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 "MCTargetDesc/AMDGPUMCTargetDesc.h"
13 #include "SIDefines.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::printS13ImmDecOperand(const MCInst *MI, unsigned OpNo,
77 O << formatDec(SignExtend32<13>(MI->getOperand(OpNo).getImm()));
80 void AMDGPUInstPrinter::printU32ImmOperand(const MCInst *MI, unsigned OpNo,
81 const MCSubtargetInfo &STI,
83 O << formatHex(MI->getOperand(OpNo).getImm() & 0xffffffff);
86 void AMDGPUInstPrinter::printNamedBit(const MCInst *MI, unsigned OpNo,
87 raw_ostream &O, StringRef BitName) {
88 if (MI->getOperand(OpNo).getImm()) {
93 void AMDGPUInstPrinter::printOffen(const MCInst *MI, unsigned OpNo,
95 printNamedBit(MI, OpNo, O, "offen");
98 void AMDGPUInstPrinter::printIdxen(const MCInst *MI, unsigned OpNo,
100 printNamedBit(MI, OpNo, O, "idxen");
103 void AMDGPUInstPrinter::printAddr64(const MCInst *MI, unsigned OpNo,
105 printNamedBit(MI, OpNo, O, "addr64");
108 void AMDGPUInstPrinter::printMBUFOffset(const MCInst *MI, unsigned OpNo,
110 if (MI->getOperand(OpNo).getImm()) {
112 printU16ImmDecOperand(MI, OpNo, O);
116 void AMDGPUInstPrinter::printOffset(const MCInst *MI, unsigned OpNo,
117 const MCSubtargetInfo &STI,
119 uint16_t Imm = MI->getOperand(OpNo).getImm();
121 O << ((OpNo == 0)? "offset:" : " offset:");
122 printU16ImmDecOperand(MI, OpNo, O);
126 void AMDGPUInstPrinter::printOffsetS13(const MCInst *MI, unsigned OpNo,
127 const MCSubtargetInfo &STI,
129 uint16_t Imm = MI->getOperand(OpNo).getImm();
131 O << ((OpNo == 0)? "offset:" : " offset:");
132 printS13ImmDecOperand(MI, OpNo, O);
136 void AMDGPUInstPrinter::printOffset0(const MCInst *MI, unsigned OpNo,
137 const MCSubtargetInfo &STI,
139 if (MI->getOperand(OpNo).getImm()) {
141 printU8ImmDecOperand(MI, OpNo, O);
145 void AMDGPUInstPrinter::printOffset1(const MCInst *MI, unsigned OpNo,
146 const MCSubtargetInfo &STI,
148 if (MI->getOperand(OpNo).getImm()) {
150 printU8ImmDecOperand(MI, OpNo, O);
154 void AMDGPUInstPrinter::printSMRDOffset8(const MCInst *MI, unsigned OpNo,
155 const MCSubtargetInfo &STI,
157 printU32ImmOperand(MI, OpNo, STI, O);
160 void AMDGPUInstPrinter::printSMRDOffset20(const MCInst *MI, unsigned OpNo,
161 const MCSubtargetInfo &STI,
163 printU32ImmOperand(MI, OpNo, STI, O);
166 void AMDGPUInstPrinter::printSMRDLiteralOffset(const MCInst *MI, unsigned OpNo,
167 const MCSubtargetInfo &STI,
169 printU32ImmOperand(MI, OpNo, STI, O);
172 void AMDGPUInstPrinter::printGDS(const MCInst *MI, unsigned OpNo,
173 const MCSubtargetInfo &STI, raw_ostream &O) {
174 printNamedBit(MI, OpNo, O, "gds");
177 void AMDGPUInstPrinter::printGLC(const MCInst *MI, unsigned OpNo,
178 const MCSubtargetInfo &STI, raw_ostream &O) {
179 printNamedBit(MI, OpNo, O, "glc");
182 void AMDGPUInstPrinter::printSLC(const MCInst *MI, unsigned OpNo,
183 const MCSubtargetInfo &STI, raw_ostream &O) {
184 printNamedBit(MI, OpNo, O, "slc");
187 void AMDGPUInstPrinter::printTFE(const MCInst *MI, unsigned OpNo,
188 const MCSubtargetInfo &STI, raw_ostream &O) {
189 printNamedBit(MI, OpNo, O, "tfe");
192 void AMDGPUInstPrinter::printDMask(const MCInst *MI, unsigned OpNo,
193 const MCSubtargetInfo &STI, raw_ostream &O) {
194 if (MI->getOperand(OpNo).getImm()) {
196 printU16ImmOperand(MI, OpNo, STI, O);
200 void AMDGPUInstPrinter::printUNorm(const MCInst *MI, unsigned OpNo,
201 const MCSubtargetInfo &STI, raw_ostream &O) {
202 printNamedBit(MI, OpNo, O, "unorm");
205 void AMDGPUInstPrinter::printDA(const MCInst *MI, unsigned OpNo,
206 const MCSubtargetInfo &STI, raw_ostream &O) {
207 printNamedBit(MI, OpNo, O, "da");
210 void AMDGPUInstPrinter::printR128(const MCInst *MI, unsigned OpNo,
211 const MCSubtargetInfo &STI, raw_ostream &O) {
212 printNamedBit(MI, OpNo, O, "r128");
215 void AMDGPUInstPrinter::printLWE(const MCInst *MI, unsigned OpNo,
216 const MCSubtargetInfo &STI, raw_ostream &O) {
217 printNamedBit(MI, OpNo, O, "lwe");
220 void AMDGPUInstPrinter::printD16(const MCInst *MI, unsigned OpNo,
221 const MCSubtargetInfo &STI, raw_ostream &O) {
222 printNamedBit(MI, OpNo, O, "d16");
225 void AMDGPUInstPrinter::printExpCompr(const MCInst *MI, unsigned OpNo,
226 const MCSubtargetInfo &STI,
228 if (MI->getOperand(OpNo).getImm())
232 void AMDGPUInstPrinter::printExpVM(const MCInst *MI, unsigned OpNo,
233 const MCSubtargetInfo &STI,
235 if (MI->getOperand(OpNo).getImm())
239 void AMDGPUInstPrinter::printDFMT(const MCInst *MI, unsigned OpNo,
240 const MCSubtargetInfo &STI,
242 if (MI->getOperand(OpNo).getImm()) {
244 printU8ImmDecOperand(MI, OpNo, O);
248 void AMDGPUInstPrinter::printNFMT(const MCInst *MI, unsigned OpNo,
249 const MCSubtargetInfo &STI,
251 if (MI->getOperand(OpNo).getImm()) {
253 printU8ImmDecOperand(MI, OpNo, O);
257 void AMDGPUInstPrinter::printRegOperand(unsigned RegNo, raw_ostream &O,
258 const MCRegisterInfo &MRI) {
272 case AMDGPU::FLAT_SCR:
275 case AMDGPU::XNACK_MASK:
296 case AMDGPU::EXEC_LO:
299 case AMDGPU::EXEC_HI:
302 case AMDGPU::FLAT_SCR_LO:
303 O << "flat_scratch_lo";
305 case AMDGPU::FLAT_SCR_HI:
306 O << "flat_scratch_hi";
308 case AMDGPU::XNACK_MASK_LO:
309 O << "xnack_mask_lo";
311 case AMDGPU::XNACK_MASK_HI:
312 O << "xnack_mask_hi";
316 case AMDGPU::SCRATCH_WAVE_OFFSET_REG:
317 case AMDGPU::PRIVATE_RSRC_REG:
318 llvm_unreachable("pseudo-register should not ever be emitted");
323 // The low 8 bits of the encoding value is the register index, for both VGPRs
325 unsigned RegIdx = MRI.getEncodingValue(RegNo) & ((1 << 8) - 1);
328 if (MRI.getRegClass(AMDGPU::VGPR_32RegClassID).contains(RegNo)) {
331 } else if (MRI.getRegClass(AMDGPU::SGPR_32RegClassID).contains(RegNo)) {
334 } else if (MRI.getRegClass(AMDGPU::VReg_64RegClassID).contains(RegNo)) {
337 } else if (MRI.getRegClass(AMDGPU::SGPR_64RegClassID).contains(RegNo)) {
340 } else if (MRI.getRegClass(AMDGPU::VReg_128RegClassID).contains(RegNo)) {
343 } else if (MRI.getRegClass(AMDGPU::SGPR_128RegClassID).contains(RegNo)) {
346 } else if (MRI.getRegClass(AMDGPU::VReg_96RegClassID).contains(RegNo)) {
349 } else if (MRI.getRegClass(AMDGPU::VReg_256RegClassID).contains(RegNo)) {
352 } else if (MRI.getRegClass(AMDGPU::SGPR_256RegClassID).contains(RegNo)) {
355 } else if (MRI.getRegClass(AMDGPU::VReg_512RegClassID).contains(RegNo)) {
358 } else if (MRI.getRegClass(AMDGPU::SGPR_512RegClassID).contains(RegNo)) {
362 O << getRegisterName(RegNo);
371 O << '[' << RegIdx << ':' << (RegIdx + NumRegs - 1) << ']';
374 void AMDGPUInstPrinter::printVOPDst(const MCInst *MI, unsigned OpNo,
375 const MCSubtargetInfo &STI, raw_ostream &O) {
376 if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3)
378 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::DPP)
380 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::SDWA)
385 printOperand(MI, OpNo, STI, O);
388 void AMDGPUInstPrinter::printVINTRPDst(const MCInst *MI, unsigned OpNo,
389 const MCSubtargetInfo &STI, raw_ostream &O) {
390 if (AMDGPU::isSI(STI) || AMDGPU::isCI(STI))
395 printOperand(MI, OpNo, STI, O);
398 void AMDGPUInstPrinter::printImmediate16(uint32_t Imm,
399 const MCSubtargetInfo &STI,
401 int16_t SImm = static_cast<int16_t>(Imm);
402 if (SImm >= -16 && SImm <= 64) {
409 else if (Imm == 0xBC00)
411 else if (Imm == 0x3800)
413 else if (Imm == 0xB800)
415 else if (Imm == 0x4000)
417 else if (Imm == 0xC000)
419 else if (Imm == 0x4400)
421 else if (Imm == 0xC400)
423 else if (Imm == 0x3118) {
424 assert(STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]);
427 O << formatHex(static_cast<uint64_t>(Imm));
430 void AMDGPUInstPrinter::printImmediateV216(uint32_t Imm,
431 const MCSubtargetInfo &STI,
433 uint16_t Lo16 = static_cast<uint16_t>(Imm);
434 printImmediate16(Lo16, STI, O);
437 void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
438 const MCSubtargetInfo &STI,
440 int32_t SImm = static_cast<int32_t>(Imm);
441 if (SImm >= -16 && SImm <= 64) {
446 if (Imm == FloatToBits(0.0f))
448 else if (Imm == FloatToBits(1.0f))
450 else if (Imm == FloatToBits(-1.0f))
452 else if (Imm == FloatToBits(0.5f))
454 else if (Imm == FloatToBits(-0.5f))
456 else if (Imm == FloatToBits(2.0f))
458 else if (Imm == FloatToBits(-2.0f))
460 else if (Imm == FloatToBits(4.0f))
462 else if (Imm == FloatToBits(-4.0f))
464 else if (Imm == 0x3e22f983 &&
465 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
468 O << formatHex(static_cast<uint64_t>(Imm));
471 void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
472 const MCSubtargetInfo &STI,
474 int64_t SImm = static_cast<int64_t>(Imm);
475 if (SImm >= -16 && SImm <= 64) {
480 if (Imm == DoubleToBits(0.0))
482 else if (Imm == DoubleToBits(1.0))
484 else if (Imm == DoubleToBits(-1.0))
486 else if (Imm == DoubleToBits(0.5))
488 else if (Imm == DoubleToBits(-0.5))
490 else if (Imm == DoubleToBits(2.0))
492 else if (Imm == DoubleToBits(-2.0))
494 else if (Imm == DoubleToBits(4.0))
496 else if (Imm == DoubleToBits(-4.0))
498 else if (Imm == 0x3fc45f306dc9c882 &&
499 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
502 assert(isUInt<32>(Imm) || Imm == 0x3fc45f306dc9c882);
504 // In rare situations, we will have a 32-bit literal in a 64-bit
505 // operand. This is technically allowed for the encoding of s_mov_b64.
506 O << formatHex(static_cast<uint64_t>(Imm));
510 void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
511 const MCSubtargetInfo &STI,
513 if (OpNo >= MI->getNumOperands()) {
514 O << "/*Missing OP" << OpNo << "*/";
518 const MCOperand &Op = MI->getOperand(OpNo);
520 printRegOperand(Op.getReg(), O, MRI);
521 } else if (Op.isImm()) {
522 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
523 switch (Desc.OpInfo[OpNo].OperandType) {
524 case AMDGPU::OPERAND_REG_IMM_INT32:
525 case AMDGPU::OPERAND_REG_IMM_FP32:
526 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
527 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
528 case MCOI::OPERAND_IMMEDIATE:
529 printImmediate32(Op.getImm(), STI, O);
531 case AMDGPU::OPERAND_REG_IMM_INT64:
532 case AMDGPU::OPERAND_REG_IMM_FP64:
533 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
534 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
535 printImmediate64(Op.getImm(), STI, O);
537 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
538 case AMDGPU::OPERAND_REG_INLINE_C_FP16:
539 case AMDGPU::OPERAND_REG_IMM_INT16:
540 case AMDGPU::OPERAND_REG_IMM_FP16:
541 printImmediate16(Op.getImm(), STI, O);
543 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
544 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
545 printImmediateV216(Op.getImm(), STI, O);
547 case MCOI::OPERAND_UNKNOWN:
548 case MCOI::OPERAND_PCREL:
549 O << formatDec(Op.getImm());
551 case MCOI::OPERAND_REGISTER:
552 // FIXME: This should be removed and handled somewhere else. Seems to come
553 // from a disassembler bug.
554 O << "/*invalid immediate*/";
557 // We hit this for the immediate instruction bits that don't yet have a
559 llvm_unreachable("unexpected immediate operand type");
561 } else if (Op.isFPImm()) {
562 // We special case 0.0 because otherwise it will be printed as an integer.
563 if (Op.getFPImm() == 0.0)
566 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
567 int RCID = Desc.OpInfo[OpNo].RegClass;
568 unsigned RCBits = AMDGPU::getRegBitWidth(MRI.getRegClass(RCID));
570 printImmediate32(FloatToBits(Op.getFPImm()), STI, O);
571 else if (RCBits == 64)
572 printImmediate64(DoubleToBits(Op.getFPImm()), STI, O);
574 llvm_unreachable("Invalid register class size");
576 } else if (Op.isExpr()) {
577 const MCExpr *Exp = Op.getExpr();
584 void AMDGPUInstPrinter::printOperandAndFPInputMods(const MCInst *MI,
586 const MCSubtargetInfo &STI,
588 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
590 // Use 'neg(...)' instead of '-' to avoid ambiguity.
591 // This is important for integer literals because
592 // -1 is not the same value as neg(1).
593 bool NegMnemo = false;
595 if (InputModifiers & SISrcMods::NEG) {
596 if (OpNo + 1 < MI->getNumOperands() &&
597 (InputModifiers & SISrcMods::ABS) == 0) {
598 const MCOperand &Op = MI->getOperand(OpNo + 1);
599 NegMnemo = Op.isImm() || Op.isFPImm();
608 if (InputModifiers & SISrcMods::ABS)
610 printOperand(MI, OpNo + 1, STI, O);
611 if (InputModifiers & SISrcMods::ABS)
619 void AMDGPUInstPrinter::printOperandAndIntInputMods(const MCInst *MI,
621 const MCSubtargetInfo &STI,
623 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
624 if (InputModifiers & SISrcMods::SEXT)
626 printOperand(MI, OpNo + 1, STI, O);
627 if (InputModifiers & SISrcMods::SEXT)
631 void AMDGPUInstPrinter::printDPPCtrl(const MCInst *MI, unsigned OpNo,
632 const MCSubtargetInfo &STI,
634 using namespace AMDGPU::DPP;
636 unsigned Imm = MI->getOperand(OpNo).getImm();
637 if (Imm <= DppCtrl::QUAD_PERM_LAST) {
639 O << formatDec(Imm & 0x3) << ',';
640 O << formatDec((Imm & 0xc) >> 2) << ',';
641 O << formatDec((Imm & 0x30) >> 4) << ',';
642 O << formatDec((Imm & 0xc0) >> 6) << ']';
643 } else if ((Imm >= DppCtrl::ROW_SHL_FIRST) &&
644 (Imm <= DppCtrl::ROW_SHL_LAST)) {
646 printU4ImmDecOperand(MI, OpNo, O);
647 } else if ((Imm >= DppCtrl::ROW_SHR_FIRST) &&
648 (Imm <= DppCtrl::ROW_SHR_LAST)) {
650 printU4ImmDecOperand(MI, OpNo, O);
651 } else if ((Imm >= DppCtrl::ROW_ROR_FIRST) &&
652 (Imm <= DppCtrl::ROW_ROR_LAST)) {
654 printU4ImmDecOperand(MI, OpNo, O);
655 } else if (Imm == DppCtrl::WAVE_SHL1) {
657 } else if (Imm == DppCtrl::WAVE_ROL1) {
659 } else if (Imm == DppCtrl::WAVE_SHR1) {
661 } else if (Imm == DppCtrl::WAVE_ROR1) {
663 } else if (Imm == DppCtrl::ROW_MIRROR) {
665 } else if (Imm == DppCtrl::ROW_HALF_MIRROR) {
666 O << " row_half_mirror";
667 } else if (Imm == DppCtrl::BCAST15) {
668 O << " row_bcast:15";
669 } else if (Imm == DppCtrl::BCAST31) {
670 O << " row_bcast:31";
672 O << " /* Invalid dpp_ctrl value */";
676 void AMDGPUInstPrinter::printRowMask(const MCInst *MI, unsigned OpNo,
677 const MCSubtargetInfo &STI,
680 printU4ImmOperand(MI, OpNo, STI, O);
683 void AMDGPUInstPrinter::printBankMask(const MCInst *MI, unsigned OpNo,
684 const MCSubtargetInfo &STI,
687 printU4ImmOperand(MI, OpNo, STI, O);
690 void AMDGPUInstPrinter::printBoundCtrl(const MCInst *MI, unsigned OpNo,
691 const MCSubtargetInfo &STI,
693 unsigned Imm = MI->getOperand(OpNo).getImm();
695 O << " bound_ctrl:0"; // XXX - this syntax is used in sp3
699 void AMDGPUInstPrinter::printSDWASel(const MCInst *MI, unsigned OpNo,
701 using namespace llvm::AMDGPU::SDWA;
703 unsigned Imm = MI->getOperand(OpNo).getImm();
705 case SdwaSel::BYTE_0: O << "BYTE_0"; break;
706 case SdwaSel::BYTE_1: O << "BYTE_1"; break;
707 case SdwaSel::BYTE_2: O << "BYTE_2"; break;
708 case SdwaSel::BYTE_3: O << "BYTE_3"; break;
709 case SdwaSel::WORD_0: O << "WORD_0"; break;
710 case SdwaSel::WORD_1: O << "WORD_1"; break;
711 case SdwaSel::DWORD: O << "DWORD"; break;
712 default: llvm_unreachable("Invalid SDWA data select operand");
716 void AMDGPUInstPrinter::printSDWADstSel(const MCInst *MI, unsigned OpNo,
717 const MCSubtargetInfo &STI,
720 printSDWASel(MI, OpNo, O);
723 void AMDGPUInstPrinter::printSDWASrc0Sel(const MCInst *MI, unsigned OpNo,
724 const MCSubtargetInfo &STI,
727 printSDWASel(MI, OpNo, O);
730 void AMDGPUInstPrinter::printSDWASrc1Sel(const MCInst *MI, unsigned OpNo,
731 const MCSubtargetInfo &STI,
734 printSDWASel(MI, OpNo, O);
737 void AMDGPUInstPrinter::printSDWADstUnused(const MCInst *MI, unsigned OpNo,
738 const MCSubtargetInfo &STI,
740 using namespace llvm::AMDGPU::SDWA;
743 unsigned Imm = MI->getOperand(OpNo).getImm();
745 case DstUnused::UNUSED_PAD: O << "UNUSED_PAD"; break;
746 case DstUnused::UNUSED_SEXT: O << "UNUSED_SEXT"; break;
747 case DstUnused::UNUSED_PRESERVE: O << "UNUSED_PRESERVE"; break;
748 default: llvm_unreachable("Invalid SDWA dest_unused operand");
752 template <unsigned N>
753 void AMDGPUInstPrinter::printExpSrcN(const MCInst *MI, unsigned OpNo,
754 const MCSubtargetInfo &STI,
756 unsigned Opc = MI->getOpcode();
757 int EnIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::en);
758 unsigned En = MI->getOperand(EnIdx).getImm();
760 int ComprIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::compr);
762 // If compr is set, print as src0, src0, src1, src1
763 if (MI->getOperand(ComprIdx).getImm()) {
764 if (N == 1 || N == 2)
771 printRegOperand(MI->getOperand(OpNo).getReg(), O, MRI);
776 void AMDGPUInstPrinter::printExpSrc0(const MCInst *MI, unsigned OpNo,
777 const MCSubtargetInfo &STI,
779 printExpSrcN<0>(MI, OpNo, STI, O);
782 void AMDGPUInstPrinter::printExpSrc1(const MCInst *MI, unsigned OpNo,
783 const MCSubtargetInfo &STI,
785 printExpSrcN<1>(MI, OpNo, STI, O);
788 void AMDGPUInstPrinter::printExpSrc2(const MCInst *MI, unsigned OpNo,
789 const MCSubtargetInfo &STI,
791 printExpSrcN<2>(MI, OpNo, STI, O);
794 void AMDGPUInstPrinter::printExpSrc3(const MCInst *MI, unsigned OpNo,
795 const MCSubtargetInfo &STI,
797 printExpSrcN<3>(MI, OpNo, STI, O);
800 void AMDGPUInstPrinter::printExpTgt(const MCInst *MI, unsigned OpNo,
801 const MCSubtargetInfo &STI,
803 // This is really a 6 bit field.
804 uint32_t Tgt = MI->getOperand(OpNo).getImm() & ((1 << 6) - 1);
812 else if (Tgt >= 12 && Tgt <= 15)
813 O << " pos" << Tgt - 12;
814 else if (Tgt >= 32 && Tgt <= 63)
815 O << " param" << Tgt - 32;
817 // Reserved values 10, 11
818 O << " invalid_target_" << Tgt;
822 static bool allOpsDefaultValue(const int* Ops, int NumOps, int Mod,
823 bool IsPacked, bool HasDstSel) {
824 int DefaultValue = IsPacked && (Mod == SISrcMods::OP_SEL_1);
826 for (int I = 0; I < NumOps; ++I) {
827 if (!!(Ops[I] & Mod) != DefaultValue)
831 if (HasDstSel && (Ops[0] & SISrcMods::DST_OP_SEL) != 0)
837 void AMDGPUInstPrinter::printPackedModifier(const MCInst *MI,
841 unsigned Opc = MI->getOpcode();
845 for (int OpName : { AMDGPU::OpName::src0_modifiers,
846 AMDGPU::OpName::src1_modifiers,
847 AMDGPU::OpName::src2_modifiers }) {
848 int Idx = AMDGPU::getNamedOperandIdx(Opc, OpName);
852 Ops[NumOps++] = MI->getOperand(Idx).getImm();
855 const bool HasDstSel =
857 Mod == SISrcMods::OP_SEL_0 &&
858 MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3_OPSEL;
860 const bool IsPacked =
861 MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::IsPacked;
863 if (allOpsDefaultValue(Ops, NumOps, Mod, IsPacked, HasDstSel))
867 for (int I = 0; I < NumOps; ++I) {
871 O << !!(Ops[I] & Mod);
875 O << ',' << !!(Ops[0] & SISrcMods::DST_OP_SEL);
881 void AMDGPUInstPrinter::printOpSel(const MCInst *MI, unsigned,
882 const MCSubtargetInfo &STI,
884 printPackedModifier(MI, " op_sel:[", SISrcMods::OP_SEL_0, O);
887 void AMDGPUInstPrinter::printOpSelHi(const MCInst *MI, unsigned OpNo,
888 const MCSubtargetInfo &STI,
890 printPackedModifier(MI, " op_sel_hi:[", SISrcMods::OP_SEL_1, O);
893 void AMDGPUInstPrinter::printNegLo(const MCInst *MI, unsigned OpNo,
894 const MCSubtargetInfo &STI,
896 printPackedModifier(MI, " neg_lo:[", SISrcMods::NEG, O);
899 void AMDGPUInstPrinter::printNegHi(const MCInst *MI, unsigned OpNo,
900 const MCSubtargetInfo &STI,
902 printPackedModifier(MI, " neg_hi:[", SISrcMods::NEG_HI, O);
905 void AMDGPUInstPrinter::printInterpSlot(const MCInst *MI, unsigned OpNum,
906 const MCSubtargetInfo &STI,
908 unsigned Imm = MI->getOperand(OpNum).getImm();
920 O << "invalid_param_" << Imm;
924 void AMDGPUInstPrinter::printInterpAttr(const MCInst *MI, unsigned OpNum,
925 const MCSubtargetInfo &STI,
927 unsigned Attr = MI->getOperand(OpNum).getImm();
931 void AMDGPUInstPrinter::printInterpAttrChan(const MCInst *MI, unsigned OpNum,
932 const MCSubtargetInfo &STI,
934 unsigned Chan = MI->getOperand(OpNum).getImm();
935 O << '.' << "xyzw"[Chan & 0x3];
938 void AMDGPUInstPrinter::printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
939 const MCSubtargetInfo &STI,
941 unsigned Val = MI->getOperand(OpNo).getImm();
947 if (Val & VGPRIndexMode::DST_ENABLE)
950 if (Val & VGPRIndexMode::SRC0_ENABLE)
953 if (Val & VGPRIndexMode::SRC1_ENABLE)
956 if (Val & VGPRIndexMode::SRC2_ENABLE)
960 void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
961 const MCSubtargetInfo &STI,
963 printOperand(MI, OpNo, STI, O);
965 printOperand(MI, OpNo + 1, STI, O);
968 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
969 raw_ostream &O, StringRef Asm,
971 const MCOperand &Op = MI->getOperand(OpNo);
973 if (Op.getImm() == 1) {
980 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
981 raw_ostream &O, char Asm) {
982 const MCOperand &Op = MI->getOperand(OpNo);
984 if (Op.getImm() == 1)
988 void AMDGPUInstPrinter::printHigh(const MCInst *MI, unsigned OpNo,
989 const MCSubtargetInfo &STI,
991 if (MI->getOperand(OpNo).getImm())
995 void AMDGPUInstPrinter::printClampSI(const MCInst *MI, unsigned OpNo,
996 const MCSubtargetInfo &STI,
998 if (MI->getOperand(OpNo).getImm())
1002 void AMDGPUInstPrinter::printOModSI(const MCInst *MI, unsigned OpNo,
1003 const MCSubtargetInfo &STI,
1005 int Imm = MI->getOperand(OpNo).getImm();
1006 if (Imm == SIOutMods::MUL2)
1008 else if (Imm == SIOutMods::MUL4)
1010 else if (Imm == SIOutMods::DIV2)
1014 void AMDGPUInstPrinter::printSendMsg(const MCInst *MI, unsigned OpNo,
1015 const MCSubtargetInfo &STI,
1017 using namespace llvm::AMDGPU::SendMsg;
1019 const unsigned SImm16 = MI->getOperand(OpNo).getImm();
1020 const unsigned Id = SImm16 & ID_MASK_;
1022 if (Id == ID_INTERRUPT) {
1023 if ((SImm16 & ~ID_MASK_) != 0) // Unused/unknown bits must be 0.
1025 O << "sendmsg(" << IdSymbolic[Id] << ')';
1028 if (Id == ID_GS || Id == ID_GS_DONE) {
1029 if ((SImm16 & ~(ID_MASK_|OP_GS_MASK_|STREAM_ID_MASK_)) != 0) // Unused/unknown bits must be 0.
1031 const unsigned OpGs = (SImm16 & OP_GS_MASK_) >> OP_SHIFT_;
1032 const unsigned StreamId = (SImm16 & STREAM_ID_MASK_) >> STREAM_ID_SHIFT_;
1033 if (OpGs == OP_GS_NOP && Id != ID_GS_DONE) // NOP to be used for GS_DONE only.
1035 if (OpGs == OP_GS_NOP && StreamId != 0) // NOP does not use/define stream id bits.
1037 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpGsSymbolic[OpGs];
1038 if (OpGs != OP_GS_NOP) { O << ", " << StreamId; }
1042 if (Id == ID_SYSMSG) {
1043 if ((SImm16 & ~(ID_MASK_|OP_SYS_MASK_)) != 0) // Unused/unknown bits must be 0.
1045 const unsigned OpSys = (SImm16 & OP_SYS_MASK_) >> OP_SHIFT_;
1046 if (! (OP_SYS_FIRST_ <= OpSys && OpSys < OP_SYS_LAST_)) // Unused/unknown.
1048 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpSysSymbolic[OpSys] << ')';
1052 O << SImm16; // Unknown simm16 code.
1055 static void printSwizzleBitmask(const uint16_t AndMask,
1056 const uint16_t OrMask,
1057 const uint16_t XorMask,
1059 using namespace llvm::AMDGPU::Swizzle;
1061 uint16_t Probe0 = ((0 & AndMask) | OrMask) ^ XorMask;
1062 uint16_t Probe1 = ((BITMASK_MASK & AndMask) | OrMask) ^ XorMask;
1066 for (unsigned Mask = 1 << (BITMASK_WIDTH - 1); Mask > 0; Mask >>= 1) {
1067 uint16_t p0 = Probe0 & Mask;
1068 uint16_t p1 = Probe1 & Mask;
1088 void AMDGPUInstPrinter::printSwizzle(const MCInst *MI, unsigned OpNo,
1089 const MCSubtargetInfo &STI,
1091 using namespace llvm::AMDGPU::Swizzle;
1093 uint16_t Imm = MI->getOperand(OpNo).getImm();
1100 if ((Imm & QUAD_PERM_ENC_MASK) == QUAD_PERM_ENC) {
1102 O << "swizzle(" << IdSymbolic[ID_QUAD_PERM];
1103 for (auto i = 0; i < LANE_NUM; ++i) {
1105 O << formatDec(Imm & LANE_MASK);
1110 } else if ((Imm & BITMASK_PERM_ENC_MASK) == BITMASK_PERM_ENC) {
1112 uint16_t AndMask = (Imm >> BITMASK_AND_SHIFT) & BITMASK_MASK;
1113 uint16_t OrMask = (Imm >> BITMASK_OR_SHIFT) & BITMASK_MASK;
1114 uint16_t XorMask = (Imm >> BITMASK_XOR_SHIFT) & BITMASK_MASK;
1116 if (AndMask == BITMASK_MAX &&
1118 countPopulation(XorMask) == 1) {
1120 O << "swizzle(" << IdSymbolic[ID_SWAP];
1122 O << formatDec(XorMask);
1125 } else if (AndMask == BITMASK_MAX &&
1126 OrMask == 0 && XorMask > 0 &&
1127 isPowerOf2_64(XorMask + 1)) {
1129 O << "swizzle(" << IdSymbolic[ID_REVERSE];
1131 O << formatDec(XorMask + 1);
1136 uint16_t GroupSize = BITMASK_MAX - AndMask + 1;
1137 if (GroupSize > 1 &&
1138 isPowerOf2_64(GroupSize) &&
1139 OrMask < GroupSize &&
1142 O << "swizzle(" << IdSymbolic[ID_BROADCAST];
1144 O << formatDec(GroupSize);
1146 O << formatDec(OrMask);
1150 O << "swizzle(" << IdSymbolic[ID_BITMASK_PERM];
1152 printSwizzleBitmask(AndMask, OrMask, XorMask, O);
1157 printU16ImmDecOperand(MI, OpNo, O);
1161 void AMDGPUInstPrinter::printWaitFlag(const MCInst *MI, unsigned OpNo,
1162 const MCSubtargetInfo &STI,
1164 AMDGPU::IsaInfo::IsaVersion ISA =
1165 AMDGPU::IsaInfo::getIsaVersion(STI.getFeatureBits());
1167 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1168 unsigned Vmcnt, Expcnt, Lgkmcnt;
1169 decodeWaitcnt(ISA, SImm16, Vmcnt, Expcnt, Lgkmcnt);
1171 bool NeedSpace = false;
1173 if (Vmcnt != getVmcntBitMask(ISA)) {
1174 O << "vmcnt(" << Vmcnt << ')';
1178 if (Expcnt != getExpcntBitMask(ISA)) {
1181 O << "expcnt(" << Expcnt << ')';
1185 if (Lgkmcnt != getLgkmcntBitMask(ISA)) {
1188 O << "lgkmcnt(" << Lgkmcnt << ')';
1192 void AMDGPUInstPrinter::printHwreg(const MCInst *MI, unsigned OpNo,
1193 const MCSubtargetInfo &STI, raw_ostream &O) {
1194 using namespace llvm::AMDGPU::Hwreg;
1196 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1197 const unsigned Id = (SImm16 & ID_MASK_) >> ID_SHIFT_;
1198 const unsigned Offset = (SImm16 & OFFSET_MASK_) >> OFFSET_SHIFT_;
1199 const unsigned Width = ((SImm16 & WIDTH_M1_MASK_) >> WIDTH_M1_SHIFT_) + 1;
1202 unsigned Last = ID_SYMBOLIC_LAST_;
1203 if (AMDGPU::isSI(STI) || AMDGPU::isCI(STI) || AMDGPU::isVI(STI))
1204 Last = ID_SYMBOLIC_FIRST_GFX9_;
1205 if (ID_SYMBOLIC_FIRST_ <= Id && Id < Last && IdSymbolic[Id]) {
1206 O << IdSymbolic[Id];
1210 if (Width != WIDTH_M1_DEFAULT_ + 1 || Offset != OFFSET_DEFAULT_) {
1211 O << ", " << Offset << ", " << Width;
1216 #include "AMDGPUGenAsmWriter.inc"
1218 void R600InstPrinter::printInst(const MCInst *MI, raw_ostream &O,
1219 StringRef Annot, const MCSubtargetInfo &STI) {
1221 printInstruction(MI, O);
1222 printAnnotation(O, Annot);
1225 void R600InstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
1227 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '|');
1230 void R600InstPrinter::printBankSwizzle(const MCInst *MI, unsigned OpNo,
1232 int BankSwizzle = MI->getOperand(OpNo).getImm();
1233 switch (BankSwizzle) {
1235 O << "BS:VEC_021/SCL_122";
1238 O << "BS:VEC_120/SCL_212";
1241 O << "BS:VEC_102/SCL_221";
1254 void R600InstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
1256 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "_SAT");
1259 void R600InstPrinter::printCT(const MCInst *MI, unsigned OpNo,
1261 unsigned CT = MI->getOperand(OpNo).getImm();
1274 void R600InstPrinter::printKCache(const MCInst *MI, unsigned OpNo,
1276 int KCacheMode = MI->getOperand(OpNo).getImm();
1277 if (KCacheMode > 0) {
1278 int KCacheBank = MI->getOperand(OpNo - 2).getImm();
1279 O << "CB" << KCacheBank << ':';
1280 int KCacheAddr = MI->getOperand(OpNo + 2).getImm();
1281 int LineSize = (KCacheMode == 1) ? 16 : 32;
1282 O << KCacheAddr * 16 << '-' << KCacheAddr * 16 + LineSize;
1286 void R600InstPrinter::printLast(const MCInst *MI, unsigned OpNo,
1288 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "*", " ");
1291 void R600InstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
1293 const MCOperand &Op = MI->getOperand(OpNo);
1294 assert(Op.isImm() || Op.isExpr());
1296 int64_t Imm = Op.getImm();
1297 O << Imm << '(' << BitsToFloat(Imm) << ')';
1300 Op.getExpr()->print(O << '@', &MAI);
1304 void R600InstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
1306 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '-');
1309 void R600InstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
1311 switch (MI->getOperand(OpNo).getImm()) {
1325 void R600InstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
1327 printOperand(MI, OpNo, O);
1329 printOperand(MI, OpNo + 1, O);
1332 void R600InstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
1334 if (OpNo >= MI->getNumOperands()) {
1335 O << "/*Missing OP" << OpNo << "*/";
1339 const MCOperand &Op = MI->getOperand(OpNo);
1341 switch (Op.getReg()) {
1342 // This is the default predicate state, so we don't need to print it.
1343 case R600::PRED_SEL_OFF:
1347 O << getRegisterName(Op.getReg());
1350 } else if (Op.isImm()) {
1352 } else if (Op.isFPImm()) {
1353 // We special case 0.0 because otherwise it will be printed as an integer.
1354 if (Op.getFPImm() == 0.0)
1359 } else if (Op.isExpr()) {
1360 const MCExpr *Exp = Op.getExpr();
1361 Exp->print(O, &MAI);
1367 void R600InstPrinter::printRel(const MCInst *MI, unsigned OpNo,
1369 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '+');
1372 void R600InstPrinter::printRSel(const MCInst *MI, unsigned OpNo,
1374 unsigned Sel = MI->getOperand(OpNo).getImm();
1402 void R600InstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
1404 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "ExecMask,");
1407 void R600InstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
1409 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "Pred,");
1412 void R600InstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
1414 const MCOperand &Op = MI->getOperand(OpNo);
1415 if (Op.getImm() == 0) {
1420 #include "R600GenAsmWriter.inc"