1//===- CriticalAntiDepBreaker.cpp - Anti-dep breaker ----------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the CriticalAntiDepBreaker class, which
10// implements register anti-dependence breaking along a blocks
11// critical path during post-RA scheduler.
12//
13//===----------------------------------------------------------------------===//
14
15#include "CriticalAntiDepBreaker.h"
16#include "llvm/ADT/ArrayRef.h"
17#include "llvm/ADT/DenseMap.h"
18#include "llvm/ADT/SmallVector.h"
19#include "llvm/CodeGen/MachineBasicBlock.h"
20#include "llvm/CodeGen/MachineFrameInfo.h"
21#include "llvm/CodeGen/MachineFunction.h"
22#include "llvm/CodeGen/MachineInstr.h"
23#include "llvm/CodeGen/MachineOperand.h"
24#include "llvm/CodeGen/MachineRegisterInfo.h"
25#include "llvm/CodeGen/RegisterClassInfo.h"
26#include "llvm/CodeGen/ScheduleDAG.h"
27#include "llvm/CodeGen/TargetInstrInfo.h"
28#include "llvm/CodeGen/TargetRegisterInfo.h"
29#include "llvm/CodeGen/TargetSubtargetInfo.h"
30#include "llvm/MC/MCInstrDesc.h"
31#include "llvm/MC/MCRegisterInfo.h"
32#include "llvm/Support/Debug.h"
33#include "llvm/Support/raw_ostream.h"
34#include <cassert>
35#include <utility>
36
37using namespace llvm;
38
39#define DEBUG_TYPE "post-RA-sched"
40
41CriticalAntiDepBreaker::CriticalAntiDepBreaker(MachineFunction &MFi,
42 const RegisterClassInfo &RCI)
43 : MF(MFi), MRI(MF.getRegInfo()), TII(MF.getSubtarget().getInstrInfo()),
44 TRI(MF.getSubtarget().getRegisterInfo()), RegClassInfo(RCI),
45 Classes(TRI->getNumRegs(), nullptr), KillIndices(TRI->getNumRegs(), 0),
46 DefIndices(TRI->getNumRegs(), 0), KeepRegs(TRI->getNumRegs(), false) {}
47
48CriticalAntiDepBreaker::~CriticalAntiDepBreaker() = default;
49
50void CriticalAntiDepBreaker::StartBlock(MachineBasicBlock *BB) {
51 const unsigned BBSize = BB->size();
52 for (unsigned i = 1, e = TRI->getNumRegs(); i != e; ++i) {
53 // Clear out the register class data.
54 Classes[i] = nullptr;
55
56 // Initialize the indices to indicate that no registers are live.
57 KillIndices[i] = ~0u;
58 DefIndices[i] = BBSize;
59 }
60
61 // Clear "do not change" set.
62 KeepRegs.reset();
63
64 bool IsReturnBlock = BB->isReturnBlock();
65
66 // Examine the live-in regs of all successors.
67 for (const MachineBasicBlock *Succ : BB->successors())
68 for (const auto &LI : Succ->liveins()) {
69 for (MCRegAliasIterator AI(LI.PhysReg, TRI, true); AI.isValid(); ++AI) {
70 MCRegister Reg = *AI;
71 Classes[Reg.id()] = reinterpret_cast<TargetRegisterClass *>(-1);
72 KillIndices[Reg.id()] = BBSize;
73 DefIndices[Reg.id()] = ~0u;
74 }
75 }
76
77 // Mark live-out callee-saved registers. In a return block this is
78 // all callee-saved registers. In non-return this is any
79 // callee-saved register that is not saved in the prolog.
80 const MachineFrameInfo &MFI = MF.getFrameInfo();
81 BitVector Pristine = MFI.getPristineRegs(MF);
82 for (const MCPhysReg *I = MF.getRegInfo().getCalleeSavedRegs(); *I;
83 ++I) {
84 unsigned Reg = *I;
85 if (!IsReturnBlock && !Pristine.test(Idx: Reg))
86 continue;
87 for (MCRegAliasIterator AI(*I, TRI, true); AI.isValid(); ++AI) {
88 MCRegister Reg = *AI;
89 Classes[Reg.id()] = reinterpret_cast<TargetRegisterClass *>(-1);
90 KillIndices[Reg.id()] = BBSize;
91 DefIndices[Reg.id()] = ~0u;
92 }
93 }
94}
95
96void CriticalAntiDepBreaker::FinishBlock() {
97 RegRefs.clear();
98 KeepRegs.reset();
99}
100
101void CriticalAntiDepBreaker::Observe(MachineInstr &MI, unsigned Count,
102 unsigned InsertPosIndex) {
103 // Kill instructions can define registers but are really nops, and there might
104 // be a real definition earlier that needs to be paired with uses dominated by
105 // this kill.
106
107 // FIXME: It may be possible to remove the isKill() restriction once PR18663
108 // has been properly fixed. There can be value in processing kills as seen in
109 // the AggressiveAntiDepBreaker class.
110 if (MI.isDebugInstr() || MI.isKill())
111 return;
112 assert(Count < InsertPosIndex && "Instruction index out of expected range!");
113
114 for (unsigned Reg = 1; Reg != TRI->getNumRegs(); ++Reg) {
115 if (KillIndices[Reg] != ~0u) {
116 // If Reg is currently live, then mark that it can't be renamed as
117 // we don't know the extent of its live-range anymore (now that it
118 // has been scheduled).
119 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
120 KillIndices[Reg] = Count;
121 } else if (DefIndices[Reg] < InsertPosIndex && DefIndices[Reg] >= Count) {
122 // Any register which was defined within the previous scheduling region
123 // may have been rescheduled and its lifetime may overlap with registers
124 // in ways not reflected in our current liveness state. For each such
125 // register, adjust the liveness state to be conservatively correct.
126 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
127
128 // Move the def index to the end of the previous region, to reflect
129 // that the def could theoretically have been scheduled at the end.
130 DefIndices[Reg] = InsertPosIndex;
131 }
132 }
133
134 PrescanInstruction(MI);
135 ScanInstruction(MI, Count);
136}
137
138/// CriticalPathStep - Return the next SUnit after SU on the bottom-up
139/// critical path.
140static const SDep *CriticalPathStep(const SUnit *SU) {
141 const SDep *Next = nullptr;
142 unsigned NextDepth = 0;
143 // Find the predecessor edge with the greatest depth.
144 for (const SDep &P : SU->Preds) {
145 const SUnit *PredSU = P.getSUnit();
146 unsigned PredLatency = P.getLatency();
147 unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
148 // In the case of a latency tie, prefer an anti-dependency edge over
149 // other types of edges.
150 if (NextDepth < PredTotalLatency ||
151 (NextDepth == PredTotalLatency && P.getKind() == SDep::Anti)) {
152 NextDepth = PredTotalLatency;
153 Next = &P;
154 }
155 }
156 return Next;
157}
158
159void CriticalAntiDepBreaker::PrescanInstruction(MachineInstr &MI) {
160 // It's not safe to change register allocation for source operands of
161 // instructions that have special allocation requirements. Also assume all
162 // registers used in a call must not be changed (ABI).
163 // FIXME: The issue with predicated instruction is more complex. We are being
164 // conservative here because the kill markers cannot be trusted after
165 // if-conversion:
166 // %r6 = LDR %sp, %reg0, 92, 14, %reg0; mem:LD4[FixedStack14]
167 // ...
168 // STR %r0, killed %r6, %reg0, 0, 0, %cpsr; mem:ST4[%395]
169 // %r6 = LDR %sp, %reg0, 100, 0, %cpsr; mem:LD4[FixedStack12]
170 // STR %r0, killed %r6, %reg0, 0, 14, %reg0; mem:ST4[%396](align=8)
171 //
172 // The first R6 kill is not really a kill since it's killed by a predicated
173 // instruction which may not be executed. The second R6 def may or may not
174 // re-define R6 so it's not safe to change it since the last R6 use cannot be
175 // changed.
176 bool Special =
177 MI.isCall() || MI.hasExtraSrcRegAllocReq() || TII->isPredicated(MI);
178
179 // Scan the register operands for this instruction and update
180 // Classes and RegRefs.
181 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
182 MachineOperand &MO = MI.getOperand(i);
183 if (!MO.isReg()) continue;
184 Register Reg = MO.getReg();
185 if (!Reg)
186 continue;
187 const TargetRegisterClass *NewRC = nullptr;
188
189 if (i < MI.getDesc().getNumOperands())
190 NewRC = TII->getRegClass(MCID: MI.getDesc(), OpNum: i);
191
192 // For now, only allow the register to be changed if its register
193 // class is consistent across all uses.
194 if (!Classes[Reg.id()] && NewRC)
195 Classes[Reg.id()] = NewRC;
196 else if (!NewRC || Classes[Reg.id()] != NewRC)
197 Classes[Reg.id()] = reinterpret_cast<TargetRegisterClass *>(-1);
198
199 // Now check for aliases.
200 for (MCRegAliasIterator AI(Reg, TRI, false); AI.isValid(); ++AI) {
201 // If an alias of the reg is used during the live range, give up.
202 // Note that this allows us to skip checking if AntiDepReg
203 // overlaps with any of the aliases, among other things.
204 unsigned AliasReg = (*AI).id();
205 if (Classes[AliasReg]) {
206 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
207 Classes[Reg.id()] = reinterpret_cast<TargetRegisterClass *>(-1);
208 }
209 }
210
211 // If we're still willing to consider this register, note the reference.
212 if (Classes[Reg.id()] != reinterpret_cast<TargetRegisterClass *>(-1))
213 RegRefs.emplace(args&: Reg, args: &MO);
214
215 if (MO.isUse() && Special) {
216 if (!KeepRegs.test(Idx: Reg.id())) {
217 for (MCPhysReg SubReg : TRI->subregs_inclusive(Reg))
218 KeepRegs.set(SubReg);
219 }
220 }
221 }
222
223 for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
224 const MachineOperand &MO = MI.getOperand(i: I);
225 if (!MO.isReg()) continue;
226 Register Reg = MO.getReg();
227 if (!Reg.isValid())
228 continue;
229 // If this reg is tied and live (Classes[Reg] is set to -1), we can't change
230 // it or any of its sub or super regs. We need to use KeepRegs to mark the
231 // reg because not all uses of the same reg within an instruction are
232 // necessarily tagged as tied.
233 // Example: an x86 "xor %eax, %eax" will have one source operand tied to the
234 // def register but not the second (see PR20020 for details).
235 // FIXME: can this check be relaxed to account for undef uses
236 // of a register? In the above 'xor' example, the uses of %eax are undef, so
237 // earlier instructions could still replace %eax even though the 'xor'
238 // itself can't be changed.
239 if (MI.isRegTiedToUseOperand(DefOpIdx: I) &&
240 Classes[Reg.id()] == reinterpret_cast<TargetRegisterClass *>(-1)) {
241 for (MCPhysReg SubReg : TRI->subregs_inclusive(Reg)) {
242 KeepRegs.set(SubReg);
243 }
244 for (MCPhysReg SuperReg : TRI->superregs(Reg)) {
245 KeepRegs.set(SuperReg);
246 }
247 }
248 }
249}
250
251void CriticalAntiDepBreaker::ScanInstruction(MachineInstr &MI, unsigned Count) {
252 // Update liveness.
253 // Proceeding upwards, registers that are defed but not used in this
254 // instruction are now dead.
255 assert(!MI.isKill() && "Attempting to scan a kill instruction");
256
257 if (!TII->isPredicated(MI)) {
258 // Predicated defs are modeled as read + write, i.e. similar to two
259 // address updates.
260 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
261 MachineOperand &MO = MI.getOperand(i);
262
263 if (MO.isRegMask()) {
264 auto ClobbersPhysRegAndSubRegs = [&](unsigned PhysReg) {
265 return all_of(Range: TRI->subregs_inclusive(Reg: PhysReg),
266 P: [&](MCPhysReg SR) { return MO.clobbersPhysReg(PhysReg: SR); });
267 };
268
269 for (unsigned i = 1, e = TRI->getNumRegs(); i != e; ++i) {
270 if (ClobbersPhysRegAndSubRegs(i)) {
271 DefIndices[i] = Count;
272 KillIndices[i] = ~0u;
273 KeepRegs.reset(Idx: i);
274 Classes[i] = nullptr;
275 RegRefs.erase(x: i);
276 }
277 }
278 }
279
280 if (!MO.isReg()) continue;
281 Register Reg = MO.getReg();
282 if (!Reg)
283 continue;
284 if (!MO.isDef()) continue;
285
286 // Ignore two-addr defs.
287 if (MI.isRegTiedToUseOperand(DefOpIdx: i))
288 continue;
289
290 // If we've already marked this reg as unchangeable, don't remove
291 // it or any of its subregs from KeepRegs.
292 bool Keep = KeepRegs.test(Idx: Reg.id());
293
294 // For the reg itself and all subregs: update the def to current;
295 // reset the kill state, any restrictions, and references.
296 for (MCPhysReg SubregReg : TRI->subregs_inclusive(Reg)) {
297 DefIndices[SubregReg] = Count;
298 KillIndices[SubregReg] = ~0u;
299 Classes[SubregReg] = nullptr;
300 RegRefs.erase(x: SubregReg);
301 if (!Keep)
302 KeepRegs.reset(Idx: SubregReg);
303 }
304 // Conservatively mark super-registers as unusable.
305 for (MCPhysReg SR : TRI->superregs(Reg))
306 Classes[SR] = reinterpret_cast<TargetRegisterClass *>(-1);
307 }
308 }
309 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
310 MachineOperand &MO = MI.getOperand(i);
311 if (!MO.isReg()) continue;
312 Register Reg = MO.getReg();
313 if (!Reg)
314 continue;
315 if (!MO.isUse()) continue;
316
317 const TargetRegisterClass *NewRC = nullptr;
318 if (i < MI.getDesc().getNumOperands())
319 NewRC = TII->getRegClass(MCID: MI.getDesc(), OpNum: i);
320
321 // For now, only allow the register to be changed if its register
322 // class is consistent across all uses.
323 if (!Classes[Reg.id()] && NewRC)
324 Classes[Reg.id()] = NewRC;
325 else if (!NewRC || Classes[Reg.id()] != NewRC)
326 Classes[Reg.id()] = reinterpret_cast<TargetRegisterClass *>(-1);
327
328 RegRefs.emplace(args&: Reg, args: &MO);
329
330 // It wasn't previously live but now it is, this is a kill.
331 // Repeat for all aliases.
332 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) {
333 MCRegister AliasReg = *AI;
334 if (KillIndices[AliasReg.id()] == ~0u) {
335 KillIndices[AliasReg.id()] = Count;
336 DefIndices[AliasReg.id()] = ~0u;
337 }
338 }
339 }
340}
341
342// Check all machine operands that reference the antidependent register and must
343// be replaced by NewReg. Return true if any of their parent instructions may
344// clobber the new register.
345//
346// Note: AntiDepReg may be referenced by a two-address instruction such that
347// it's use operand is tied to a def operand. We guard against the case in which
348// the two-address instruction also defines NewReg, as may happen with
349// pre/postincrement loads. In this case, both the use and def operands are in
350// RegRefs because the def is inserted by PrescanInstruction and not erased
351// during ScanInstruction. So checking for an instruction with definitions of
352// both NewReg and AntiDepReg covers it.
353bool CriticalAntiDepBreaker::isNewRegClobberedByRefs(RegRefIter RegRefBegin,
354 RegRefIter RegRefEnd,
355 MCRegister NewReg) {
356 for (RegRefIter I = RegRefBegin; I != RegRefEnd; ++I ) {
357 MachineOperand *RefOper = I->second;
358
359 // Don't allow the instruction defining AntiDepReg to earlyclobber its
360 // operands, in case they may be assigned to NewReg. In this case antidep
361 // breaking must fail, but it's too rare to bother optimizing.
362 if (RefOper->isDef() && RefOper->isEarlyClobber())
363 return true;
364
365 // Handle cases in which this instruction defines NewReg.
366 MachineInstr *MI = RefOper->getParent();
367 for (const MachineOperand &CheckOper : MI->operands()) {
368 if (CheckOper.isRegMask() && CheckOper.clobbersPhysReg(PhysReg: NewReg))
369 return true;
370
371 if (!CheckOper.isReg() || !CheckOper.isDef() ||
372 CheckOper.getReg() != NewReg)
373 continue;
374
375 // Don't allow the instruction to define NewReg and AntiDepReg.
376 // When AntiDepReg is renamed it will be an illegal op.
377 if (RefOper->isDef())
378 return true;
379
380 // Don't allow an instruction using AntiDepReg to be earlyclobbered by
381 // NewReg.
382 if (CheckOper.isEarlyClobber())
383 return true;
384
385 // Don't allow inline asm to define NewReg at all. Who knows what it's
386 // doing with it.
387 if (MI->isInlineAsm())
388 return true;
389 }
390 }
391 return false;
392}
393
394MCRegister CriticalAntiDepBreaker::findSuitableFreeRegister(
395 RegRefIter RegRefBegin, RegRefIter RegRefEnd, MCRegister AntiDepReg,
396 MCRegister LastNewReg, const TargetRegisterClass *RC,
397 const SmallVectorImpl<Register> &Forbid) {
398 ArrayRef<MCPhysReg> Order = RegClassInfo.getOrder(RC);
399 for (MCRegister NewReg : Order) {
400 // Don't replace a register with itself.
401 if (NewReg == AntiDepReg) continue;
402 // Don't replace a register with one that was recently used to repair
403 // an anti-dependence with this AntiDepReg, because that would
404 // re-introduce that anti-dependence.
405 if (NewReg == LastNewReg) continue;
406 // If any instructions that define AntiDepReg also define the NewReg, it's
407 // not suitable. For example, Instruction with multiple definitions can
408 // result in this condition.
409 if (isNewRegClobberedByRefs(RegRefBegin, RegRefEnd, NewReg)) continue;
410 // If NewReg is dead and NewReg's most recent def is not before
411 // AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
412 assert(((KillIndices[AntiDepReg.id()] == ~0u) !=
413 (DefIndices[AntiDepReg.id()] == ~0u)) &&
414 "Kill and Def maps aren't consistent for AntiDepReg!");
415 assert(((KillIndices[NewReg.id()] == ~0u) !=
416 (DefIndices[NewReg.id()] == ~0u)) &&
417 "Kill and Def maps aren't consistent for NewReg!");
418 if (KillIndices[NewReg.id()] != ~0u ||
419 Classes[NewReg.id()] == reinterpret_cast<TargetRegisterClass *>(-1) ||
420 KillIndices[AntiDepReg.id()] > DefIndices[NewReg.id()])
421 continue;
422 // If NewReg overlaps any of the forbidden registers, we can't use it.
423 bool Forbidden = false;
424 for (Register R : Forbid)
425 if (TRI->regsOverlap(RegA: NewReg, RegB: R)) {
426 Forbidden = true;
427 break;
428 }
429 if (Forbidden) continue;
430 return NewReg;
431 }
432
433 // No registers are free and available!
434 return MCRegister();
435}
436
437unsigned CriticalAntiDepBreaker::
438BreakAntiDependencies(const std::vector<SUnit> &SUnits,
439 MachineBasicBlock::iterator Begin,
440 MachineBasicBlock::iterator End,
441 unsigned InsertPosIndex,
442 DbgValueVector &DbgValues) {
443 // The code below assumes that there is at least one instruction,
444 // so just duck out immediately if the block is empty.
445 if (SUnits.empty()) return 0;
446
447 // Keep a map of the MachineInstr*'s back to the SUnit representing them.
448 // This is used for updating debug information.
449 //
450 // FIXME: Replace this with the existing map in ScheduleDAGInstrs::MISUnitMap
451 DenseMap<MachineInstr *, const SUnit *> MISUnitMap;
452
453 // Find the node at the bottom of the critical path.
454 const SUnit *Max = nullptr;
455 for (const SUnit &SU : SUnits) {
456 MISUnitMap[SU.getInstr()] = &SU;
457 if (!Max || SU.getDepth() + SU.Latency > Max->getDepth() + Max->Latency)
458 Max = &SU;
459 }
460 assert(Max && "Failed to find bottom of the critical path");
461
462#ifndef NDEBUG
463 {
464 LLVM_DEBUG(dbgs() << "Critical path has total latency "
465 << (Max->getDepth() + Max->Latency) << "\n");
466 LLVM_DEBUG(dbgs() << "Available regs:");
467 for (unsigned Reg = 1; Reg < TRI->getNumRegs(); ++Reg) {
468 if (KillIndices[Reg] == ~0u)
469 LLVM_DEBUG(dbgs() << " " << printReg(Reg, TRI));
470 }
471 LLVM_DEBUG(dbgs() << '\n');
472 }
473#endif
474
475 // Track progress along the critical path through the SUnit graph as we walk
476 // the instructions.
477 const SUnit *CriticalPathSU = Max;
478 MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
479
480 // Consider this pattern:
481 // A = ...
482 // ... = A
483 // A = ...
484 // ... = A
485 // A = ...
486 // ... = A
487 // A = ...
488 // ... = A
489 // There are three anti-dependencies here, and without special care,
490 // we'd break all of them using the same register:
491 // A = ...
492 // ... = A
493 // B = ...
494 // ... = B
495 // B = ...
496 // ... = B
497 // B = ...
498 // ... = B
499 // because at each anti-dependence, B is the first register that
500 // isn't A which is free. This re-introduces anti-dependencies
501 // at all but one of the original anti-dependencies that we were
502 // trying to break. To avoid this, keep track of the most recent
503 // register that each register was replaced with, avoid
504 // using it to repair an anti-dependence on the same register.
505 // This lets us produce this:
506 // A = ...
507 // ... = A
508 // B = ...
509 // ... = B
510 // C = ...
511 // ... = C
512 // B = ...
513 // ... = B
514 // This still has an anti-dependence on B, but at least it isn't on the
515 // original critical path.
516 //
517 // TODO: If we tracked more than one register here, we could potentially
518 // fix that remaining critical edge too. This is a little more involved,
519 // because unlike the most recent register, less recent registers should
520 // still be considered, though only if no other registers are available.
521 std::vector<MCRegister> LastNewReg(TRI->getNumRegs(), MCRegister());
522
523 // Attempt to break anti-dependence edges on the critical path. Walk the
524 // instructions from the bottom up, tracking information about liveness
525 // as we go to help determine which registers are available.
526 unsigned Broken = 0;
527 unsigned Count = InsertPosIndex - 1;
528 for (MachineBasicBlock::iterator I = End, E = Begin; I != E; --Count) {
529 MachineInstr &MI = *--I;
530 // Kill instructions can define registers but are really nops, and there
531 // might be a real definition earlier that needs to be paired with uses
532 // dominated by this kill.
533
534 // FIXME: It may be possible to remove the isKill() restriction once PR18663
535 // has been properly fixed. There can be value in processing kills as seen
536 // in the AggressiveAntiDepBreaker class.
537 if (MI.isDebugInstr() || MI.isKill())
538 continue;
539
540 // Check if this instruction has a dependence on the critical path that
541 // is an anti-dependence that we may be able to break. If it is, set
542 // AntiDepReg to the non-zero register associated with the anti-dependence.
543 //
544 // We limit our attention to the critical path as a heuristic to avoid
545 // breaking anti-dependence edges that aren't going to significantly
546 // impact the overall schedule. There are a limited number of registers
547 // and we want to save them for the important edges.
548 //
549 // TODO: Instructions with multiple defs could have multiple
550 // anti-dependencies. The current code here only knows how to break one
551 // edge per instruction. Note that we'd have to be able to break all of
552 // the anti-dependencies in an instruction in order to be effective.
553 MCRegister AntiDepReg;
554 if (&MI == CriticalPathMI) {
555 if (const SDep *Edge = CriticalPathStep(SU: CriticalPathSU)) {
556 const SUnit *NextSU = Edge->getSUnit();
557
558 // Only consider anti-dependence edges.
559 if (Edge->getKind() == SDep::Anti) {
560 AntiDepReg = Edge->getReg().asMCReg();
561 assert(AntiDepReg && "Anti-dependence on reg0?");
562 if (!MRI.isAllocatable(PhysReg: AntiDepReg))
563 // Don't break anti-dependencies on non-allocatable registers.
564 AntiDepReg = MCRegister();
565 else if (KeepRegs.test(Idx: AntiDepReg.id()))
566 // Don't break anti-dependencies if a use down below requires
567 // this exact register.
568 AntiDepReg = MCRegister();
569 else {
570 // If the SUnit has other dependencies on the SUnit that it
571 // anti-depends on, don't bother breaking the anti-dependency
572 // since those edges would prevent such units from being
573 // scheduled past each other regardless.
574 //
575 // Also, if there are dependencies on other SUnits with the
576 // same register as the anti-dependency, don't attempt to
577 // break it.
578 for (const SDep &P : CriticalPathSU->Preds)
579 if (P.getSUnit() == NextSU
580 ? (P.getKind() != SDep::Anti || P.getReg() != AntiDepReg)
581 : (P.getKind() == SDep::Data &&
582 P.getReg() == AntiDepReg)) {
583 AntiDepReg = MCRegister();
584 break;
585 }
586 }
587 }
588 CriticalPathSU = NextSU;
589 CriticalPathMI = CriticalPathSU->getInstr();
590 } else {
591 // We've reached the end of the critical path.
592 CriticalPathSU = nullptr;
593 CriticalPathMI = nullptr;
594 }
595 }
596
597 PrescanInstruction(MI);
598
599 SmallVector<Register, 2> ForbidRegs;
600
601 // If MI's defs have a special allocation requirement, don't allow
602 // any def registers to be changed. Also assume all registers
603 // defined in a call must not be changed (ABI).
604 if (MI.isCall() || MI.hasExtraDefRegAllocReq() || TII->isPredicated(MI))
605 // If this instruction's defs have special allocation requirement, don't
606 // break this anti-dependency.
607 AntiDepReg = MCRegister();
608 else if (AntiDepReg) {
609 // If this instruction has a use of AntiDepReg, breaking it
610 // is invalid. If the instruction defines other registers,
611 // save a list of them so that we don't pick a new register
612 // that overlaps any of them.
613 for (const MachineOperand &MO : MI.operands()) {
614 if (!MO.isReg()) continue;
615 Register Reg = MO.getReg();
616 if (!Reg)
617 continue;
618 if (MO.isUse() && TRI->regsOverlap(RegA: AntiDepReg, RegB: Reg)) {
619 AntiDepReg = MCRegister();
620 break;
621 }
622 if (MO.isDef() && Reg != AntiDepReg)
623 ForbidRegs.push_back(Elt: Reg);
624 }
625 }
626
627 // Determine AntiDepReg's register class, if it is live and is
628 // consistently used within a single class.
629 const TargetRegisterClass *RC =
630 AntiDepReg ? Classes[AntiDepReg.id()] : nullptr;
631 assert((!AntiDepReg || RC != nullptr) &&
632 "Register should be live if it's causing an anti-dependence!");
633 if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
634 AntiDepReg = MCRegister();
635
636 // Look for a suitable register to use to break the anti-dependence.
637 //
638 // TODO: Instead of picking the first free register, consider which might
639 // be the best.
640 if (AntiDepReg) {
641 std::pair<std::multimap<MCRegister, MachineOperand *>::iterator,
642 std::multimap<MCRegister, MachineOperand *>::iterator>
643 Range = RegRefs.equal_range(x: AntiDepReg);
644 if (MCRegister NewReg = findSuitableFreeRegister(
645 RegRefBegin: Range.first, RegRefEnd: Range.second, AntiDepReg,
646 LastNewReg: LastNewReg[AntiDepReg.id()], RC, Forbid: ForbidRegs)) {
647 LLVM_DEBUG(dbgs() << "Breaking anti-dependence edge on "
648 << printReg(AntiDepReg, TRI) << " with "
649 << RegRefs.count(AntiDepReg) << " references"
650 << " using " << printReg(NewReg, TRI) << "!\n");
651
652 // Update the references to the old register to refer to the new
653 // register.
654 for (std::multimap<MCRegister, MachineOperand *>::iterator
655 Q = Range.first,
656 QE = Range.second;
657 Q != QE; ++Q) {
658 Q->second->setReg(NewReg);
659 // If the SU for the instruction being updated has debug information
660 // related to the anti-dependency register, make sure to update that
661 // as well.
662 const SUnit *SU = MISUnitMap[Q->second->getParent()];
663 if (!SU) continue;
664 UpdateDbgValues(DbgValues, ParentMI: Q->second->getParent(),
665 OldReg: AntiDepReg, NewReg);
666 }
667
668 // We just went back in time and modified history; the
669 // liveness information for the anti-dependence reg is now
670 // inconsistent. Set the state as if it were dead.
671 Classes[NewReg.id()] = Classes[AntiDepReg.id()];
672 DefIndices[NewReg.id()] = DefIndices[AntiDepReg.id()];
673 KillIndices[NewReg.id()] = KillIndices[AntiDepReg.id()];
674 assert(((KillIndices[NewReg.id()] == ~0u) !=
675 (DefIndices[NewReg.id()] == ~0u)) &&
676 "Kill and Def maps aren't consistent for NewReg!");
677
678 Classes[AntiDepReg.id()] = nullptr;
679 DefIndices[AntiDepReg.id()] = KillIndices[AntiDepReg.id()];
680 KillIndices[AntiDepReg.id()] = ~0u;
681 assert(((KillIndices[AntiDepReg.id()] == ~0u) !=
682 (DefIndices[AntiDepReg.id()] == ~0u)) &&
683 "Kill and Def maps aren't consistent for AntiDepReg!");
684
685 RegRefs.erase(x: AntiDepReg);
686 LastNewReg[AntiDepReg.id()] = NewReg;
687 ++Broken;
688 }
689 }
690
691 ScanInstruction(MI, Count);
692 }
693
694 return Broken;
695}
696
697AntiDepBreaker *
698llvm::createCriticalAntiDepBreaker(MachineFunction &MFi,
699 const RegisterClassInfo &RCI) {
700 return new CriticalAntiDepBreaker(MFi, RCI);
701}
702