1 //===-- R600InstrInfo.cpp - R600 Instruction Information ------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief R600 Implementation of TargetInstrInfo.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "R600InstrInfo.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUSubtarget.h"
18 #include "AMDGPUTargetMachine.h"
19 #include "R600Defines.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25
26 using namespace llvm;
27
28 #define GET_INSTRINFO_CTOR_DTOR
29 #include "AMDGPUGenDFAPacketizer.inc"
30
R600InstrInfo(const AMDGPUSubtarget & st)31 R600InstrInfo::R600InstrInfo(const AMDGPUSubtarget &st)
32 : AMDGPUInstrInfo(st), RI() {}
33
getRegisterInfo() const34 const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
35 return RI;
36 }
37
isTrig(const MachineInstr & MI) const38 bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
39 return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
40 }
41
isVector(const MachineInstr & MI) const42 bool R600InstrInfo::isVector(const MachineInstr &MI) const {
43 return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
44 }
45
46 void
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,DebugLoc DL,unsigned DestReg,unsigned SrcReg,bool KillSrc) const47 R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
48 MachineBasicBlock::iterator MI, DebugLoc DL,
49 unsigned DestReg, unsigned SrcReg,
50 bool KillSrc) const {
51 unsigned VectorComponents = 0;
52 if ((AMDGPU::R600_Reg128RegClass.contains(DestReg) ||
53 AMDGPU::R600_Reg128VerticalRegClass.contains(DestReg)) &&
54 (AMDGPU::R600_Reg128RegClass.contains(SrcReg) ||
55 AMDGPU::R600_Reg128VerticalRegClass.contains(SrcReg))) {
56 VectorComponents = 4;
57 } else if((AMDGPU::R600_Reg64RegClass.contains(DestReg) ||
58 AMDGPU::R600_Reg64VerticalRegClass.contains(DestReg)) &&
59 (AMDGPU::R600_Reg64RegClass.contains(SrcReg) ||
60 AMDGPU::R600_Reg64VerticalRegClass.contains(SrcReg))) {
61 VectorComponents = 2;
62 }
63
64 if (VectorComponents > 0) {
65 for (unsigned I = 0; I < VectorComponents; I++) {
66 unsigned SubRegIndex = RI.getSubRegFromChannel(I);
67 buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
68 RI.getSubReg(DestReg, SubRegIndex),
69 RI.getSubReg(SrcReg, SubRegIndex))
70 .addReg(DestReg,
71 RegState::Define | RegState::Implicit);
72 }
73 } else {
74 MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
75 DestReg, SrcReg);
76 NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
77 .setIsKill(KillSrc);
78 }
79 }
80
81 /// \returns true if \p MBBI can be moved into a new basic.
isLegalToSplitMBBAt(MachineBasicBlock & MBB,MachineBasicBlock::iterator MBBI) const82 bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
83 MachineBasicBlock::iterator MBBI) const {
84 for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(),
85 E = MBBI->operands_end(); I != E; ++I) {
86 if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) &&
87 I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg()))
88 return false;
89 }
90 return true;
91 }
92
isMov(unsigned Opcode) const93 bool R600InstrInfo::isMov(unsigned Opcode) const {
94
95
96 switch(Opcode) {
97 default: return false;
98 case AMDGPU::MOV:
99 case AMDGPU::MOV_IMM_F32:
100 case AMDGPU::MOV_IMM_I32:
101 return true;
102 }
103 }
104
105 // Some instructions act as place holders to emulate operations that the GPU
106 // hardware does automatically. This function can be used to check if
107 // an opcode falls into this category.
isPlaceHolderOpcode(unsigned Opcode) const108 bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
109 switch (Opcode) {
110 default: return false;
111 case AMDGPU::RETURN:
112 return true;
113 }
114 }
115
isReductionOp(unsigned Opcode) const116 bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
117 return false;
118 }
119
isCubeOp(unsigned Opcode) const120 bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
121 switch(Opcode) {
122 default: return false;
123 case AMDGPU::CUBE_r600_pseudo:
124 case AMDGPU::CUBE_r600_real:
125 case AMDGPU::CUBE_eg_pseudo:
126 case AMDGPU::CUBE_eg_real:
127 return true;
128 }
129 }
130
isALUInstr(unsigned Opcode) const131 bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
132 unsigned TargetFlags = get(Opcode).TSFlags;
133
134 return (TargetFlags & R600_InstFlag::ALU_INST);
135 }
136
hasInstrModifiers(unsigned Opcode) const137 bool R600InstrInfo::hasInstrModifiers(unsigned Opcode) const {
138 unsigned TargetFlags = get(Opcode).TSFlags;
139
140 return ((TargetFlags & R600_InstFlag::OP1) |
141 (TargetFlags & R600_InstFlag::OP2) |
142 (TargetFlags & R600_InstFlag::OP3));
143 }
144
isLDSInstr(unsigned Opcode) const145 bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
146 unsigned TargetFlags = get(Opcode).TSFlags;
147
148 return ((TargetFlags & R600_InstFlag::LDS_1A) |
149 (TargetFlags & R600_InstFlag::LDS_1A1D) |
150 (TargetFlags & R600_InstFlag::LDS_1A2D));
151 }
152
isLDSNoRetInstr(unsigned Opcode) const153 bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const {
154 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1;
155 }
156
isLDSRetInstr(unsigned Opcode) const157 bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const {
158 return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1;
159 }
160
canBeConsideredALU(const MachineInstr * MI) const161 bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const {
162 if (isALUInstr(MI->getOpcode()))
163 return true;
164 if (isVector(*MI) || isCubeOp(MI->getOpcode()))
165 return true;
166 switch (MI->getOpcode()) {
167 case AMDGPU::PRED_X:
168 case AMDGPU::INTERP_PAIR_XY:
169 case AMDGPU::INTERP_PAIR_ZW:
170 case AMDGPU::INTERP_VEC_LOAD:
171 case AMDGPU::COPY:
172 case AMDGPU::DOT_4:
173 return true;
174 default:
175 return false;
176 }
177 }
178
isTransOnly(unsigned Opcode) const179 bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
180 if (ST.hasCaymanISA())
181 return false;
182 return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
183 }
184
isTransOnly(const MachineInstr * MI) const185 bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
186 return isTransOnly(MI->getOpcode());
187 }
188
isVectorOnly(unsigned Opcode) const189 bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
190 return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
191 }
192
isVectorOnly(const MachineInstr * MI) const193 bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
194 return isVectorOnly(MI->getOpcode());
195 }
196
isExport(unsigned Opcode) const197 bool R600InstrInfo::isExport(unsigned Opcode) const {
198 return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
199 }
200
usesVertexCache(unsigned Opcode) const201 bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
202 return ST.hasVertexCache() && IS_VTX(get(Opcode));
203 }
204
usesVertexCache(const MachineInstr * MI) const205 bool R600InstrInfo::usesVertexCache(const MachineInstr *MI) const {
206 const MachineFunction *MF = MI->getParent()->getParent();
207 const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
208 return MFI->getShaderType() != ShaderType::COMPUTE &&
209 usesVertexCache(MI->getOpcode());
210 }
211
usesTextureCache(unsigned Opcode) const212 bool R600InstrInfo::usesTextureCache(unsigned Opcode) const {
213 return (!ST.hasVertexCache() && IS_VTX(get(Opcode))) || IS_TEX(get(Opcode));
214 }
215
usesTextureCache(const MachineInstr * MI) const216 bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
217 const MachineFunction *MF = MI->getParent()->getParent();
218 const R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
219 return (MFI->getShaderType() == ShaderType::COMPUTE &&
220 usesVertexCache(MI->getOpcode())) ||
221 usesTextureCache(MI->getOpcode());
222 }
223
mustBeLastInClause(unsigned Opcode) const224 bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
225 switch (Opcode) {
226 case AMDGPU::KILLGT:
227 case AMDGPU::GROUP_BARRIER:
228 return true;
229 default:
230 return false;
231 }
232 }
233
usesAddressRegister(MachineInstr * MI) const234 bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const {
235 return MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1;
236 }
237
definesAddressRegister(MachineInstr * MI) const238 bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const {
239 return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1;
240 }
241
readsLDSSrcReg(const MachineInstr * MI) const242 bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
243 if (!isALUInstr(MI->getOpcode())) {
244 return false;
245 }
246 for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
247 E = MI->operands_end(); I != E; ++I) {
248 if (!I->isReg() || !I->isUse() ||
249 TargetRegisterInfo::isVirtualRegister(I->getReg()))
250 continue;
251
252 if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
253 return true;
254 }
255 return false;
256 }
257
getSrcIdx(unsigned Opcode,unsigned SrcNum) const258 int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
259 static const unsigned OpTable[] = {
260 AMDGPU::OpName::src0,
261 AMDGPU::OpName::src1,
262 AMDGPU::OpName::src2
263 };
264
265 assert (SrcNum < 3);
266 return getOperandIdx(Opcode, OpTable[SrcNum]);
267 }
268
getSelIdx(unsigned Opcode,unsigned SrcIdx) const269 int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
270 static const unsigned SrcSelTable[][2] = {
271 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
272 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
273 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
274 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
275 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
276 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
277 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
278 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
279 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
280 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
281 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
282 };
283
284 for (const auto &Row : SrcSelTable) {
285 if (getOperandIdx(Opcode, Row[0]) == (int)SrcIdx) {
286 return getOperandIdx(Opcode, Row[1]);
287 }
288 }
289 return -1;
290 }
291
292 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
getSrcs(MachineInstr * MI) const293 R600InstrInfo::getSrcs(MachineInstr *MI) const {
294 SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
295
296 if (MI->getOpcode() == AMDGPU::DOT_4) {
297 static const unsigned OpTable[8][2] = {
298 {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
299 {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
300 {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
301 {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
302 {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
303 {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
304 {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
305 {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W},
306 };
307
308 for (unsigned j = 0; j < 8; j++) {
309 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
310 OpTable[j][0]));
311 unsigned Reg = MO.getReg();
312 if (Reg == AMDGPU::ALU_CONST) {
313 unsigned Sel = MI->getOperand(getOperandIdx(MI->getOpcode(),
314 OpTable[j][1])).getImm();
315 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
316 continue;
317 }
318
319 }
320 return Result;
321 }
322
323 static const unsigned OpTable[3][2] = {
324 {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
325 {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
326 {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
327 };
328
329 for (unsigned j = 0; j < 3; j++) {
330 int SrcIdx = getOperandIdx(MI->getOpcode(), OpTable[j][0]);
331 if (SrcIdx < 0)
332 break;
333 MachineOperand &MO = MI->getOperand(SrcIdx);
334 unsigned Reg = MI->getOperand(SrcIdx).getReg();
335 if (Reg == AMDGPU::ALU_CONST) {
336 unsigned Sel = MI->getOperand(
337 getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
338 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Sel));
339 continue;
340 }
341 if (Reg == AMDGPU::ALU_LITERAL_X) {
342 unsigned Imm = MI->getOperand(
343 getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal)).getImm();
344 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, Imm));
345 continue;
346 }
347 Result.push_back(std::pair<MachineOperand *, int64_t>(&MO, 0));
348 }
349 return Result;
350 }
351
352 std::vector<std::pair<int, unsigned> >
ExtractSrcs(MachineInstr * MI,const DenseMap<unsigned,unsigned> & PV,unsigned & ConstCount) const353 R600InstrInfo::ExtractSrcs(MachineInstr *MI,
354 const DenseMap<unsigned, unsigned> &PV,
355 unsigned &ConstCount) const {
356 ConstCount = 0;
357 const SmallVector<std::pair<MachineOperand *, int64_t>, 3> Srcs = getSrcs(MI);
358 const std::pair<int, unsigned> DummyPair(-1, 0);
359 std::vector<std::pair<int, unsigned> > Result;
360 unsigned i = 0;
361 for (unsigned n = Srcs.size(); i < n; ++i) {
362 unsigned Reg = Srcs[i].first->getReg();
363 unsigned Index = RI.getEncodingValue(Reg) & 0xff;
364 if (Reg == AMDGPU::OQAP) {
365 Result.push_back(std::pair<int, unsigned>(Index, 0));
366 }
367 if (PV.find(Reg) != PV.end()) {
368 // 255 is used to tells its a PS/PV reg
369 Result.push_back(std::pair<int, unsigned>(255, 0));
370 continue;
371 }
372 if (Index > 127) {
373 ConstCount++;
374 Result.push_back(DummyPair);
375 continue;
376 }
377 unsigned Chan = RI.getHWRegChan(Reg);
378 Result.push_back(std::pair<int, unsigned>(Index, Chan));
379 }
380 for (; i < 3; ++i)
381 Result.push_back(DummyPair);
382 return Result;
383 }
384
385 static std::vector<std::pair<int, unsigned> >
Swizzle(std::vector<std::pair<int,unsigned>> Src,R600InstrInfo::BankSwizzle Swz)386 Swizzle(std::vector<std::pair<int, unsigned> > Src,
387 R600InstrInfo::BankSwizzle Swz) {
388 if (Src[0] == Src[1])
389 Src[1].first = -1;
390 switch (Swz) {
391 case R600InstrInfo::ALU_VEC_012_SCL_210:
392 break;
393 case R600InstrInfo::ALU_VEC_021_SCL_122:
394 std::swap(Src[1], Src[2]);
395 break;
396 case R600InstrInfo::ALU_VEC_102_SCL_221:
397 std::swap(Src[0], Src[1]);
398 break;
399 case R600InstrInfo::ALU_VEC_120_SCL_212:
400 std::swap(Src[0], Src[1]);
401 std::swap(Src[0], Src[2]);
402 break;
403 case R600InstrInfo::ALU_VEC_201:
404 std::swap(Src[0], Src[2]);
405 std::swap(Src[0], Src[1]);
406 break;
407 case R600InstrInfo::ALU_VEC_210:
408 std::swap(Src[0], Src[2]);
409 break;
410 }
411 return Src;
412 }
413
414 static unsigned
getTransSwizzle(R600InstrInfo::BankSwizzle Swz,unsigned Op)415 getTransSwizzle(R600InstrInfo::BankSwizzle Swz, unsigned Op) {
416 switch (Swz) {
417 case R600InstrInfo::ALU_VEC_012_SCL_210: {
418 unsigned Cycles[3] = { 2, 1, 0};
419 return Cycles[Op];
420 }
421 case R600InstrInfo::ALU_VEC_021_SCL_122: {
422 unsigned Cycles[3] = { 1, 2, 2};
423 return Cycles[Op];
424 }
425 case R600InstrInfo::ALU_VEC_120_SCL_212: {
426 unsigned Cycles[3] = { 2, 1, 2};
427 return Cycles[Op];
428 }
429 case R600InstrInfo::ALU_VEC_102_SCL_221: {
430 unsigned Cycles[3] = { 2, 2, 1};
431 return Cycles[Op];
432 }
433 default:
434 llvm_unreachable("Wrong Swizzle for Trans Slot");
435 return 0;
436 }
437 }
438
439 /// returns how many MIs (whose inputs are represented by IGSrcs) can be packed
440 /// in the same Instruction Group while meeting read port limitations given a
441 /// Swz swizzle sequence.
isLegalUpTo(const std::vector<std::vector<std::pair<int,unsigned>>> & IGSrcs,const std::vector<R600InstrInfo::BankSwizzle> & Swz,const std::vector<std::pair<int,unsigned>> & TransSrcs,R600InstrInfo::BankSwizzle TransSwz) const442 unsigned R600InstrInfo::isLegalUpTo(
443 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
444 const std::vector<R600InstrInfo::BankSwizzle> &Swz,
445 const std::vector<std::pair<int, unsigned> > &TransSrcs,
446 R600InstrInfo::BankSwizzle TransSwz) const {
447 int Vector[4][3];
448 memset(Vector, -1, sizeof(Vector));
449 for (unsigned i = 0, e = IGSrcs.size(); i < e; i++) {
450 const std::vector<std::pair<int, unsigned> > &Srcs =
451 Swizzle(IGSrcs[i], Swz[i]);
452 for (unsigned j = 0; j < 3; j++) {
453 const std::pair<int, unsigned> &Src = Srcs[j];
454 if (Src.first < 0 || Src.first == 255)
455 continue;
456 if (Src.first == GET_REG_INDEX(RI.getEncodingValue(AMDGPU::OQAP))) {
457 if (Swz[i] != R600InstrInfo::ALU_VEC_012_SCL_210 &&
458 Swz[i] != R600InstrInfo::ALU_VEC_021_SCL_122) {
459 // The value from output queue A (denoted by register OQAP) can
460 // only be fetched during the first cycle.
461 return false;
462 }
463 // OQAP does not count towards the normal read port restrictions
464 continue;
465 }
466 if (Vector[Src.second][j] < 0)
467 Vector[Src.second][j] = Src.first;
468 if (Vector[Src.second][j] != Src.first)
469 return i;
470 }
471 }
472 // Now check Trans Alu
473 for (unsigned i = 0, e = TransSrcs.size(); i < e; ++i) {
474 const std::pair<int, unsigned> &Src = TransSrcs[i];
475 unsigned Cycle = getTransSwizzle(TransSwz, i);
476 if (Src.first < 0)
477 continue;
478 if (Src.first == 255)
479 continue;
480 if (Vector[Src.second][Cycle] < 0)
481 Vector[Src.second][Cycle] = Src.first;
482 if (Vector[Src.second][Cycle] != Src.first)
483 return IGSrcs.size() - 1;
484 }
485 return IGSrcs.size();
486 }
487
488 /// Given a swizzle sequence SwzCandidate and an index Idx, returns the next
489 /// (in lexicographic term) swizzle sequence assuming that all swizzles after
490 /// Idx can be skipped
491 static bool
NextPossibleSolution(std::vector<R600InstrInfo::BankSwizzle> & SwzCandidate,unsigned Idx)492 NextPossibleSolution(
493 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
494 unsigned Idx) {
495 assert(Idx < SwzCandidate.size());
496 int ResetIdx = Idx;
497 while (ResetIdx > -1 && SwzCandidate[ResetIdx] == R600InstrInfo::ALU_VEC_210)
498 ResetIdx --;
499 for (unsigned i = ResetIdx + 1, e = SwzCandidate.size(); i < e; i++) {
500 SwzCandidate[i] = R600InstrInfo::ALU_VEC_012_SCL_210;
501 }
502 if (ResetIdx == -1)
503 return false;
504 int NextSwizzle = SwzCandidate[ResetIdx] + 1;
505 SwzCandidate[ResetIdx] = (R600InstrInfo::BankSwizzle)NextSwizzle;
506 return true;
507 }
508
509 /// Enumerate all possible Swizzle sequence to find one that can meet all
510 /// read port requirements.
FindSwizzleForVectorSlot(const std::vector<std::vector<std::pair<int,unsigned>>> & IGSrcs,std::vector<R600InstrInfo::BankSwizzle> & SwzCandidate,const std::vector<std::pair<int,unsigned>> & TransSrcs,R600InstrInfo::BankSwizzle TransSwz) const511 bool R600InstrInfo::FindSwizzleForVectorSlot(
512 const std::vector<std::vector<std::pair<int, unsigned> > > &IGSrcs,
513 std::vector<R600InstrInfo::BankSwizzle> &SwzCandidate,
514 const std::vector<std::pair<int, unsigned> > &TransSrcs,
515 R600InstrInfo::BankSwizzle TransSwz) const {
516 unsigned ValidUpTo = 0;
517 do {
518 ValidUpTo = isLegalUpTo(IGSrcs, SwzCandidate, TransSrcs, TransSwz);
519 if (ValidUpTo == IGSrcs.size())
520 return true;
521 } while (NextPossibleSolution(SwzCandidate, ValidUpTo));
522 return false;
523 }
524
525 /// Instructions in Trans slot can't read gpr at cycle 0 if they also read
526 /// a const, and can't read a gpr at cycle 1 if they read 2 const.
527 static bool
isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,const std::vector<std::pair<int,unsigned>> & TransOps,unsigned ConstCount)528 isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
529 const std::vector<std::pair<int, unsigned> > &TransOps,
530 unsigned ConstCount) {
531 // TransALU can't read 3 constants
532 if (ConstCount > 2)
533 return false;
534 for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
535 const std::pair<int, unsigned> &Src = TransOps[i];
536 unsigned Cycle = getTransSwizzle(TransSwz, i);
537 if (Src.first < 0)
538 continue;
539 if (ConstCount > 0 && Cycle == 0)
540 return false;
541 if (ConstCount > 1 && Cycle == 1)
542 return false;
543 }
544 return true;
545 }
546
547 bool
fitsReadPortLimitations(const std::vector<MachineInstr * > & IG,const DenseMap<unsigned,unsigned> & PV,std::vector<BankSwizzle> & ValidSwizzle,bool isLastAluTrans) const548 R600InstrInfo::fitsReadPortLimitations(const std::vector<MachineInstr *> &IG,
549 const DenseMap<unsigned, unsigned> &PV,
550 std::vector<BankSwizzle> &ValidSwizzle,
551 bool isLastAluTrans)
552 const {
553 //Todo : support shared src0 - src1 operand
554
555 std::vector<std::vector<std::pair<int, unsigned> > > IGSrcs;
556 ValidSwizzle.clear();
557 unsigned ConstCount;
558 BankSwizzle TransBS = ALU_VEC_012_SCL_210;
559 for (unsigned i = 0, e = IG.size(); i < e; ++i) {
560 IGSrcs.push_back(ExtractSrcs(IG[i], PV, ConstCount));
561 unsigned Op = getOperandIdx(IG[i]->getOpcode(),
562 AMDGPU::OpName::bank_swizzle);
563 ValidSwizzle.push_back( (R600InstrInfo::BankSwizzle)
564 IG[i]->getOperand(Op).getImm());
565 }
566 std::vector<std::pair<int, unsigned> > TransOps;
567 if (!isLastAluTrans)
568 return FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps, TransBS);
569
570 TransOps = std::move(IGSrcs.back());
571 IGSrcs.pop_back();
572 ValidSwizzle.pop_back();
573
574 static const R600InstrInfo::BankSwizzle TransSwz[] = {
575 ALU_VEC_012_SCL_210,
576 ALU_VEC_021_SCL_122,
577 ALU_VEC_120_SCL_212,
578 ALU_VEC_102_SCL_221
579 };
580 for (unsigned i = 0; i < 4; i++) {
581 TransBS = TransSwz[i];
582 if (!isConstCompatible(TransBS, TransOps, ConstCount))
583 continue;
584 bool Result = FindSwizzleForVectorSlot(IGSrcs, ValidSwizzle, TransOps,
585 TransBS);
586 if (Result) {
587 ValidSwizzle.push_back(TransBS);
588 return true;
589 }
590 }
591
592 return false;
593 }
594
595
596 bool
fitsConstReadLimitations(const std::vector<unsigned> & Consts) const597 R600InstrInfo::fitsConstReadLimitations(const std::vector<unsigned> &Consts)
598 const {
599 assert (Consts.size() <= 12 && "Too many operands in instructions group");
600 unsigned Pair1 = 0, Pair2 = 0;
601 for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
602 unsigned ReadConstHalf = Consts[i] & 2;
603 unsigned ReadConstIndex = Consts[i] & (~3);
604 unsigned ReadHalfConst = ReadConstIndex | ReadConstHalf;
605 if (!Pair1) {
606 Pair1 = ReadHalfConst;
607 continue;
608 }
609 if (Pair1 == ReadHalfConst)
610 continue;
611 if (!Pair2) {
612 Pair2 = ReadHalfConst;
613 continue;
614 }
615 if (Pair2 != ReadHalfConst)
616 return false;
617 }
618 return true;
619 }
620
621 bool
fitsConstReadLimitations(const std::vector<MachineInstr * > & MIs) const622 R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
623 const {
624 std::vector<unsigned> Consts;
625 SmallSet<int64_t, 4> Literals;
626 for (unsigned i = 0, n = MIs.size(); i < n; i++) {
627 MachineInstr *MI = MIs[i];
628 if (!isALUInstr(MI->getOpcode()))
629 continue;
630
631 const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
632 getSrcs(MI);
633
634 for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
635 std::pair<MachineOperand *, unsigned> Src = Srcs[j];
636 if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
637 Literals.insert(Src.second);
638 if (Literals.size() > 4)
639 return false;
640 if (Src.first->getReg() == AMDGPU::ALU_CONST)
641 Consts.push_back(Src.second);
642 if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
643 AMDGPU::R600_KC1RegClass.contains(Src.first->getReg())) {
644 unsigned Index = RI.getEncodingValue(Src.first->getReg()) & 0xff;
645 unsigned Chan = RI.getHWRegChan(Src.first->getReg());
646 Consts.push_back((Index << 2) | Chan);
647 }
648 }
649 }
650 return fitsConstReadLimitations(Consts);
651 }
652
653 DFAPacketizer *
CreateTargetScheduleState(const TargetSubtargetInfo & STI) const654 R600InstrInfo::CreateTargetScheduleState(const TargetSubtargetInfo &STI) const {
655 const InstrItineraryData *II = STI.getInstrItineraryData();
656 return static_cast<const AMDGPUSubtarget &>(STI).createDFAPacketizer(II);
657 }
658
659 static bool
isPredicateSetter(unsigned Opcode)660 isPredicateSetter(unsigned Opcode) {
661 switch (Opcode) {
662 case AMDGPU::PRED_X:
663 return true;
664 default:
665 return false;
666 }
667 }
668
669 static MachineInstr *
findFirstPredicateSetterFrom(MachineBasicBlock & MBB,MachineBasicBlock::iterator I)670 findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
671 MachineBasicBlock::iterator I) {
672 while (I != MBB.begin()) {
673 --I;
674 MachineInstr *MI = I;
675 if (isPredicateSetter(MI->getOpcode()))
676 return MI;
677 }
678
679 return nullptr;
680 }
681
682 static
isJump(unsigned Opcode)683 bool isJump(unsigned Opcode) {
684 return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
685 }
686
isBranch(unsigned Opcode)687 static bool isBranch(unsigned Opcode) {
688 return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 ||
689 Opcode == AMDGPU::BRANCH_COND_f32;
690 }
691
692 bool
AnalyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const693 R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
694 MachineBasicBlock *&TBB,
695 MachineBasicBlock *&FBB,
696 SmallVectorImpl<MachineOperand> &Cond,
697 bool AllowModify) const {
698 // Most of the following comes from the ARM implementation of AnalyzeBranch
699
700 // If the block has no terminators, it just falls into the block after it.
701 MachineBasicBlock::iterator I = MBB.end();
702 if (I == MBB.begin())
703 return false;
704 --I;
705 while (I->isDebugValue()) {
706 if (I == MBB.begin())
707 return false;
708 --I;
709 }
710 // AMDGPU::BRANCH* instructions are only available after isel and are not
711 // handled
712 if (isBranch(I->getOpcode()))
713 return true;
714 if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
715 return false;
716 }
717
718 // Remove successive JUMP
719 while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) {
720 MachineBasicBlock::iterator PriorI = std::prev(I);
721 if (AllowModify)
722 I->removeFromParent();
723 I = PriorI;
724 }
725 MachineInstr *LastInst = I;
726
727 // If there is only one terminator instruction, process it.
728 unsigned LastOpc = LastInst->getOpcode();
729 if (I == MBB.begin() ||
730 !isJump(static_cast<MachineInstr *>(--I)->getOpcode())) {
731 if (LastOpc == AMDGPU::JUMP) {
732 TBB = LastInst->getOperand(0).getMBB();
733 return false;
734 } else if (LastOpc == AMDGPU::JUMP_COND) {
735 MachineInstr *predSet = I;
736 while (!isPredicateSetter(predSet->getOpcode())) {
737 predSet = --I;
738 }
739 TBB = LastInst->getOperand(0).getMBB();
740 Cond.push_back(predSet->getOperand(1));
741 Cond.push_back(predSet->getOperand(2));
742 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
743 return false;
744 }
745 return true; // Can't handle indirect branch.
746 }
747
748 // Get the instruction before it if it is a terminator.
749 MachineInstr *SecondLastInst = I;
750 unsigned SecondLastOpc = SecondLastInst->getOpcode();
751
752 // If the block ends with a B and a Bcc, handle it.
753 if (SecondLastOpc == AMDGPU::JUMP_COND && LastOpc == AMDGPU::JUMP) {
754 MachineInstr *predSet = --I;
755 while (!isPredicateSetter(predSet->getOpcode())) {
756 predSet = --I;
757 }
758 TBB = SecondLastInst->getOperand(0).getMBB();
759 FBB = LastInst->getOperand(0).getMBB();
760 Cond.push_back(predSet->getOperand(1));
761 Cond.push_back(predSet->getOperand(2));
762 Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
763 return false;
764 }
765
766 // Otherwise, can't handle this.
767 return true;
768 }
769
770 static
FindLastAluClause(MachineBasicBlock & MBB)771 MachineBasicBlock::iterator FindLastAluClause(MachineBasicBlock &MBB) {
772 for (MachineBasicBlock::reverse_iterator It = MBB.rbegin(), E = MBB.rend();
773 It != E; ++It) {
774 if (It->getOpcode() == AMDGPU::CF_ALU ||
775 It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
776 return std::prev(It.base());
777 }
778 return MBB.end();
779 }
780
781 unsigned
InsertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,const SmallVectorImpl<MachineOperand> & Cond,DebugLoc DL) const782 R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
783 MachineBasicBlock *TBB,
784 MachineBasicBlock *FBB,
785 const SmallVectorImpl<MachineOperand> &Cond,
786 DebugLoc DL) const {
787 assert(TBB && "InsertBranch must not be told to insert a fallthrough");
788
789 if (!FBB) {
790 if (Cond.empty()) {
791 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB);
792 return 1;
793 } else {
794 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
795 assert(PredSet && "No previous predicate !");
796 addFlag(PredSet, 0, MO_FLAG_PUSH);
797 PredSet->getOperand(2).setImm(Cond[1].getImm());
798
799 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
800 .addMBB(TBB)
801 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
802 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
803 if (CfAlu == MBB.end())
804 return 1;
805 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
806 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
807 return 1;
808 }
809 } else {
810 MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
811 assert(PredSet && "No previous predicate !");
812 addFlag(PredSet, 0, MO_FLAG_PUSH);
813 PredSet->getOperand(2).setImm(Cond[1].getImm());
814 BuildMI(&MBB, DL, get(AMDGPU::JUMP_COND))
815 .addMBB(TBB)
816 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
817 BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB);
818 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
819 if (CfAlu == MBB.end())
820 return 2;
821 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU);
822 CfAlu->setDesc(get(AMDGPU::CF_ALU_PUSH_BEFORE));
823 return 2;
824 }
825 }
826
827 unsigned
RemoveBranch(MachineBasicBlock & MBB) const828 R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
829
830 // Note : we leave PRED* instructions there.
831 // They may be needed when predicating instructions.
832
833 MachineBasicBlock::iterator I = MBB.end();
834
835 if (I == MBB.begin()) {
836 return 0;
837 }
838 --I;
839 switch (I->getOpcode()) {
840 default:
841 return 0;
842 case AMDGPU::JUMP_COND: {
843 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
844 clearFlag(predSet, 0, MO_FLAG_PUSH);
845 I->eraseFromParent();
846 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
847 if (CfAlu == MBB.end())
848 break;
849 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
850 CfAlu->setDesc(get(AMDGPU::CF_ALU));
851 break;
852 }
853 case AMDGPU::JUMP:
854 I->eraseFromParent();
855 break;
856 }
857 I = MBB.end();
858
859 if (I == MBB.begin()) {
860 return 1;
861 }
862 --I;
863 switch (I->getOpcode()) {
864 // FIXME: only one case??
865 default:
866 return 1;
867 case AMDGPU::JUMP_COND: {
868 MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
869 clearFlag(predSet, 0, MO_FLAG_PUSH);
870 I->eraseFromParent();
871 MachineBasicBlock::iterator CfAlu = FindLastAluClause(MBB);
872 if (CfAlu == MBB.end())
873 break;
874 assert (CfAlu->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE);
875 CfAlu->setDesc(get(AMDGPU::CF_ALU));
876 break;
877 }
878 case AMDGPU::JUMP:
879 I->eraseFromParent();
880 break;
881 }
882 return 2;
883 }
884
885 bool
isPredicated(const MachineInstr * MI) const886 R600InstrInfo::isPredicated(const MachineInstr *MI) const {
887 int idx = MI->findFirstPredOperandIdx();
888 if (idx < 0)
889 return false;
890
891 unsigned Reg = MI->getOperand(idx).getReg();
892 switch (Reg) {
893 default: return false;
894 case AMDGPU::PRED_SEL_ONE:
895 case AMDGPU::PRED_SEL_ZERO:
896 case AMDGPU::PREDICATE_BIT:
897 return true;
898 }
899 }
900
901 bool
isPredicable(MachineInstr * MI) const902 R600InstrInfo::isPredicable(MachineInstr *MI) const {
903 // XXX: KILL* instructions can be predicated, but they must be the last
904 // instruction in a clause, so this means any instructions after them cannot
905 // be predicated. Until we have proper support for instruction clauses in the
906 // backend, we will mark KILL* instructions as unpredicable.
907
908 if (MI->getOpcode() == AMDGPU::KILLGT) {
909 return false;
910 } else if (MI->getOpcode() == AMDGPU::CF_ALU) {
911 // If the clause start in the middle of MBB then the MBB has more
912 // than a single clause, unable to predicate several clauses.
913 if (MI->getParent()->begin() != MachineBasicBlock::iterator(MI))
914 return false;
915 // TODO: We don't support KC merging atm
916 if (MI->getOperand(3).getImm() != 0 || MI->getOperand(4).getImm() != 0)
917 return false;
918 return true;
919 } else if (isVector(*MI)) {
920 return false;
921 } else {
922 return AMDGPUInstrInfo::isPredicable(MI);
923 }
924 }
925
926
927 bool
isProfitableToIfCvt(MachineBasicBlock & MBB,unsigned NumCyles,unsigned ExtraPredCycles,const BranchProbability & Probability) const928 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
929 unsigned NumCyles,
930 unsigned ExtraPredCycles,
931 const BranchProbability &Probability) const{
932 return true;
933 }
934
935 bool
isProfitableToIfCvt(MachineBasicBlock & TMBB,unsigned NumTCycles,unsigned ExtraTCycles,MachineBasicBlock & FMBB,unsigned NumFCycles,unsigned ExtraFCycles,const BranchProbability & Probability) const936 R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
937 unsigned NumTCycles,
938 unsigned ExtraTCycles,
939 MachineBasicBlock &FMBB,
940 unsigned NumFCycles,
941 unsigned ExtraFCycles,
942 const BranchProbability &Probability) const {
943 return true;
944 }
945
946 bool
isProfitableToDupForIfCvt(MachineBasicBlock & MBB,unsigned NumCyles,const BranchProbability & Probability) const947 R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
948 unsigned NumCyles,
949 const BranchProbability &Probability)
950 const {
951 return true;
952 }
953
954 bool
isProfitableToUnpredicate(MachineBasicBlock & TMBB,MachineBasicBlock & FMBB) const955 R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
956 MachineBasicBlock &FMBB) const {
957 return false;
958 }
959
960
961 bool
ReverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const962 R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
963 MachineOperand &MO = Cond[1];
964 switch (MO.getImm()) {
965 case OPCODE_IS_ZERO_INT:
966 MO.setImm(OPCODE_IS_NOT_ZERO_INT);
967 break;
968 case OPCODE_IS_NOT_ZERO_INT:
969 MO.setImm(OPCODE_IS_ZERO_INT);
970 break;
971 case OPCODE_IS_ZERO:
972 MO.setImm(OPCODE_IS_NOT_ZERO);
973 break;
974 case OPCODE_IS_NOT_ZERO:
975 MO.setImm(OPCODE_IS_ZERO);
976 break;
977 default:
978 return true;
979 }
980
981 MachineOperand &MO2 = Cond[2];
982 switch (MO2.getReg()) {
983 case AMDGPU::PRED_SEL_ZERO:
984 MO2.setReg(AMDGPU::PRED_SEL_ONE);
985 break;
986 case AMDGPU::PRED_SEL_ONE:
987 MO2.setReg(AMDGPU::PRED_SEL_ZERO);
988 break;
989 default:
990 return true;
991 }
992 return false;
993 }
994
995 bool
DefinesPredicate(MachineInstr * MI,std::vector<MachineOperand> & Pred) const996 R600InstrInfo::DefinesPredicate(MachineInstr *MI,
997 std::vector<MachineOperand> &Pred) const {
998 return isPredicateSetter(MI->getOpcode());
999 }
1000
1001
1002 bool
SubsumesPredicate(const SmallVectorImpl<MachineOperand> & Pred1,const SmallVectorImpl<MachineOperand> & Pred2) const1003 R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
1004 const SmallVectorImpl<MachineOperand> &Pred2) const {
1005 return false;
1006 }
1007
1008
1009 bool
PredicateInstruction(MachineInstr * MI,const SmallVectorImpl<MachineOperand> & Pred) const1010 R600InstrInfo::PredicateInstruction(MachineInstr *MI,
1011 const SmallVectorImpl<MachineOperand> &Pred) const {
1012 int PIdx = MI->findFirstPredOperandIdx();
1013
1014 if (MI->getOpcode() == AMDGPU::CF_ALU) {
1015 MI->getOperand(8).setImm(0);
1016 return true;
1017 }
1018
1019 if (MI->getOpcode() == AMDGPU::DOT_4) {
1020 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X))
1021 .setReg(Pred[2].getReg());
1022 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y))
1023 .setReg(Pred[2].getReg());
1024 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z))
1025 .setReg(Pred[2].getReg());
1026 MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W))
1027 .setReg(Pred[2].getReg());
1028 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1029 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1030 return true;
1031 }
1032
1033 if (PIdx != -1) {
1034 MachineOperand &PMO = MI->getOperand(PIdx);
1035 PMO.setReg(Pred[2].getReg());
1036 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
1037 MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
1038 return true;
1039 }
1040
1041 return false;
1042 }
1043
getPredicationCost(const MachineInstr *) const1044 unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const {
1045 return 2;
1046 }
1047
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr * MI,unsigned * PredCost) const1048 unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1049 const MachineInstr *MI,
1050 unsigned *PredCost) const {
1051 if (PredCost)
1052 *PredCost = 2;
1053 return 2;
1054 }
1055
expandPostRAPseudo(MachineBasicBlock::iterator MI) const1056 bool R600InstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const {
1057
1058 switch(MI->getOpcode()) {
1059 default: return AMDGPUInstrInfo::expandPostRAPseudo(MI);
1060 case AMDGPU::R600_EXTRACT_ELT_V2:
1061 case AMDGPU::R600_EXTRACT_ELT_V4:
1062 buildIndirectRead(MI->getParent(), MI, MI->getOperand(0).getReg(),
1063 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address
1064 MI->getOperand(2).getReg(),
1065 RI.getHWRegChan(MI->getOperand(1).getReg()));
1066 break;
1067 case AMDGPU::R600_INSERT_ELT_V2:
1068 case AMDGPU::R600_INSERT_ELT_V4:
1069 buildIndirectWrite(MI->getParent(), MI, MI->getOperand(2).getReg(), // Value
1070 RI.getHWRegIndex(MI->getOperand(1).getReg()), // Address
1071 MI->getOperand(3).getReg(), // Offset
1072 RI.getHWRegChan(MI->getOperand(1).getReg())); // Channel
1073 break;
1074 }
1075 MI->eraseFromParent();
1076 return true;
1077 }
1078
reserveIndirectRegisters(BitVector & Reserved,const MachineFunction & MF) const1079 void R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
1080 const MachineFunction &MF) const {
1081 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
1082 MF.getSubtarget().getFrameLowering());
1083
1084 unsigned StackWidth = TFL->getStackWidth(MF);
1085 int End = getIndirectIndexEnd(MF);
1086
1087 if (End == -1)
1088 return;
1089
1090 for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
1091 unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
1092 Reserved.set(SuperReg);
1093 for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
1094 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
1095 Reserved.set(Reg);
1096 }
1097 }
1098 }
1099
calculateIndirectAddress(unsigned RegIndex,unsigned Channel) const1100 unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
1101 unsigned Channel) const {
1102 // XXX: Remove when we support a stack width > 2
1103 assert(Channel == 0);
1104 return RegIndex;
1105 }
1106
getIndirectAddrRegClass() const1107 const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const {
1108 return &AMDGPU::R600_TReg32_XRegClass;
1109 }
1110
buildIndirectWrite(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg) const1111 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1112 MachineBasicBlock::iterator I,
1113 unsigned ValueReg, unsigned Address,
1114 unsigned OffsetReg) const {
1115 return buildIndirectWrite(MBB, I, ValueReg, Address, OffsetReg, 0);
1116 }
1117
buildIndirectWrite(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg,unsigned AddrChan) const1118 MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
1119 MachineBasicBlock::iterator I,
1120 unsigned ValueReg, unsigned Address,
1121 unsigned OffsetReg,
1122 unsigned AddrChan) const {
1123 unsigned AddrReg;
1124 switch (AddrChan) {
1125 default: llvm_unreachable("Invalid Channel");
1126 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1127 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1128 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1129 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1130 }
1131 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1132 AMDGPU::AR_X, OffsetReg);
1133 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1134
1135 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1136 AddrReg, ValueReg)
1137 .addReg(AMDGPU::AR_X,
1138 RegState::Implicit | RegState::Kill);
1139 setImmOperand(Mov, AMDGPU::OpName::dst_rel, 1);
1140 return Mov;
1141 }
1142
buildIndirectRead(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg) const1143 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1144 MachineBasicBlock::iterator I,
1145 unsigned ValueReg, unsigned Address,
1146 unsigned OffsetReg) const {
1147 return buildIndirectRead(MBB, I, ValueReg, Address, OffsetReg, 0);
1148 }
1149
buildIndirectRead(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned ValueReg,unsigned Address,unsigned OffsetReg,unsigned AddrChan) const1150 MachineInstrBuilder R600InstrInfo::buildIndirectRead(MachineBasicBlock *MBB,
1151 MachineBasicBlock::iterator I,
1152 unsigned ValueReg, unsigned Address,
1153 unsigned OffsetReg,
1154 unsigned AddrChan) const {
1155 unsigned AddrReg;
1156 switch (AddrChan) {
1157 default: llvm_unreachable("Invalid Channel");
1158 case 0: AddrReg = AMDGPU::R600_AddrRegClass.getRegister(Address); break;
1159 case 1: AddrReg = AMDGPU::R600_Addr_YRegClass.getRegister(Address); break;
1160 case 2: AddrReg = AMDGPU::R600_Addr_ZRegClass.getRegister(Address); break;
1161 case 3: AddrReg = AMDGPU::R600_Addr_WRegClass.getRegister(Address); break;
1162 }
1163 MachineInstr *MOVA = buildDefaultInstruction(*MBB, I, AMDGPU::MOVA_INT_eg,
1164 AMDGPU::AR_X,
1165 OffsetReg);
1166 setImmOperand(MOVA, AMDGPU::OpName::write, 0);
1167 MachineInstrBuilder Mov = buildDefaultInstruction(*MBB, I, AMDGPU::MOV,
1168 ValueReg,
1169 AddrReg)
1170 .addReg(AMDGPU::AR_X,
1171 RegState::Implicit | RegState::Kill);
1172 setImmOperand(Mov, AMDGPU::OpName::src0_rel, 1);
1173
1174 return Mov;
1175 }
1176
getMaxAlusPerClause() const1177 unsigned R600InstrInfo::getMaxAlusPerClause() const {
1178 return 115;
1179 }
1180
buildDefaultInstruction(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,unsigned Opcode,unsigned DstReg,unsigned Src0Reg,unsigned Src1Reg) const1181 MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
1182 MachineBasicBlock::iterator I,
1183 unsigned Opcode,
1184 unsigned DstReg,
1185 unsigned Src0Reg,
1186 unsigned Src1Reg) const {
1187 MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
1188 DstReg); // $dst
1189
1190 if (Src1Reg) {
1191 MIB.addImm(0) // $update_exec_mask
1192 .addImm(0); // $update_predicate
1193 }
1194 MIB.addImm(1) // $write
1195 .addImm(0) // $omod
1196 .addImm(0) // $dst_rel
1197 .addImm(0) // $dst_clamp
1198 .addReg(Src0Reg) // $src0
1199 .addImm(0) // $src0_neg
1200 .addImm(0) // $src0_rel
1201 .addImm(0) // $src0_abs
1202 .addImm(-1); // $src0_sel
1203
1204 if (Src1Reg) {
1205 MIB.addReg(Src1Reg) // $src1
1206 .addImm(0) // $src1_neg
1207 .addImm(0) // $src1_rel
1208 .addImm(0) // $src1_abs
1209 .addImm(-1); // $src1_sel
1210 }
1211
1212 //XXX: The r600g finalizer expects this to be 1, once we've moved the
1213 //scheduling to the backend, we can change the default to 0.
1214 MIB.addImm(1) // $last
1215 .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
1216 .addImm(0) // $literal
1217 .addImm(0); // $bank_swizzle
1218
1219 return MIB;
1220 }
1221
1222 #define OPERAND_CASE(Label) \
1223 case Label: { \
1224 static const unsigned Ops[] = \
1225 { \
1226 Label##_X, \
1227 Label##_Y, \
1228 Label##_Z, \
1229 Label##_W \
1230 }; \
1231 return Ops[Slot]; \
1232 }
1233
getSlotedOps(unsigned Op,unsigned Slot)1234 static unsigned getSlotedOps(unsigned Op, unsigned Slot) {
1235 switch (Op) {
1236 OPERAND_CASE(AMDGPU::OpName::update_exec_mask)
1237 OPERAND_CASE(AMDGPU::OpName::update_pred)
1238 OPERAND_CASE(AMDGPU::OpName::write)
1239 OPERAND_CASE(AMDGPU::OpName::omod)
1240 OPERAND_CASE(AMDGPU::OpName::dst_rel)
1241 OPERAND_CASE(AMDGPU::OpName::clamp)
1242 OPERAND_CASE(AMDGPU::OpName::src0)
1243 OPERAND_CASE(AMDGPU::OpName::src0_neg)
1244 OPERAND_CASE(AMDGPU::OpName::src0_rel)
1245 OPERAND_CASE(AMDGPU::OpName::src0_abs)
1246 OPERAND_CASE(AMDGPU::OpName::src0_sel)
1247 OPERAND_CASE(AMDGPU::OpName::src1)
1248 OPERAND_CASE(AMDGPU::OpName::src1_neg)
1249 OPERAND_CASE(AMDGPU::OpName::src1_rel)
1250 OPERAND_CASE(AMDGPU::OpName::src1_abs)
1251 OPERAND_CASE(AMDGPU::OpName::src1_sel)
1252 OPERAND_CASE(AMDGPU::OpName::pred_sel)
1253 default:
1254 llvm_unreachable("Wrong Operand");
1255 }
1256 }
1257
1258 #undef OPERAND_CASE
1259
buildSlotOfVectorInstruction(MachineBasicBlock & MBB,MachineInstr * MI,unsigned Slot,unsigned DstReg) const1260 MachineInstr *R600InstrInfo::buildSlotOfVectorInstruction(
1261 MachineBasicBlock &MBB, MachineInstr *MI, unsigned Slot, unsigned DstReg)
1262 const {
1263 assert (MI->getOpcode() == AMDGPU::DOT_4 && "Not Implemented");
1264 unsigned Opcode;
1265 if (ST.getGeneration() <= AMDGPUSubtarget::R700)
1266 Opcode = AMDGPU::DOT4_r600;
1267 else
1268 Opcode = AMDGPU::DOT4_eg;
1269 MachineBasicBlock::iterator I = MI;
1270 MachineOperand &Src0 = MI->getOperand(
1271 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src0, Slot)));
1272 MachineOperand &Src1 = MI->getOperand(
1273 getOperandIdx(MI->getOpcode(), getSlotedOps(AMDGPU::OpName::src1, Slot)));
1274 MachineInstr *MIB = buildDefaultInstruction(
1275 MBB, I, Opcode, DstReg, Src0.getReg(), Src1.getReg());
1276 static const unsigned Operands[14] = {
1277 AMDGPU::OpName::update_exec_mask,
1278 AMDGPU::OpName::update_pred,
1279 AMDGPU::OpName::write,
1280 AMDGPU::OpName::omod,
1281 AMDGPU::OpName::dst_rel,
1282 AMDGPU::OpName::clamp,
1283 AMDGPU::OpName::src0_neg,
1284 AMDGPU::OpName::src0_rel,
1285 AMDGPU::OpName::src0_abs,
1286 AMDGPU::OpName::src0_sel,
1287 AMDGPU::OpName::src1_neg,
1288 AMDGPU::OpName::src1_rel,
1289 AMDGPU::OpName::src1_abs,
1290 AMDGPU::OpName::src1_sel,
1291 };
1292
1293 MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
1294 getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
1295 MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
1296 .setReg(MO.getReg());
1297
1298 for (unsigned i = 0; i < 14; i++) {
1299 MachineOperand &MO = MI->getOperand(
1300 getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
1301 assert (MO.isImm());
1302 setImmOperand(MIB, Operands[i], MO.getImm());
1303 }
1304 MIB->getOperand(20).setImm(0);
1305 return MIB;
1306 }
1307
buildMovImm(MachineBasicBlock & BB,MachineBasicBlock::iterator I,unsigned DstReg,uint64_t Imm) const1308 MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
1309 MachineBasicBlock::iterator I,
1310 unsigned DstReg,
1311 uint64_t Imm) const {
1312 MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
1313 AMDGPU::ALU_LITERAL_X);
1314 setImmOperand(MovImm, AMDGPU::OpName::literal, Imm);
1315 return MovImm;
1316 }
1317
buildMovInstr(MachineBasicBlock * MBB,MachineBasicBlock::iterator I,unsigned DstReg,unsigned SrcReg) const1318 MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB,
1319 MachineBasicBlock::iterator I,
1320 unsigned DstReg, unsigned SrcReg) const {
1321 return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg);
1322 }
1323
getOperandIdx(const MachineInstr & MI,unsigned Op) const1324 int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
1325 return getOperandIdx(MI.getOpcode(), Op);
1326 }
1327
getOperandIdx(unsigned Opcode,unsigned Op) const1328 int R600InstrInfo::getOperandIdx(unsigned Opcode, unsigned Op) const {
1329 return AMDGPU::getNamedOperandIdx(Opcode, Op);
1330 }
1331
setImmOperand(MachineInstr * MI,unsigned Op,int64_t Imm) const1332 void R600InstrInfo::setImmOperand(MachineInstr *MI, unsigned Op,
1333 int64_t Imm) const {
1334 int Idx = getOperandIdx(*MI, Op);
1335 assert(Idx != -1 && "Operand not supported for this instruction.");
1336 assert(MI->getOperand(Idx).isImm());
1337 MI->getOperand(Idx).setImm(Imm);
1338 }
1339
1340 //===----------------------------------------------------------------------===//
1341 // Instruction flag getters/setters
1342 //===----------------------------------------------------------------------===//
1343
hasFlagOperand(const MachineInstr & MI) const1344 bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
1345 return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
1346 }
1347
getFlagOp(MachineInstr * MI,unsigned SrcIdx,unsigned Flag) const1348 MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
1349 unsigned Flag) const {
1350 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1351 int FlagIndex = 0;
1352 if (Flag != 0) {
1353 // If we pass something other than the default value of Flag to this
1354 // function, it means we are want to set a flag on an instruction
1355 // that uses native encoding.
1356 assert(HAS_NATIVE_OPERANDS(TargetFlags));
1357 bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
1358 switch (Flag) {
1359 case MO_FLAG_CLAMP:
1360 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::clamp);
1361 break;
1362 case MO_FLAG_MASK:
1363 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::write);
1364 break;
1365 case MO_FLAG_NOT_LAST:
1366 case MO_FLAG_LAST:
1367 FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::last);
1368 break;
1369 case MO_FLAG_NEG:
1370 switch (SrcIdx) {
1371 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_neg); break;
1372 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_neg); break;
1373 case 2: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src2_neg); break;
1374 }
1375 break;
1376
1377 case MO_FLAG_ABS:
1378 assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
1379 "instructions.");
1380 (void)IsOP3;
1381 switch (SrcIdx) {
1382 case 0: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src0_abs); break;
1383 case 1: FlagIndex = getOperandIdx(*MI, AMDGPU::OpName::src1_abs); break;
1384 }
1385 break;
1386
1387 default:
1388 FlagIndex = -1;
1389 break;
1390 }
1391 assert(FlagIndex != -1 && "Flag not supported for this instruction");
1392 } else {
1393 FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
1394 assert(FlagIndex != 0 &&
1395 "Instruction flags not supported for this instruction");
1396 }
1397
1398 MachineOperand &FlagOp = MI->getOperand(FlagIndex);
1399 assert(FlagOp.isImm());
1400 return FlagOp;
1401 }
1402
addFlag(MachineInstr * MI,unsigned Operand,unsigned Flag) const1403 void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
1404 unsigned Flag) const {
1405 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1406 if (Flag == 0) {
1407 return;
1408 }
1409 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1410 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1411 if (Flag == MO_FLAG_NOT_LAST) {
1412 clearFlag(MI, Operand, MO_FLAG_LAST);
1413 } else if (Flag == MO_FLAG_MASK) {
1414 clearFlag(MI, Operand, Flag);
1415 } else {
1416 FlagOp.setImm(1);
1417 }
1418 } else {
1419 MachineOperand &FlagOp = getFlagOp(MI, Operand);
1420 FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
1421 }
1422 }
1423
clearFlag(MachineInstr * MI,unsigned Operand,unsigned Flag) const1424 void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
1425 unsigned Flag) const {
1426 unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
1427 if (HAS_NATIVE_OPERANDS(TargetFlags)) {
1428 MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
1429 FlagOp.setImm(0);
1430 } else {
1431 MachineOperand &FlagOp = getFlagOp(MI);
1432 unsigned InstFlags = FlagOp.getImm();
1433 InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
1434 FlagOp.setImm(InstFlags);
1435 }
1436 }
1437