Imm;
858 void AMDGPUInstPrinter::printInterpAttr(const MCInst *MI, unsigned OpNum,
859 const MCSubtargetInfo &STI,
861 unsigned Attr = MI->getOperand(OpNum).getImm();
865 void AMDGPUInstPrinter::printInterpAttrChan(const MCInst *MI, unsigned OpNum,
866 const MCSubtargetInfo &STI,
868 unsigned Chan = MI->getOperand(OpNum).getImm();
869 O << '.' << "xyzw"[Chan & 0x3];
872 void AMDGPUInstPrinter::printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
873 const MCSubtargetInfo &STI,
875 unsigned Val = MI->getOperand(OpNo).getImm();
881 if (Val & VGPRIndexMode::DST_ENABLE)
884 if (Val & VGPRIndexMode::SRC0_ENABLE)
887 if (Val & VGPRIndexMode::SRC1_ENABLE)
890 if (Val & VGPRIndexMode::SRC2_ENABLE)
894 void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
895 const MCSubtargetInfo &STI,
897 printOperand(MI, OpNo, STI, O);
899 printOperand(MI, OpNo + 1, STI, O);
902 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
903 raw_ostream &O, StringRef Asm,
905 const MCOperand &Op = MI->getOperand(OpNo);
907 if (Op.getImm() == 1) {
914 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
915 raw_ostream &O, char Asm) {
916 const MCOperand &Op = MI->getOperand(OpNo);
918 if (Op.getImm() == 1)
922 void AMDGPUInstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
923 const MCSubtargetInfo &STI, raw_ostream &O) {
924 printIfSet(MI, OpNo, O, '|');
927 void AMDGPUInstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
928 const MCSubtargetInfo &STI, raw_ostream &O) {
929 printIfSet(MI, OpNo, O, "_SAT");
932 void AMDGPUInstPrinter::printClampSI(const MCInst *MI, unsigned OpNo,
933 const MCSubtargetInfo &STI,
935 if (MI->getOperand(OpNo).getImm())
939 void AMDGPUInstPrinter::printOModSI(const MCInst *MI, unsigned OpNo,
940 const MCSubtargetInfo &STI,
942 int Imm = MI->getOperand(OpNo).getImm();
943 if (Imm == SIOutMods::MUL2)
945 else if (Imm == SIOutMods::MUL4)
947 else if (Imm == SIOutMods::DIV2)
951 void AMDGPUInstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
952 const MCSubtargetInfo &STI,
954 const MCOperand &Op = MI->getOperand(OpNo);
955 assert(Op.isImm() || Op.isExpr());
957 int64_t Imm = Op.getImm();
958 O << Imm << '(' << BitsToFloat(Imm) << ')';
961 Op.getExpr()->print(O << '@', &MAI);
965 void AMDGPUInstPrinter::printLast(const MCInst *MI, unsigned OpNo,
966 const MCSubtargetInfo &STI, raw_ostream &O) {
967 printIfSet(MI, OpNo, O, "*", " ");
970 void AMDGPUInstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
971 const MCSubtargetInfo &STI, raw_ostream &O) {
972 printIfSet(MI, OpNo, O, '-');
975 void AMDGPUInstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
976 const MCSubtargetInfo &STI, raw_ostream &O) {
977 switch (MI->getOperand(OpNo).getImm()) {
991 void AMDGPUInstPrinter::printRel(const MCInst *MI, unsigned OpNo,
992 const MCSubtargetInfo &STI, raw_ostream &O) {
993 printIfSet(MI, OpNo, O, '+');
996 void AMDGPUInstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
997 const MCSubtargetInfo &STI,
999 printIfSet(MI, OpNo, O, "ExecMask,");
1002 void AMDGPUInstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
1003 const MCSubtargetInfo &STI,
1005 printIfSet(MI, OpNo, O, "Pred,");
1008 void AMDGPUInstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
1009 const MCSubtargetInfo &STI, raw_ostream &O) {
1010 const MCOperand &Op = MI->getOperand(OpNo);
1011 if (Op.getImm() == 0) {
1016 void AMDGPUInstPrinter::printSel(const MCInst *MI, unsigned OpNo,
1018 const char * chans = "XYZW";
1019 int sel = MI->getOperand(OpNo).getImm();
1028 O << cb << '[' << sel << ']';
1029 } else if (sel >= 448) {
1032 } else if (sel >= 0){
1037 O << '.' << chans[chan];
1040 void AMDGPUInstPrinter::printBankSwizzle(const MCInst *MI, unsigned OpNo,
1041 const MCSubtargetInfo &STI,
1043 int BankSwizzle = MI->getOperand(OpNo).getImm();
1044 switch (BankSwizzle) {
1046 O << "BS:VEC_021/SCL_122";
1049 O << "BS:VEC_120/SCL_212";
1052 O << "BS:VEC_102/SCL_221";
1065 void AMDGPUInstPrinter::printRSel(const MCInst *MI, unsigned OpNo,
1066 const MCSubtargetInfo &STI, raw_ostream &O) {
1067 unsigned Sel = MI->getOperand(OpNo).getImm();
1095 void AMDGPUInstPrinter::printCT(const MCInst *MI, unsigned OpNo,
1096 const MCSubtargetInfo &STI, raw_ostream &O) {
1097 unsigned CT = MI->getOperand(OpNo).getImm();
1110 void AMDGPUInstPrinter::printKCache(const MCInst *MI, unsigned OpNo,
1111 const MCSubtargetInfo &STI, raw_ostream &O) {
1112 int KCacheMode = MI->getOperand(OpNo).getImm();
1113 if (KCacheMode > 0) {
1114 int KCacheBank = MI->getOperand(OpNo - 2).getImm();
1115 O << "CB" << KCacheBank << ':';
1116 int KCacheAddr = MI->getOperand(OpNo + 2).getImm();
1117 int LineSize = (KCacheMode == 1) ? 16 : 32;
1118 O << KCacheAddr * 16 << '-' << KCacheAddr * 16 + LineSize;
1122 void AMDGPUInstPrinter::printSendMsg(const MCInst *MI, unsigned OpNo,
1123 const MCSubtargetInfo &STI,
1125 using namespace llvm::AMDGPU::SendMsg;
1127 const unsigned SImm16 = MI->getOperand(OpNo).getImm();
1128 const unsigned Id = SImm16 & ID_MASK_;
1130 if (Id == ID_INTERRUPT) {
1131 if ((SImm16 & ~ID_MASK_) != 0) // Unused/unknown bits must be 0.
1133 O << "sendmsg(" << IdSymbolic[Id] << ')';
1136 if (Id == ID_GS || Id == ID_GS_DONE) {
1137 if ((SImm16 & ~(ID_MASK_|OP_GS_MASK_|STREAM_ID_MASK_)) != 0) // Unused/unknown bits must be 0.
1139 const unsigned OpGs = (SImm16 & OP_GS_MASK_) >> OP_SHIFT_;
1140 const unsigned StreamId = (SImm16 & STREAM_ID_MASK_) >> STREAM_ID_SHIFT_;
1141 if (OpGs == OP_GS_NOP && Id != ID_GS_DONE) // NOP to be used for GS_DONE only.
1143 if (OpGs == OP_GS_NOP && StreamId != 0) // NOP does not use/define stream id bits.
1145 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpGsSymbolic[OpGs];
1146 if (OpGs != OP_GS_NOP) { O << ", " << StreamId; }
1150 if (Id == ID_SYSMSG) {
1151 if ((SImm16 & ~(ID_MASK_|OP_SYS_MASK_)) != 0) // Unused/unknown bits must be 0.
1153 const unsigned OpSys = (SImm16 & OP_SYS_MASK_) >> OP_SHIFT_;
1154 if (! (OP_SYS_FIRST_ <= OpSys && OpSys < OP_SYS_LAST_)) // Unused/unknown.
1156 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpSysSymbolic[OpSys] << ')';
1160 O << SImm16; // Unknown simm16 code.
1163 static void printSwizzleBitmask(const uint16_t AndMask,
1164 const uint16_t OrMask,
1165 const uint16_t XorMask,
1167 using namespace llvm::AMDGPU::Swizzle;
1169 uint16_t Probe0 = ((0 & AndMask) | OrMask) ^ XorMask;
1170 uint16_t Probe1 = ((BITMASK_MASK & AndMask) | OrMask) ^ XorMask;
1174 for (unsigned Mask = 1 << (BITMASK_WIDTH - 1); Mask > 0; Mask >>= 1) {
1175 uint16_t p0 = Probe0 & Mask;
1176 uint16_t p1 = Probe1 & Mask;
1196 void AMDGPUInstPrinter::printSwizzle(const MCInst *MI, unsigned OpNo,
1197 const MCSubtargetInfo &STI,
1199 using namespace llvm::AMDGPU::Swizzle;
1201 uint16_t Imm = MI->getOperand(OpNo).getImm();
1208 if ((Imm & QUAD_PERM_ENC_MASK) == QUAD_PERM_ENC) {
1210 O << "swizzle(" << IdSymbolic[ID_QUAD_PERM];
1211 for (auto i = 0; i < LANE_NUM; ++i) {
1213 O << formatDec(Imm & LANE_MASK);
1218 } else if ((Imm & BITMASK_PERM_ENC_MASK) == BITMASK_PERM_ENC) {
1220 uint16_t AndMask = (Imm >> BITMASK_AND_SHIFT) & BITMASK_MASK;
1221 uint16_t OrMask = (Imm >> BITMASK_OR_SHIFT) & BITMASK_MASK;
1222 uint16_t XorMask = (Imm >> BITMASK_XOR_SHIFT) & BITMASK_MASK;
1224 if (AndMask == BITMASK_MAX &&
1226 countPopulation(XorMask) == 1) {
1228 O << "swizzle(" << IdSymbolic[ID_SWAP];
1230 O << formatDec(XorMask);
1233 } else if (AndMask == BITMASK_MAX &&
1234 OrMask == 0 && XorMask > 0 &&
1235 isPowerOf2_64(XorMask + 1)) {
1237 O << "swizzle(" << IdSymbolic[ID_REVERSE];
1239 O << formatDec(XorMask + 1);
1244 uint16_t GroupSize = BITMASK_MAX - AndMask + 1;
1245 if (GroupSize > 1 &&
1246 isPowerOf2_64(GroupSize) &&
1247 OrMask < GroupSize &&
1250 O << "swizzle(" << IdSymbolic[ID_BROADCAST];
1252 O << formatDec(GroupSize);
1254 O << formatDec(OrMask);
1258 O << "swizzle(" << IdSymbolic[ID_BITMASK_PERM];
1260 printSwizzleBitmask(AndMask, OrMask, XorMask, O);
1265 printU16ImmDecOperand(MI, OpNo, O);
1269 void AMDGPUInstPrinter::printWaitFlag(const MCInst *MI, unsigned OpNo,
1270 const MCSubtargetInfo &STI,
1272 AMDGPU::IsaInfo::IsaVersion ISA =
1273 AMDGPU::IsaInfo::getIsaVersion(STI.getFeatureBits());
1275 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1276 unsigned Vmcnt, Expcnt, Lgkmcnt;
1277 decodeWaitcnt(ISA, SImm16, Vmcnt, Expcnt, Lgkmcnt);
1279 bool NeedSpace = false;
1281 if (Vmcnt != getVmcntBitMask(ISA)) {
1282 O << "vmcnt(" << Vmcnt << ')';
1286 if (Expcnt != getExpcntBitMask(ISA)) {
1289 O << "expcnt(" << Expcnt << ')';
1293 if (Lgkmcnt != getLgkmcntBitMask(ISA)) {
1296 O << "lgkmcnt(" << Lgkmcnt << ')';
1300 void AMDGPUInstPrinter::printHwreg(const MCInst *MI, unsigned OpNo,
1301 const MCSubtargetInfo &STI, raw_ostream &O) {
1302 using namespace llvm::AMDGPU::Hwreg;
1304 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1305 const unsigned Id = (SImm16 & ID_MASK_) >> ID_SHIFT_;
1306 const unsigned Offset = (SImm16 & OFFSET_MASK_) >> OFFSET_SHIFT_;
1307 const unsigned Width = ((SImm16 & WIDTH_M1_MASK_) >> WIDTH_M1_SHIFT_) + 1;
1310 if (ID_SYMBOLIC_FIRST_ <= Id && Id < ID_SYMBOLIC_LAST_) {
1311 O << IdSymbolic[Id];
1315 if (Width != WIDTH_M1_DEFAULT_ + 1 || Offset != OFFSET_DEFAULT_) {
1316 O << ", " << Offset << ", " << Width;
1321 #include "AMDGPUGenAsmWriter.inc"