[x86] fix assert with horizontal math + broadcast of vector (PR43402)
[llvm-core.git] / lib / Target / X86 / X86DomainReassignment.cpp
blobb4cf5cafbc6e5549ed31e60e85d304177cdcb547
1 //===--- X86DomainReassignment.cpp - Selectively switch register classes---===//
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 pass attempts to find instruction chains (closures) in one domain,
10 // and convert them to equivalent instructions in a different domain,
11 // if profitable.
13 //===----------------------------------------------------------------------===//
15 #include "X86.h"
16 #include "X86InstrInfo.h"
17 #include "X86Subtarget.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DenseMapInfo.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/TargetRegisterInfo.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/Printable.h"
29 #include <bitset>
31 using namespace llvm;
33 #define DEBUG_TYPE "x86-domain-reassignment"
35 STATISTIC(NumClosuresConverted, "Number of closures converted by the pass");
37 static cl::opt<bool> DisableX86DomainReassignment(
38 "disable-x86-domain-reassignment", cl::Hidden,
39 cl::desc("X86: Disable Virtual Register Reassignment."), cl::init(false));
41 namespace {
42 enum RegDomain { NoDomain = -1, GPRDomain, MaskDomain, OtherDomain, NumDomains };
44 static bool isGPR(const TargetRegisterClass *RC) {
45 return X86::GR64RegClass.hasSubClassEq(RC) ||
46 X86::GR32RegClass.hasSubClassEq(RC) ||
47 X86::GR16RegClass.hasSubClassEq(RC) ||
48 X86::GR8RegClass.hasSubClassEq(RC);
51 static bool isMask(const TargetRegisterClass *RC,
52 const TargetRegisterInfo *TRI) {
53 return X86::VK16RegClass.hasSubClassEq(RC);
56 static RegDomain getDomain(const TargetRegisterClass *RC,
57 const TargetRegisterInfo *TRI) {
58 if (isGPR(RC))
59 return GPRDomain;
60 if (isMask(RC, TRI))
61 return MaskDomain;
62 return OtherDomain;
65 /// Return a register class equivalent to \p SrcRC, in \p Domain.
66 static const TargetRegisterClass *getDstRC(const TargetRegisterClass *SrcRC,
67 RegDomain Domain) {
68 assert(Domain == MaskDomain && "add domain");
69 if (X86::GR8RegClass.hasSubClassEq(SrcRC))
70 return &X86::VK8RegClass;
71 if (X86::GR16RegClass.hasSubClassEq(SrcRC))
72 return &X86::VK16RegClass;
73 if (X86::GR32RegClass.hasSubClassEq(SrcRC))
74 return &X86::VK32RegClass;
75 if (X86::GR64RegClass.hasSubClassEq(SrcRC))
76 return &X86::VK64RegClass;
77 llvm_unreachable("add register class");
78 return nullptr;
81 /// Abstract Instruction Converter class.
82 class InstrConverterBase {
83 protected:
84 unsigned SrcOpcode;
86 public:
87 InstrConverterBase(unsigned SrcOpcode) : SrcOpcode(SrcOpcode) {}
89 virtual ~InstrConverterBase() {}
91 /// \returns true if \p MI is legal to convert.
92 virtual bool isLegal(const MachineInstr *MI,
93 const TargetInstrInfo *TII) const {
94 assert(MI->getOpcode() == SrcOpcode &&
95 "Wrong instruction passed to converter");
96 return true;
99 /// Applies conversion to \p MI.
101 /// \returns true if \p MI is no longer need, and can be deleted.
102 virtual bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
103 MachineRegisterInfo *MRI) const = 0;
105 /// \returns the cost increment incurred by converting \p MI.
106 virtual double getExtraCost(const MachineInstr *MI,
107 MachineRegisterInfo *MRI) const = 0;
110 /// An Instruction Converter which ignores the given instruction.
111 /// For example, PHI instructions can be safely ignored since only the registers
112 /// need to change.
113 class InstrIgnore : public InstrConverterBase {
114 public:
115 InstrIgnore(unsigned SrcOpcode) : InstrConverterBase(SrcOpcode) {}
117 bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
118 MachineRegisterInfo *MRI) const override {
119 assert(isLegal(MI, TII) && "Cannot convert instruction");
120 return false;
123 double getExtraCost(const MachineInstr *MI,
124 MachineRegisterInfo *MRI) const override {
125 return 0;
129 /// An Instruction Converter which replaces an instruction with another.
130 class InstrReplacer : public InstrConverterBase {
131 public:
132 /// Opcode of the destination instruction.
133 unsigned DstOpcode;
135 InstrReplacer(unsigned SrcOpcode, unsigned DstOpcode)
136 : InstrConverterBase(SrcOpcode), DstOpcode(DstOpcode) {}
138 bool isLegal(const MachineInstr *MI,
139 const TargetInstrInfo *TII) const override {
140 if (!InstrConverterBase::isLegal(MI, TII))
141 return false;
142 // It's illegal to replace an instruction that implicitly defines a register
143 // with an instruction that doesn't, unless that register dead.
144 for (auto &MO : MI->implicit_operands())
145 if (MO.isReg() && MO.isDef() && !MO.isDead() &&
146 !TII->get(DstOpcode).hasImplicitDefOfPhysReg(MO.getReg()))
147 return false;
148 return true;
151 bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
152 MachineRegisterInfo *MRI) const override {
153 assert(isLegal(MI, TII) && "Cannot convert instruction");
154 MachineInstrBuilder Bld =
155 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII->get(DstOpcode));
156 // Transfer explicit operands from original instruction. Implicit operands
157 // are handled by BuildMI.
158 for (auto &Op : MI->explicit_operands())
159 Bld.add(Op);
160 return true;
163 double getExtraCost(const MachineInstr *MI,
164 MachineRegisterInfo *MRI) const override {
165 // Assuming instructions have the same cost.
166 return 0;
170 /// An Instruction Converter which replaces an instruction with another, and
171 /// adds a COPY from the new instruction's destination to the old one's.
172 class InstrReplacerDstCOPY : public InstrConverterBase {
173 public:
174 unsigned DstOpcode;
176 InstrReplacerDstCOPY(unsigned SrcOpcode, unsigned DstOpcode)
177 : InstrConverterBase(SrcOpcode), DstOpcode(DstOpcode) {}
179 bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
180 MachineRegisterInfo *MRI) const override {
181 assert(isLegal(MI, TII) && "Cannot convert instruction");
182 MachineBasicBlock *MBB = MI->getParent();
183 auto &DL = MI->getDebugLoc();
185 Register Reg = MRI->createVirtualRegister(
186 TII->getRegClass(TII->get(DstOpcode), 0, MRI->getTargetRegisterInfo(),
187 *MBB->getParent()));
188 MachineInstrBuilder Bld = BuildMI(*MBB, MI, DL, TII->get(DstOpcode), Reg);
189 for (unsigned Idx = 1, End = MI->getNumOperands(); Idx < End; ++Idx)
190 Bld.add(MI->getOperand(Idx));
192 BuildMI(*MBB, MI, DL, TII->get(TargetOpcode::COPY))
193 .add(MI->getOperand(0))
194 .addReg(Reg);
196 return true;
199 double getExtraCost(const MachineInstr *MI,
200 MachineRegisterInfo *MRI) const override {
201 // Assuming instructions have the same cost, and that COPY is in the same
202 // domain so it will be eliminated.
203 return 0;
207 /// An Instruction Converter for replacing COPY instructions.
208 class InstrCOPYReplacer : public InstrReplacer {
209 public:
210 RegDomain DstDomain;
212 InstrCOPYReplacer(unsigned SrcOpcode, RegDomain DstDomain, unsigned DstOpcode)
213 : InstrReplacer(SrcOpcode, DstOpcode), DstDomain(DstDomain) {}
215 bool isLegal(const MachineInstr *MI,
216 const TargetInstrInfo *TII) const override {
217 if (!InstrConverterBase::isLegal(MI, TII))
218 return false;
220 // Don't allow copies to/flow GR8/GR16 physical registers.
221 // FIXME: Is there some better way to support this?
222 Register DstReg = MI->getOperand(0).getReg();
223 if (Register::isPhysicalRegister(DstReg) &&
224 (X86::GR8RegClass.contains(DstReg) ||
225 X86::GR16RegClass.contains(DstReg)))
226 return false;
227 Register SrcReg = MI->getOperand(1).getReg();
228 if (Register::isPhysicalRegister(SrcReg) &&
229 (X86::GR8RegClass.contains(SrcReg) ||
230 X86::GR16RegClass.contains(SrcReg)))
231 return false;
233 return true;
236 double getExtraCost(const MachineInstr *MI,
237 MachineRegisterInfo *MRI) const override {
238 assert(MI->getOpcode() == TargetOpcode::COPY && "Expected a COPY");
240 for (auto &MO : MI->operands()) {
241 // Physical registers will not be converted. Assume that converting the
242 // COPY to the destination domain will eventually result in a actual
243 // instruction.
244 if (Register::isPhysicalRegister(MO.getReg()))
245 return 1;
247 RegDomain OpDomain = getDomain(MRI->getRegClass(MO.getReg()),
248 MRI->getTargetRegisterInfo());
249 // Converting a cross domain COPY to a same domain COPY should eliminate
250 // an insturction
251 if (OpDomain == DstDomain)
252 return -1;
254 return 0;
258 /// An Instruction Converter which replaces an instruction with a COPY.
259 class InstrReplaceWithCopy : public InstrConverterBase {
260 public:
261 // Source instruction operand Index, to be used as the COPY source.
262 unsigned SrcOpIdx;
264 InstrReplaceWithCopy(unsigned SrcOpcode, unsigned SrcOpIdx)
265 : InstrConverterBase(SrcOpcode), SrcOpIdx(SrcOpIdx) {}
267 bool convertInstr(MachineInstr *MI, const TargetInstrInfo *TII,
268 MachineRegisterInfo *MRI) const override {
269 assert(isLegal(MI, TII) && "Cannot convert instruction");
270 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
271 TII->get(TargetOpcode::COPY))
272 .add({MI->getOperand(0), MI->getOperand(SrcOpIdx)});
273 return true;
276 double getExtraCost(const MachineInstr *MI,
277 MachineRegisterInfo *MRI) const override {
278 return 0;
282 // Key type to be used by the Instruction Converters map.
283 // A converter is identified by <destination domain, source opcode>
284 typedef std::pair<int, unsigned> InstrConverterBaseKeyTy;
286 typedef DenseMap<InstrConverterBaseKeyTy, InstrConverterBase *>
287 InstrConverterBaseMap;
289 /// A closure is a set of virtual register representing all of the edges in
290 /// the closure, as well as all of the instructions connected by those edges.
292 /// A closure may encompass virtual registers in the same register bank that
293 /// have different widths. For example, it may contain 32-bit GPRs as well as
294 /// 64-bit GPRs.
296 /// A closure that computes an address (i.e. defines a virtual register that is
297 /// used in a memory operand) excludes the instructions that contain memory
298 /// operands using the address. Such an instruction will be included in a
299 /// different closure that manipulates the loaded or stored value.
300 class Closure {
301 private:
302 /// Virtual registers in the closure.
303 DenseSet<unsigned> Edges;
305 /// Instructions in the closure.
306 SmallVector<MachineInstr *, 8> Instrs;
308 /// Domains which this closure can legally be reassigned to.
309 std::bitset<NumDomains> LegalDstDomains;
311 /// An ID to uniquely identify this closure, even when it gets
312 /// moved around
313 unsigned ID;
315 public:
316 Closure(unsigned ID, std::initializer_list<RegDomain> LegalDstDomainList) : ID(ID) {
317 for (RegDomain D : LegalDstDomainList)
318 LegalDstDomains.set(D);
321 /// Mark this closure as illegal for reassignment to all domains.
322 void setAllIllegal() { LegalDstDomains.reset(); }
324 /// \returns true if this closure has domains which are legal to reassign to.
325 bool hasLegalDstDomain() const { return LegalDstDomains.any(); }
327 /// \returns true if is legal to reassign this closure to domain \p RD.
328 bool isLegal(RegDomain RD) const { return LegalDstDomains[RD]; }
330 /// Mark this closure as illegal for reassignment to domain \p RD.
331 void setIllegal(RegDomain RD) { LegalDstDomains[RD] = false; }
333 bool empty() const { return Edges.empty(); }
335 bool insertEdge(unsigned Reg) {
336 return Edges.insert(Reg).second;
339 using const_edge_iterator = DenseSet<unsigned>::const_iterator;
340 iterator_range<const_edge_iterator> edges() const {
341 return iterator_range<const_edge_iterator>(Edges.begin(), Edges.end());
344 void addInstruction(MachineInstr *I) {
345 Instrs.push_back(I);
348 ArrayRef<MachineInstr *> instructions() const {
349 return Instrs;
352 LLVM_DUMP_METHOD void dump(const MachineRegisterInfo *MRI) const {
353 dbgs() << "Registers: ";
354 bool First = true;
355 for (unsigned Reg : Edges) {
356 if (!First)
357 dbgs() << ", ";
358 First = false;
359 dbgs() << printReg(Reg, MRI->getTargetRegisterInfo(), 0, MRI);
361 dbgs() << "\n" << "Instructions:";
362 for (MachineInstr *MI : Instrs) {
363 dbgs() << "\n ";
364 MI->print(dbgs());
366 dbgs() << "\n";
369 unsigned getID() const {
370 return ID;
375 class X86DomainReassignment : public MachineFunctionPass {
376 const X86Subtarget *STI;
377 MachineRegisterInfo *MRI;
378 const X86InstrInfo *TII;
380 /// All edges that are included in some closure
381 DenseSet<unsigned> EnclosedEdges;
383 /// All instructions that are included in some closure.
384 DenseMap<MachineInstr *, unsigned> EnclosedInstrs;
386 public:
387 static char ID;
389 X86DomainReassignment() : MachineFunctionPass(ID) { }
391 bool runOnMachineFunction(MachineFunction &MF) override;
393 void getAnalysisUsage(AnalysisUsage &AU) const override {
394 AU.setPreservesCFG();
395 MachineFunctionPass::getAnalysisUsage(AU);
398 StringRef getPassName() const override {
399 return "X86 Domain Reassignment Pass";
402 private:
403 /// A map of available Instruction Converters.
404 InstrConverterBaseMap Converters;
406 /// Initialize Converters map.
407 void initConverters();
409 /// Starting from \Reg, expand the closure as much as possible.
410 void buildClosure(Closure &, unsigned Reg);
412 /// Enqueue \p Reg to be considered for addition to the closure.
413 void visitRegister(Closure &, unsigned Reg, RegDomain &Domain,
414 SmallVectorImpl<unsigned> &Worklist);
416 /// Reassign the closure to \p Domain.
417 void reassign(const Closure &C, RegDomain Domain) const;
419 /// Add \p MI to the closure.
420 void encloseInstr(Closure &C, MachineInstr *MI);
422 /// /returns true if it is profitable to reassign the closure to \p Domain.
423 bool isReassignmentProfitable(const Closure &C, RegDomain Domain) const;
425 /// Calculate the total cost of reassigning the closure to \p Domain.
426 double calculateCost(const Closure &C, RegDomain Domain) const;
429 char X86DomainReassignment::ID = 0;
431 } // End anonymous namespace.
433 void X86DomainReassignment::visitRegister(Closure &C, unsigned Reg,
434 RegDomain &Domain,
435 SmallVectorImpl<unsigned> &Worklist) {
436 if (EnclosedEdges.count(Reg))
437 return;
439 if (!Register::isVirtualRegister(Reg))
440 return;
442 if (!MRI->hasOneDef(Reg))
443 return;
445 RegDomain RD = getDomain(MRI->getRegClass(Reg), MRI->getTargetRegisterInfo());
446 // First edge in closure sets the domain.
447 if (Domain == NoDomain)
448 Domain = RD;
450 if (Domain != RD)
451 return;
453 Worklist.push_back(Reg);
456 void X86DomainReassignment::encloseInstr(Closure &C, MachineInstr *MI) {
457 auto I = EnclosedInstrs.find(MI);
458 if (I != EnclosedInstrs.end()) {
459 if (I->second != C.getID())
460 // Instruction already belongs to another closure, avoid conflicts between
461 // closure and mark this closure as illegal.
462 C.setAllIllegal();
463 return;
466 EnclosedInstrs[MI] = C.getID();
467 C.addInstruction(MI);
469 // Mark closure as illegal for reassignment to domains, if there is no
470 // converter for the instruction or if the converter cannot convert the
471 // instruction.
472 for (int i = 0; i != NumDomains; ++i) {
473 if (C.isLegal((RegDomain)i)) {
474 InstrConverterBase *IC = Converters.lookup({i, MI->getOpcode()});
475 if (!IC || !IC->isLegal(MI, TII))
476 C.setIllegal((RegDomain)i);
481 double X86DomainReassignment::calculateCost(const Closure &C,
482 RegDomain DstDomain) const {
483 assert(C.isLegal(DstDomain) && "Cannot calculate cost for illegal closure");
485 double Cost = 0.0;
486 for (auto *MI : C.instructions())
487 Cost +=
488 Converters.lookup({DstDomain, MI->getOpcode()})->getExtraCost(MI, MRI);
489 return Cost;
492 bool X86DomainReassignment::isReassignmentProfitable(const Closure &C,
493 RegDomain Domain) const {
494 return calculateCost(C, Domain) < 0.0;
497 void X86DomainReassignment::reassign(const Closure &C, RegDomain Domain) const {
498 assert(C.isLegal(Domain) && "Cannot convert illegal closure");
500 // Iterate all instructions in the closure, convert each one using the
501 // appropriate converter.
502 SmallVector<MachineInstr *, 8> ToErase;
503 for (auto *MI : C.instructions())
504 if (Converters.lookup({Domain, MI->getOpcode()})
505 ->convertInstr(MI, TII, MRI))
506 ToErase.push_back(MI);
508 // Iterate all registers in the closure, replace them with registers in the
509 // destination domain.
510 for (unsigned Reg : C.edges()) {
511 MRI->setRegClass(Reg, getDstRC(MRI->getRegClass(Reg), Domain));
512 for (auto &MO : MRI->use_operands(Reg)) {
513 if (MO.isReg())
514 // Remove all subregister references as they are not valid in the
515 // destination domain.
516 MO.setSubReg(0);
520 for (auto MI : ToErase)
521 MI->eraseFromParent();
524 /// \returns true when \p Reg is used as part of an address calculation in \p
525 /// MI.
526 static bool usedAsAddr(const MachineInstr &MI, unsigned Reg,
527 const TargetInstrInfo *TII) {
528 if (!MI.mayLoadOrStore())
529 return false;
531 const MCInstrDesc &Desc = TII->get(MI.getOpcode());
532 int MemOpStart = X86II::getMemoryOperandNo(Desc.TSFlags);
533 if (MemOpStart == -1)
534 return false;
536 MemOpStart += X86II::getOperandBias(Desc);
537 for (unsigned MemOpIdx = MemOpStart,
538 MemOpEnd = MemOpStart + X86::AddrNumOperands;
539 MemOpIdx < MemOpEnd; ++MemOpIdx) {
540 auto &Op = MI.getOperand(MemOpIdx);
541 if (Op.isReg() && Op.getReg() == Reg)
542 return true;
544 return false;
547 void X86DomainReassignment::buildClosure(Closure &C, unsigned Reg) {
548 SmallVector<unsigned, 4> Worklist;
549 RegDomain Domain = NoDomain;
550 visitRegister(C, Reg, Domain, Worklist);
551 while (!Worklist.empty()) {
552 unsigned CurReg = Worklist.pop_back_val();
554 // Register already in this closure.
555 if (!C.insertEdge(CurReg))
556 continue;
557 EnclosedEdges.insert(Reg);
559 MachineInstr *DefMI = MRI->getVRegDef(CurReg);
560 encloseInstr(C, DefMI);
562 // Add register used by the defining MI to the worklist.
563 // Do not add registers which are used in address calculation, they will be
564 // added to a different closure.
565 int OpEnd = DefMI->getNumOperands();
566 const MCInstrDesc &Desc = DefMI->getDesc();
567 int MemOp = X86II::getMemoryOperandNo(Desc.TSFlags);
568 if (MemOp != -1)
569 MemOp += X86II::getOperandBias(Desc);
570 for (int OpIdx = 0; OpIdx < OpEnd; ++OpIdx) {
571 if (OpIdx == MemOp) {
572 // skip address calculation.
573 OpIdx += (X86::AddrNumOperands - 1);
574 continue;
576 auto &Op = DefMI->getOperand(OpIdx);
577 if (!Op.isReg() || !Op.isUse())
578 continue;
579 visitRegister(C, Op.getReg(), Domain, Worklist);
582 // Expand closure through register uses.
583 for (auto &UseMI : MRI->use_nodbg_instructions(CurReg)) {
584 // We would like to avoid converting closures which calculare addresses,
585 // as this should remain in GPRs.
586 if (usedAsAddr(UseMI, CurReg, TII)) {
587 C.setAllIllegal();
588 continue;
590 encloseInstr(C, &UseMI);
592 for (auto &DefOp : UseMI.defs()) {
593 if (!DefOp.isReg())
594 continue;
596 Register DefReg = DefOp.getReg();
597 if (!Register::isVirtualRegister(DefReg)) {
598 C.setAllIllegal();
599 continue;
601 visitRegister(C, DefReg, Domain, Worklist);
607 void X86DomainReassignment::initConverters() {
608 Converters[{MaskDomain, TargetOpcode::PHI}] =
609 new InstrIgnore(TargetOpcode::PHI);
611 Converters[{MaskDomain, TargetOpcode::IMPLICIT_DEF}] =
612 new InstrIgnore(TargetOpcode::IMPLICIT_DEF);
614 Converters[{MaskDomain, TargetOpcode::INSERT_SUBREG}] =
615 new InstrReplaceWithCopy(TargetOpcode::INSERT_SUBREG, 2);
617 Converters[{MaskDomain, TargetOpcode::COPY}] =
618 new InstrCOPYReplacer(TargetOpcode::COPY, MaskDomain, TargetOpcode::COPY);
620 auto createReplacerDstCOPY = [&](unsigned From, unsigned To) {
621 Converters[{MaskDomain, From}] = new InstrReplacerDstCOPY(From, To);
624 createReplacerDstCOPY(X86::MOVZX32rm16, X86::KMOVWkm);
625 createReplacerDstCOPY(X86::MOVZX64rm16, X86::KMOVWkm);
627 createReplacerDstCOPY(X86::MOVZX32rr16, X86::KMOVWkk);
628 createReplacerDstCOPY(X86::MOVZX64rr16, X86::KMOVWkk);
630 if (STI->hasDQI()) {
631 createReplacerDstCOPY(X86::MOVZX16rm8, X86::KMOVBkm);
632 createReplacerDstCOPY(X86::MOVZX32rm8, X86::KMOVBkm);
633 createReplacerDstCOPY(X86::MOVZX64rm8, X86::KMOVBkm);
635 createReplacerDstCOPY(X86::MOVZX16rr8, X86::KMOVBkk);
636 createReplacerDstCOPY(X86::MOVZX32rr8, X86::KMOVBkk);
637 createReplacerDstCOPY(X86::MOVZX64rr8, X86::KMOVBkk);
640 auto createReplacer = [&](unsigned From, unsigned To) {
641 Converters[{MaskDomain, From}] = new InstrReplacer(From, To);
644 createReplacer(X86::MOV16rm, X86::KMOVWkm);
645 createReplacer(X86::MOV16mr, X86::KMOVWmk);
646 createReplacer(X86::MOV16rr, X86::KMOVWkk);
647 createReplacer(X86::SHR16ri, X86::KSHIFTRWri);
648 createReplacer(X86::SHL16ri, X86::KSHIFTLWri);
649 createReplacer(X86::NOT16r, X86::KNOTWrr);
650 createReplacer(X86::OR16rr, X86::KORWrr);
651 createReplacer(X86::AND16rr, X86::KANDWrr);
652 createReplacer(X86::XOR16rr, X86::KXORWrr);
654 if (STI->hasBWI()) {
655 createReplacer(X86::MOV32rm, X86::KMOVDkm);
656 createReplacer(X86::MOV64rm, X86::KMOVQkm);
658 createReplacer(X86::MOV32mr, X86::KMOVDmk);
659 createReplacer(X86::MOV64mr, X86::KMOVQmk);
661 createReplacer(X86::MOV32rr, X86::KMOVDkk);
662 createReplacer(X86::MOV64rr, X86::KMOVQkk);
664 createReplacer(X86::SHR32ri, X86::KSHIFTRDri);
665 createReplacer(X86::SHR64ri, X86::KSHIFTRQri);
667 createReplacer(X86::SHL32ri, X86::KSHIFTLDri);
668 createReplacer(X86::SHL64ri, X86::KSHIFTLQri);
670 createReplacer(X86::ADD32rr, X86::KADDDrr);
671 createReplacer(X86::ADD64rr, X86::KADDQrr);
673 createReplacer(X86::NOT32r, X86::KNOTDrr);
674 createReplacer(X86::NOT64r, X86::KNOTQrr);
676 createReplacer(X86::OR32rr, X86::KORDrr);
677 createReplacer(X86::OR64rr, X86::KORQrr);
679 createReplacer(X86::AND32rr, X86::KANDDrr);
680 createReplacer(X86::AND64rr, X86::KANDQrr);
682 createReplacer(X86::ANDN32rr, X86::KANDNDrr);
683 createReplacer(X86::ANDN64rr, X86::KANDNQrr);
685 createReplacer(X86::XOR32rr, X86::KXORDrr);
686 createReplacer(X86::XOR64rr, X86::KXORQrr);
688 // TODO: KTEST is not a replacement for TEST due to flag differences. Need
689 // to prove only Z flag is used.
690 //createReplacer(X86::TEST32rr, X86::KTESTDrr);
691 //createReplacer(X86::TEST64rr, X86::KTESTQrr);
694 if (STI->hasDQI()) {
695 createReplacer(X86::ADD8rr, X86::KADDBrr);
696 createReplacer(X86::ADD16rr, X86::KADDWrr);
698 createReplacer(X86::AND8rr, X86::KANDBrr);
700 createReplacer(X86::MOV8rm, X86::KMOVBkm);
701 createReplacer(X86::MOV8mr, X86::KMOVBmk);
702 createReplacer(X86::MOV8rr, X86::KMOVBkk);
704 createReplacer(X86::NOT8r, X86::KNOTBrr);
706 createReplacer(X86::OR8rr, X86::KORBrr);
708 createReplacer(X86::SHR8ri, X86::KSHIFTRBri);
709 createReplacer(X86::SHL8ri, X86::KSHIFTLBri);
711 // TODO: KTEST is not a replacement for TEST due to flag differences. Need
712 // to prove only Z flag is used.
713 //createReplacer(X86::TEST8rr, X86::KTESTBrr);
714 //createReplacer(X86::TEST16rr, X86::KTESTWrr);
716 createReplacer(X86::XOR8rr, X86::KXORBrr);
720 bool X86DomainReassignment::runOnMachineFunction(MachineFunction &MF) {
721 if (skipFunction(MF.getFunction()))
722 return false;
723 if (DisableX86DomainReassignment)
724 return false;
726 LLVM_DEBUG(
727 dbgs() << "***** Machine Function before Domain Reassignment *****\n");
728 LLVM_DEBUG(MF.print(dbgs()));
730 STI = &MF.getSubtarget<X86Subtarget>();
731 // GPR->K is the only transformation currently supported, bail out early if no
732 // AVX512.
733 // TODO: We're also bailing of AVX512BW isn't supported since we use VK32 and
734 // VK64 for GR32/GR64, but those aren't legal classes on KNL. If the register
735 // coalescer doesn't clean it up and we generate a spill we will crash.
736 if (!STI->hasAVX512() || !STI->hasBWI())
737 return false;
739 MRI = &MF.getRegInfo();
740 assert(MRI->isSSA() && "Expected MIR to be in SSA form");
742 TII = STI->getInstrInfo();
743 initConverters();
744 bool Changed = false;
746 EnclosedEdges.clear();
747 EnclosedInstrs.clear();
749 std::vector<Closure> Closures;
751 // Go over all virtual registers and calculate a closure.
752 unsigned ClosureID = 0;
753 for (unsigned Idx = 0; Idx < MRI->getNumVirtRegs(); ++Idx) {
754 unsigned Reg = Register::index2VirtReg(Idx);
756 // GPR only current source domain supported.
757 if (!isGPR(MRI->getRegClass(Reg)))
758 continue;
760 // Register already in closure.
761 if (EnclosedEdges.count(Reg))
762 continue;
764 // Calculate closure starting with Reg.
765 Closure C(ClosureID++, {MaskDomain});
766 buildClosure(C, Reg);
768 // Collect all closures that can potentially be converted.
769 if (!C.empty() && C.isLegal(MaskDomain))
770 Closures.push_back(std::move(C));
773 for (Closure &C : Closures) {
774 LLVM_DEBUG(C.dump(MRI));
775 if (isReassignmentProfitable(C, MaskDomain)) {
776 reassign(C, MaskDomain);
777 ++NumClosuresConverted;
778 Changed = true;
782 DeleteContainerSeconds(Converters);
784 LLVM_DEBUG(
785 dbgs() << "***** Machine Function after Domain Reassignment *****\n");
786 LLVM_DEBUG(MF.print(dbgs()));
788 return Changed;
791 INITIALIZE_PASS(X86DomainReassignment, "x86-domain-reassignment",
792 "X86 Domain Reassignment Pass", false, false)
794 /// Returns an instance of the Domain Reassignment pass.
795 FunctionPass *llvm::createX86DomainReassignmentPass() {
796 return new X86DomainReassignment();