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 //===----------------------------------------------------------------------===//
28 #include "ARMBaseInstrInfo.h"
29 #include "ARMBaseRegisterInfo.h"
30 #include "ARMSubtarget.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineInstr.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/TargetRegisterInfo.h"
38 #include "llvm/CodeGen/TargetSubtargetInfo.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
46 #define DEBUG_TYPE "a15-sd-optimizer"
49 struct A15SDOptimizer : public MachineFunctionPass {
51 A15SDOptimizer() : MachineFunctionPass(ID) {}
53 bool runOnMachineFunction(MachineFunction &Fn) override;
55 StringRef getPassName() const override { return "ARM A15 S->D optimizer"; }
58 const ARMBaseInstrInfo *TII;
59 const TargetRegisterInfo *TRI;
60 MachineRegisterInfo *MRI;
62 bool runOnInstruction(MachineInstr *MI);
65 // Instruction builder helpers
67 unsigned createDupLane(MachineBasicBlock &MBB,
68 MachineBasicBlock::iterator InsertBefore,
69 const DebugLoc &DL, unsigned Reg, unsigned Lane,
72 unsigned createExtractSubreg(MachineBasicBlock &MBB,
73 MachineBasicBlock::iterator InsertBefore,
74 const DebugLoc &DL, unsigned DReg,
75 unsigned Lane, const TargetRegisterClass *TRC);
77 unsigned createVExt(MachineBasicBlock &MBB,
78 MachineBasicBlock::iterator InsertBefore,
79 const DebugLoc &DL, unsigned Ssub0, unsigned Ssub1);
81 unsigned createRegSequence(MachineBasicBlock &MBB,
82 MachineBasicBlock::iterator InsertBefore,
83 const DebugLoc &DL, unsigned Reg1,
86 unsigned createInsertSubreg(MachineBasicBlock &MBB,
87 MachineBasicBlock::iterator InsertBefore,
88 const DebugLoc &DL, unsigned DReg,
89 unsigned Lane, unsigned ToInsert);
91 unsigned createImplicitDef(MachineBasicBlock &MBB,
92 MachineBasicBlock::iterator InsertBefore,
96 // Various property checkers
98 bool usesRegClass(MachineOperand &MO, const TargetRegisterClass *TRC);
99 bool hasPartialWrite(MachineInstr *MI);
100 SmallVector<unsigned, 8> getReadDPRs(MachineInstr *MI);
101 unsigned getDPRLaneFromSPR(unsigned SReg);
104 // Methods used for getting the definitions of partial registers
107 MachineInstr *elideCopies(MachineInstr *MI);
108 void elideCopiesAndPHIs(MachineInstr *MI,
109 SmallVectorImpl<MachineInstr*> &Outs);
112 // Pattern optimization methods
114 unsigned optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg);
115 unsigned optimizeSDPattern(MachineInstr *MI);
116 unsigned getPrefSPRLane(unsigned SReg);
119 // Sanitizing method - used to make sure if don't leave dead code around.
121 void eraseInstrWithNoUses(MachineInstr *MI);
124 // A map used to track the changes done by this pass.
126 std::map<MachineInstr*, unsigned> Replacements;
127 std::set<MachineInstr *> DeadInstr;
129 char A15SDOptimizer::ID = 0;
130 } // end anonymous namespace
132 // Returns true if this is a use of a SPR register.
133 bool A15SDOptimizer::usesRegClass(MachineOperand &MO,
134 const TargetRegisterClass *TRC) {
137 unsigned Reg = MO.getReg();
139 if (TargetRegisterInfo::isVirtualRegister(Reg))
140 return MRI->getRegClass(Reg)->hasSuperClassEq(TRC);
142 return TRC->contains(Reg);
145 unsigned A15SDOptimizer::getDPRLaneFromSPR(unsigned SReg) {
146 unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1,
148 if (DReg != ARM::NoRegister) return ARM::ssub_1;
152 // Get the subreg type that is most likely to be coalesced
153 // for an SPR register that will be used in VDUP32d pseudo.
154 unsigned A15SDOptimizer::getPrefSPRLane(unsigned SReg) {
155 if (!TRI->isVirtualRegister(SReg))
156 return getDPRLaneFromSPR(SReg);
158 MachineInstr *MI = MRI->getVRegDef(SReg);
159 if (!MI) return ARM::ssub_0;
160 MachineOperand *MO = MI->findRegisterDefOperand(SReg);
162 assert(MO->isReg() && "Non-register operand found!");
163 if (!MO) return ARM::ssub_0;
165 if (MI->isCopy() && usesRegClass(MI->getOperand(1),
166 &ARM::SPRRegClass)) {
167 SReg = MI->getOperand(1).getReg();
170 if (TargetRegisterInfo::isVirtualRegister(SReg)) {
171 if (MO->getSubReg() == ARM::ssub_1) return ARM::ssub_1;
174 return getDPRLaneFromSPR(SReg);
177 // MI is known to be dead. Figure out what instructions
178 // are also made dead by this and mark them for removal.
179 void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) {
180 SmallVector<MachineInstr *, 8> Front;
181 DeadInstr.insert(MI);
183 LLVM_DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n");
186 while (Front.size() != 0) {
190 // MI is already known to be dead. We need to see
191 // if other instructions can also be removed.
192 for (MachineOperand &MO : MI->operands()) {
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 (MachineOperand &MODef : Def->operands()) {
215 if ((!MODef.isReg()) || (!MODef.isDef()))
217 unsigned DefReg = MODef.getReg();
218 if (!TRI->isVirtualRegister(DefReg)) {
222 for (MachineInstr &Use : MRI->use_instructions(Reg)) {
223 // We don't care about self references.
226 if (DeadInstr.find(&Use) == DeadInstr.end()) {
233 if (!IsDead) continue;
235 LLVM_DEBUG(dbgs() << "Deleting instruction " << *Def << "\n");
236 DeadInstr.insert(Def);
241 // Creates the more optimized patterns and generally does all the code
242 // transformations in this pass.
243 unsigned A15SDOptimizer::optimizeSDPattern(MachineInstr *MI) {
245 return optimizeAllLanesPattern(MI, MI->getOperand(1).getReg());
248 if (MI->isInsertSubreg()) {
249 unsigned DPRReg = MI->getOperand(1).getReg();
250 unsigned SPRReg = MI->getOperand(2).getReg();
252 if (TRI->isVirtualRegister(DPRReg) && TRI->isVirtualRegister(SPRReg)) {
253 MachineInstr *DPRMI = MRI->getVRegDef(MI->getOperand(1).getReg());
254 MachineInstr *SPRMI = MRI->getVRegDef(MI->getOperand(2).getReg());
256 if (DPRMI && SPRMI) {
257 // See if the first operand of this insert_subreg is IMPLICIT_DEF
258 MachineInstr *ECDef = elideCopies(DPRMI);
259 if (ECDef && ECDef->isImplicitDef()) {
260 // Another corner case - if we're inserting something that is purely
261 // a subreg copy of a DPR, just use that DPR.
263 MachineInstr *EC = elideCopies(SPRMI);
264 // Is it a subreg copy of ssub_0?
265 if (EC && EC->isCopy() &&
266 EC->getOperand(1).getSubReg() == ARM::ssub_0) {
267 LLVM_DEBUG(dbgs() << "Found a subreg copy: " << *SPRMI);
269 // Find the thing we're subreg copying out of - is it of the same
270 // regclass as DPRMI? (i.e. a DPR or QPR).
271 unsigned FullReg = SPRMI->getOperand(1).getReg();
272 const TargetRegisterClass *TRC =
273 MRI->getRegClass(MI->getOperand(1).getReg());
274 if (TRC->hasSuperClassEq(MRI->getRegClass(FullReg))) {
275 LLVM_DEBUG(dbgs() << "Subreg copy is compatible - returning ");
276 LLVM_DEBUG(dbgs() << printReg(FullReg) << "\n");
277 eraseInstrWithNoUses(MI);
282 return optimizeAllLanesPattern(MI, MI->getOperand(2).getReg());
286 return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
289 if (MI->isRegSequence() && usesRegClass(MI->getOperand(1),
290 &ARM::SPRRegClass)) {
291 // See if all bar one of the operands are IMPLICIT_DEF and insert the
292 // optimizer pattern accordingly.
293 unsigned NumImplicit = 0, NumTotal = 0;
294 unsigned NonImplicitReg = ~0U;
296 for (unsigned I = 1; I < MI->getNumExplicitOperands(); ++I) {
297 if (!MI->getOperand(I).isReg())
300 unsigned OpReg = MI->getOperand(I).getReg();
302 if (!TRI->isVirtualRegister(OpReg))
305 MachineInstr *Def = MRI->getVRegDef(OpReg);
308 if (Def->isImplicitDef())
311 NonImplicitReg = MI->getOperand(I).getReg();
314 if (NumImplicit == NumTotal - 1)
315 return optimizeAllLanesPattern(MI, NonImplicitReg);
317 return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
320 llvm_unreachable("Unhandled update pattern!");
323 // Return true if this MachineInstr inserts a scalar (SPR) value into
324 // a D or Q register.
325 bool A15SDOptimizer::hasPartialWrite(MachineInstr *MI) {
326 // The only way we can do a partial register update is through a COPY,
327 // INSERT_SUBREG or REG_SEQUENCE.
328 if (MI->isCopy() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
331 if (MI->isInsertSubreg() && usesRegClass(MI->getOperand(2),
335 if (MI->isRegSequence() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
341 // Looks through full copies to get the instruction that defines the input
343 MachineInstr *A15SDOptimizer::elideCopies(MachineInstr *MI) {
344 if (!MI->isFullCopy())
346 if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
348 MachineInstr *Def = MRI->getVRegDef(MI->getOperand(1).getReg());
351 return elideCopies(Def);
354 // Look through full copies and PHIs to get the set of non-copy MachineInstrs
355 // that can produce MI.
356 void A15SDOptimizer::elideCopiesAndPHIs(MachineInstr *MI,
357 SmallVectorImpl<MachineInstr*> &Outs) {
358 // Looking through PHIs may create loops so we need to track what
359 // instructions we have visited before.
360 std::set<MachineInstr *> Reached;
361 SmallVector<MachineInstr *, 8> Front;
363 while (Front.size() != 0) {
367 // If we have already explored this MachineInstr, ignore it.
368 if (Reached.find(MI) != Reached.end())
372 for (unsigned I = 1, E = MI->getNumOperands(); I != E; I += 2) {
373 unsigned Reg = MI->getOperand(I).getReg();
374 if (!TRI->isVirtualRegister(Reg)) {
377 MachineInstr *NewMI = MRI->getVRegDef(Reg);
380 Front.push_back(NewMI);
382 } else if (MI->isFullCopy()) {
383 if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
385 MachineInstr *NewMI = MRI->getVRegDef(MI->getOperand(1).getReg());
388 Front.push_back(NewMI);
390 LLVM_DEBUG(dbgs() << "Found partial copy" << *MI << "\n");
396 // Return the DPR virtual registers that are read by this machine instruction
398 SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) {
399 if (MI->isCopyLike() || MI->isInsertSubreg() || MI->isRegSequence() ||
401 return SmallVector<unsigned, 8>();
403 SmallVector<unsigned, 8> Defs;
404 for (MachineOperand &MO : MI->operands()) {
405 if (!MO.isReg() || !MO.isUse())
407 if (!usesRegClass(MO, &ARM::DPRRegClass) &&
408 !usesRegClass(MO, &ARM::QPRRegClass) &&
409 !usesRegClass(MO, &ARM::DPairRegClass)) // Treat DPair as QPR
412 Defs.push_back(MO.getReg());
417 // Creates a DPR register from an SPR one by using a VDUP.
418 unsigned A15SDOptimizer::createDupLane(MachineBasicBlock &MBB,
419 MachineBasicBlock::iterator InsertBefore,
420 const DebugLoc &DL, unsigned Reg,
421 unsigned Lane, bool QPR) {
422 unsigned Out = MRI->createVirtualRegister(QPR ? &ARM::QPRRegClass :
424 BuildMI(MBB, InsertBefore, DL,
425 TII->get(QPR ? ARM::VDUPLN32q : ARM::VDUPLN32d), Out)
428 .add(predOps(ARMCC::AL));
433 // Creates a SPR register from a DPR by copying the value in lane 0.
434 unsigned A15SDOptimizer::createExtractSubreg(
435 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
436 const DebugLoc &DL, unsigned DReg, unsigned Lane,
437 const TargetRegisterClass *TRC) {
438 unsigned Out = MRI->createVirtualRegister(TRC);
442 TII->get(TargetOpcode::COPY), Out)
443 .addReg(DReg, 0, Lane);
448 // Takes two SPR registers and creates a DPR by using a REG_SEQUENCE.
449 unsigned A15SDOptimizer::createRegSequence(
450 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
451 const DebugLoc &DL, unsigned Reg1, unsigned Reg2) {
452 unsigned Out = MRI->createVirtualRegister(&ARM::QPRRegClass);
456 TII->get(TargetOpcode::REG_SEQUENCE), Out)
460 .addImm(ARM::dsub_1);
464 // Takes two DPR registers that have previously been VDUPed (Ssub0 and Ssub1)
465 // and merges them into one DPR register.
466 unsigned A15SDOptimizer::createVExt(MachineBasicBlock &MBB,
467 MachineBasicBlock::iterator InsertBefore,
468 const DebugLoc &DL, unsigned Ssub0,
470 unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
471 BuildMI(MBB, InsertBefore, DL, TII->get(ARM::VEXTd32), Out)
475 .add(predOps(ARMCC::AL));
479 unsigned A15SDOptimizer::createInsertSubreg(
480 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore,
481 const DebugLoc &DL, unsigned DReg, unsigned Lane, unsigned ToInsert) {
482 unsigned Out = MRI->createVirtualRegister(&ARM::DPR_VFP2RegClass);
486 TII->get(TargetOpcode::INSERT_SUBREG), Out)
495 A15SDOptimizer::createImplicitDef(MachineBasicBlock &MBB,
496 MachineBasicBlock::iterator InsertBefore,
497 const DebugLoc &DL) {
498 unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
502 TII->get(TargetOpcode::IMPLICIT_DEF), Out);
506 // This function inserts instructions in order to optimize interactions between
507 // SPR registers and DPR/QPR registers. It does so by performing VDUPs on all
508 // lanes, and the using VEXT instructions to recompose the result.
510 A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
511 MachineBasicBlock::iterator InsertPt(MI);
512 DebugLoc DL = MI->getDebugLoc();
513 MachineBasicBlock &MBB = *MI->getParent();
517 // DPair has the same length as QPR and also has two DPRs as subreg.
518 // Treat DPair as QPR.
519 if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass) ||
520 MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPairRegClass)) {
521 unsigned DSub0 = createExtractSubreg(MBB, InsertPt, DL, Reg,
522 ARM::dsub_0, &ARM::DPRRegClass);
523 unsigned DSub1 = createExtractSubreg(MBB, InsertPt, DL, Reg,
524 ARM::dsub_1, &ARM::DPRRegClass);
526 unsigned Out1 = createDupLane(MBB, InsertPt, DL, DSub0, 0);
527 unsigned Out2 = createDupLane(MBB, InsertPt, DL, DSub0, 1);
528 Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
530 unsigned Out3 = createDupLane(MBB, InsertPt, DL, DSub1, 0);
531 unsigned Out4 = createDupLane(MBB, InsertPt, DL, DSub1, 1);
532 Out2 = createVExt(MBB, InsertPt, DL, Out3, Out4);
534 Out = createRegSequence(MBB, InsertPt, DL, Out, Out2);
536 } else if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPRRegClass)) {
537 unsigned Out1 = createDupLane(MBB, InsertPt, DL, Reg, 0);
538 unsigned Out2 = createDupLane(MBB, InsertPt, DL, Reg, 1);
539 Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
542 assert(MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::SPRRegClass) &&
543 "Found unexpected regclass!");
545 unsigned PrefLane = getPrefSPRLane(Reg);
548 case ARM::ssub_0: Lane = 0; break;
549 case ARM::ssub_1: Lane = 1; break;
550 default: llvm_unreachable("Unknown preferred lane!");
553 // Treat DPair as QPR
554 bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass) ||
555 usesRegClass(MI->getOperand(0), &ARM::DPairRegClass);
557 Out = createImplicitDef(MBB, InsertPt, DL);
558 Out = createInsertSubreg(MBB, InsertPt, DL, Out, PrefLane, Reg);
559 Out = createDupLane(MBB, InsertPt, DL, Out, Lane, UsesQPR);
560 eraseInstrWithNoUses(MI);
565 bool A15SDOptimizer::runOnInstruction(MachineInstr *MI) {
566 // We look for instructions that write S registers that are then read as
567 // D/Q registers. These can only be caused by COPY, INSERT_SUBREG and
568 // REG_SEQUENCE pseudos that insert an SPR value into a DPR register or
569 // merge two SPR values to form a DPR register. In order avoid false
570 // positives we make sure that there is an SPR producer so we look past
571 // COPY and PHI nodes to find it.
573 // The best code pattern for when an SPR producer is going to be used by a
574 // DPR or QPR consumer depends on whether the other lanes of the
575 // corresponding DPR/QPR are currently defined.
577 // We can handle these efficiently, depending on the type of
578 // pseudo-instruction that is producing the pattern
580 // * COPY: * VDUP all lanes and merge the results together
583 // * INSERT_SUBREG: * If the SPR value was originally in another DPR/QPR
584 // lane, and the other lane(s) of the DPR/QPR register
585 // that we are inserting in are undefined, use the
586 // original DPR/QPR value.
587 // * Otherwise, fall back on the same stategy as COPY.
589 // * REG_SEQUENCE: * If all except one of the input operands are
590 // IMPLICIT_DEFs, insert the VDUP pattern for just the
591 // defined input operand
592 // * Otherwise, fall back on the same stategy as COPY.
595 // First, get all the reads of D-registers done by this instruction.
596 SmallVector<unsigned, 8> Defs = getReadDPRs(MI);
597 bool Modified = false;
599 for (SmallVectorImpl<unsigned>::iterator I = Defs.begin(), E = Defs.end();
601 // Follow the def-use chain for this DPR through COPYs, and also through
602 // PHIs (which are essentially multi-way COPYs). It is because of PHIs that
603 // we can end up with multiple defs of this DPR.
605 SmallVector<MachineInstr *, 8> DefSrcs;
606 if (!TRI->isVirtualRegister(*I))
608 MachineInstr *Def = MRI->getVRegDef(*I);
612 elideCopiesAndPHIs(Def, DefSrcs);
614 for (MachineInstr *MI : DefSrcs) {
615 // If we've already analyzed and replaced this operand, don't do
617 if (Replacements.find(MI) != Replacements.end())
620 // Now, work out if the instruction causes a SPR->DPR dependency.
621 if (!hasPartialWrite(MI))
624 // Collect all the uses of this MI's DPR def for updating later.
625 SmallVector<MachineOperand*, 8> Uses;
626 unsigned DPRDefReg = MI->getOperand(0).getReg();
627 for (MachineRegisterInfo::use_iterator I = MRI->use_begin(DPRDefReg),
628 E = MRI->use_end(); I != E; ++I)
631 // We can optimize this.
632 unsigned NewReg = optimizeSDPattern(MI);
636 for (SmallVectorImpl<MachineOperand *>::const_iterator I = Uses.begin(),
637 E = Uses.end(); I != E; ++I) {
638 // Make sure to constrain the register class of the new register to
639 // match what we're replacing. Otherwise we can optimize a DPR_VFP2
640 // reference into a plain DPR, and that will end poorly. NewReg is
641 // always virtual here, so there will always be a matching subclass
643 MRI->constrainRegClass(NewReg, MRI->getRegClass((*I)->getReg()));
645 LLVM_DEBUG(dbgs() << "Replacing operand " << **I << " with "
646 << printReg(NewReg) << "\n");
647 (*I)->substVirtReg(NewReg, 0, *TRI);
650 Replacements[MI] = NewReg;
656 bool A15SDOptimizer::runOnMachineFunction(MachineFunction &Fn) {
657 if (skipFunction(Fn.getFunction()))
660 const ARMSubtarget &STI = Fn.getSubtarget<ARMSubtarget>();
661 // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be
662 // enabled when NEON is available.
663 if (!(STI.useSplatVFPToNeon() && STI.hasNEON()))
666 TII = STI.getInstrInfo();
667 TRI = STI.getRegisterInfo();
668 MRI = &Fn.getRegInfo();
669 bool Modified = false;
671 LLVM_DEBUG(dbgs() << "Running on function " << Fn.getName() << "\n");
674 Replacements.clear();
676 for (MachineBasicBlock &MBB : Fn) {
677 for (MachineInstr &MI : MBB) {
678 Modified |= runOnInstruction(&MI);
682 for (MachineInstr *MI : DeadInstr) {
683 MI->eraseFromParent();
689 FunctionPass *llvm::createA15SDOptimizerPass() {
690 return new A15SDOptimizer();