1 //===- R600MCCodeEmitter.cpp - Code Emitter for R600->Cayman GPU families -===//
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 //===----------------------------------------------------------------------===//
12 /// This code emitter outputs bytecode that is understood by the r600g driver
13 /// in the Mesa [1] project. The bytecode is very similar to the hardware's ISA,
14 /// but it still needs to be run through a finalizer in order to be executed
17 /// [1] http://www.mesa3d.org/
19 //===----------------------------------------------------------------------===//
21 #include "R600Defines.h"
22 #include "MCTargetDesc/AMDGPUMCCodeEmitter.h"
23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
24 #include "llvm/MC/MCCodeEmitter.h"
25 #include "llvm/MC/MCContext.h"
26 #include "llvm/MC/MCInst.h"
27 #include "llvm/MC/MCInstrInfo.h"
28 #include "llvm/MC/MCRegisterInfo.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/Support/raw_ostream.h"
33 #define SRC_BYTE_COUNT 11
34 #define DST_BYTE_COUNT 5
40 class R600MCCodeEmitter : public AMDGPUMCCodeEmitter {
41 R600MCCodeEmitter(const R600MCCodeEmitter &) LLVM_DELETED_FUNCTION;
42 void operator=(const R600MCCodeEmitter &) LLVM_DELETED_FUNCTION;
43 const MCInstrInfo &MCII;
44 const MCRegisterInfo &MRI;
45 const MCSubtargetInfo &STI;
50 R600MCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
51 const MCSubtargetInfo &sti, MCContext &ctx)
52 : MCII(mcii), MRI(mri), STI(sti), Ctx(ctx) { }
54 /// \brief Encode the instruction and write it to the OS.
55 virtual void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
56 SmallVectorImpl<MCFixup> &Fixups) const;
58 /// \returns the encoding for an MCOperand.
59 virtual uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
60 SmallVectorImpl<MCFixup> &Fixups) const;
63 void EmitALUInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
64 raw_ostream &OS) const;
65 void EmitSrc(const MCInst &MI, unsigned OpIdx, raw_ostream &OS) const;
66 void EmitSrcISA(const MCInst &MI, unsigned RegOpIdx, unsigned SelOpIdx,
67 raw_ostream &OS) const;
68 void EmitDst(const MCInst &MI, raw_ostream &OS) const;
69 void EmitFCInstr(const MCInst &MI, raw_ostream &OS) const;
71 void EmitNullBytes(unsigned int byteCount, raw_ostream &OS) const;
73 void EmitByte(unsigned int byte, raw_ostream &OS) const;
75 void EmitTwoBytes(uint32_t bytes, raw_ostream &OS) const;
77 void Emit(uint32_t value, raw_ostream &OS) const;
78 void Emit(uint64_t value, raw_ostream &OS) const;
80 unsigned getHWRegChan(unsigned reg) const;
81 unsigned getHWReg(unsigned regNo) const;
83 bool isFCOp(unsigned opcode) const;
84 bool isTexOp(unsigned opcode) const;
85 bool isFlagSet(const MCInst &MI, unsigned Operand, unsigned Flag) const;
89 } // End anonymous namespace
129 TEXTURE_SHADOW1D_ARRAY,
130 TEXTURE_SHADOW2D_ARRAY
133 MCCodeEmitter *llvm::createR600MCCodeEmitter(const MCInstrInfo &MCII,
134 const MCRegisterInfo &MRI,
135 const MCSubtargetInfo &STI,
137 return new R600MCCodeEmitter(MCII, MRI, STI, Ctx);
140 void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
141 SmallVectorImpl<MCFixup> &Fixups) const {
142 if (isFCOp(MI.getOpcode())){
144 } else if (MI.getOpcode() == AMDGPU::RETURN ||
145 MI.getOpcode() == AMDGPU::BUNDLE ||
146 MI.getOpcode() == AMDGPU::KILL) {
149 switch(MI.getOpcode()) {
150 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
151 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
152 uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
153 EmitByte(INSTR_NATIVE, OS);
157 case AMDGPU::CONSTANT_LOAD_eg:
158 case AMDGPU::VTX_READ_PARAM_8_eg:
159 case AMDGPU::VTX_READ_PARAM_16_eg:
160 case AMDGPU::VTX_READ_PARAM_32_eg:
161 case AMDGPU::VTX_READ_PARAM_128_eg:
162 case AMDGPU::VTX_READ_GLOBAL_8_eg:
163 case AMDGPU::VTX_READ_GLOBAL_32_eg:
164 case AMDGPU::VTX_READ_GLOBAL_128_eg:
165 case AMDGPU::TEX_VTX_CONSTBUF:
166 case AMDGPU::TEX_VTX_TEXBUF : {
167 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
168 uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
170 EmitByte(INSTR_VTX, OS);
171 Emit(InstWord01, OS);
176 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
177 case AMDGPU::TEX_SAMPLE:
178 case AMDGPU::TEX_SAMPLE_C:
179 case AMDGPU::TEX_SAMPLE_L:
180 case AMDGPU::TEX_SAMPLE_C_L:
181 case AMDGPU::TEX_SAMPLE_LB:
182 case AMDGPU::TEX_SAMPLE_C_LB:
183 case AMDGPU::TEX_SAMPLE_G:
184 case AMDGPU::TEX_SAMPLE_C_G:
185 case AMDGPU::TEX_GET_GRADIENTS_H:
186 case AMDGPU::TEX_GET_GRADIENTS_V:
187 case AMDGPU::TEX_SET_GRADIENTS_H:
188 case AMDGPU::TEX_SET_GRADIENTS_V: {
189 unsigned Opcode = MI.getOpcode();
190 bool HasOffsets = (Opcode == AMDGPU::TEX_LD);
191 unsigned OpOffset = HasOffsets ? 3 : 0;
192 int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
193 int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
195 uint32_t SrcSelect[4] = {0, 1, 2, 3};
196 uint32_t Offsets[3] = {0, 0, 0};
197 uint64_t CoordType[4] = {1, 1, 1, 1};
200 for (unsigned i = 0; i < 3; i++)
201 Offsets[i] = MI.getOperand(i + 2).getImm();
203 if (TextureType == TEXTURE_RECT ||
204 TextureType == TEXTURE_SHADOWRECT) {
205 CoordType[ELEMENT_X] = 0;
206 CoordType[ELEMENT_Y] = 0;
209 if (TextureType == TEXTURE_1D_ARRAY ||
210 TextureType == TEXTURE_SHADOW1D_ARRAY) {
211 if (Opcode == AMDGPU::TEX_SAMPLE_C_L ||
212 Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
213 CoordType[ELEMENT_Y] = 0;
215 CoordType[ELEMENT_Z] = 0;
216 SrcSelect[ELEMENT_Z] = ELEMENT_Y;
218 } else if (TextureType == TEXTURE_2D_ARRAY ||
219 TextureType == TEXTURE_SHADOW2D_ARRAY) {
220 CoordType[ELEMENT_Z] = 0;
224 if ((TextureType == TEXTURE_SHADOW1D ||
225 TextureType == TEXTURE_SHADOW2D ||
226 TextureType == TEXTURE_SHADOWRECT ||
227 TextureType == TEXTURE_SHADOW1D_ARRAY) &&
228 Opcode != AMDGPU::TEX_SAMPLE_C_L &&
229 Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
230 SrcSelect[ELEMENT_W] = ELEMENT_Z;
233 uint64_t Word01 = getBinaryCodeForInstr(MI, Fixups) |
234 CoordType[ELEMENT_X] << 60 | CoordType[ELEMENT_Y] << 61 |
235 CoordType[ELEMENT_Z] << 62 | CoordType[ELEMENT_W] << 63;
236 uint32_t Word2 = Sampler << 15 | SrcSelect[ELEMENT_X] << 20 |
237 SrcSelect[ELEMENT_Y] << 23 | SrcSelect[ELEMENT_Z] << 26 |
238 SrcSelect[ELEMENT_W] << 29 | Offsets[0] << 0 | Offsets[1] << 5 |
241 EmitByte(INSTR_TEX, OS);
246 case AMDGPU::EG_ExportSwz:
247 case AMDGPU::R600_ExportSwz:
248 case AMDGPU::EG_ExportBuf:
249 case AMDGPU::R600_ExportBuf: {
250 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
251 EmitByte(INSTR_EXPORT, OS);
256 case AMDGPU::CF_ALU_PUSH_BEFORE: {
257 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
258 EmitByte(INSTR_CFALU, OS);
264 EmitALUInstr(MI, Fixups, OS);
270 void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
271 SmallVectorImpl<MCFixup> &Fixups,
272 raw_ostream &OS) const {
273 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
275 // Emit instruction type
276 EmitByte(INSTR_ALU, OS);
278 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
280 //older alu have different encoding for instructions with one or two src
282 if ((STI.getFeatureBits() & AMDGPU::FeatureR600ALUInst) &&
283 !(MCDesc.TSFlags & R600_InstFlag::OP3)) {
284 uint64_t ISAOpCode = InstWord01 & (0x3FFULL << 39);
285 InstWord01 &= ~(0x3FFULL << 39);
286 InstWord01 |= ISAOpCode << 1;
289 unsigned SrcNum = MCDesc.TSFlags & R600_InstFlag::OP3 ? 3 :
290 MCDesc.TSFlags & R600_InstFlag::OP2 ? 2 : 1;
292 EmitByte(SrcNum, OS);
294 const unsigned SrcOps[3][2] = {
295 {R600Operands::SRC0, R600Operands::SRC0_SEL},
296 {R600Operands::SRC1, R600Operands::SRC1_SEL},
297 {R600Operands::SRC2, R600Operands::SRC2_SEL}
300 for (unsigned SrcIdx = 0; SrcIdx < SrcNum; ++SrcIdx) {
301 unsigned RegOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][0]];
302 unsigned SelOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][1]];
303 EmitSrcISA(MI, RegOpIdx, SelOpIdx, OS);
306 Emit(InstWord01, OS);
310 void R600MCCodeEmitter::EmitSrc(const MCInst &MI, unsigned OpIdx,
311 raw_ostream &OS) const {
312 const MCOperand &MO = MI.getOperand(OpIdx);
318 // Emit the source select (2 bytes). For GPRs, this is the register index.
319 // For other potential instruction operands, (e.g. constant registers) the
320 // value of the source select is defined in the r600isa docs.
322 unsigned reg = MO.getReg();
323 EmitTwoBytes(getHWReg(reg), OS);
324 if (reg == AMDGPU::ALU_LITERAL_X) {
325 unsigned ImmOpIndex = MI.getNumOperands() - 1;
326 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
327 if (ImmOp.isFPImm()) {
328 Value.f = ImmOp.getFPImm();
330 assert(ImmOp.isImm());
331 Value.i = ImmOp.getImm();
335 // XXX: Handle other operand types.
339 // Emit the source channel (1 byte)
341 EmitByte(getHWRegChan(MO.getReg()), OS);
346 // XXX: Emit isNegated (1 byte)
347 if ((!(isFlagSet(MI, OpIdx, MO_FLAG_ABS)))
348 && (isFlagSet(MI, OpIdx, MO_FLAG_NEG) ||
350 (MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
356 // Emit isAbsolute (1 byte)
357 if (isFlagSet(MI, OpIdx, MO_FLAG_ABS)) {
363 // XXX: Emit relative addressing mode (1 byte)
366 // Emit kc_bank, This will be adjusted later by r600_asm
369 // Emit the literal value, if applicable (4 bytes).
374 void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
375 unsigned SelOpIdx, raw_ostream &OS) const {
376 const MCOperand &RegMO = MI.getOperand(RegOpIdx);
377 const MCOperand &SelMO = MI.getOperand(SelOpIdx);
383 InlineConstant.i = 0;
384 // Emit source type (1 byte) and source select (4 bytes). For GPRs type is 0
385 // and select is 0 (GPR index is encoded in the instr encoding. For constants
386 // type is 1 and select is the original const select passed from the driver.
387 unsigned Reg = RegMO.getReg();
388 if (Reg == AMDGPU::ALU_CONST) {
390 uint32_t Sel = SelMO.getImm();
394 Emit((uint32_t)0, OS);
397 if (Reg == AMDGPU::ALU_LITERAL_X) {
398 unsigned ImmOpIndex = MI.getNumOperands() - 1;
399 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
400 if (ImmOp.isFPImm()) {
401 InlineConstant.f = ImmOp.getFPImm();
403 assert(ImmOp.isImm());
404 InlineConstant.i = ImmOp.getImm();
408 // Emit the literal value, if applicable (4 bytes).
409 Emit(InlineConstant.i, OS);
412 void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
414 // Emit instruction type
415 EmitByte(INSTR_FC, OS);
418 unsigned NumOperands = MI.getNumOperands();
419 if (NumOperands > 0) {
420 assert(NumOperands == 1);
423 EmitNullBytes(SRC_BYTE_COUNT, OS);
426 // Emit FC Instruction
428 switch (MI.getOpcode()) {
429 case AMDGPU::PREDICATED_BREAK:
430 instr = FC_BREAK_PREDICATE;
432 case AMDGPU::CONTINUE:
435 case AMDGPU::IF_PREDICATE_SET:
436 instr = FC_IF_PREDICATE;
444 case AMDGPU::ENDLOOP:
447 case AMDGPU::WHILELOOP:
457 void R600MCCodeEmitter::EmitNullBytes(unsigned int ByteCount,
458 raw_ostream &OS) const {
460 for (unsigned int i = 0; i < ByteCount; i++) {
465 void R600MCCodeEmitter::EmitByte(unsigned int Byte, raw_ostream &OS) const {
466 OS.write((uint8_t) Byte & 0xff);
469 void R600MCCodeEmitter::EmitTwoBytes(unsigned int Bytes,
470 raw_ostream &OS) const {
471 OS.write((uint8_t) (Bytes & 0xff));
472 OS.write((uint8_t) ((Bytes >> 8) & 0xff));
475 void R600MCCodeEmitter::Emit(uint32_t Value, raw_ostream &OS) const {
476 for (unsigned i = 0; i < 4; i++) {
477 OS.write((uint8_t) ((Value >> (8 * i)) & 0xff));
481 void R600MCCodeEmitter::Emit(uint64_t Value, raw_ostream &OS) const {
482 for (unsigned i = 0; i < 8; i++) {
483 EmitByte((Value >> (8 * i)) & 0xff, OS);
487 unsigned R600MCCodeEmitter::getHWRegChan(unsigned reg) const {
488 return MRI.getEncodingValue(reg) >> HW_CHAN_SHIFT;
491 unsigned R600MCCodeEmitter::getHWReg(unsigned RegNo) const {
492 return MRI.getEncodingValue(RegNo) & HW_REG_MASK;
495 uint64_t R600MCCodeEmitter::getMachineOpValue(const MCInst &MI,
497 SmallVectorImpl<MCFixup> &Fixup) const {
499 if (HAS_NATIVE_OPERANDS(MCII.get(MI.getOpcode()).TSFlags)) {
500 return MRI.getEncodingValue(MO.getReg());
502 return getHWReg(MO.getReg());
504 } else if (MO.isImm()) {
512 //===----------------------------------------------------------------------===//
513 // Encoding helper functions
514 //===----------------------------------------------------------------------===//
516 bool R600MCCodeEmitter::isFCOp(unsigned opcode) const {
518 default: return false;
519 case AMDGPU::PREDICATED_BREAK:
520 case AMDGPU::CONTINUE:
521 case AMDGPU::IF_PREDICATE_SET:
524 case AMDGPU::ENDLOOP:
525 case AMDGPU::WHILELOOP:
530 bool R600MCCodeEmitter::isTexOp(unsigned opcode) const {
532 default: return false;
534 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
535 case AMDGPU::TEX_SAMPLE:
536 case AMDGPU::TEX_SAMPLE_C:
537 case AMDGPU::TEX_SAMPLE_L:
538 case AMDGPU::TEX_SAMPLE_C_L:
539 case AMDGPU::TEX_SAMPLE_LB:
540 case AMDGPU::TEX_SAMPLE_C_LB:
541 case AMDGPU::TEX_SAMPLE_G:
542 case AMDGPU::TEX_SAMPLE_C_G:
543 case AMDGPU::TEX_GET_GRADIENTS_H:
544 case AMDGPU::TEX_GET_GRADIENTS_V:
545 case AMDGPU::TEX_SET_GRADIENTS_H:
546 case AMDGPU::TEX_SET_GRADIENTS_V:
551 bool R600MCCodeEmitter::isFlagSet(const MCInst &MI, unsigned Operand,
552 unsigned Flag) const {
553 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
554 unsigned FlagIndex = GET_FLAG_OPERAND_IDX(MCDesc.TSFlags);
555 if (FlagIndex == 0) {
558 assert(MI.getOperand(FlagIndex).isImm());
559 return !!((MI.getOperand(FlagIndex).getImm() >>
560 (NUM_MO_FLAGS * Operand)) & Flag);
563 #include "AMDGPUGenMCCodeEmitter.inc"