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 /// \brief The R600 code emitter produces machine code that can be executed
13 /// directly on the GPU device.
15 //===----------------------------------------------------------------------===//
17 #include "R600Defines.h"
18 #include "MCTargetDesc/AMDGPUMCCodeEmitter.h"
19 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
20 #include "llvm/MC/MCCodeEmitter.h"
21 #include "llvm/MC/MCContext.h"
22 #include "llvm/MC/MCInst.h"
23 #include "llvm/MC/MCInstrInfo.h"
24 #include "llvm/MC/MCRegisterInfo.h"
25 #include "llvm/MC/MCSubtargetInfo.h"
26 #include "llvm/Support/raw_ostream.h"
29 #define SRC_BYTE_COUNT 11
30 #define DST_BYTE_COUNT 5
36 class R600MCCodeEmitter : public AMDGPUMCCodeEmitter {
37 R600MCCodeEmitter(const R600MCCodeEmitter &) LLVM_DELETED_FUNCTION;
38 void operator=(const R600MCCodeEmitter &) LLVM_DELETED_FUNCTION;
39 const MCInstrInfo &MCII;
40 const MCRegisterInfo &MRI;
41 const MCSubtargetInfo &STI;
46 R600MCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
47 const MCSubtargetInfo &sti, MCContext &ctx)
48 : MCII(mcii), MRI(mri), STI(sti), Ctx(ctx) { }
50 /// \brief Encode the instruction and write it to the OS.
51 virtual void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
52 SmallVectorImpl<MCFixup> &Fixups) const;
54 /// \returns the encoding for an MCOperand.
55 virtual uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
56 SmallVectorImpl<MCFixup> &Fixups) const;
59 void EmitALUInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
60 raw_ostream &OS) const;
61 void EmitSrc(const MCInst &MI, unsigned OpIdx, raw_ostream &OS) const;
62 void EmitSrcISA(const MCInst &MI, unsigned RegOpIdx, unsigned SelOpIdx,
63 raw_ostream &OS) const;
64 void EmitDst(const MCInst &MI, raw_ostream &OS) const;
65 void EmitFCInstr(const MCInst &MI, raw_ostream &OS) const;
67 void EmitNullBytes(unsigned int byteCount, raw_ostream &OS) const;
69 void EmitByte(unsigned int byte, raw_ostream &OS) const;
71 void EmitTwoBytes(uint32_t bytes, raw_ostream &OS) const;
73 void Emit(uint32_t value, raw_ostream &OS) const;
74 void Emit(uint64_t value, raw_ostream &OS) const;
76 unsigned getHWRegChan(unsigned reg) const;
77 unsigned getHWReg(unsigned regNo) const;
79 bool isFCOp(unsigned opcode) const;
80 bool isTexOp(unsigned opcode) const;
81 bool isFlagSet(const MCInst &MI, unsigned Operand, unsigned Flag) const;
85 } // End anonymous namespace
115 TEXTURE_SHADOW1D_ARRAY,
116 TEXTURE_SHADOW2D_ARRAY
119 MCCodeEmitter *llvm::createR600MCCodeEmitter(const MCInstrInfo &MCII,
120 const MCRegisterInfo &MRI,
121 const MCSubtargetInfo &STI,
123 return new R600MCCodeEmitter(MCII, MRI, STI, Ctx);
126 void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
127 SmallVectorImpl<MCFixup> &Fixups) const {
128 if (isFCOp(MI.getOpcode())){
130 } else if (MI.getOpcode() == AMDGPU::RETURN ||
131 MI.getOpcode() == AMDGPU::FETCH_CLAUSE ||
132 MI.getOpcode() == AMDGPU::ALU_CLAUSE ||
133 MI.getOpcode() == AMDGPU::BUNDLE ||
134 MI.getOpcode() == AMDGPU::KILL) {
137 switch(MI.getOpcode()) {
138 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
139 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
140 uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
144 case AMDGPU::CONSTANT_LOAD_eg:
145 case AMDGPU::VTX_READ_PARAM_8_eg:
146 case AMDGPU::VTX_READ_PARAM_16_eg:
147 case AMDGPU::VTX_READ_PARAM_32_eg:
148 case AMDGPU::VTX_READ_PARAM_128_eg:
149 case AMDGPU::VTX_READ_GLOBAL_8_eg:
150 case AMDGPU::VTX_READ_GLOBAL_32_eg:
151 case AMDGPU::VTX_READ_GLOBAL_128_eg:
152 case AMDGPU::TEX_VTX_CONSTBUF:
153 case AMDGPU::TEX_VTX_TEXBUF : {
154 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
155 uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
156 InstWord2 |= 1 << 19;
158 Emit(InstWord01, OS);
160 Emit((u_int32_t) 0, OS);
164 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
165 case AMDGPU::TEX_SAMPLE:
166 case AMDGPU::TEX_SAMPLE_C:
167 case AMDGPU::TEX_SAMPLE_L:
168 case AMDGPU::TEX_SAMPLE_C_L:
169 case AMDGPU::TEX_SAMPLE_LB:
170 case AMDGPU::TEX_SAMPLE_C_LB:
171 case AMDGPU::TEX_SAMPLE_G:
172 case AMDGPU::TEX_SAMPLE_C_G:
173 case AMDGPU::TEX_GET_GRADIENTS_H:
174 case AMDGPU::TEX_GET_GRADIENTS_V:
175 case AMDGPU::TEX_SET_GRADIENTS_H:
176 case AMDGPU::TEX_SET_GRADIENTS_V: {
177 unsigned Opcode = MI.getOpcode();
178 bool HasOffsets = (Opcode == AMDGPU::TEX_LD);
179 unsigned OpOffset = HasOffsets ? 3 : 0;
180 int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
181 int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
183 uint32_t SrcSelect[4] = {0, 1, 2, 3};
184 uint32_t Offsets[3] = {0, 0, 0};
185 uint64_t CoordType[4] = {1, 1, 1, 1};
188 for (unsigned i = 0; i < 3; i++) {
189 int SignedOffset = MI.getOperand(i + 2).getImm();
190 Offsets[i] = (SignedOffset & 0x1F);
194 if (TextureType == TEXTURE_RECT ||
195 TextureType == TEXTURE_SHADOWRECT) {
196 CoordType[ELEMENT_X] = 0;
197 CoordType[ELEMENT_Y] = 0;
200 if (TextureType == TEXTURE_1D_ARRAY ||
201 TextureType == TEXTURE_SHADOW1D_ARRAY) {
202 if (Opcode == AMDGPU::TEX_SAMPLE_C_L ||
203 Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
204 CoordType[ELEMENT_Y] = 0;
206 CoordType[ELEMENT_Z] = 0;
207 SrcSelect[ELEMENT_Z] = ELEMENT_Y;
209 } else if (TextureType == TEXTURE_2D_ARRAY ||
210 TextureType == TEXTURE_SHADOW2D_ARRAY) {
211 CoordType[ELEMENT_Z] = 0;
215 if ((TextureType == TEXTURE_SHADOW1D ||
216 TextureType == TEXTURE_SHADOW2D ||
217 TextureType == TEXTURE_SHADOWRECT ||
218 TextureType == TEXTURE_SHADOW1D_ARRAY) &&
219 Opcode != AMDGPU::TEX_SAMPLE_C_L &&
220 Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
221 SrcSelect[ELEMENT_W] = ELEMENT_Z;
224 uint64_t Word01 = getBinaryCodeForInstr(MI, Fixups) |
225 CoordType[ELEMENT_X] << 60 | CoordType[ELEMENT_Y] << 61 |
226 CoordType[ELEMENT_Z] << 62 | CoordType[ELEMENT_W] << 63;
227 uint32_t Word2 = Sampler << 15 | SrcSelect[ELEMENT_X] << 20 |
228 SrcSelect[ELEMENT_Y] << 23 | SrcSelect[ELEMENT_Z] << 26 |
229 SrcSelect[ELEMENT_W] << 29 | Offsets[0] << 0 | Offsets[1] << 5 |
234 Emit((u_int32_t) 0, OS);
238 case AMDGPU::CF_ALU_PUSH_BEFORE: {
239 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
243 case AMDGPU::CF_CALL_FS_EG:
244 case AMDGPU::CF_CALL_FS_R600:
245 case AMDGPU::CF_TC_EG:
246 case AMDGPU::CF_VC_EG:
247 case AMDGPU::CF_TC_R600:
248 case AMDGPU::CF_VC_R600:
249 case AMDGPU::WHILE_LOOP_EG:
250 case AMDGPU::END_LOOP_EG:
251 case AMDGPU::LOOP_BREAK_EG:
252 case AMDGPU::CF_CONTINUE_EG:
253 case AMDGPU::CF_JUMP_EG:
254 case AMDGPU::CF_ELSE_EG:
256 case AMDGPU::WHILE_LOOP_R600:
257 case AMDGPU::END_LOOP_R600:
258 case AMDGPU::LOOP_BREAK_R600:
259 case AMDGPU::CF_CONTINUE_R600:
260 case AMDGPU::CF_JUMP_R600:
261 case AMDGPU::CF_ELSE_R600:
262 case AMDGPU::POP_R600:
263 case AMDGPU::EG_ExportSwz:
264 case AMDGPU::R600_ExportSwz:
265 case AMDGPU::EG_ExportBuf:
266 case AMDGPU::R600_ExportBuf:
268 case AMDGPU::CF_END_R600:
269 case AMDGPU::CF_END_EG:
270 case AMDGPU::CF_END_CM: {
271 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
276 uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
283 void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
284 SmallVectorImpl<MCFixup> &Fixups,
285 raw_ostream &OS) const {
286 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
288 uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
290 //older alu have different encoding for instructions with one or two src
292 if ((STI.getFeatureBits() & AMDGPU::FeatureR600ALUInst) &&
293 !(MCDesc.TSFlags & R600_InstFlag::OP3)) {
294 uint64_t ISAOpCode = InstWord01 & (0x3FFULL << 39);
295 InstWord01 &= ~(0x3FFULL << 39);
296 InstWord01 |= ISAOpCode << 1;
299 unsigned SrcNum = MCDesc.TSFlags & R600_InstFlag::OP3 ? 3 :
300 MCDesc.TSFlags & R600_InstFlag::OP2 ? 2 : 1;
302 const unsigned SrcOps[3][2] = {
303 {R600Operands::SRC0, R600Operands::SRC0_SEL},
304 {R600Operands::SRC1, R600Operands::SRC1_SEL},
305 {R600Operands::SRC2, R600Operands::SRC2_SEL}
308 for (unsigned SrcIdx = 0; SrcIdx < SrcNum; ++SrcIdx) {
309 unsigned RegOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][0]];
310 unsigned SelOpIdx = R600Operands::ALUOpTable[SrcNum-1][SrcOps[SrcIdx][1]];
313 Emit(InstWord01, OS);
317 void R600MCCodeEmitter::EmitSrc(const MCInst &MI, unsigned OpIdx,
318 raw_ostream &OS) const {
319 const MCOperand &MO = MI.getOperand(OpIdx);
325 // Emit the source select (2 bytes). For GPRs, this is the register index.
326 // For other potential instruction operands, (e.g. constant registers) the
327 // value of the source select is defined in the r600isa docs.
329 unsigned reg = MO.getReg();
330 EmitTwoBytes(getHWReg(reg), OS);
331 if (reg == AMDGPU::ALU_LITERAL_X) {
332 unsigned ImmOpIndex = MI.getNumOperands() - 1;
333 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
334 if (ImmOp.isFPImm()) {
335 Value.f = ImmOp.getFPImm();
337 assert(ImmOp.isImm());
338 Value.i = ImmOp.getImm();
342 // XXX: Handle other operand types.
346 // Emit the source channel (1 byte)
348 EmitByte(getHWRegChan(MO.getReg()), OS);
353 // XXX: Emit isNegated (1 byte)
354 if ((!(isFlagSet(MI, OpIdx, MO_FLAG_ABS)))
355 && (isFlagSet(MI, OpIdx, MO_FLAG_NEG) ||
357 (MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
363 // Emit isAbsolute (1 byte)
364 if (isFlagSet(MI, OpIdx, MO_FLAG_ABS)) {
370 // XXX: Emit relative addressing mode (1 byte)
373 // Emit kc_bank, This will be adjusted later by r600_asm
376 // Emit the literal value, if applicable (4 bytes).
381 void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned RegOpIdx,
382 unsigned SelOpIdx, raw_ostream &OS) const {
383 const MCOperand &RegMO = MI.getOperand(RegOpIdx);
384 const MCOperand &SelMO = MI.getOperand(SelOpIdx);
390 InlineConstant.i = 0;
391 // Emit source type (1 byte) and source select (4 bytes). For GPRs type is 0
392 // and select is 0 (GPR index is encoded in the instr encoding. For constants
393 // type is 1 and select is the original const select passed from the driver.
394 unsigned Reg = RegMO.getReg();
395 if (Reg == AMDGPU::ALU_CONST) {
397 uint32_t Sel = SelMO.getImm();
401 Emit((uint32_t)0, OS);
404 if (Reg == AMDGPU::ALU_LITERAL_X) {
405 unsigned ImmOpIndex = MI.getNumOperands() - 2;
406 MCOperand ImmOp = MI.getOperand(ImmOpIndex);
407 if (ImmOp.isFPImm()) {
408 InlineConstant.f = ImmOp.getFPImm();
410 assert(ImmOp.isImm());
411 InlineConstant.i = ImmOp.getImm();
415 // Emit the literal value, if applicable (4 bytes).
416 Emit(InlineConstant.i, OS);
419 void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
422 unsigned NumOperands = MI.getNumOperands();
423 if (NumOperands > 0) {
424 assert(NumOperands == 1);
427 EmitNullBytes(SRC_BYTE_COUNT, OS);
430 // Emit FC Instruction
432 switch (MI.getOpcode()) {
433 case AMDGPU::PREDICATED_BREAK:
434 instr = FC_BREAK_PREDICATE;
436 case AMDGPU::CONTINUE:
439 case AMDGPU::IF_PREDICATE_SET:
440 instr = FC_IF_PREDICATE;
448 case AMDGPU::ENDLOOP:
451 case AMDGPU::WHILELOOP:
461 void R600MCCodeEmitter::EmitNullBytes(unsigned int ByteCount,
462 raw_ostream &OS) const {
464 for (unsigned int i = 0; i < ByteCount; i++) {
469 void R600MCCodeEmitter::EmitByte(unsigned int Byte, raw_ostream &OS) const {
470 OS.write((uint8_t) Byte & 0xff);
473 void R600MCCodeEmitter::EmitTwoBytes(unsigned int Bytes,
474 raw_ostream &OS) const {
475 OS.write((uint8_t) (Bytes & 0xff));
476 OS.write((uint8_t) ((Bytes >> 8) & 0xff));
479 void R600MCCodeEmitter::Emit(uint32_t Value, raw_ostream &OS) const {
480 for (unsigned i = 0; i < 4; i++) {
481 OS.write((uint8_t) ((Value >> (8 * i)) & 0xff));
485 void R600MCCodeEmitter::Emit(uint64_t Value, raw_ostream &OS) const {
486 for (unsigned i = 0; i < 8; i++) {
487 EmitByte((Value >> (8 * i)) & 0xff, OS);
491 unsigned R600MCCodeEmitter::getHWRegChan(unsigned reg) const {
492 return MRI.getEncodingValue(reg) >> HW_CHAN_SHIFT;
495 unsigned R600MCCodeEmitter::getHWReg(unsigned RegNo) const {
496 return MRI.getEncodingValue(RegNo) & HW_REG_MASK;
499 uint64_t R600MCCodeEmitter::getMachineOpValue(const MCInst &MI,
501 SmallVectorImpl<MCFixup> &Fixup) const {
503 if (HAS_NATIVE_OPERANDS(MCII.get(MI.getOpcode()).TSFlags)) {
504 return MRI.getEncodingValue(MO.getReg());
506 return getHWReg(MO.getReg());
508 } else if (MO.isImm()) {
516 //===----------------------------------------------------------------------===//
517 // Encoding helper functions
518 //===----------------------------------------------------------------------===//
520 bool R600MCCodeEmitter::isFCOp(unsigned opcode) const {
522 default: return false;
523 case AMDGPU::PREDICATED_BREAK:
524 case AMDGPU::CONTINUE:
525 case AMDGPU::IF_PREDICATE_SET:
528 case AMDGPU::ENDLOOP:
529 case AMDGPU::WHILELOOP:
534 bool R600MCCodeEmitter::isTexOp(unsigned opcode) const {
536 default: return false;
538 case AMDGPU::TEX_GET_TEXTURE_RESINFO:
539 case AMDGPU::TEX_SAMPLE:
540 case AMDGPU::TEX_SAMPLE_C:
541 case AMDGPU::TEX_SAMPLE_L:
542 case AMDGPU::TEX_SAMPLE_C_L:
543 case AMDGPU::TEX_SAMPLE_LB:
544 case AMDGPU::TEX_SAMPLE_C_LB:
545 case AMDGPU::TEX_SAMPLE_G:
546 case AMDGPU::TEX_SAMPLE_C_G:
547 case AMDGPU::TEX_GET_GRADIENTS_H:
548 case AMDGPU::TEX_GET_GRADIENTS_V:
549 case AMDGPU::TEX_SET_GRADIENTS_H:
550 case AMDGPU::TEX_SET_GRADIENTS_V:
555 bool R600MCCodeEmitter::isFlagSet(const MCInst &MI, unsigned Operand,
556 unsigned Flag) const {
557 const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
558 unsigned FlagIndex = GET_FLAG_OPERAND_IDX(MCDesc.TSFlags);
559 if (FlagIndex == 0) {
562 assert(MI.getOperand(FlagIndex).isImm());
563 return !!((MI.getOperand(FlagIndex).getImm() >>
564 (NUM_MO_FLAGS * Operand)) & Flag);
567 #include "AMDGPUGenMCCodeEmitter.inc"