[ORC] Add std::tuple support to SimplePackedSerialization.
[llvm-project.git] / llvm / lib / Target / ARM / ARMInstructionSelector.cpp
blob8be4e3f160e309bd66e745004dd50edeec8566b7
1 //===- ARMInstructionSelector.cpp ----------------------------*- C++ -*-==//
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 /// \file
9 /// This file implements the targeting of the InstructionSelector class for ARM.
10 /// \todo This should be generated by TableGen.
11 //===----------------------------------------------------------------------===//
13 #include "ARMRegisterBankInfo.h"
14 #include "ARMSubtarget.h"
15 #include "ARMTargetMachine.h"
16 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
17 #include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
18 #include "llvm/CodeGen/MachineConstantPool.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/IR/IntrinsicsARM.h"
21 #include "llvm/Support/Debug.h"
23 #define DEBUG_TYPE "arm-isel"
25 using namespace llvm;
27 namespace {
29 #define GET_GLOBALISEL_PREDICATE_BITSET
30 #include "ARMGenGlobalISel.inc"
31 #undef GET_GLOBALISEL_PREDICATE_BITSET
33 class ARMInstructionSelector : public InstructionSelector {
34 public:
35 ARMInstructionSelector(const ARMBaseTargetMachine &TM, const ARMSubtarget &STI,
36 const ARMRegisterBankInfo &RBI);
38 bool select(MachineInstr &I) override;
39 static const char *getName() { return DEBUG_TYPE; }
41 private:
42 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
44 struct CmpConstants;
45 struct InsertInfo;
47 bool selectCmp(CmpConstants Helper, MachineInstrBuilder &MIB,
48 MachineRegisterInfo &MRI) const;
50 // Helper for inserting a comparison sequence that sets \p ResReg to either 1
51 // if \p LHSReg and \p RHSReg are in the relationship defined by \p Cond, or
52 // \p PrevRes otherwise. In essence, it computes PrevRes OR (LHS Cond RHS).
53 bool insertComparison(CmpConstants Helper, InsertInfo I, unsigned ResReg,
54 ARMCC::CondCodes Cond, unsigned LHSReg, unsigned RHSReg,
55 unsigned PrevRes) const;
57 // Set \p DestReg to \p Constant.
58 void putConstant(InsertInfo I, unsigned DestReg, unsigned Constant) const;
60 bool selectGlobal(MachineInstrBuilder &MIB, MachineRegisterInfo &MRI) const;
61 bool selectSelect(MachineInstrBuilder &MIB, MachineRegisterInfo &MRI) const;
62 bool selectShift(unsigned ShiftOpc, MachineInstrBuilder &MIB) const;
64 // Check if the types match and both operands have the expected size and
65 // register bank.
66 bool validOpRegPair(MachineRegisterInfo &MRI, unsigned LHS, unsigned RHS,
67 unsigned ExpectedSize, unsigned ExpectedRegBankID) const;
69 // Check if the register has the expected size and register bank.
70 bool validReg(MachineRegisterInfo &MRI, unsigned Reg, unsigned ExpectedSize,
71 unsigned ExpectedRegBankID) const;
73 const ARMBaseInstrInfo &TII;
74 const ARMBaseRegisterInfo &TRI;
75 const ARMBaseTargetMachine &TM;
76 const ARMRegisterBankInfo &RBI;
77 const ARMSubtarget &STI;
79 // FIXME: This is necessary because DAGISel uses "Subtarget->" and GlobalISel
80 // uses "STI." in the code generated by TableGen. If we want to reuse some of
81 // the custom C++ predicates written for DAGISel, we need to have both around.
82 const ARMSubtarget *Subtarget = &STI;
84 // Store the opcodes that we might need, so we don't have to check what kind
85 // of subtarget (ARM vs Thumb) we have all the time.
86 struct OpcodeCache {
87 unsigned ZEXT16;
88 unsigned SEXT16;
90 unsigned ZEXT8;
91 unsigned SEXT8;
93 // Used for implementing ZEXT/SEXT from i1
94 unsigned AND;
95 unsigned RSB;
97 unsigned STORE32;
98 unsigned LOAD32;
100 unsigned STORE16;
101 unsigned LOAD16;
103 unsigned STORE8;
104 unsigned LOAD8;
106 unsigned ADDrr;
107 unsigned ADDri;
109 // Used for G_ICMP
110 unsigned CMPrr;
111 unsigned MOVi;
112 unsigned MOVCCi;
114 // Used for G_SELECT
115 unsigned MOVCCr;
117 unsigned TSTri;
118 unsigned Bcc;
120 // Used for G_GLOBAL_VALUE
121 unsigned MOVi32imm;
122 unsigned ConstPoolLoad;
123 unsigned MOV_ga_pcrel;
124 unsigned LDRLIT_ga_pcrel;
125 unsigned LDRLIT_ga_abs;
127 OpcodeCache(const ARMSubtarget &STI);
128 } const Opcodes;
130 // Select the opcode for simple extensions (that translate to a single SXT/UXT
131 // instruction). Extension operations more complicated than that should not
132 // invoke this. Returns the original opcode if it doesn't know how to select a
133 // better one.
134 unsigned selectSimpleExtOpc(unsigned Opc, unsigned Size) const;
136 // Select the opcode for simple loads and stores. Returns the original opcode
137 // if it doesn't know how to select a better one.
138 unsigned selectLoadStoreOpCode(unsigned Opc, unsigned RegBank,
139 unsigned Size) const;
141 void renderVFPF32Imm(MachineInstrBuilder &New, const MachineInstr &Old,
142 int OpIdx = -1) const;
143 void renderVFPF64Imm(MachineInstrBuilder &New, const MachineInstr &Old,
144 int OpIdx = -1) const;
146 #define GET_GLOBALISEL_PREDICATES_DECL
147 #include "ARMGenGlobalISel.inc"
148 #undef GET_GLOBALISEL_PREDICATES_DECL
150 // We declare the temporaries used by selectImpl() in the class to minimize the
151 // cost of constructing placeholder values.
152 #define GET_GLOBALISEL_TEMPORARIES_DECL
153 #include "ARMGenGlobalISel.inc"
154 #undef GET_GLOBALISEL_TEMPORARIES_DECL
156 } // end anonymous namespace
158 namespace llvm {
159 InstructionSelector *
160 createARMInstructionSelector(const ARMBaseTargetMachine &TM,
161 const ARMSubtarget &STI,
162 const ARMRegisterBankInfo &RBI) {
163 return new ARMInstructionSelector(TM, STI, RBI);
167 #define GET_GLOBALISEL_IMPL
168 #include "ARMGenGlobalISel.inc"
169 #undef GET_GLOBALISEL_IMPL
171 ARMInstructionSelector::ARMInstructionSelector(const ARMBaseTargetMachine &TM,
172 const ARMSubtarget &STI,
173 const ARMRegisterBankInfo &RBI)
174 : InstructionSelector(), TII(*STI.getInstrInfo()),
175 TRI(*STI.getRegisterInfo()), TM(TM), RBI(RBI), STI(STI), Opcodes(STI),
176 #define GET_GLOBALISEL_PREDICATES_INIT
177 #include "ARMGenGlobalISel.inc"
178 #undef GET_GLOBALISEL_PREDICATES_INIT
179 #define GET_GLOBALISEL_TEMPORARIES_INIT
180 #include "ARMGenGlobalISel.inc"
181 #undef GET_GLOBALISEL_TEMPORARIES_INIT
185 static const TargetRegisterClass *guessRegClass(unsigned Reg,
186 MachineRegisterInfo &MRI,
187 const TargetRegisterInfo &TRI,
188 const RegisterBankInfo &RBI) {
189 const RegisterBank *RegBank = RBI.getRegBank(Reg, MRI, TRI);
190 assert(RegBank && "Can't get reg bank for virtual register");
192 const unsigned Size = MRI.getType(Reg).getSizeInBits();
193 assert((RegBank->getID() == ARM::GPRRegBankID ||
194 RegBank->getID() == ARM::FPRRegBankID) &&
195 "Unsupported reg bank");
197 if (RegBank->getID() == ARM::FPRRegBankID) {
198 if (Size == 32)
199 return &ARM::SPRRegClass;
200 else if (Size == 64)
201 return &ARM::DPRRegClass;
202 else if (Size == 128)
203 return &ARM::QPRRegClass;
204 else
205 llvm_unreachable("Unsupported destination size");
208 return &ARM::GPRRegClass;
211 static bool selectCopy(MachineInstr &I, const TargetInstrInfo &TII,
212 MachineRegisterInfo &MRI, const TargetRegisterInfo &TRI,
213 const RegisterBankInfo &RBI) {
214 Register DstReg = I.getOperand(0).getReg();
215 if (Register::isPhysicalRegister(DstReg))
216 return true;
218 const TargetRegisterClass *RC = guessRegClass(DstReg, MRI, TRI, RBI);
220 // No need to constrain SrcReg. It will get constrained when
221 // we hit another of its uses or its defs.
222 // Copies do not have constraints.
223 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
224 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
225 << " operand\n");
226 return false;
228 return true;
231 static bool selectMergeValues(MachineInstrBuilder &MIB,
232 const ARMBaseInstrInfo &TII,
233 MachineRegisterInfo &MRI,
234 const TargetRegisterInfo &TRI,
235 const RegisterBankInfo &RBI) {
236 assert(TII.getSubtarget().hasVFP2Base() && "Can't select merge without VFP");
238 // We only support G_MERGE_VALUES as a way to stick together two scalar GPRs
239 // into one DPR.
240 Register VReg0 = MIB.getReg(0);
241 (void)VReg0;
242 assert(MRI.getType(VReg0).getSizeInBits() == 64 &&
243 RBI.getRegBank(VReg0, MRI, TRI)->getID() == ARM::FPRRegBankID &&
244 "Unsupported operand for G_MERGE_VALUES");
245 Register VReg1 = MIB.getReg(1);
246 (void)VReg1;
247 assert(MRI.getType(VReg1).getSizeInBits() == 32 &&
248 RBI.getRegBank(VReg1, MRI, TRI)->getID() == ARM::GPRRegBankID &&
249 "Unsupported operand for G_MERGE_VALUES");
250 Register VReg2 = MIB.getReg(2);
251 (void)VReg2;
252 assert(MRI.getType(VReg2).getSizeInBits() == 32 &&
253 RBI.getRegBank(VReg2, MRI, TRI)->getID() == ARM::GPRRegBankID &&
254 "Unsupported operand for G_MERGE_VALUES");
256 MIB->setDesc(TII.get(ARM::VMOVDRR));
257 MIB.add(predOps(ARMCC::AL));
259 return true;
262 static bool selectUnmergeValues(MachineInstrBuilder &MIB,
263 const ARMBaseInstrInfo &TII,
264 MachineRegisterInfo &MRI,
265 const TargetRegisterInfo &TRI,
266 const RegisterBankInfo &RBI) {
267 assert(TII.getSubtarget().hasVFP2Base() &&
268 "Can't select unmerge without VFP");
270 // We only support G_UNMERGE_VALUES as a way to break up one DPR into two
271 // GPRs.
272 Register VReg0 = MIB.getReg(0);
273 (void)VReg0;
274 assert(MRI.getType(VReg0).getSizeInBits() == 32 &&
275 RBI.getRegBank(VReg0, MRI, TRI)->getID() == ARM::GPRRegBankID &&
276 "Unsupported operand for G_UNMERGE_VALUES");
277 Register VReg1 = MIB.getReg(1);
278 (void)VReg1;
279 assert(MRI.getType(VReg1).getSizeInBits() == 32 &&
280 RBI.getRegBank(VReg1, MRI, TRI)->getID() == ARM::GPRRegBankID &&
281 "Unsupported operand for G_UNMERGE_VALUES");
282 Register VReg2 = MIB.getReg(2);
283 (void)VReg2;
284 assert(MRI.getType(VReg2).getSizeInBits() == 64 &&
285 RBI.getRegBank(VReg2, MRI, TRI)->getID() == ARM::FPRRegBankID &&
286 "Unsupported operand for G_UNMERGE_VALUES");
288 MIB->setDesc(TII.get(ARM::VMOVRRD));
289 MIB.add(predOps(ARMCC::AL));
291 return true;
294 ARMInstructionSelector::OpcodeCache::OpcodeCache(const ARMSubtarget &STI) {
295 bool isThumb = STI.isThumb();
297 using namespace TargetOpcode;
299 #define STORE_OPCODE(VAR, OPC) VAR = isThumb ? ARM::t2##OPC : ARM::OPC
300 STORE_OPCODE(SEXT16, SXTH);
301 STORE_OPCODE(ZEXT16, UXTH);
303 STORE_OPCODE(SEXT8, SXTB);
304 STORE_OPCODE(ZEXT8, UXTB);
306 STORE_OPCODE(AND, ANDri);
307 STORE_OPCODE(RSB, RSBri);
309 STORE_OPCODE(STORE32, STRi12);
310 STORE_OPCODE(LOAD32, LDRi12);
312 // LDRH/STRH are special...
313 STORE16 = isThumb ? ARM::t2STRHi12 : ARM::STRH;
314 LOAD16 = isThumb ? ARM::t2LDRHi12 : ARM::LDRH;
316 STORE_OPCODE(STORE8, STRBi12);
317 STORE_OPCODE(LOAD8, LDRBi12);
319 STORE_OPCODE(ADDrr, ADDrr);
320 STORE_OPCODE(ADDri, ADDri);
322 STORE_OPCODE(CMPrr, CMPrr);
323 STORE_OPCODE(MOVi, MOVi);
324 STORE_OPCODE(MOVCCi, MOVCCi);
326 STORE_OPCODE(MOVCCr, MOVCCr);
328 STORE_OPCODE(TSTri, TSTri);
329 STORE_OPCODE(Bcc, Bcc);
331 STORE_OPCODE(MOVi32imm, MOVi32imm);
332 ConstPoolLoad = isThumb ? ARM::t2LDRpci : ARM::LDRi12;
333 STORE_OPCODE(MOV_ga_pcrel, MOV_ga_pcrel);
334 LDRLIT_ga_pcrel = isThumb ? ARM::tLDRLIT_ga_pcrel : ARM::LDRLIT_ga_pcrel;
335 LDRLIT_ga_abs = isThumb ? ARM::tLDRLIT_ga_abs : ARM::LDRLIT_ga_abs;
336 #undef MAP_OPCODE
339 unsigned ARMInstructionSelector::selectSimpleExtOpc(unsigned Opc,
340 unsigned Size) const {
341 using namespace TargetOpcode;
343 if (Size != 8 && Size != 16)
344 return Opc;
346 if (Opc == G_SEXT)
347 return Size == 8 ? Opcodes.SEXT8 : Opcodes.SEXT16;
349 if (Opc == G_ZEXT)
350 return Size == 8 ? Opcodes.ZEXT8 : Opcodes.ZEXT16;
352 return Opc;
355 unsigned ARMInstructionSelector::selectLoadStoreOpCode(unsigned Opc,
356 unsigned RegBank,
357 unsigned Size) const {
358 bool isStore = Opc == TargetOpcode::G_STORE;
360 if (RegBank == ARM::GPRRegBankID) {
361 switch (Size) {
362 case 1:
363 case 8:
364 return isStore ? Opcodes.STORE8 : Opcodes.LOAD8;
365 case 16:
366 return isStore ? Opcodes.STORE16 : Opcodes.LOAD16;
367 case 32:
368 return isStore ? Opcodes.STORE32 : Opcodes.LOAD32;
369 default:
370 return Opc;
374 if (RegBank == ARM::FPRRegBankID) {
375 switch (Size) {
376 case 32:
377 return isStore ? ARM::VSTRS : ARM::VLDRS;
378 case 64:
379 return isStore ? ARM::VSTRD : ARM::VLDRD;
380 default:
381 return Opc;
385 return Opc;
388 // When lowering comparisons, we sometimes need to perform two compares instead
389 // of just one. Get the condition codes for both comparisons. If only one is
390 // needed, the second member of the pair is ARMCC::AL.
391 static std::pair<ARMCC::CondCodes, ARMCC::CondCodes>
392 getComparePreds(CmpInst::Predicate Pred) {
393 std::pair<ARMCC::CondCodes, ARMCC::CondCodes> Preds = {ARMCC::AL, ARMCC::AL};
394 switch (Pred) {
395 case CmpInst::FCMP_ONE:
396 Preds = {ARMCC::GT, ARMCC::MI};
397 break;
398 case CmpInst::FCMP_UEQ:
399 Preds = {ARMCC::EQ, ARMCC::VS};
400 break;
401 case CmpInst::ICMP_EQ:
402 case CmpInst::FCMP_OEQ:
403 Preds.first = ARMCC::EQ;
404 break;
405 case CmpInst::ICMP_SGT:
406 case CmpInst::FCMP_OGT:
407 Preds.first = ARMCC::GT;
408 break;
409 case CmpInst::ICMP_SGE:
410 case CmpInst::FCMP_OGE:
411 Preds.first = ARMCC::GE;
412 break;
413 case CmpInst::ICMP_UGT:
414 case CmpInst::FCMP_UGT:
415 Preds.first = ARMCC::HI;
416 break;
417 case CmpInst::FCMP_OLT:
418 Preds.first = ARMCC::MI;
419 break;
420 case CmpInst::ICMP_ULE:
421 case CmpInst::FCMP_OLE:
422 Preds.first = ARMCC::LS;
423 break;
424 case CmpInst::FCMP_ORD:
425 Preds.first = ARMCC::VC;
426 break;
427 case CmpInst::FCMP_UNO:
428 Preds.first = ARMCC::VS;
429 break;
430 case CmpInst::FCMP_UGE:
431 Preds.first = ARMCC::PL;
432 break;
433 case CmpInst::ICMP_SLT:
434 case CmpInst::FCMP_ULT:
435 Preds.first = ARMCC::LT;
436 break;
437 case CmpInst::ICMP_SLE:
438 case CmpInst::FCMP_ULE:
439 Preds.first = ARMCC::LE;
440 break;
441 case CmpInst::FCMP_UNE:
442 case CmpInst::ICMP_NE:
443 Preds.first = ARMCC::NE;
444 break;
445 case CmpInst::ICMP_UGE:
446 Preds.first = ARMCC::HS;
447 break;
448 case CmpInst::ICMP_ULT:
449 Preds.first = ARMCC::LO;
450 break;
451 default:
452 break;
454 assert(Preds.first != ARMCC::AL && "No comparisons needed?");
455 return Preds;
458 struct ARMInstructionSelector::CmpConstants {
459 CmpConstants(unsigned CmpOpcode, unsigned FlagsOpcode, unsigned SelectOpcode,
460 unsigned OpRegBank, unsigned OpSize)
461 : ComparisonOpcode(CmpOpcode), ReadFlagsOpcode(FlagsOpcode),
462 SelectResultOpcode(SelectOpcode), OperandRegBankID(OpRegBank),
463 OperandSize(OpSize) {}
465 // The opcode used for performing the comparison.
466 const unsigned ComparisonOpcode;
468 // The opcode used for reading the flags set by the comparison. May be
469 // ARM::INSTRUCTION_LIST_END if we don't need to read the flags.
470 const unsigned ReadFlagsOpcode;
472 // The opcode used for materializing the result of the comparison.
473 const unsigned SelectResultOpcode;
475 // The assumed register bank ID for the operands.
476 const unsigned OperandRegBankID;
478 // The assumed size in bits for the operands.
479 const unsigned OperandSize;
482 struct ARMInstructionSelector::InsertInfo {
483 InsertInfo(MachineInstrBuilder &MIB)
484 : MBB(*MIB->getParent()), InsertBefore(std::next(MIB->getIterator())),
485 DbgLoc(MIB->getDebugLoc()) {}
487 MachineBasicBlock &MBB;
488 const MachineBasicBlock::instr_iterator InsertBefore;
489 const DebugLoc &DbgLoc;
492 void ARMInstructionSelector::putConstant(InsertInfo I, unsigned DestReg,
493 unsigned Constant) const {
494 (void)BuildMI(I.MBB, I.InsertBefore, I.DbgLoc, TII.get(Opcodes.MOVi))
495 .addDef(DestReg)
496 .addImm(Constant)
497 .add(predOps(ARMCC::AL))
498 .add(condCodeOp());
501 bool ARMInstructionSelector::validOpRegPair(MachineRegisterInfo &MRI,
502 unsigned LHSReg, unsigned RHSReg,
503 unsigned ExpectedSize,
504 unsigned ExpectedRegBankID) const {
505 return MRI.getType(LHSReg) == MRI.getType(RHSReg) &&
506 validReg(MRI, LHSReg, ExpectedSize, ExpectedRegBankID) &&
507 validReg(MRI, RHSReg, ExpectedSize, ExpectedRegBankID);
510 bool ARMInstructionSelector::validReg(MachineRegisterInfo &MRI, unsigned Reg,
511 unsigned ExpectedSize,
512 unsigned ExpectedRegBankID) const {
513 if (MRI.getType(Reg).getSizeInBits() != ExpectedSize) {
514 LLVM_DEBUG(dbgs() << "Unexpected size for register");
515 return false;
518 if (RBI.getRegBank(Reg, MRI, TRI)->getID() != ExpectedRegBankID) {
519 LLVM_DEBUG(dbgs() << "Unexpected register bank for register");
520 return false;
523 return true;
526 bool ARMInstructionSelector::selectCmp(CmpConstants Helper,
527 MachineInstrBuilder &MIB,
528 MachineRegisterInfo &MRI) const {
529 const InsertInfo I(MIB);
531 auto ResReg = MIB.getReg(0);
532 if (!validReg(MRI, ResReg, 1, ARM::GPRRegBankID))
533 return false;
535 auto Cond =
536 static_cast<CmpInst::Predicate>(MIB->getOperand(1).getPredicate());
537 if (Cond == CmpInst::FCMP_TRUE || Cond == CmpInst::FCMP_FALSE) {
538 putConstant(I, ResReg, Cond == CmpInst::FCMP_TRUE ? 1 : 0);
539 MIB->eraseFromParent();
540 return true;
543 auto LHSReg = MIB.getReg(2);
544 auto RHSReg = MIB.getReg(3);
545 if (!validOpRegPair(MRI, LHSReg, RHSReg, Helper.OperandSize,
546 Helper.OperandRegBankID))
547 return false;
549 auto ARMConds = getComparePreds(Cond);
550 auto ZeroReg = MRI.createVirtualRegister(&ARM::GPRRegClass);
551 putConstant(I, ZeroReg, 0);
553 if (ARMConds.second == ARMCC::AL) {
554 // Simple case, we only need one comparison and we're done.
555 if (!insertComparison(Helper, I, ResReg, ARMConds.first, LHSReg, RHSReg,
556 ZeroReg))
557 return false;
558 } else {
559 // Not so simple, we need two successive comparisons.
560 auto IntermediateRes = MRI.createVirtualRegister(&ARM::GPRRegClass);
561 if (!insertComparison(Helper, I, IntermediateRes, ARMConds.first, LHSReg,
562 RHSReg, ZeroReg))
563 return false;
564 if (!insertComparison(Helper, I, ResReg, ARMConds.second, LHSReg, RHSReg,
565 IntermediateRes))
566 return false;
569 MIB->eraseFromParent();
570 return true;
573 bool ARMInstructionSelector::insertComparison(CmpConstants Helper, InsertInfo I,
574 unsigned ResReg,
575 ARMCC::CondCodes Cond,
576 unsigned LHSReg, unsigned RHSReg,
577 unsigned PrevRes) const {
578 // Perform the comparison.
579 auto CmpI =
580 BuildMI(I.MBB, I.InsertBefore, I.DbgLoc, TII.get(Helper.ComparisonOpcode))
581 .addUse(LHSReg)
582 .addUse(RHSReg)
583 .add(predOps(ARMCC::AL));
584 if (!constrainSelectedInstRegOperands(*CmpI, TII, TRI, RBI))
585 return false;
587 // Read the comparison flags (if necessary).
588 if (Helper.ReadFlagsOpcode != ARM::INSTRUCTION_LIST_END) {
589 auto ReadI = BuildMI(I.MBB, I.InsertBefore, I.DbgLoc,
590 TII.get(Helper.ReadFlagsOpcode))
591 .add(predOps(ARMCC::AL));
592 if (!constrainSelectedInstRegOperands(*ReadI, TII, TRI, RBI))
593 return false;
596 // Select either 1 or the previous result based on the value of the flags.
597 auto Mov1I = BuildMI(I.MBB, I.InsertBefore, I.DbgLoc,
598 TII.get(Helper.SelectResultOpcode))
599 .addDef(ResReg)
600 .addUse(PrevRes)
601 .addImm(1)
602 .add(predOps(Cond, ARM::CPSR));
603 if (!constrainSelectedInstRegOperands(*Mov1I, TII, TRI, RBI))
604 return false;
606 return true;
609 bool ARMInstructionSelector::selectGlobal(MachineInstrBuilder &MIB,
610 MachineRegisterInfo &MRI) const {
611 if ((STI.isROPI() || STI.isRWPI()) && !STI.isTargetELF()) {
612 LLVM_DEBUG(dbgs() << "ROPI and RWPI only supported for ELF\n");
613 return false;
616 auto GV = MIB->getOperand(1).getGlobal();
617 if (GV->isThreadLocal()) {
618 LLVM_DEBUG(dbgs() << "TLS variables not supported yet\n");
619 return false;
622 auto &MBB = *MIB->getParent();
623 auto &MF = *MBB.getParent();
625 bool UseMovt = STI.useMovt();
627 unsigned Size = TM.getPointerSize(0);
628 const Align Alignment(4);
630 auto addOpsForConstantPoolLoad = [&MF, Alignment,
631 Size](MachineInstrBuilder &MIB,
632 const GlobalValue *GV, bool IsSBREL) {
633 assert((MIB->getOpcode() == ARM::LDRi12 ||
634 MIB->getOpcode() == ARM::t2LDRpci) &&
635 "Unsupported instruction");
636 auto ConstPool = MF.getConstantPool();
637 auto CPIndex =
638 // For SB relative entries we need a target-specific constant pool.
639 // Otherwise, just use a regular constant pool entry.
640 IsSBREL
641 ? ConstPool->getConstantPoolIndex(
642 ARMConstantPoolConstant::Create(GV, ARMCP::SBREL), Alignment)
643 : ConstPool->getConstantPoolIndex(GV, Alignment);
644 MIB.addConstantPoolIndex(CPIndex, /*Offset*/ 0, /*TargetFlags*/ 0)
645 .addMemOperand(MF.getMachineMemOperand(
646 MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
647 Size, Alignment));
648 if (MIB->getOpcode() == ARM::LDRi12)
649 MIB.addImm(0);
650 MIB.add(predOps(ARMCC::AL));
653 auto addGOTMemOperand = [this, &MF, Alignment](MachineInstrBuilder &MIB) {
654 MIB.addMemOperand(MF.getMachineMemOperand(
655 MachinePointerInfo::getGOT(MF), MachineMemOperand::MOLoad,
656 TM.getProgramPointerSize(), Alignment));
659 if (TM.isPositionIndependent()) {
660 bool Indirect = STI.isGVIndirectSymbol(GV);
662 // For ARM mode, we have different pseudoinstructions for direct accesses
663 // and indirect accesses, and the ones for indirect accesses include the
664 // load from GOT. For Thumb mode, we use the same pseudoinstruction for both
665 // direct and indirect accesses, and we need to manually generate the load
666 // from GOT.
667 bool UseOpcodeThatLoads = Indirect && !STI.isThumb();
669 // FIXME: Taking advantage of MOVT for ELF is pretty involved, so we don't
670 // support it yet. See PR28229.
671 unsigned Opc =
672 UseMovt && !STI.isTargetELF()
673 ? (UseOpcodeThatLoads ? (unsigned)ARM::MOV_ga_pcrel_ldr
674 : Opcodes.MOV_ga_pcrel)
675 : (UseOpcodeThatLoads ? (unsigned)ARM::LDRLIT_ga_pcrel_ldr
676 : Opcodes.LDRLIT_ga_pcrel);
677 MIB->setDesc(TII.get(Opc));
679 int TargetFlags = ARMII::MO_NO_FLAG;
680 if (STI.isTargetDarwin())
681 TargetFlags |= ARMII::MO_NONLAZY;
682 if (STI.isGVInGOT(GV))
683 TargetFlags |= ARMII::MO_GOT;
684 MIB->getOperand(1).setTargetFlags(TargetFlags);
686 if (Indirect) {
687 if (!UseOpcodeThatLoads) {
688 auto ResultReg = MIB.getReg(0);
689 auto AddressReg = MRI.createVirtualRegister(&ARM::GPRRegClass);
691 MIB->getOperand(0).setReg(AddressReg);
693 auto InsertBefore = std::next(MIB->getIterator());
694 auto MIBLoad = BuildMI(MBB, InsertBefore, MIB->getDebugLoc(),
695 TII.get(Opcodes.LOAD32))
696 .addDef(ResultReg)
697 .addReg(AddressReg)
698 .addImm(0)
699 .add(predOps(ARMCC::AL));
700 addGOTMemOperand(MIBLoad);
702 if (!constrainSelectedInstRegOperands(*MIBLoad, TII, TRI, RBI))
703 return false;
704 } else {
705 addGOTMemOperand(MIB);
709 return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
712 bool isReadOnly = STI.getTargetLowering()->isReadOnly(GV);
713 if (STI.isROPI() && isReadOnly) {
714 unsigned Opc = UseMovt ? Opcodes.MOV_ga_pcrel : Opcodes.LDRLIT_ga_pcrel;
715 MIB->setDesc(TII.get(Opc));
716 return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
718 if (STI.isRWPI() && !isReadOnly) {
719 auto Offset = MRI.createVirtualRegister(&ARM::GPRRegClass);
720 MachineInstrBuilder OffsetMIB;
721 if (UseMovt) {
722 OffsetMIB = BuildMI(MBB, *MIB, MIB->getDebugLoc(),
723 TII.get(Opcodes.MOVi32imm), Offset);
724 OffsetMIB.addGlobalAddress(GV, /*Offset*/ 0, ARMII::MO_SBREL);
725 } else {
726 // Load the offset from the constant pool.
727 OffsetMIB = BuildMI(MBB, *MIB, MIB->getDebugLoc(),
728 TII.get(Opcodes.ConstPoolLoad), Offset);
729 addOpsForConstantPoolLoad(OffsetMIB, GV, /*IsSBREL*/ true);
731 if (!constrainSelectedInstRegOperands(*OffsetMIB, TII, TRI, RBI))
732 return false;
734 // Add the offset to the SB register.
735 MIB->setDesc(TII.get(Opcodes.ADDrr));
736 MIB->RemoveOperand(1);
737 MIB.addReg(ARM::R9) // FIXME: don't hardcode R9
738 .addReg(Offset)
739 .add(predOps(ARMCC::AL))
740 .add(condCodeOp());
742 return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
745 if (STI.isTargetELF()) {
746 if (UseMovt) {
747 MIB->setDesc(TII.get(Opcodes.MOVi32imm));
748 } else {
749 // Load the global's address from the constant pool.
750 MIB->setDesc(TII.get(Opcodes.ConstPoolLoad));
751 MIB->RemoveOperand(1);
752 addOpsForConstantPoolLoad(MIB, GV, /*IsSBREL*/ false);
754 } else if (STI.isTargetMachO()) {
755 if (UseMovt)
756 MIB->setDesc(TII.get(Opcodes.MOVi32imm));
757 else
758 MIB->setDesc(TII.get(Opcodes.LDRLIT_ga_abs));
759 } else {
760 LLVM_DEBUG(dbgs() << "Object format not supported yet\n");
761 return false;
764 return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
767 bool ARMInstructionSelector::selectSelect(MachineInstrBuilder &MIB,
768 MachineRegisterInfo &MRI) const {
769 auto &MBB = *MIB->getParent();
770 auto InsertBefore = std::next(MIB->getIterator());
771 auto &DbgLoc = MIB->getDebugLoc();
773 // Compare the condition to 1.
774 auto CondReg = MIB.getReg(1);
775 assert(validReg(MRI, CondReg, 1, ARM::GPRRegBankID) &&
776 "Unsupported types for select operation");
777 auto CmpI = BuildMI(MBB, InsertBefore, DbgLoc, TII.get(Opcodes.TSTri))
778 .addUse(CondReg)
779 .addImm(1)
780 .add(predOps(ARMCC::AL));
781 if (!constrainSelectedInstRegOperands(*CmpI, TII, TRI, RBI))
782 return false;
784 // Move a value into the result register based on the result of the
785 // comparison.
786 auto ResReg = MIB.getReg(0);
787 auto TrueReg = MIB.getReg(2);
788 auto FalseReg = MIB.getReg(3);
789 assert(validOpRegPair(MRI, ResReg, TrueReg, 32, ARM::GPRRegBankID) &&
790 validOpRegPair(MRI, TrueReg, FalseReg, 32, ARM::GPRRegBankID) &&
791 "Unsupported types for select operation");
792 auto Mov1I = BuildMI(MBB, InsertBefore, DbgLoc, TII.get(Opcodes.MOVCCr))
793 .addDef(ResReg)
794 .addUse(TrueReg)
795 .addUse(FalseReg)
796 .add(predOps(ARMCC::EQ, ARM::CPSR));
797 if (!constrainSelectedInstRegOperands(*Mov1I, TII, TRI, RBI))
798 return false;
800 MIB->eraseFromParent();
801 return true;
804 bool ARMInstructionSelector::selectShift(unsigned ShiftOpc,
805 MachineInstrBuilder &MIB) const {
806 assert(!STI.isThumb() && "Unsupported subtarget");
807 MIB->setDesc(TII.get(ARM::MOVsr));
808 MIB.addImm(ShiftOpc);
809 MIB.add(predOps(ARMCC::AL)).add(condCodeOp());
810 return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
813 void ARMInstructionSelector::renderVFPF32Imm(
814 MachineInstrBuilder &NewInstBuilder, const MachineInstr &OldInst,
815 int OpIdx) const {
816 assert(OldInst.getOpcode() == TargetOpcode::G_FCONSTANT &&
817 OpIdx == -1 && "Expected G_FCONSTANT");
819 APFloat FPImmValue = OldInst.getOperand(1).getFPImm()->getValueAPF();
820 int FPImmEncoding = ARM_AM::getFP32Imm(FPImmValue);
821 assert(FPImmEncoding != -1 && "Invalid immediate value");
823 NewInstBuilder.addImm(FPImmEncoding);
826 void ARMInstructionSelector::renderVFPF64Imm(
827 MachineInstrBuilder &NewInstBuilder, const MachineInstr &OldInst, int OpIdx) const {
828 assert(OldInst.getOpcode() == TargetOpcode::G_FCONSTANT &&
829 OpIdx == -1 && "Expected G_FCONSTANT");
831 APFloat FPImmValue = OldInst.getOperand(1).getFPImm()->getValueAPF();
832 int FPImmEncoding = ARM_AM::getFP64Imm(FPImmValue);
833 assert(FPImmEncoding != -1 && "Invalid immediate value");
835 NewInstBuilder.addImm(FPImmEncoding);
838 bool ARMInstructionSelector::select(MachineInstr &I) {
839 assert(I.getParent() && "Instruction should be in a basic block!");
840 assert(I.getParent()->getParent() && "Instruction should be in a function!");
842 auto &MBB = *I.getParent();
843 auto &MF = *MBB.getParent();
844 auto &MRI = MF.getRegInfo();
846 if (!isPreISelGenericOpcode(I.getOpcode())) {
847 if (I.isCopy())
848 return selectCopy(I, TII, MRI, TRI, RBI);
850 return true;
853 using namespace TargetOpcode;
855 if (selectImpl(I, *CoverageInfo))
856 return true;
858 MachineInstrBuilder MIB{MF, I};
859 bool isSExt = false;
861 switch (I.getOpcode()) {
862 case G_SEXT:
863 isSExt = true;
864 LLVM_FALLTHROUGH;
865 case G_ZEXT: {
866 assert(MRI.getType(I.getOperand(0).getReg()).getSizeInBits() <= 32 &&
867 "Unsupported destination size for extension");
869 LLT SrcTy = MRI.getType(I.getOperand(1).getReg());
870 unsigned SrcSize = SrcTy.getSizeInBits();
871 switch (SrcSize) {
872 case 1: {
873 // ZExt boils down to & 0x1; for SExt we also subtract that from 0
874 I.setDesc(TII.get(Opcodes.AND));
875 MIB.addImm(1).add(predOps(ARMCC::AL)).add(condCodeOp());
877 if (isSExt) {
878 Register SExtResult = I.getOperand(0).getReg();
880 // Use a new virtual register for the result of the AND
881 Register AndResult = MRI.createVirtualRegister(&ARM::GPRRegClass);
882 I.getOperand(0).setReg(AndResult);
884 auto InsertBefore = std::next(I.getIterator());
885 auto SubI =
886 BuildMI(MBB, InsertBefore, I.getDebugLoc(), TII.get(Opcodes.RSB))
887 .addDef(SExtResult)
888 .addUse(AndResult)
889 .addImm(0)
890 .add(predOps(ARMCC::AL))
891 .add(condCodeOp());
892 if (!constrainSelectedInstRegOperands(*SubI, TII, TRI, RBI))
893 return false;
895 break;
897 case 8:
898 case 16: {
899 unsigned NewOpc = selectSimpleExtOpc(I.getOpcode(), SrcSize);
900 if (NewOpc == I.getOpcode())
901 return false;
902 I.setDesc(TII.get(NewOpc));
903 MIB.addImm(0).add(predOps(ARMCC::AL));
904 break;
906 default:
907 LLVM_DEBUG(dbgs() << "Unsupported source size for extension");
908 return false;
910 break;
912 case G_ANYEXT:
913 case G_TRUNC: {
914 // The high bits are undefined, so there's nothing special to do, just
915 // treat it as a copy.
916 auto SrcReg = I.getOperand(1).getReg();
917 auto DstReg = I.getOperand(0).getReg();
919 const auto &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
920 const auto &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);
922 if (SrcRegBank.getID() == ARM::FPRRegBankID) {
923 // This should only happen in the obscure case where we have put a 64-bit
924 // integer into a D register. Get it out of there and keep only the
925 // interesting part.
926 assert(I.getOpcode() == G_TRUNC && "Unsupported operand for G_ANYEXT");
927 assert(DstRegBank.getID() == ARM::GPRRegBankID &&
928 "Unsupported combination of register banks");
929 assert(MRI.getType(SrcReg).getSizeInBits() == 64 && "Unsupported size");
930 assert(MRI.getType(DstReg).getSizeInBits() <= 32 && "Unsupported size");
932 Register IgnoredBits = MRI.createVirtualRegister(&ARM::GPRRegClass);
933 auto InsertBefore = std::next(I.getIterator());
934 auto MovI =
935 BuildMI(MBB, InsertBefore, I.getDebugLoc(), TII.get(ARM::VMOVRRD))
936 .addDef(DstReg)
937 .addDef(IgnoredBits)
938 .addUse(SrcReg)
939 .add(predOps(ARMCC::AL));
940 if (!constrainSelectedInstRegOperands(*MovI, TII, TRI, RBI))
941 return false;
943 MIB->eraseFromParent();
944 return true;
947 if (SrcRegBank.getID() != DstRegBank.getID()) {
948 LLVM_DEBUG(
949 dbgs() << "G_TRUNC/G_ANYEXT operands on different register banks\n");
950 return false;
953 if (SrcRegBank.getID() != ARM::GPRRegBankID) {
954 LLVM_DEBUG(dbgs() << "G_TRUNC/G_ANYEXT on non-GPR not supported yet\n");
955 return false;
958 I.setDesc(TII.get(COPY));
959 return selectCopy(I, TII, MRI, TRI, RBI);
961 case G_CONSTANT: {
962 if (!MRI.getType(I.getOperand(0).getReg()).isPointer()) {
963 // Non-pointer constants should be handled by TableGen.
964 LLVM_DEBUG(dbgs() << "Unsupported constant type\n");
965 return false;
968 auto &Val = I.getOperand(1);
969 if (Val.isCImm()) {
970 if (!Val.getCImm()->isZero()) {
971 LLVM_DEBUG(dbgs() << "Unsupported pointer constant value\n");
972 return false;
974 Val.ChangeToImmediate(0);
975 } else {
976 assert(Val.isImm() && "Unexpected operand for G_CONSTANT");
977 if (Val.getImm() != 0) {
978 LLVM_DEBUG(dbgs() << "Unsupported pointer constant value\n");
979 return false;
983 assert(!STI.isThumb() && "Unsupported subtarget");
984 I.setDesc(TII.get(ARM::MOVi));
985 MIB.add(predOps(ARMCC::AL)).add(condCodeOp());
986 break;
988 case G_FCONSTANT: {
989 // Load from constant pool
990 unsigned Size = MRI.getType(I.getOperand(0).getReg()).getSizeInBits() / 8;
991 Align Alignment(Size);
993 assert((Size == 4 || Size == 8) && "Unsupported FP constant type");
994 auto LoadOpcode = Size == 4 ? ARM::VLDRS : ARM::VLDRD;
996 auto ConstPool = MF.getConstantPool();
997 auto CPIndex =
998 ConstPool->getConstantPoolIndex(I.getOperand(1).getFPImm(), Alignment);
999 MIB->setDesc(TII.get(LoadOpcode));
1000 MIB->RemoveOperand(1);
1001 MIB.addConstantPoolIndex(CPIndex, /*Offset*/ 0, /*TargetFlags*/ 0)
1002 .addMemOperand(
1003 MF.getMachineMemOperand(MachinePointerInfo::getConstantPool(MF),
1004 MachineMemOperand::MOLoad, Size, Alignment))
1005 .addImm(0)
1006 .add(predOps(ARMCC::AL));
1007 break;
1009 case G_INTTOPTR:
1010 case G_PTRTOINT: {
1011 auto SrcReg = I.getOperand(1).getReg();
1012 auto DstReg = I.getOperand(0).getReg();
1014 const auto &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
1015 const auto &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);
1017 if (SrcRegBank.getID() != DstRegBank.getID()) {
1018 LLVM_DEBUG(
1019 dbgs()
1020 << "G_INTTOPTR/G_PTRTOINT operands on different register banks\n");
1021 return false;
1024 if (SrcRegBank.getID() != ARM::GPRRegBankID) {
1025 LLVM_DEBUG(
1026 dbgs() << "G_INTTOPTR/G_PTRTOINT on non-GPR not supported yet\n");
1027 return false;
1030 I.setDesc(TII.get(COPY));
1031 return selectCopy(I, TII, MRI, TRI, RBI);
1033 case G_SELECT:
1034 return selectSelect(MIB, MRI);
1035 case G_ICMP: {
1036 CmpConstants Helper(Opcodes.CMPrr, ARM::INSTRUCTION_LIST_END,
1037 Opcodes.MOVCCi, ARM::GPRRegBankID, 32);
1038 return selectCmp(Helper, MIB, MRI);
1040 case G_FCMP: {
1041 assert(STI.hasVFP2Base() && "Can't select fcmp without VFP");
1043 Register OpReg = I.getOperand(2).getReg();
1044 unsigned Size = MRI.getType(OpReg).getSizeInBits();
1046 if (Size == 64 && !STI.hasFP64()) {
1047 LLVM_DEBUG(dbgs() << "Subtarget only supports single precision");
1048 return false;
1050 if (Size != 32 && Size != 64) {
1051 LLVM_DEBUG(dbgs() << "Unsupported size for G_FCMP operand");
1052 return false;
1055 CmpConstants Helper(Size == 32 ? ARM::VCMPS : ARM::VCMPD, ARM::FMSTAT,
1056 Opcodes.MOVCCi, ARM::FPRRegBankID, Size);
1057 return selectCmp(Helper, MIB, MRI);
1059 case G_LSHR:
1060 return selectShift(ARM_AM::ShiftOpc::lsr, MIB);
1061 case G_ASHR:
1062 return selectShift(ARM_AM::ShiftOpc::asr, MIB);
1063 case G_SHL: {
1064 return selectShift(ARM_AM::ShiftOpc::lsl, MIB);
1066 case G_PTR_ADD:
1067 I.setDesc(TII.get(Opcodes.ADDrr));
1068 MIB.add(predOps(ARMCC::AL)).add(condCodeOp());
1069 break;
1070 case G_FRAME_INDEX:
1071 // Add 0 to the given frame index and hope it will eventually be folded into
1072 // the user(s).
1073 I.setDesc(TII.get(Opcodes.ADDri));
1074 MIB.addImm(0).add(predOps(ARMCC::AL)).add(condCodeOp());
1075 break;
1076 case G_GLOBAL_VALUE:
1077 return selectGlobal(MIB, MRI);
1078 case G_STORE:
1079 case G_LOAD: {
1080 const auto &MemOp = **I.memoperands_begin();
1081 if (MemOp.isAtomic()) {
1082 LLVM_DEBUG(dbgs() << "Atomic load/store not supported yet\n");
1083 return false;
1086 Register Reg = I.getOperand(0).getReg();
1087 unsigned RegBank = RBI.getRegBank(Reg, MRI, TRI)->getID();
1089 LLT ValTy = MRI.getType(Reg);
1090 const auto ValSize = ValTy.getSizeInBits();
1092 assert((ValSize != 64 || STI.hasVFP2Base()) &&
1093 "Don't know how to load/store 64-bit value without VFP");
1095 const auto NewOpc = selectLoadStoreOpCode(I.getOpcode(), RegBank, ValSize);
1096 if (NewOpc == G_LOAD || NewOpc == G_STORE)
1097 return false;
1099 I.setDesc(TII.get(NewOpc));
1101 if (NewOpc == ARM::LDRH || NewOpc == ARM::STRH)
1102 // LDRH has a funny addressing mode (there's already a FIXME for it).
1103 MIB.addReg(0);
1104 MIB.addImm(0).add(predOps(ARMCC::AL));
1105 break;
1107 case G_MERGE_VALUES: {
1108 if (!selectMergeValues(MIB, TII, MRI, TRI, RBI))
1109 return false;
1110 break;
1112 case G_UNMERGE_VALUES: {
1113 if (!selectUnmergeValues(MIB, TII, MRI, TRI, RBI))
1114 return false;
1115 break;
1117 case G_BRCOND: {
1118 if (!validReg(MRI, I.getOperand(0).getReg(), 1, ARM::GPRRegBankID)) {
1119 LLVM_DEBUG(dbgs() << "Unsupported condition register for G_BRCOND");
1120 return false;
1123 // Set the flags.
1124 auto Test =
1125 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcodes.TSTri))
1126 .addReg(I.getOperand(0).getReg())
1127 .addImm(1)
1128 .add(predOps(ARMCC::AL));
1129 if (!constrainSelectedInstRegOperands(*Test, TII, TRI, RBI))
1130 return false;
1132 // Branch conditionally.
1133 auto Branch =
1134 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcodes.Bcc))
1135 .add(I.getOperand(1))
1136 .add(predOps(ARMCC::NE, ARM::CPSR));
1137 if (!constrainSelectedInstRegOperands(*Branch, TII, TRI, RBI))
1138 return false;
1139 I.eraseFromParent();
1140 return true;
1142 case G_PHI: {
1143 I.setDesc(TII.get(PHI));
1145 Register DstReg = I.getOperand(0).getReg();
1146 const TargetRegisterClass *RC = guessRegClass(DstReg, MRI, TRI, RBI);
1147 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
1148 break;
1151 return true;
1153 default:
1154 return false;
1157 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);