1 //===-- R600MachineScheduler.cpp - R600 Scheduler Interface -*- C++ -*-----===//
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 //===----------------------------------------------------------------------===//
11 /// \brief R600 Machine Scheduler interface
13 //===----------------------------------------------------------------------===//
15 #include "R600MachineScheduler.h"
16 #include "AMDGPUSubtarget.h"
17 #include "llvm/CodeGen/MachineRegisterInfo.h"
18 #include "llvm/Pass.h"
19 #include "llvm/PassManager.h"
20 #include "llvm/Support/raw_ostream.h"
24 #define DEBUG_TYPE "misched"
26 void R600SchedStrategy::initialize(ScheduleDAGMI
*dag
) {
27 assert(dag
->hasVRegLiveness() && "R600SchedStrategy needs vreg liveness");
28 DAG
= static_cast<ScheduleDAGMILive
*>(dag
);
29 TII
= static_cast<const R600InstrInfo
*>(DAG
->TII
);
30 TRI
= static_cast<const R600RegisterInfo
*>(DAG
->TRI
);
31 VLIW5
= !DAG
->MF
.getTarget().getSubtarget
<AMDGPUSubtarget
>().hasCaymanISA();
33 CurInstKind
= IDOther
;
35 OccupedSlotsMask
= 31;
36 InstKindLimit
[IDAlu
] = TII
->getMaxAlusPerClause();
37 InstKindLimit
[IDOther
] = 32;
39 const AMDGPUSubtarget
&ST
= DAG
->TM
.getSubtarget
<AMDGPUSubtarget
>();
40 InstKindLimit
[IDFetch
] = ST
.getTexVTXClauseSize();
45 void R600SchedStrategy::MoveUnits(std::vector
<SUnit
*> &QSrc
,
46 std::vector
<SUnit
*> &QDst
)
48 QDst
.insert(QDst
.end(), QSrc
.begin(), QSrc
.end());
53 unsigned getWFCountLimitedByGPR(unsigned GPRCount
) {
54 assert (GPRCount
&& "GPRCount cannot be 0");
55 return 248 / GPRCount
;
58 SUnit
* R600SchedStrategy::pickNode(bool &IsTopNode
) {
60 NextInstKind
= IDOther
;
64 // check if we might want to switch current clause type
65 bool AllowSwitchToAlu
= (CurEmitted
>= InstKindLimit
[CurInstKind
]) ||
66 (Available
[CurInstKind
].empty());
67 bool AllowSwitchFromAlu
= (CurEmitted
>= InstKindLimit
[CurInstKind
]) &&
68 (!Available
[IDFetch
].empty() || !Available
[IDOther
].empty());
70 if (CurInstKind
== IDAlu
&& !Available
[IDFetch
].empty()) {
71 // We use the heuristic provided by AMD Accelerated Parallel Processing
72 // OpenCL Programming Guide :
73 // The approx. number of WF that allows TEX inst to hide ALU inst is :
74 // 500 (cycles for TEX) / (AluFetchRatio * 8 (cycles for ALU))
75 float ALUFetchRationEstimate
=
76 (AluInstCount
+ AvailablesAluCount() + Pending
[IDAlu
].size()) /
77 (FetchInstCount
+ Available
[IDFetch
].size());
78 if (ALUFetchRationEstimate
== 0) {
79 AllowSwitchFromAlu
= true;
81 unsigned NeededWF
= 62.5f
/ ALUFetchRationEstimate
;
82 DEBUG( dbgs() << NeededWF
<< " approx. Wavefronts Required\n" );
83 // We assume the local GPR requirements to be "dominated" by the requirement
84 // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
85 // after TEX are indeed likely to consume or generate values from/for the
87 // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
88 // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
89 // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
90 // (TODO : use RegisterPressure)
91 // If we are going too use too many GPR, we flush Fetch instruction to lower
92 // register pressure on 128 bits regs.
93 unsigned NearRegisterRequirement
= 2 * Available
[IDFetch
].size();
94 if (NeededWF
> getWFCountLimitedByGPR(NearRegisterRequirement
))
95 AllowSwitchFromAlu
= true;
99 if (!SU
&& ((AllowSwitchToAlu
&& CurInstKind
!= IDAlu
) ||
100 (!AllowSwitchFromAlu
&& CurInstKind
== IDAlu
))) {
103 if (!SU
&& !PhysicalRegCopy
.empty()) {
104 SU
= PhysicalRegCopy
.front();
105 PhysicalRegCopy
.erase(PhysicalRegCopy
.begin());
108 if (CurEmitted
>= InstKindLimit
[IDAlu
])
110 NextInstKind
= IDAlu
;
116 SU
= pickOther(IDFetch
);
118 NextInstKind
= IDFetch
;
123 SU
= pickOther(IDOther
);
125 NextInstKind
= IDOther
;
130 dbgs() << " ** Pick node **\n";
133 dbgs() << "NO NODE \n";
134 for (unsigned i
= 0; i
< DAG
->SUnits
.size(); i
++) {
135 const SUnit
&S
= DAG
->SUnits
[i
];
145 void R600SchedStrategy::schedNode(SUnit
*SU
, bool IsTopNode
) {
146 if (NextInstKind
!= CurInstKind
) {
147 DEBUG(dbgs() << "Instruction Type Switch\n");
148 if (NextInstKind
!= IDAlu
)
149 OccupedSlotsMask
|= 31;
151 CurInstKind
= NextInstKind
;
154 if (CurInstKind
== IDAlu
) {
156 switch (getAluKind(SU
)) {
164 for (MachineInstr::mop_iterator It
= SU
->getInstr()->operands_begin(),
165 E
= SU
->getInstr()->operands_end(); It
!= E
; ++It
) {
166 MachineOperand
&MO
= *It
;
167 if (MO
.isReg() && MO
.getReg() == AMDGPU::ALU_LITERAL_X
)
177 DEBUG(dbgs() << CurEmitted
<< " Instructions Emitted in this clause\n");
179 if (CurInstKind
!= IDFetch
) {
180 MoveUnits(Pending
[IDFetch
], Available
[IDFetch
]);
186 isPhysicalRegCopy(MachineInstr
*MI
) {
187 if (MI
->getOpcode() != AMDGPU::COPY
)
190 return !TargetRegisterInfo::isVirtualRegister(MI
->getOperand(1).getReg());
193 void R600SchedStrategy::releaseTopNode(SUnit
*SU
) {
194 DEBUG(dbgs() << "Top Releasing ";SU
->dump(DAG
););
197 void R600SchedStrategy::releaseBottomNode(SUnit
*SU
) {
198 DEBUG(dbgs() << "Bottom Releasing ";SU
->dump(DAG
););
199 if (isPhysicalRegCopy(SU
->getInstr())) {
200 PhysicalRegCopy
.push_back(SU
);
204 int IK
= getInstKind(SU
);
206 // There is no export clause, we can schedule one as soon as its ready
208 Available
[IDOther
].push_back(SU
);
210 Pending
[IK
].push_back(SU
);
214 bool R600SchedStrategy::regBelongsToClass(unsigned Reg
,
215 const TargetRegisterClass
*RC
) const {
216 if (!TargetRegisterInfo::isVirtualRegister(Reg
)) {
217 return RC
->contains(Reg
);
219 return MRI
->getRegClass(Reg
) == RC
;
223 R600SchedStrategy::AluKind
R600SchedStrategy::getAluKind(SUnit
*SU
) const {
224 MachineInstr
*MI
= SU
->getInstr();
226 if (TII
->isTransOnly(MI
))
229 switch (MI
->getOpcode()) {
232 case AMDGPU::INTERP_PAIR_XY
:
233 case AMDGPU::INTERP_PAIR_ZW
:
234 case AMDGPU::INTERP_VEC_LOAD
:
238 if (MI
->getOperand(1).isUndef()) {
239 // MI will become a KILL, don't considers it in scheduling
246 // Does the instruction take a whole IG ?
247 // XXX: Is it possible to add a helper function in R600InstrInfo that can
248 // be used here and in R600PacketizerList::isSoloInstruction() ?
249 if(TII
->isVector(*MI
) ||
250 TII
->isCubeOp(MI
->getOpcode()) ||
251 TII
->isReductionOp(MI
->getOpcode()) ||
252 MI
->getOpcode() == AMDGPU::GROUP_BARRIER
) {
256 if (TII
->isLDSInstr(MI
->getOpcode())) {
260 // Is the result already assigned to a channel ?
261 unsigned DestSubReg
= MI
->getOperand(0).getSubReg();
262 switch (DestSubReg
) {
275 // Is the result already member of a X/Y/Z/W class ?
276 unsigned DestReg
= MI
->getOperand(0).getReg();
277 if (regBelongsToClass(DestReg
, &AMDGPU::R600_TReg32_XRegClass
) ||
278 regBelongsToClass(DestReg
, &AMDGPU::R600_AddrRegClass
))
280 if (regBelongsToClass(DestReg
, &AMDGPU::R600_TReg32_YRegClass
))
282 if (regBelongsToClass(DestReg
, &AMDGPU::R600_TReg32_ZRegClass
))
284 if (regBelongsToClass(DestReg
, &AMDGPU::R600_TReg32_WRegClass
))
286 if (regBelongsToClass(DestReg
, &AMDGPU::R600_Reg128RegClass
))
289 // LDS src registers cannot be used in the Trans slot.
290 if (TII
->readsLDSSrcReg(MI
))
297 int R600SchedStrategy::getInstKind(SUnit
* SU
) {
298 int Opcode
= SU
->getInstr()->getOpcode();
300 if (TII
->usesTextureCache(Opcode
) || TII
->usesVertexCache(Opcode
))
303 if (TII
->isALUInstr(Opcode
)) {
310 case AMDGPU::CONST_COPY
:
311 case AMDGPU::INTERP_PAIR_XY
:
312 case AMDGPU::INTERP_PAIR_ZW
:
313 case AMDGPU::INTERP_VEC_LOAD
:
321 SUnit
*R600SchedStrategy::PopInst(std::vector
<SUnit
*> &Q
, bool AnyALU
) {
324 for (std::vector
<SUnit
*>::reverse_iterator It
= Q
.rbegin(), E
= Q
.rend();
327 InstructionsGroupCandidate
.push_back(SU
->getInstr());
328 if (TII
->fitsConstReadLimitations(InstructionsGroupCandidate
)
329 && (!AnyALU
|| !TII
->isVectorOnly(SU
->getInstr()))
331 InstructionsGroupCandidate
.pop_back();
332 Q
.erase((It
+ 1).base());
335 InstructionsGroupCandidate
.pop_back();
341 void R600SchedStrategy::LoadAlu() {
342 std::vector
<SUnit
*> &QSrc
= Pending
[IDAlu
];
343 for (unsigned i
= 0, e
= QSrc
.size(); i
< e
; ++i
) {
344 AluKind AK
= getAluKind(QSrc
[i
]);
345 AvailableAlus
[AK
].push_back(QSrc
[i
]);
350 void R600SchedStrategy::PrepareNextSlot() {
351 DEBUG(dbgs() << "New Slot\n");
352 assert (OccupedSlotsMask
&& "Slot wasn't filled");
353 OccupedSlotsMask
= 0;
354 // if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
355 // OccupedSlotsMask |= 16;
356 InstructionsGroupCandidate
.clear();
360 void R600SchedStrategy::AssignSlot(MachineInstr
* MI
, unsigned Slot
) {
361 int DstIndex
= TII
->getOperandIdx(MI
->getOpcode(), AMDGPU::OpName::dst
);
362 if (DstIndex
== -1) {
365 unsigned DestReg
= MI
->getOperand(DstIndex
).getReg();
366 // PressureRegister crashes if an operand is def and used in the same inst
367 // and we try to constraint its regclass
368 for (MachineInstr::mop_iterator It
= MI
->operands_begin(),
369 E
= MI
->operands_end(); It
!= E
; ++It
) {
370 MachineOperand
&MO
= *It
;
371 if (MO
.isReg() && !MO
.isDef() &&
372 MO
.getReg() == DestReg
)
375 // Constrains the regclass of DestReg to assign it to Slot
378 MRI
->constrainRegClass(DestReg
, &AMDGPU::R600_TReg32_XRegClass
);
381 MRI
->constrainRegClass(DestReg
, &AMDGPU::R600_TReg32_YRegClass
);
384 MRI
->constrainRegClass(DestReg
, &AMDGPU::R600_TReg32_ZRegClass
);
387 MRI
->constrainRegClass(DestReg
, &AMDGPU::R600_TReg32_WRegClass
);
392 SUnit
*R600SchedStrategy::AttemptFillSlot(unsigned Slot
, bool AnyAlu
) {
393 static const AluKind IndexToID
[] = {AluT_X
, AluT_Y
, AluT_Z
, AluT_W
};
394 SUnit
*SlotedSU
= PopInst(AvailableAlus
[IndexToID
[Slot
]], AnyAlu
);
397 SUnit
*UnslotedSU
= PopInst(AvailableAlus
[AluAny
], AnyAlu
);
399 AssignSlot(UnslotedSU
->getInstr(), Slot
);
403 unsigned R600SchedStrategy::AvailablesAluCount() const {
404 return AvailableAlus
[AluAny
].size() + AvailableAlus
[AluT_XYZW
].size() +
405 AvailableAlus
[AluT_X
].size() + AvailableAlus
[AluT_Y
].size() +
406 AvailableAlus
[AluT_Z
].size() + AvailableAlus
[AluT_W
].size() +
407 AvailableAlus
[AluTrans
].size() + AvailableAlus
[AluDiscarded
].size() +
408 AvailableAlus
[AluPredX
].size();
411 SUnit
* R600SchedStrategy::pickAlu() {
412 while (AvailablesAluCount() || !Pending
[IDAlu
].empty()) {
413 if (!OccupedSlotsMask
) {
414 // Bottom up scheduling : predX must comes first
415 if (!AvailableAlus
[AluPredX
].empty()) {
416 OccupedSlotsMask
|= 31;
417 return PopInst(AvailableAlus
[AluPredX
], false);
419 // Flush physical reg copies (RA will discard them)
420 if (!AvailableAlus
[AluDiscarded
].empty()) {
421 OccupedSlotsMask
|= 31;
422 return PopInst(AvailableAlus
[AluDiscarded
], false);
424 // If there is a T_XYZW alu available, use it
425 if (!AvailableAlus
[AluT_XYZW
].empty()) {
426 OccupedSlotsMask
|= 15;
427 return PopInst(AvailableAlus
[AluT_XYZW
], false);
430 bool TransSlotOccuped
= OccupedSlotsMask
& 16;
431 if (!TransSlotOccuped
&& VLIW5
) {
432 if (!AvailableAlus
[AluTrans
].empty()) {
433 OccupedSlotsMask
|= 16;
434 return PopInst(AvailableAlus
[AluTrans
], false);
436 SUnit
*SU
= AttemptFillSlot(3, true);
438 OccupedSlotsMask
|= 16;
442 for (int Chan
= 3; Chan
> -1; --Chan
) {
443 bool isOccupied
= OccupedSlotsMask
& (1 << Chan
);
445 SUnit
*SU
= AttemptFillSlot(Chan
, false);
447 OccupedSlotsMask
|= (1 << Chan
);
448 InstructionsGroupCandidate
.push_back(SU
->getInstr());
458 SUnit
* R600SchedStrategy::pickOther(int QID
) {
460 std::vector
<SUnit
*> &AQ
= Available
[QID
];
463 MoveUnits(Pending
[QID
], AQ
);
467 AQ
.resize(AQ
.size() - 1);