1 //===-- AArch64A57FPLoadBalancing.cpp - Balance FP ops statically on A57---===//
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 //===----------------------------------------------------------------------===//
9 // For best-case performance on Cortex-A57, we should try to use a balanced
10 // mix of odd and even D-registers when performing a critical sequence of
11 // independent, non-quadword FP/ASIMD floating-point multiply or
12 // multiply-accumulate operations.
14 // This pass attempts to detect situations where the register allocation may
15 // adversely affect this load balancing and to change the registers used so as
16 // to better utilize the CPU.
18 // Ideally we'd just take each multiply or multiply-accumulate in turn and
19 // allocate it alternating even or odd registers. However, multiply-accumulates
20 // are most efficiently performed in the same functional unit as their
21 // accumulation operand. Therefore this pass tries to find maximal sequences
22 // ("Chains") of multiply-accumulates linked via their accumulation operand,
23 // and assign them all the same "color" (oddness/evenness).
25 // This optimization affects S-register and D-register floating point
26 // multiplies and FMADD/FMAs, as well as vector (floating point only) muls and
27 // FMADD/FMA. Q register instructions (and 128-bit vector instructions) are
29 //===----------------------------------------------------------------------===//
32 #include "AArch64InstrInfo.h"
33 #include "AArch64Subtarget.h"
34 #include "llvm/ADT/BitVector.h"
35 #include "llvm/ADT/EquivalenceClasses.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineFunctionPass.h"
38 #include "llvm/CodeGen/MachineInstr.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineRegisterInfo.h"
41 #include "llvm/CodeGen/RegisterScavenging.h"
42 #include "llvm/CodeGen/RegisterClassInfo.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
49 #define DEBUG_TYPE "aarch64-a57-fp-load-balancing"
51 // Enforce the algorithm to use the scavenged register even when the original
52 // destination register is the correct color. Used for testing.
54 TransformAll("aarch64-a57-fp-load-balancing-force-all",
55 cl::desc("Always modify dest registers regardless of color"),
56 cl::init(false), cl::Hidden);
58 // Never use the balance information obtained from chains - return a specific
59 // color always. Used for testing.
60 static cl::opt<unsigned>
61 OverrideBalance("aarch64-a57-fp-load-balancing-override",
62 cl::desc("Ignore balance information, always return "
63 "(1: Even, 2: Odd)."),
64 cl::init(0), cl::Hidden);
66 //===----------------------------------------------------------------------===//
69 // Is the instruction a type of multiply on 64-bit (or 32-bit) FPRs?
70 static bool isMul(MachineInstr *MI) {
71 switch (MI->getOpcode()) {
72 case AArch64::FMULSrr:
73 case AArch64::FNMULSrr:
74 case AArch64::FMULDrr:
75 case AArch64::FNMULDrr:
77 case AArch64::FMULv2f32:
84 // Is the instruction a type of FP multiply-accumulate on 64-bit (or 32-bit) FPRs?
85 static bool isMla(MachineInstr *MI) {
86 switch (MI->getOpcode()) {
87 case AArch64::FMSUBSrrr:
88 case AArch64::FMADDSrrr:
89 case AArch64::FNMSUBSrrr:
90 case AArch64::FNMADDSrrr:
91 case AArch64::FMSUBDrrr:
92 case AArch64::FMADDDrrr:
93 case AArch64::FNMSUBDrrr:
94 case AArch64::FNMADDDrrr:
96 case AArch64::FMLAv2f32:
97 case AArch64::FMLSv2f32:
104 //===----------------------------------------------------------------------===//
107 /// A "color", which is either even or odd. Yes, these aren't really colors
108 /// but the algorithm is conceptually doing two-color graph coloring.
109 enum class Color { Even, Odd };
110 static const char *ColorNames[2] = { "Even", "Odd" };
114 class AArch64A57FPLoadBalancing : public MachineFunctionPass {
115 const AArch64InstrInfo *TII;
116 MachineRegisterInfo *MRI;
117 const TargetRegisterInfo *TRI;
118 RegisterClassInfo RCI;
122 explicit AArch64A57FPLoadBalancing() : MachineFunctionPass(ID) {}
124 bool runOnMachineFunction(MachineFunction &F) override;
126 const char *getPassName() const override {
127 return "A57 FP Anti-dependency breaker";
130 void getAnalysisUsage(AnalysisUsage &AU) const override {
131 AU.setPreservesCFG();
132 MachineFunctionPass::getAnalysisUsage(AU);
136 bool runOnBasicBlock(MachineBasicBlock &MBB);
137 bool colorChainSet(std::vector<Chain*> GV, MachineBasicBlock &MBB,
139 bool colorChain(Chain *G, Color C, MachineBasicBlock &MBB);
140 int scavengeRegister(Chain *G, Color C, MachineBasicBlock &MBB);
141 void scanInstruction(MachineInstr *MI, unsigned Idx,
142 std::map<unsigned, Chain*> &Chains,
143 std::set<Chain*> &ChainSet);
144 void maybeKillChain(MachineOperand &MO, unsigned Idx,
145 std::map<unsigned, Chain*> &RegChains);
146 Color getColor(unsigned Register);
147 Chain *getAndEraseNext(Color PreferredColor, std::vector<Chain*> &L);
149 char AArch64A57FPLoadBalancing::ID = 0;
151 /// A Chain is a sequence of instructions that are linked together by
152 /// an accumulation operand. For example:
155 /// fmla d1<def>, ?, ?, d0<kill>
156 /// fmla d2<def>, ?, ?, d1<kill>
158 /// There may be other instructions interleaved in the sequence that
159 /// do not belong to the chain. These other instructions must not use
160 /// the "chain" register at any point.
162 /// We currently only support chains where the "chain" operand is killed
163 /// at each link in the chain for simplicity.
164 /// A chain has three important instructions - Start, Last and Kill.
165 /// * The start instruction is the first instruction in the chain.
166 /// * Last is the final instruction in the chain.
167 /// * Kill may or may not be defined. If defined, Kill is the instruction
168 /// where the outgoing value of the Last instruction is killed.
169 /// This information is important as if we know the outgoing value is
170 /// killed with no intervening uses, we can safely change its register.
172 /// Without a kill instruction, we must assume the outgoing value escapes
173 /// beyond our model and either must not change its register or must
174 /// create a fixup FMOV to keep the old register value consistent.
178 /// The important (marker) instructions.
179 MachineInstr *StartInst, *LastInst, *KillInst;
180 /// The index, from the start of the basic block, that each marker
181 /// appears. These are stored so we can do quick interval tests.
182 unsigned StartInstIdx, LastInstIdx, KillInstIdx;
183 /// All instructions in the chain.
184 std::set<MachineInstr*> Insts;
185 /// True if KillInst cannot be modified. If this is true,
186 /// we cannot change LastInst's outgoing register.
187 /// This will be true for tied values and regmasks.
188 bool KillIsImmutable;
189 /// The "color" of LastInst. This will be the preferred chain color,
190 /// as changing intermediate nodes is easy but changing the last
191 /// instruction can be more tricky.
194 Chain(MachineInstr *MI, unsigned Idx, Color C) :
195 StartInst(MI), LastInst(MI), KillInst(NULL),
196 StartInstIdx(Idx), LastInstIdx(Idx), KillInstIdx(0),
201 /// Add a new instruction into the chain. The instruction's dest operand
202 /// has the given color.
203 void add(MachineInstr *MI, unsigned Idx, Color C) {
211 /// Return true if MI is a member of the chain.
212 bool contains(MachineInstr *MI) { return Insts.count(MI) > 0; }
214 /// Return the number of instructions in the chain.
215 unsigned size() const {
219 /// Inform the chain that its last active register (the dest register of
220 /// LastInst) is killed by MI with no intervening uses or defs.
221 void setKill(MachineInstr *MI, unsigned Idx, bool Immutable) {
224 KillIsImmutable = Immutable;
227 /// Return the first instruction in the chain.
228 MachineInstr *getStart() const { return StartInst; }
229 /// Return the last instruction in the chain.
230 MachineInstr *getLast() const { return LastInst; }
231 /// Return the "kill" instruction (as set with setKill()) or NULL.
232 MachineInstr *getKill() const { return KillInst; }
233 /// Return an instruction that can be used as an iterator for the end
234 /// of the chain. This is the maximum of KillInst (if set) and LastInst.
235 MachineInstr *getEnd() const {
236 return ++MachineBasicBlock::iterator(KillInst ? KillInst : LastInst);
239 /// Can the Kill instruction (assuming one exists) be modified?
240 bool isKillImmutable() const { return KillIsImmutable; }
242 /// Return the preferred color of this chain.
243 Color getPreferredColor() {
244 if (OverrideBalance != 0)
245 return OverrideBalance == 1 ? Color::Even : Color::Odd;
249 /// Return true if this chain (StartInst..KillInst) overlaps with Other.
250 bool rangeOverlapsWith(Chain *Other) {
251 unsigned End = KillInst ? KillInstIdx : LastInstIdx;
252 unsigned OtherEnd = Other->KillInst ?
253 Other->KillInstIdx : Other->LastInstIdx;
255 return StartInstIdx <= OtherEnd && Other->StartInstIdx <= End;
258 /// Return true if this chain starts before Other.
259 bool startsBefore(Chain *Other) {
260 return StartInstIdx < Other->StartInstIdx;
263 /// Return true if the group will require a fixup MOV at the end.
264 bool requiresFixup() const {
265 return (getKill() && isKillImmutable()) || !getKill();
268 /// Return a simple string representation of the chain.
269 std::string str() const {
271 raw_string_ostream OS(S);
274 StartInst->print(OS, NULL, true);
276 LastInst->print(OS, NULL, true);
279 KillInst->print(OS, NULL, true);
289 } // end anonymous namespace
291 //===----------------------------------------------------------------------===//
293 bool AArch64A57FPLoadBalancing::runOnMachineFunction(MachineFunction &F) {
294 bool Changed = false;
295 DEBUG(dbgs() << "***** AArch64A57FPLoadBalancing *****\n");
297 const TargetMachine &TM = F.getTarget();
298 MRI = &F.getRegInfo();
299 TRI = F.getRegInfo().getTargetRegisterInfo();
300 TII = TM.getSubtarget<AArch64Subtarget>().getInstrInfo();
301 RCI.runOnMachineFunction(F);
303 for (auto &MBB : F) {
304 Changed |= runOnBasicBlock(MBB);
310 bool AArch64A57FPLoadBalancing::runOnBasicBlock(MachineBasicBlock &MBB) {
311 bool Changed = false;
312 DEBUG(dbgs() << "Running on MBB: " << MBB << " - scanning instructions...\n");
314 // First, scan the basic block producing a set of chains.
316 // The currently "active" chains - chains that can be added to and haven't
317 // been killed yet. This is keyed by register - all chains can only have one
318 // "link" register between each inst in the chain.
319 std::map<unsigned, Chain*> ActiveChains;
320 std::set<Chain*> AllChains;
323 scanInstruction(&MI, Idx++, ActiveChains, AllChains);
325 DEBUG(dbgs() << "Scan complete, "<< AllChains.size() << " chains created.\n");
327 // Group the chains into disjoint sets based on their liveness range. This is
328 // a poor-man's version of graph coloring. Ideally we'd create an interference
329 // graph and perform full-on graph coloring on that, but;
330 // (a) That's rather heavyweight for only two colors.
331 // (b) We expect multiple disjoint interference regions - in practice the live
332 // range of chains is quite small and they are clustered between loads
334 EquivalenceClasses<Chain*> EC;
335 for (auto *I : AllChains)
338 for (auto *I : AllChains) {
339 for (auto *J : AllChains) {
340 if (I != J && I->rangeOverlapsWith(J))
344 DEBUG(dbgs() << "Created " << EC.getNumClasses() << " disjoint sets.\n");
346 // Now we assume that every member of an equivalence class interferes
347 // with every other member of that class, and with no members of other classes.
349 // Convert the EquivalenceClasses to a simpler set of sets.
350 std::vector<std::vector<Chain*> > V;
351 for (auto I = EC.begin(), E = EC.end(); I != E; ++I) {
352 std::vector<Chain*> Cs(EC.member_begin(I), EC.member_end());
353 if (Cs.empty()) continue;
357 // Now we have a set of sets, order them by start address so
358 // we can iterate over them sequentially.
359 std::sort(V.begin(), V.end(),
360 [](const std::vector<Chain*> &A,
361 const std::vector<Chain*> &B) {
362 return A.front()->startsBefore(B.front());
365 // As we only have two colors, we can track the global (BB-level) balance of
366 // odds versus evens. We aim to keep this near zero to keep both execution
368 // Positive means we're even-heavy, negative we're odd-heavy.
370 // FIXME: If chains have interdependencies, for example:
373 // We do not model this and may color each one differently, assuming we'll
374 // get ILP when we obviously can't. This hasn't been seen to be a problem
375 // in practice so far, so we simplify the algorithm by ignoring it.
379 Changed |= colorChainSet(I, MBB, Parity);
381 for (auto *C : AllChains)
387 Chain *AArch64A57FPLoadBalancing::getAndEraseNext(Color PreferredColor,
388 std::vector<Chain*> &L) {
392 // We try and get the best candidate from L to color next, given that our
393 // preferred color is "PreferredColor". L is ordered from larger to smaller
394 // chains. It is beneficial to color the large chains before the small chains,
395 // but if we can't find a chain of the maximum length with the preferred color,
396 // we fuzz the size and look for slightly smaller chains before giving up and
397 // returning a chain that must be recolored.
399 // FIXME: Does this need to be configurable?
400 const unsigned SizeFuzz = 1;
401 unsigned MinSize = L.front()->size() - SizeFuzz;
402 for (auto I = L.begin(), E = L.end(); I != E; ++I) {
403 if ((*I)->size() <= MinSize) {
404 // We've gone past the size limit. Return the previous item.
410 if ((*I)->getPreferredColor() == PreferredColor) {
417 // Bailout case - just return the first item.
418 Chain *Ch = L.front();
423 bool AArch64A57FPLoadBalancing::colorChainSet(std::vector<Chain*> GV,
424 MachineBasicBlock &MBB,
426 bool Changed = false;
427 DEBUG(dbgs() << "colorChainSet(): #sets=" << GV.size() << "\n");
429 // Sort by descending size order so that we allocate the most important
431 // Tie-break equivalent sizes by sorting chains requiring fixups before
432 // those without fixups. The logic here is that we should look at the
433 // chains that we cannot change before we look at those we can,
434 // so the parity counter is updated and we know what color we should
436 std::sort(GV.begin(), GV.end(), [](const Chain *G1, const Chain *G2) {
437 if (G1->size() != G2->size())
438 return G1->size() > G2->size();
439 return G1->requiresFixup() > G2->requiresFixup();
442 Color PreferredColor = Parity < 0 ? Color::Even : Color::Odd;
443 while (Chain *G = getAndEraseNext(PreferredColor, GV)) {
444 // Start off by assuming we'll color to our own preferred color.
445 Color C = PreferredColor;
447 // But if we really don't care, use the chain's preferred color.
448 C = G->getPreferredColor();
450 DEBUG(dbgs() << " - Parity=" << Parity << ", Color="
451 << ColorNames[(int)C] << "\n");
453 // If we'll need a fixup FMOV, don't bother. Testing has shown that this
454 // happens infrequently and when it does it has at least a 50% chance of
455 // slowing code down instead of speeding it up.
456 if (G->requiresFixup() && C != G->getPreferredColor()) {
457 C = G->getPreferredColor();
458 DEBUG(dbgs() << " - " << G->str() << " - not worthwhile changing; "
459 "color remains " << ColorNames[(int)C] << "\n");
462 Changed |= colorChain(G, C, MBB);
464 Parity += (C == Color::Even) ? G->size() : -G->size();
465 PreferredColor = Parity < 0 ? Color::Even : Color::Odd;
471 int AArch64A57FPLoadBalancing::scavengeRegister(Chain *G, Color C,
472 MachineBasicBlock &MBB) {
474 RS.enterBasicBlock(&MBB);
475 RS.forward(MachineBasicBlock::iterator(G->getStart()));
477 // Can we find an appropriate register that is available throughout the life
479 unsigned RegClassID = G->getStart()->getDesc().OpInfo[0].RegClass;
480 BitVector AvailableRegs = RS.getRegsAvailable(TRI->getRegClass(RegClassID));
481 for (MachineBasicBlock::iterator I = G->getStart(), E = G->getEnd();
484 AvailableRegs &= RS.getRegsAvailable(TRI->getRegClass(RegClassID));
486 // Remove any registers clobbered by a regmask.
487 for (auto J : I->operands()) {
489 AvailableRegs.clearBitsNotInMask(J.getRegMask());
493 // Make sure we allocate in-order, to get the cheapest registers first.
494 auto Ord = RCI.getOrder(TRI->getRegClass(RegClassID));
495 for (auto Reg : Ord) {
496 if (!AvailableRegs[Reg])
498 if ((C == Color::Even && (Reg % 2) == 0) ||
499 (C == Color::Odd && (Reg % 2) == 1))
506 bool AArch64A57FPLoadBalancing::colorChain(Chain *G, Color C,
507 MachineBasicBlock &MBB) {
508 bool Changed = false;
509 DEBUG(dbgs() << " - colorChain(" << G->str() << ", "
510 << ColorNames[(int)C] << ")\n");
512 // Try and obtain a free register of the right class. Without a register
513 // to play with we cannot continue.
514 int Reg = scavengeRegister(G, C, MBB);
516 DEBUG(dbgs() << "Scavenging (thus coloring) failed!\n");
519 DEBUG(dbgs() << " - Scavenged register: " << TRI->getName(Reg) << "\n");
521 std::map<unsigned, unsigned> Substs;
522 for (MachineBasicBlock::iterator I = G->getStart(), E = G->getEnd();
524 if (!G->contains(I) &&
525 (&*I != G->getKill() || G->isKillImmutable()))
528 // I is a member of G, or I is a mutable instruction that kills G.
530 std::vector<unsigned> ToErase;
531 for (auto &U : I->operands()) {
532 if (U.isReg() && U.isUse() && Substs.find(U.getReg()) != Substs.end()) {
533 unsigned OrigReg = U.getReg();
534 U.setReg(Substs[OrigReg]);
536 // Don't erase straight away, because there may be other operands
537 // that also reference this substitution!
538 ToErase.push_back(OrigReg);
539 } else if (U.isRegMask()) {
540 for (auto J : Substs) {
541 if (U.clobbersPhysReg(J.first))
542 ToErase.push_back(J.first);
546 // Now it's safe to remove the substs identified earlier.
547 for (auto J : ToErase)
550 // Only change the def if this isn't the last instruction.
551 if (&*I != G->getKill()) {
552 MachineOperand &MO = I->getOperand(0);
554 bool Change = TransformAll || getColor(MO.getReg()) != C;
555 if (G->requiresFixup() && &*I == G->getLast())
559 Substs[MO.getReg()] = Reg;
561 MRI->setPhysRegUsed(Reg);
567 assert(Substs.size() == 0 && "No substitutions should be left active!");
570 DEBUG(dbgs() << " - Kill instruction seen.\n");
572 // We didn't have a kill instruction, but we didn't seem to need to change
573 // the destination register anyway.
574 DEBUG(dbgs() << " - Destination register not changed.\n");
579 void AArch64A57FPLoadBalancing::
580 scanInstruction(MachineInstr *MI, unsigned Idx,
581 std::map<unsigned, Chain*> &ActiveChains,
582 std::set<Chain*> &AllChains) {
583 // Inspect "MI", updating ActiveChains and AllChains.
587 for (auto &I : MI->operands())
588 maybeKillChain(I, Idx, ActiveChains);
590 // Create a new chain. Multiplies don't require forwarding so can go on any
592 unsigned DestReg = MI->getOperand(0).getReg();
594 DEBUG(dbgs() << "New chain started for register "
595 << TRI->getName(DestReg) << " at " << *MI);
597 Chain *G = new Chain(MI, Idx, getColor(DestReg));
598 ActiveChains[DestReg] = G;
601 } else if (isMla(MI)) {
603 // It is beneficial to keep MLAs on the same functional unit as their
604 // accumulator operand.
605 unsigned DestReg = MI->getOperand(0).getReg();
606 unsigned AccumReg = MI->getOperand(3).getReg();
608 maybeKillChain(MI->getOperand(1), Idx, ActiveChains);
609 maybeKillChain(MI->getOperand(2), Idx, ActiveChains);
610 if (DestReg != AccumReg)
611 maybeKillChain(MI->getOperand(0), Idx, ActiveChains);
613 if (ActiveChains.find(AccumReg) != ActiveChains.end()) {
614 DEBUG(dbgs() << "Chain found for accumulator register "
615 << TRI->getName(AccumReg) << " in MI " << *MI);
617 // For simplicity we only chain together sequences of MULs/MLAs where the
618 // accumulator register is killed on each instruction. This means we don't
619 // need to track other uses of the registers we want to rewrite.
621 // FIXME: We could extend to handle the non-kill cases for more coverage.
622 if (MI->getOperand(3).isKill()) {
624 DEBUG(dbgs() << "Instruction was successfully added to chain.\n");
625 ActiveChains[AccumReg]->add(MI, Idx, getColor(DestReg));
626 // Handle cases where the destination is not the same as the accumulator.
627 ActiveChains[DestReg] = ActiveChains[AccumReg];
631 DEBUG(dbgs() << "Cannot add to chain because accumulator operand wasn't "
632 << "marked <kill>!\n");
633 maybeKillChain(MI->getOperand(3), Idx, ActiveChains);
636 DEBUG(dbgs() << "Creating new chain for dest register "
637 << TRI->getName(DestReg) << "\n");
638 Chain *G = new Chain(MI, Idx, getColor(DestReg));
639 ActiveChains[DestReg] = G;
644 // Non-MUL or MLA instruction. Invalidate any chain in the uses or defs
646 for (auto &I : MI->operands())
647 maybeKillChain(I, Idx, ActiveChains);
652 void AArch64A57FPLoadBalancing::
653 maybeKillChain(MachineOperand &MO, unsigned Idx,
654 std::map<unsigned, Chain*> &ActiveChains) {
655 // Given an operand and the set of active chains (keyed by register),
656 // determine if a chain should be ended and remove from ActiveChains.
657 MachineInstr *MI = MO.getParent();
661 // If this is a KILL of a current chain, record it.
662 if (MO.isKill() && ActiveChains.find(MO.getReg()) != ActiveChains.end()) {
663 DEBUG(dbgs() << "Kill seen for chain " << TRI->getName(MO.getReg())
665 ActiveChains[MO.getReg()]->setKill(MI, Idx, /*Immutable=*/MO.isTied());
667 ActiveChains.erase(MO.getReg());
669 } else if (MO.isRegMask()) {
671 for (auto I = ActiveChains.begin(), E = ActiveChains.end();
673 if (MO.clobbersPhysReg(I->first)) {
674 DEBUG(dbgs() << "Kill (regmask) seen for chain "
675 << TRI->getName(I->first) << "\n");
676 I->second->setKill(MI, Idx, /*Immutable=*/true);
677 ActiveChains.erase(I);
684 Color AArch64A57FPLoadBalancing::getColor(unsigned Reg) {
685 if ((TRI->getEncodingValue(Reg) % 2) == 0)
691 // Factory function used by AArch64TargetMachine to add the pass to the passmanager.
692 FunctionPass *llvm::createAArch64A57FPLoadBalancing() {
693 return new AArch64A57FPLoadBalancing();