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::FETCH_CLAUSE ||
146 MI.getOpcode() == AMDGPU::ALU_CLAUSE ||
147 MI.getOpcode() == AMDGPU::BUNDLE ||
148 MI.getOpcode() == AMDGPU::KILL) {
151 switch(MI.getOpcode()) {
152 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
153 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
154 uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
155 EmitByte(INSTR_NATIVE, OS);
159 case AMDGPU::CONSTANT_LOAD_eg:
160 case AMDGPU::VTX_READ_PARAM_8_eg:
161 case AMDGPU::VTX_READ_PARAM_16_eg:
162 case AMDGPU::VTX_READ_PARAM_32_eg:
163 case AMDGPU::VTX_READ_PARAM_128_eg:
164 case AMDGPU::VTX_READ_GLOBAL_8_eg:
165 case AMDGPU::VTX_READ_GLOBAL_32_eg:
166 case AMDGPU::VTX_READ_GLOBAL_128_eg:
167 case AMDGPU::TEX_VTX_CONSTBUF:
168 case AMDGPU::TEX_VTX_TEXBUF : {
169 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
170 uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
171 InstWord2 |= 1 << 19;
173 EmitByte(INSTR_NATIVE, OS);
174 Emit(InstWord01, OS);
175 EmitByte(INSTR_NATIVE, OS);
177 Emit((u_int32_t) 0, OS);
181 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
182 case AMDGPU::TEX_SAMPLE:
183 case AMDGPU::TEX_SAMPLE_C:
184 case AMDGPU::TEX_SAMPLE_L:
185 case AMDGPU::TEX_SAMPLE_C_L:
186 case AMDGPU::TEX_SAMPLE_LB:
187 case AMDGPU::TEX_SAMPLE_C_LB:
188 case AMDGPU::TEX_SAMPLE_G:
189 case AMDGPU::TEX_SAMPLE_C_G:
190 case AMDGPU::TEX_GET_GRADIENTS_H:
191 case AMDGPU::TEX_GET_GRADIENTS_V:
192 case AMDGPU::TEX_SET_GRADIENTS_H:
193 case AMDGPU::TEX_SET_GRADIENTS_V: {
194 unsigned Opcode = MI.getOpcode();
195 bool HasOffsets = (Opcode == AMDGPU::TEX_LD);
196 unsigned OpOffset = HasOffsets ? 3 : 0;
197 int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
198 int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
200 uint32_t SrcSelect[4] = {0, 1, 2, 3};
201 uint32_t Offsets[3] = {0, 0, 0};
202 uint64_t CoordType[4] = {1, 1, 1, 1};
205 for (unsigned i = 0; i < 3; i++) {
206 int SignedOffset = MI.getOperand(i + 2).getImm();
207 Offsets[i] = (SignedOffset & 0x1F);
211 if (TextureType == TEXTURE_RECT ||
212 TextureType == TEXTURE_SHADOWRECT) {
213 CoordType[ELEMENT_X] = 0;
214 CoordType[ELEMENT_Y] = 0;
217 if (TextureType == TEXTURE_1D_ARRAY ||
218 TextureType == TEXTURE_SHADOW1D_ARRAY) {
219 if (Opcode == AMDGPU::TEX_SAMPLE_C_L ||
220 Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
221 CoordType[ELEMENT_Y] = 0;
223 CoordType[ELEMENT_Z] = 0;
224 SrcSelect[ELEMENT_Z] = ELEMENT_Y;
226 } else if (TextureType == TEXTURE_2D_ARRAY ||
227 TextureType == TEXTURE_SHADOW2D_ARRAY) {
228 CoordType[ELEMENT_Z] = 0;
232 if ((TextureType == TEXTURE_SHADOW1D ||
233 TextureType == TEXTURE_SHADOW2D ||
234 TextureType == TEXTURE_SHADOWRECT ||
235 TextureType == TEXTURE_SHADOW1D_ARRAY) &&
236 Opcode != AMDGPU::TEX_SAMPLE_C_L &&
237 Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
238 SrcSelect[ELEMENT_W] = ELEMENT_Z;
241 uint64_t Word01 = getBinaryCodeForInstr(MI, Fixups) |
242 CoordType[ELEMENT_X] << 60 | CoordType[ELEMENT_Y] << 61 |
243 CoordType[ELEMENT_Z] << 62 | CoordType[ELEMENT_W] << 63;
244 uint32_t Word2 = Sampler << 15 | SrcSelect[ELEMENT_X] << 20 |
245 SrcSelect[ELEMENT_Y] << 23 | SrcSelect[ELEMENT_Z] << 26 |
246 SrcSelect[ELEMENT_W] << 29 | Offsets[0] << 0 | Offsets[1] << 5 |
249 EmitByte(INSTR_NATIVE, OS);
251 EmitByte(INSTR_NATIVE, OS);
253 Emit((u_int32_t) 0, OS);
257 case AMDGPU::CF_ALU_PUSH_BEFORE: {
258 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
259 EmitByte(INSTR_NATIVE, OS);
263 case AMDGPU::CF_CALL_FS_EG:
264 case AMDGPU::CF_CALL_FS_R600:
265 case AMDGPU::CF_TC_EG:
266 case AMDGPU::CF_VC_EG:
267 case AMDGPU::CF_TC_R600:
268 case AMDGPU::CF_VC_R600:
269 case AMDGPU::WHILE_LOOP_EG:
270 case AMDGPU::END_LOOP_EG:
271 case AMDGPU::LOOP_BREAK_EG:
272 case AMDGPU::CF_CONTINUE_EG:
273 case AMDGPU::CF_JUMP_EG:
274 case AMDGPU::CF_ELSE_EG:
276 case AMDGPU::WHILE_LOOP_R600:
277 case AMDGPU::END_LOOP_R600:
278 case AMDGPU::LOOP_BREAK_R600:
279 case AMDGPU::CF_CONTINUE_R600:
280 case AMDGPU::CF_JUMP_R600:
281 case AMDGPU::CF_ELSE_R600:
282 case AMDGPU::POP_R600:
283 case AMDGPU::EG_ExportSwz:
284 case AMDGPU::R600_ExportSwz:
285 case AMDGPU::EG_ExportBuf:
286 case AMDGPU::R600_ExportBuf:
288 case AMDGPU::CF_END_R600:
289 case AMDGPU::CF_END_EG:
290 case AMDGPU::CF_END_CM: {
291 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
292 EmitByte(INSTR_NATIVE, OS);
297 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
298 EmitByte(INSTR_NATIVE, OS);
305 void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
306 SmallVectorImpl<MCFixup> &Fixups,
307 raw_ostream &OS) const {
308 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
310 // Emit instruction type
311 EmitByte(INSTR_ALU, OS);
313 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
315 //older alu have different encoding for instructions with one or two src
317 if ((STI.getFeatureBits() & AMDGPU::FeatureR600ALUInst) &&
318 !(MCDesc.TSFlags & R600_InstFlag::OP3)) {
319 uint64_t ISAOpCode = InstWord01 & (0x3FFULL << 39);
320 InstWord01 &= ~(0x3FFULL << 39);
321 InstWord01 |= ISAOpCode << 1;
324 unsigned SrcNum = MCDesc.TSFlags & R600_InstFlag::OP3 ? 3 :
325 MCDesc.TSFlags & R600_InstFlag::OP2 ? 2 : 1;
327 EmitByte(SrcNum, OS);
329 const unsigned SrcOps[3][2] = {
330 {R600Operands::SRC0, R600Operands::SRC0_SEL},
331 {R600Operands::SRC1, R600Operands::SRC1_SEL},
332 {R600Operands::SRC2, R600Operands::SRC2_SEL}
335 for (unsigned SrcIdx = 0; SrcIdx < SrcNum; ++SrcIdx) {
336 unsigned RegOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][0]];
337 unsigned SelOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][1]];
338 EmitSrcISA(MI, RegOpIdx, SelOpIdx, OS);
341 Emit(InstWord01, OS);
345 void R600MCCodeEmitter::EmitSrc(const MCInst &MI, unsigned OpIdx,
346 raw_ostream &OS) const {
347 const MCOperand &MO = MI.getOperand(OpIdx);
353 // Emit the source select (2 bytes). For GPRs, this is the register index.
354 // For other potential instruction operands, (e.g. constant registers) the
355 // value of the source select is defined in the r600isa docs.
357 unsigned reg = MO.getReg();
358 EmitTwoBytes(getHWReg(reg), OS);
359 if (reg == AMDGPU::ALU_LITERAL_X) {
360 unsigned ImmOpIndex = MI.getNumOperands() - 1;
361 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
362 if (ImmOp.isFPImm()) {
363 Value.f = ImmOp.getFPImm();
365 assert(ImmOp.isImm());
366 Value.i = ImmOp.getImm();
370 // XXX: Handle other operand types.
374 // Emit the source channel (1 byte)
376 EmitByte(getHWRegChan(MO.getReg()), OS);
381 // XXX: Emit isNegated (1 byte)
382 if ((!(isFlagSet(MI, OpIdx, MO_FLAG_ABS)))
383 && (isFlagSet(MI, OpIdx, MO_FLAG_NEG) ||
385 (MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
391 // Emit isAbsolute (1 byte)
392 if (isFlagSet(MI, OpIdx, MO_FLAG_ABS)) {
398 // XXX: Emit relative addressing mode (1 byte)
401 // Emit kc_bank, This will be adjusted later by r600_asm
404 // Emit the literal value, if applicable (4 bytes).
409 void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
410 unsigned SelOpIdx, raw_ostream &OS) const {
411 const MCOperand &RegMO = MI.getOperand(RegOpIdx);
412 const MCOperand &SelMO = MI.getOperand(SelOpIdx);
418 InlineConstant.i = 0;
419 // Emit source type (1 byte) and source select (4 bytes). For GPRs type is 0
420 // and select is 0 (GPR index is encoded in the instr encoding. For constants
421 // type is 1 and select is the original const select passed from the driver.
422 unsigned Reg = RegMO.getReg();
423 if (Reg == AMDGPU::ALU_CONST) {
425 uint32_t Sel = SelMO.getImm();
429 Emit((uint32_t)0, OS);
432 if (Reg == AMDGPU::ALU_LITERAL_X) {
433 unsigned ImmOpIndex = MI.getNumOperands() - 2;
434 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
435 if (ImmOp.isFPImm()) {
436 InlineConstant.f = ImmOp.getFPImm();
438 assert(ImmOp.isImm());
439 InlineConstant.i = ImmOp.getImm();
443 // Emit the literal value, if applicable (4 bytes).
444 Emit(InlineConstant.i, OS);
447 void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
449 // Emit instruction type
450 EmitByte(INSTR_FC, OS);
453 unsigned NumOperands = MI.getNumOperands();
454 if (NumOperands > 0) {
455 assert(NumOperands == 1);
458 EmitNullBytes(SRC_BYTE_COUNT, OS);
461 // Emit FC Instruction
463 switch (MI.getOpcode()) {
464 case AMDGPU::PREDICATED_BREAK:
465 instr = FC_BREAK_PREDICATE;
467 case AMDGPU::CONTINUE:
470 case AMDGPU::IF_PREDICATE_SET:
471 instr = FC_IF_PREDICATE;
479 case AMDGPU::ENDLOOP:
482 case AMDGPU::WHILELOOP:
492 void R600MCCodeEmitter::EmitNullBytes(unsigned int ByteCount,
493 raw_ostream &OS) const {
495 for (unsigned int i = 0; i < ByteCount; i++) {
500 void R600MCCodeEmitter::EmitByte(unsigned int Byte, raw_ostream &OS) const {
501 OS.write((uint8_t) Byte & 0xff);
504 void R600MCCodeEmitter::EmitTwoBytes(unsigned int Bytes,
505 raw_ostream &OS) const {
506 OS.write((uint8_t) (Bytes & 0xff));
507 OS.write((uint8_t) ((Bytes >> 8) & 0xff));
510 void R600MCCodeEmitter::Emit(uint32_t Value, raw_ostream &OS) const {
511 for (unsigned i = 0; i < 4; i++) {
512 OS.write((uint8_t) ((Value >> (8 * i)) & 0xff));
516 void R600MCCodeEmitter::Emit(uint64_t Value, raw_ostream &OS) const {
517 for (unsigned i = 0; i < 8; i++) {
518 EmitByte((Value >> (8 * i)) & 0xff, OS);
522 unsigned R600MCCodeEmitter::getHWRegChan(unsigned reg) const {
523 return MRI.getEncodingValue(reg) >> HW_CHAN_SHIFT;
526 unsigned R600MCCodeEmitter::getHWReg(unsigned RegNo) const {
527 return MRI.getEncodingValue(RegNo) & HW_REG_MASK;
530 uint64_t R600MCCodeEmitter::getMachineOpValue(const MCInst &MI,
532 SmallVectorImpl<MCFixup> &Fixup) const {
534 if (HAS_NATIVE_OPERANDS(MCII.get(MI.getOpcode()).TSFlags)) {
535 return MRI.getEncodingValue(MO.getReg());
537 return getHWReg(MO.getReg());
539 } else if (MO.isImm()) {
547 //===----------------------------------------------------------------------===//
548 // Encoding helper functions
549 //===----------------------------------------------------------------------===//
551 bool R600MCCodeEmitter::isFCOp(unsigned opcode) const {
553 default: return false;
554 case AMDGPU::PREDICATED_BREAK:
555 case AMDGPU::CONTINUE:
556 case AMDGPU::IF_PREDICATE_SET:
559 case AMDGPU::ENDLOOP:
560 case AMDGPU::WHILELOOP:
565 bool R600MCCodeEmitter::isTexOp(unsigned opcode) const {
567 default: return false;
569 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
570 case AMDGPU::TEX_SAMPLE:
571 case AMDGPU::TEX_SAMPLE_C:
572 case AMDGPU::TEX_SAMPLE_L:
573 case AMDGPU::TEX_SAMPLE_C_L:
574 case AMDGPU::TEX_SAMPLE_LB:
575 case AMDGPU::TEX_SAMPLE_C_LB:
576 case AMDGPU::TEX_SAMPLE_G:
577 case AMDGPU::TEX_SAMPLE_C_G:
578 case AMDGPU::TEX_GET_GRADIENTS_H:
579 case AMDGPU::TEX_GET_GRADIENTS_V:
580 case AMDGPU::TEX_SET_GRADIENTS_H:
581 case AMDGPU::TEX_SET_GRADIENTS_V:
586 bool R600MCCodeEmitter::isFlagSet(const MCInst &MI, unsigned Operand,
587 unsigned Flag) const {
588 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
589 unsigned FlagIndex = GET_FLAG_OPERAND_IDX(MCDesc.TSFlags);
590 if (FlagIndex == 0) {
593 assert(MI.getOperand(FlagIndex).isImm());
594 return !!((MI.getOperand(FlagIndex).getImm() >>
595 (NUM_MO_FLAGS * Operand)) & Flag);
598 #include "AMDGPUGenMCCodeEmitter.inc"