1 //=== A15SDOptimizerPass.cpp - Optimize DPR and SPR register accesses on A15==//
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 //===----------------------------------------------------------------------===//
10 // The Cortex-A15 processor employs a tracking scheme in its register renaming
11 // in order to process each instruction's micro-ops speculatively and
12 // out-of-order with appropriate forwarding. The ARM architecture allows VFP
13 // instructions to read and write 32-bit S-registers. Each S-register
14 // corresponds to one half (upper or lower) of an overlaid 64-bit D-register.
16 // There are several instruction patterns which can be used to provide this
17 // capability which can provide higher performance than other, potentially more
18 // direct patterns, specifically around when one micro-op reads a D-register
19 // operand that has recently been written as one or more S-register results.
21 // This file defines a pre-regalloc pass which looks for SPR producers which
22 // are going to be used by a DPR (or QPR) consumers and creates the more
23 // optimized access pattern.
25 //===----------------------------------------------------------------------===//
27 #define DEBUG_TYPE "a15-sd-optimizer"
29 #include "ARMBaseInstrInfo.h"
30 #include "ARMBaseRegisterInfo.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/CodeGen/MachineFunctionPass.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
43 struct A15SDOptimizer : public MachineFunctionPass {
45 A15SDOptimizer() : MachineFunctionPass(ID) {}
47 bool runOnMachineFunction(MachineFunction &Fn) override;
49 const char *getPassName() const override {
50 return "ARM A15 S->D optimizer";
54 const ARMBaseInstrInfo *TII;
55 const TargetRegisterInfo *TRI;
56 MachineRegisterInfo *MRI;
58 bool runOnInstruction(MachineInstr *MI);
61 // Instruction builder helpers
63 unsigned createDupLane(MachineBasicBlock &MBB,
64 MachineBasicBlock::iterator InsertBefore,
66 unsigned Reg, unsigned Lane,
69 unsigned createExtractSubreg(MachineBasicBlock &MBB,
70 MachineBasicBlock::iterator InsertBefore,
72 unsigned DReg, unsigned Lane,
73 const TargetRegisterClass *TRC);
75 unsigned createVExt(MachineBasicBlock &MBB,
76 MachineBasicBlock::iterator InsertBefore,
78 unsigned Ssub0, unsigned Ssub1);
80 unsigned createRegSequence(MachineBasicBlock &MBB,
81 MachineBasicBlock::iterator InsertBefore,
83 unsigned Reg1, unsigned Reg2);
85 unsigned createInsertSubreg(MachineBasicBlock &MBB,
86 MachineBasicBlock::iterator InsertBefore,
87 DebugLoc DL, unsigned DReg, unsigned Lane,
90 unsigned createImplicitDef(MachineBasicBlock &MBB,
91 MachineBasicBlock::iterator InsertBefore,
95 // Various property checkers
97 bool usesRegClass(MachineOperand &MO, const TargetRegisterClass *TRC);
98 bool hasPartialWrite(MachineInstr *MI);
99 SmallVector<unsigned, 8> getReadDPRs(MachineInstr *MI);
100 unsigned getDPRLaneFromSPR(unsigned SReg);
103 // Methods used for getting the definitions of partial registers
106 MachineInstr *elideCopies(MachineInstr *MI);
107 void elideCopiesAndPHIs(MachineInstr *MI,
108 SmallVectorImpl<MachineInstr*> &Outs);
111 // Pattern optimization methods
113 unsigned optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg);
114 unsigned optimizeSDPattern(MachineInstr *MI);
115 unsigned getPrefSPRLane(unsigned SReg);
118 // Sanitizing method - used to make sure if don't leave dead code around.
120 void eraseInstrWithNoUses(MachineInstr *MI);
123 // A map used to track the changes done by this pass.
125 std::map<MachineInstr*, unsigned> Replacements;
126 std::set<MachineInstr *> DeadInstr;
128 char A15SDOptimizer::ID = 0;
129 } // end anonymous namespace
131 // Returns true if this is a use of a SPR register.
132 bool A15SDOptimizer::usesRegClass(MachineOperand &MO,
133 const TargetRegisterClass *TRC) {
136 unsigned Reg = MO.getReg();
138 if (TargetRegisterInfo::isVirtualRegister(Reg))
139 return MRI->getRegClass(Reg)->hasSuperClassEq(TRC);
141 return TRC->contains(Reg);
144 unsigned A15SDOptimizer::getDPRLaneFromSPR(unsigned SReg) {
145 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1,
147 if (DReg != ARM::NoRegister) return ARM::ssub_1;
151 // Get the subreg type that is most likely to be coalesced
152 // for an SPR register that will be used in VDUP32d pseudo.
153 unsigned A15SDOptimizer::getPrefSPRLane(unsigned SReg) {
154 if (!TRI->isVirtualRegister(SReg))
155 return getDPRLaneFromSPR(SReg);
157 MachineInstr *MI = MRI->getVRegDef(SReg);
158 if (!MI) return ARM::ssub_0;
159 MachineOperand *MO = MI->findRegisterDefOperand(SReg);
161 assert(MO->isReg() && "Non-register operand found!");
162 if (!MO) return ARM::ssub_0;
164 if (MI->isCopy() && usesRegClass(MI->getOperand(1),
165 &ARM::SPRRegClass)) {
166 SReg = MI->getOperand(1).getReg();
169 if (TargetRegisterInfo::isVirtualRegister(SReg)) {
170 if (MO->getSubReg() == ARM::ssub_1) return ARM::ssub_1;
173 return getDPRLaneFromSPR(SReg);
176 // MI is known to be dead. Figure out what instructions
177 // are also made dead by this and mark them for removal.
178 void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) {
179 SmallVector<MachineInstr *, 8> Front;
180 DeadInstr.insert(MI);
182 DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n");
185 while (Front.size() != 0) {
189 // MI is already known to be dead. We need to see
190 // if other instructions can also be removed.
191 for (unsigned int i = 0; i < MI->getNumOperands(); ++i) {
192 MachineOperand &MO = MI->getOperand(i);
193 if ((!MO.isReg()) || (!MO.isUse()))
195 unsigned Reg = MO.getReg();
196 if (!TRI->isVirtualRegister(Reg))
198 MachineOperand *Op = MI->findRegisterDefOperand(Reg);
203 MachineInstr *Def = Op->getParent();
205 // We don't need to do anything if we have already marked
206 // this instruction as being dead.
207 if (DeadInstr.find(Def) != DeadInstr.end())
210 // Check if all the uses of this instruction are marked as
211 // dead. If so, we can also mark this instruction as being
214 for (unsigned int j = 0; j < Def->getNumOperands(); ++j) {
215 MachineOperand &MODef = Def->getOperand(j);
216 if ((!MODef.isReg()) || (!MODef.isDef()))
218 unsigned DefReg = MODef.getReg();
219 if (!TRI->isVirtualRegister(DefReg)) {
223 for (MachineRegisterInfo::use_instr_iterator
224 II = MRI->use_instr_begin(Reg), EE = MRI->use_instr_end();
226 // We don't care about self references.
229 if (DeadInstr.find(&*II) == DeadInstr.end()) {
236 if (!IsDead) continue;
238 DEBUG(dbgs() << "Deleting instruction " << *Def << "\n");
239 DeadInstr.insert(Def);
244 // Creates the more optimized patterns and generally does all the code
245 // transformations in this pass.
246 unsigned A15SDOptimizer::optimizeSDPattern(MachineInstr *MI) {
248 return optimizeAllLanesPattern(MI, MI->getOperand(1).getReg());
251 if (MI->isInsertSubreg()) {
252 unsigned DPRReg = MI->getOperand(1).getReg();
253 unsigned SPRReg = MI->getOperand(2).getReg();
255 if (TRI->isVirtualRegister(DPRReg) && TRI->isVirtualRegister(SPRReg)) {
256 MachineInstr *DPRMI = MRI->getVRegDef(MI->getOperand(1).getReg());
257 MachineInstr *SPRMI = MRI->getVRegDef(MI->getOperand(2).getReg());
259 if (DPRMI && SPRMI) {
260 // See if the first operand of this insert_subreg is IMPLICIT_DEF
261 MachineInstr *ECDef = elideCopies(DPRMI);
262 if (ECDef != 0 && ECDef->isImplicitDef()) {
263 // Another corner case - if we're inserting something that is purely
264 // a subreg copy of a DPR, just use that DPR.
266 MachineInstr *EC = elideCopies(SPRMI);
267 // Is it a subreg copy of ssub_0?
268 if (EC && EC->isCopy() &&
269 EC->getOperand(1).getSubReg() == ARM::ssub_0) {
270 DEBUG(dbgs() << "Found a subreg copy: " << *SPRMI);
272 // Find the thing we're subreg copying out of - is it of the same
273 // regclass as DPRMI? (i.e. a DPR or QPR).
274 unsigned FullReg = SPRMI->getOperand(1).getReg();
275 const TargetRegisterClass *TRC =
276 MRI->getRegClass(MI->getOperand(1).getReg());
277 if (TRC->hasSuperClassEq(MRI->getRegClass(FullReg))) {
278 DEBUG(dbgs() << "Subreg copy is compatible - returning ");
279 DEBUG(dbgs() << PrintReg(FullReg) << "\n");
280 eraseInstrWithNoUses(MI);
285 return optimizeAllLanesPattern(MI, MI->getOperand(2).getReg());
289 return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
292 if (MI->isRegSequence() && usesRegClass(MI->getOperand(1),
293 &ARM::SPRRegClass)) {
294 // See if all bar one of the operands are IMPLICIT_DEF and insert the
295 // optimizer pattern accordingly.
296 unsigned NumImplicit = 0, NumTotal = 0;
297 unsigned NonImplicitReg = ~0U;
299 for (unsigned I = 1; I < MI->getNumExplicitOperands(); ++I) {
300 if (!MI->getOperand(I).isReg())
303 unsigned OpReg = MI->getOperand(I).getReg();
305 if (!TRI->isVirtualRegister(OpReg))
308 MachineInstr *Def = MRI->getVRegDef(OpReg);
311 if (Def->isImplicitDef())
314 NonImplicitReg = MI->getOperand(I).getReg();
317 if (NumImplicit == NumTotal - 1)
318 return optimizeAllLanesPattern(MI, NonImplicitReg);
320 return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
323 assert(0 && "Unhandled update pattern!");
327 // Return true if this MachineInstr inserts a scalar (SPR) value into
328 // a D or Q register.
329 bool A15SDOptimizer::hasPartialWrite(MachineInstr *MI) {
330 // The only way we can do a partial register update is through a COPY,
331 // INSERT_SUBREG or REG_SEQUENCE.
332 if (MI->isCopy() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
335 if (MI->isInsertSubreg() && usesRegClass(MI->getOperand(2),
339 if (MI->isRegSequence() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
345 // Looks through full copies to get the instruction that defines the input
347 MachineInstr *A15SDOptimizer::elideCopies(MachineInstr *MI) {
348 if (!MI->isFullCopy())
350 if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
352 MachineInstr *Def = MRI->getVRegDef(MI->getOperand(1).getReg());
355 return elideCopies(Def);
358 // Look through full copies and PHIs to get the set of non-copy MachineInstrs
359 // that can produce MI.
360 void A15SDOptimizer::elideCopiesAndPHIs(MachineInstr *MI,
361 SmallVectorImpl<MachineInstr*> &Outs) {
362 // Looking through PHIs may create loops so we need to track what
363 // instructions we have visited before.
364 std::set<MachineInstr *> Reached;
365 SmallVector<MachineInstr *, 8> Front;
367 while (Front.size() != 0) {
371 // If we have already explored this MachineInstr, ignore it.
372 if (Reached.find(MI) != Reached.end())
376 for (unsigned I = 1, E = MI->getNumOperands(); I != E; I += 2) {
377 unsigned Reg = MI->getOperand(I).getReg();
378 if (!TRI->isVirtualRegister(Reg)) {
381 MachineInstr *NewMI = MRI->getVRegDef(Reg);
384 Front.push_back(NewMI);
386 } else if (MI->isFullCopy()) {
387 if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
389 MachineInstr *NewMI = MRI->getVRegDef(MI->getOperand(1).getReg());
392 Front.push_back(NewMI);
394 DEBUG(dbgs() << "Found partial copy" << *MI <<"\n");
400 // Return the DPR virtual registers that are read by this machine instruction
402 SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) {
403 if (MI->isCopyLike() || MI->isInsertSubreg() || MI->isRegSequence() ||
405 return SmallVector<unsigned, 8>();
407 SmallVector<unsigned, 8> Defs;
408 for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
409 MachineOperand &MO = MI->getOperand(i);
411 if (!MO.isReg() || !MO.isUse())
413 if (!usesRegClass(MO, &ARM::DPRRegClass) &&
414 !usesRegClass(MO, &ARM::QPRRegClass) &&
415 !usesRegClass(MO, &ARM::DPairRegClass)) // Treat DPair as QPR
418 Defs.push_back(MO.getReg());
423 // Creates a DPR register from an SPR one by using a VDUP.
425 A15SDOptimizer::createDupLane(MachineBasicBlock &MBB,
426 MachineBasicBlock::iterator InsertBefore,
428 unsigned Reg, unsigned Lane, bool QPR) {
429 unsigned Out = MRI->createVirtualRegister(QPR ? &ARM::QPRRegClass :
431 AddDefaultPred(BuildMI(MBB,
434 TII->get(QPR ? ARM::VDUPLN32q : ARM::VDUPLN32d),
442 // Creates a SPR register from a DPR by copying the value in lane 0.
444 A15SDOptimizer::createExtractSubreg(MachineBasicBlock &MBB,
445 MachineBasicBlock::iterator InsertBefore,
447 unsigned DReg, unsigned Lane,
448 const TargetRegisterClass *TRC) {
449 unsigned Out = MRI->createVirtualRegister(TRC);
453 TII->get(TargetOpcode::COPY), Out)
454 .addReg(DReg, 0, Lane);
459 // Takes two SPR registers and creates a DPR by using a REG_SEQUENCE.
461 A15SDOptimizer::createRegSequence(MachineBasicBlock &MBB,
462 MachineBasicBlock::iterator InsertBefore,
464 unsigned Reg1, unsigned Reg2) {
465 unsigned Out = MRI->createVirtualRegister(&ARM::QPRRegClass);
469 TII->get(TargetOpcode::REG_SEQUENCE), Out)
473 .addImm(ARM::dsub_1);
477 // Takes two DPR registers that have previously been VDUPed (Ssub0 and Ssub1)
478 // and merges them into one DPR register.
480 A15SDOptimizer::createVExt(MachineBasicBlock &MBB,
481 MachineBasicBlock::iterator InsertBefore,
483 unsigned Ssub0, unsigned Ssub1) {
484 unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
485 AddDefaultPred(BuildMI(MBB,
488 TII->get(ARM::VEXTd32), Out)
496 A15SDOptimizer::createInsertSubreg(MachineBasicBlock &MBB,
497 MachineBasicBlock::iterator InsertBefore,
498 DebugLoc DL, unsigned DReg, unsigned Lane,
500 unsigned Out = MRI->createVirtualRegister(&ARM::DPR_VFP2RegClass);
504 TII->get(TargetOpcode::INSERT_SUBREG), Out)
513 A15SDOptimizer::createImplicitDef(MachineBasicBlock &MBB,
514 MachineBasicBlock::iterator InsertBefore,
516 unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
520 TII->get(TargetOpcode::IMPLICIT_DEF), Out);
524 // This function inserts instructions in order to optimize interactions between
525 // SPR registers and DPR/QPR registers. It does so by performing VDUPs on all
526 // lanes, and the using VEXT instructions to recompose the result.
528 A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
529 MachineBasicBlock::iterator InsertPt(MI);
530 DebugLoc DL = MI->getDebugLoc();
531 MachineBasicBlock &MBB = *MI->getParent();
535 // DPair has the same length as QPR and also has two DPRs as subreg.
536 // Treat DPair as QPR.
537 if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass) ||
538 MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPairRegClass)) {
539 unsigned DSub0 = createExtractSubreg(MBB, InsertPt, DL, Reg,
540 ARM::dsub_0, &ARM::DPRRegClass);
541 unsigned DSub1 = createExtractSubreg(MBB, InsertPt, DL, Reg,
542 ARM::dsub_1, &ARM::DPRRegClass);
544 unsigned Out1 = createDupLane(MBB, InsertPt, DL, DSub0, 0);
545 unsigned Out2 = createDupLane(MBB, InsertPt, DL, DSub0, 1);
546 Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
548 unsigned Out3 = createDupLane(MBB, InsertPt, DL, DSub1, 0);
549 unsigned Out4 = createDupLane(MBB, InsertPt, DL, DSub1, 1);
550 Out2 = createVExt(MBB, InsertPt, DL, Out3, Out4);
552 Out = createRegSequence(MBB, InsertPt, DL, Out, Out2);
554 } else if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPRRegClass)) {
555 unsigned Out1 = createDupLane(MBB, InsertPt, DL, Reg, 0);
556 unsigned Out2 = createDupLane(MBB, InsertPt, DL, Reg, 1);
557 Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
560 assert(MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::SPRRegClass) &&
561 "Found unexpected regclass!");
563 unsigned PrefLane = getPrefSPRLane(Reg);
566 case ARM::ssub_0: Lane = 0; break;
567 case ARM::ssub_1: Lane = 1; break;
568 default: llvm_unreachable("Unknown preferred lane!");
571 // Treat DPair as QPR
572 bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass) ||
573 usesRegClass(MI->getOperand(0), &ARM::DPairRegClass);
575 Out = createImplicitDef(MBB, InsertPt, DL);
576 Out = createInsertSubreg(MBB, InsertPt, DL, Out, PrefLane, Reg);
577 Out = createDupLane(MBB, InsertPt, DL, Out, Lane, UsesQPR);
578 eraseInstrWithNoUses(MI);
583 bool A15SDOptimizer::runOnInstruction(MachineInstr *MI) {
584 // We look for instructions that write S registers that are then read as
585 // D/Q registers. These can only be caused by COPY, INSERT_SUBREG and
586 // REG_SEQUENCE pseudos that insert an SPR value into a DPR register or
587 // merge two SPR values to form a DPR register. In order avoid false
588 // positives we make sure that there is an SPR producer so we look past
589 // COPY and PHI nodes to find it.
591 // The best code pattern for when an SPR producer is going to be used by a
592 // DPR or QPR consumer depends on whether the other lanes of the
593 // corresponding DPR/QPR are currently defined.
595 // We can handle these efficiently, depending on the type of
596 // pseudo-instruction that is producing the pattern
598 // * COPY: * VDUP all lanes and merge the results together
601 // * INSERT_SUBREG: * If the SPR value was originally in another DPR/QPR
602 // lane, and the other lane(s) of the DPR/QPR register
603 // that we are inserting in are undefined, use the
604 // original DPR/QPR value.
605 // * Otherwise, fall back on the same stategy as COPY.
607 // * REG_SEQUENCE: * If all except one of the input operands are
608 // IMPLICIT_DEFs, insert the VDUP pattern for just the
609 // defined input operand
610 // * Otherwise, fall back on the same stategy as COPY.
613 // First, get all the reads of D-registers done by this instruction.
614 SmallVector<unsigned, 8> Defs = getReadDPRs(MI);
615 bool Modified = false;
617 for (SmallVectorImpl<unsigned>::iterator I = Defs.begin(), E = Defs.end();
619 // Follow the def-use chain for this DPR through COPYs, and also through
620 // PHIs (which are essentially multi-way COPYs). It is because of PHIs that
621 // we can end up with multiple defs of this DPR.
623 SmallVector<MachineInstr *, 8> DefSrcs;
624 if (!TRI->isVirtualRegister(*I))
626 MachineInstr *Def = MRI->getVRegDef(*I);
630 elideCopiesAndPHIs(Def, DefSrcs);
632 for (SmallVectorImpl<MachineInstr *>::iterator II = DefSrcs.begin(),
633 EE = DefSrcs.end(); II != EE; ++II) {
634 MachineInstr *MI = *II;
636 // If we've already analyzed and replaced this operand, don't do
638 if (Replacements.find(MI) != Replacements.end())
641 // Now, work out if the instruction causes a SPR->DPR dependency.
642 if (!hasPartialWrite(MI))
645 // Collect all the uses of this MI's DPR def for updating later.
646 SmallVector<MachineOperand*, 8> Uses;
647 unsigned DPRDefReg = MI->getOperand(0).getReg();
648 for (MachineRegisterInfo::use_iterator I = MRI->use_begin(DPRDefReg),
649 E = MRI->use_end(); I != E; ++I)
652 // We can optimize this.
653 unsigned NewReg = optimizeSDPattern(MI);
657 for (SmallVectorImpl<MachineOperand *>::const_iterator I = Uses.begin(),
658 E = Uses.end(); I != E; ++I) {
659 // Make sure to constrain the register class of the new register to
660 // match what we're replacing. Otherwise we can optimize a DPR_VFP2
661 // reference into a plain DPR, and that will end poorly. NewReg is
662 // always virtual here, so there will always be a matching subclass
664 MRI->constrainRegClass(NewReg, MRI->getRegClass((*I)->getReg()));
666 DEBUG(dbgs() << "Replacing operand "
668 << PrintReg(NewReg) << "\n");
669 (*I)->substVirtReg(NewReg, 0, *TRI);
672 Replacements[MI] = NewReg;
678 bool A15SDOptimizer::runOnMachineFunction(MachineFunction &Fn) {
679 TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo());
680 TRI = Fn.getTarget().getRegisterInfo();
681 MRI = &Fn.getRegInfo();
682 bool Modified = false;
684 DEBUG(dbgs() << "Running on function " << Fn.getName()<< "\n");
687 Replacements.clear();
689 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
692 for (MachineBasicBlock::iterator MI = MFI->begin(), ME = MFI->end();
694 Modified |= runOnInstruction(MI++);
699 for (std::set<MachineInstr *>::iterator I = DeadInstr.begin(),
702 (*I)->eraseFromParent();
708 FunctionPass *llvm::createA15SDOptimizerPass() {
709 return new A15SDOptimizer();