1 //===- X86InstructionSelector.cpp -----------------------------------------===//
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
10 /// X86.
11 /// \todo This should be generated by TableGen.
12 //===----------------------------------------------------------------------===//
13
14 #include "MCTargetDesc/X86BaseInfo.h"
15 #include "X86.h"
16 #include "X86InstrBuilder.h"
17 #include "X86InstrInfo.h"
18 #include "X86RegisterBankInfo.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
23 #include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
24 #include "llvm/CodeGen/GlobalISel/Utils.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineConstantPool.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstr.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineMemOperand.h"
31 #include "llvm/CodeGen/MachineOperand.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/CodeGen/RegisterBank.h"
34 #include "llvm/CodeGen/TargetOpcodes.h"
35 #include "llvm/CodeGen/TargetRegisterInfo.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/InstrTypes.h"
38 #include "llvm/IR/IntrinsicsX86.h"
39 #include "llvm/Support/AtomicOrdering.h"
40 #include "llvm/Support/CodeGen.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/LowLevelTypeImpl.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include <cassert>
47 #include <cstdint>
48 #include <tuple>
49
50 #define DEBUG_TYPE "X86-isel"
51
52 using namespace llvm;
53
54 namespace {
55
56 #define GET_GLOBALISEL_PREDICATE_BITSET
57 #include "X86GenGlobalISel.inc"
58 #undef GET_GLOBALISEL_PREDICATE_BITSET
59
60 class X86InstructionSelector : public InstructionSelector {
61 public:
62 X86InstructionSelector(const X86TargetMachine &TM, const X86Subtarget &STI,
63 const X86RegisterBankInfo &RBI);
64
65 bool select(MachineInstr &I) override;
getName()66 static const char *getName() { return DEBUG_TYPE; }
67
68 private:
69 /// tblgen-erated 'select' implementation, used as the initial selector for
70 /// the patterns that don't require complex C++.
71 bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;
72
73 // TODO: remove after supported by Tablegen-erated instruction selection.
74 unsigned getLoadStoreOp(const LLT &Ty, const RegisterBank &RB, unsigned Opc,
75 Align Alignment) const;
76
77 bool selectLoadStoreOp(MachineInstr &I, MachineRegisterInfo &MRI,
78 MachineFunction &MF) const;
79 bool selectFrameIndexOrGep(MachineInstr &I, MachineRegisterInfo &MRI,
80 MachineFunction &MF) const;
81 bool selectGlobalValue(MachineInstr &I, MachineRegisterInfo &MRI,
82 MachineFunction &MF) const;
83 bool selectConstant(MachineInstr &I, MachineRegisterInfo &MRI,
84 MachineFunction &MF) const;
85 bool selectTruncOrPtrToInt(MachineInstr &I, MachineRegisterInfo &MRI,
86 MachineFunction &MF) const;
87 bool selectZext(MachineInstr &I, MachineRegisterInfo &MRI,
88 MachineFunction &MF) const;
89 bool selectAnyext(MachineInstr &I, MachineRegisterInfo &MRI,
90 MachineFunction &MF) const;
91 bool selectCmp(MachineInstr &I, MachineRegisterInfo &MRI,
92 MachineFunction &MF) const;
93 bool selectFCmp(MachineInstr &I, MachineRegisterInfo &MRI,
94 MachineFunction &MF) const;
95 bool selectUadde(MachineInstr &I, MachineRegisterInfo &MRI,
96 MachineFunction &MF) const;
97 bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const;
98 bool selectUnmergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
99 MachineFunction &MF);
100 bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
101 MachineFunction &MF);
102 bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI,
103 MachineFunction &MF) const;
104 bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI,
105 MachineFunction &MF) const;
106 bool selectCondBranch(MachineInstr &I, MachineRegisterInfo &MRI,
107 MachineFunction &MF) const;
108 bool selectTurnIntoCOPY(MachineInstr &I, MachineRegisterInfo &MRI,
109 const unsigned DstReg,
110 const TargetRegisterClass *DstRC,
111 const unsigned SrcReg,
112 const TargetRegisterClass *SrcRC) const;
113 bool materializeFP(MachineInstr &I, MachineRegisterInfo &MRI,
114 MachineFunction &MF) const;
115 bool selectImplicitDefOrPHI(MachineInstr &I, MachineRegisterInfo &MRI) const;
116 bool selectDivRem(MachineInstr &I, MachineRegisterInfo &MRI,
117 MachineFunction &MF) const;
118 bool selectIntrinsicWSideEffects(MachineInstr &I, MachineRegisterInfo &MRI,
119 MachineFunction &MF) const;
120
121 // emit insert subreg instruction and insert it before MachineInstr &I
122 bool emitInsertSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
123 MachineRegisterInfo &MRI, MachineFunction &MF) const;
124 // emit extract subreg instruction and insert it before MachineInstr &I
125 bool emitExtractSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
126 MachineRegisterInfo &MRI, MachineFunction &MF) const;
127
128 const TargetRegisterClass *getRegClass(LLT Ty, const RegisterBank &RB) const;
129 const TargetRegisterClass *getRegClass(LLT Ty, unsigned Reg,
130 MachineRegisterInfo &MRI) const;
131
132 const X86TargetMachine &TM;
133 const X86Subtarget &STI;
134 const X86InstrInfo &TII;
135 const X86RegisterInfo &TRI;
136 const X86RegisterBankInfo &RBI;
137
138 #define GET_GLOBALISEL_PREDICATES_DECL
139 #include "X86GenGlobalISel.inc"
140 #undef GET_GLOBALISEL_PREDICATES_DECL
141
142 #define GET_GLOBALISEL_TEMPORARIES_DECL
143 #include "X86GenGlobalISel.inc"
144 #undef GET_GLOBALISEL_TEMPORARIES_DECL
145 };
146
147 } // end anonymous namespace
148
149 #define GET_GLOBALISEL_IMPL
150 #include "X86GenGlobalISel.inc"
151 #undef GET_GLOBALISEL_IMPL
152
X86InstructionSelector(const X86TargetMachine & TM,const X86Subtarget & STI,const X86RegisterBankInfo & RBI)153 X86InstructionSelector::X86InstructionSelector(const X86TargetMachine &TM,
154 const X86Subtarget &STI,
155 const X86RegisterBankInfo &RBI)
156 : TM(TM), STI(STI), TII(*STI.getInstrInfo()), TRI(*STI.getRegisterInfo()),
157 RBI(RBI),
158 #define GET_GLOBALISEL_PREDICATES_INIT
159 #include "X86GenGlobalISel.inc"
160 #undef GET_GLOBALISEL_PREDICATES_INIT
161 #define GET_GLOBALISEL_TEMPORARIES_INIT
162 #include "X86GenGlobalISel.inc"
163 #undef GET_GLOBALISEL_TEMPORARIES_INIT
164 {
165 }
166
167 // FIXME: This should be target-independent, inferred from the types declared
168 // for each class in the bank.
169 const TargetRegisterClass *
getRegClass(LLT Ty,const RegisterBank & RB) const170 X86InstructionSelector::getRegClass(LLT Ty, const RegisterBank &RB) const {
171 if (RB.getID() == X86::GPRRegBankID) {
172 if (Ty.getSizeInBits() <= 8)
173 return &X86::GR8RegClass;
174 if (Ty.getSizeInBits() == 16)
175 return &X86::GR16RegClass;
176 if (Ty.getSizeInBits() == 32)
177 return &X86::GR32RegClass;
178 if (Ty.getSizeInBits() == 64)
179 return &X86::GR64RegClass;
180 }
181 if (RB.getID() == X86::VECRRegBankID) {
182 if (Ty.getSizeInBits() == 16)
183 return STI.hasAVX512() ? &X86::FR16XRegClass : &X86::FR16RegClass;
184 if (Ty.getSizeInBits() == 32)
185 return STI.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass;
186 if (Ty.getSizeInBits() == 64)
187 return STI.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass;
188 if (Ty.getSizeInBits() == 128)
189 return STI.hasAVX512() ? &X86::VR128XRegClass : &X86::VR128RegClass;
190 if (Ty.getSizeInBits() == 256)
191 return STI.hasAVX512() ? &X86::VR256XRegClass : &X86::VR256RegClass;
192 if (Ty.getSizeInBits() == 512)
193 return &X86::VR512RegClass;
194 }
195
196 llvm_unreachable("Unknown RegBank!");
197 }
198
199 const TargetRegisterClass *
getRegClass(LLT Ty,unsigned Reg,MachineRegisterInfo & MRI) const200 X86InstructionSelector::getRegClass(LLT Ty, unsigned Reg,
201 MachineRegisterInfo &MRI) const {
202 const RegisterBank &RegBank = *RBI.getRegBank(Reg, MRI, TRI);
203 return getRegClass(Ty, RegBank);
204 }
205
getSubRegIndex(const TargetRegisterClass * RC)206 static unsigned getSubRegIndex(const TargetRegisterClass *RC) {
207 unsigned SubIdx = X86::NoSubRegister;
208 if (RC == &X86::GR32RegClass) {
209 SubIdx = X86::sub_32bit;
210 } else if (RC == &X86::GR16RegClass) {
211 SubIdx = X86::sub_16bit;
212 } else if (RC == &X86::GR8RegClass) {
213 SubIdx = X86::sub_8bit;
214 }
215
216 return SubIdx;
217 }
218
getRegClassFromGRPhysReg(Register Reg)219 static const TargetRegisterClass *getRegClassFromGRPhysReg(Register Reg) {
220 assert(Reg.isPhysical());
221 if (X86::GR64RegClass.contains(Reg))
222 return &X86::GR64RegClass;
223 if (X86::GR32RegClass.contains(Reg))
224 return &X86::GR32RegClass;
225 if (X86::GR16RegClass.contains(Reg))
226 return &X86::GR16RegClass;
227 if (X86::GR8RegClass.contains(Reg))
228 return &X86::GR8RegClass;
229
230 llvm_unreachable("Unknown RegClass for PhysReg!");
231 }
232
233 // Set X86 Opcode and constrain DestReg.
selectCopy(MachineInstr & I,MachineRegisterInfo & MRI) const234 bool X86InstructionSelector::selectCopy(MachineInstr &I,
235 MachineRegisterInfo &MRI) const {
236 Register DstReg = I.getOperand(0).getReg();
237 const unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI);
238 const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);
239
240 Register SrcReg = I.getOperand(1).getReg();
241 const unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI);
242 const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
243
244 if (DstReg.isPhysical()) {
245 assert(I.isCopy() && "Generic operators do not allow physical registers");
246
247 if (DstSize > SrcSize && SrcRegBank.getID() == X86::GPRRegBankID &&
248 DstRegBank.getID() == X86::GPRRegBankID) {
249
250 const TargetRegisterClass *SrcRC =
251 getRegClass(MRI.getType(SrcReg), SrcRegBank);
252 const TargetRegisterClass *DstRC = getRegClassFromGRPhysReg(DstReg);
253
254 if (SrcRC != DstRC) {
255 // This case can be generated by ABI lowering, performe anyext
256 Register ExtSrc = MRI.createVirtualRegister(DstRC);
257 BuildMI(*I.getParent(), I, I.getDebugLoc(),
258 TII.get(TargetOpcode::SUBREG_TO_REG))
259 .addDef(ExtSrc)
260 .addImm(0)
261 .addReg(SrcReg)
262 .addImm(getSubRegIndex(SrcRC));
263
264 I.getOperand(1).setReg(ExtSrc);
265 }
266 }
267
268 return true;
269 }
270
271 assert((!SrcReg.isPhysical() || I.isCopy()) &&
272 "No phys reg on generic operators");
273 assert((DstSize == SrcSize ||
274 // Copies are a mean to setup initial types, the number of
275 // bits may not exactly match.
276 (SrcReg.isPhysical() &&
277 DstSize <= RBI.getSizeInBits(SrcReg, MRI, TRI))) &&
278 "Copy with different width?!");
279
280 const TargetRegisterClass *DstRC =
281 getRegClass(MRI.getType(DstReg), DstRegBank);
282
283 if (SrcRegBank.getID() == X86::GPRRegBankID &&
284 DstRegBank.getID() == X86::GPRRegBankID && SrcSize > DstSize &&
285 SrcReg.isPhysical()) {
286 // Change the physical register to performe truncate.
287
288 const TargetRegisterClass *SrcRC = getRegClassFromGRPhysReg(SrcReg);
289
290 if (DstRC != SrcRC) {
291 I.getOperand(1).setSubReg(getSubRegIndex(DstRC));
292 I.getOperand(1).substPhysReg(SrcReg, TRI);
293 }
294 }
295
296 // No need to constrain SrcReg. It will get constrained when
297 // we hit another of its use or its defs.
298 // Copies do not have constraints.
299 const TargetRegisterClass *OldRC = MRI.getRegClassOrNull(DstReg);
300 if (!OldRC || !DstRC->hasSubClassEq(OldRC)) {
301 if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
302 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
303 << " operand\n");
304 return false;
305 }
306 }
307 I.setDesc(TII.get(X86::COPY));
308 return true;
309 }
310
select(MachineInstr & I)311 bool X86InstructionSelector::select(MachineInstr &I) {
312 assert(I.getParent() && "Instruction should be in a basic block!");
313 assert(I.getParent()->getParent() && "Instruction should be in a function!");
314
315 MachineBasicBlock &MBB = *I.getParent();
316 MachineFunction &MF = *MBB.getParent();
317 MachineRegisterInfo &MRI = MF.getRegInfo();
318
319 unsigned Opcode = I.getOpcode();
320 if (!isPreISelGenericOpcode(Opcode)) {
321 // Certain non-generic instructions also need some special handling.
322
323 if (Opcode == TargetOpcode::LOAD_STACK_GUARD)
324 return false;
325
326 if (I.isCopy())
327 return selectCopy(I, MRI);
328
329 return true;
330 }
331
332 assert(I.getNumOperands() == I.getNumExplicitOperands() &&
333 "Generic instruction has unexpected implicit operands\n");
334
335 if (selectImpl(I, *CoverageInfo))
336 return true;
337
338 LLVM_DEBUG(dbgs() << " C++ instruction selection: "; I.print(dbgs()));
339
340 // TODO: This should be implemented by tblgen.
341 switch (I.getOpcode()) {
342 default:
343 return false;
344 case TargetOpcode::G_STORE:
345 case TargetOpcode::G_LOAD:
346 return selectLoadStoreOp(I, MRI, MF);
347 case TargetOpcode::G_PTR_ADD:
348 case TargetOpcode::G_FRAME_INDEX:
349 return selectFrameIndexOrGep(I, MRI, MF);
350 case TargetOpcode::G_GLOBAL_VALUE:
351 return selectGlobalValue(I, MRI, MF);
352 case TargetOpcode::G_CONSTANT:
353 return selectConstant(I, MRI, MF);
354 case TargetOpcode::G_FCONSTANT:
355 return materializeFP(I, MRI, MF);
356 case TargetOpcode::G_PTRTOINT:
357 case TargetOpcode::G_TRUNC:
358 return selectTruncOrPtrToInt(I, MRI, MF);
359 case TargetOpcode::G_INTTOPTR:
360 return selectCopy(I, MRI);
361 case TargetOpcode::G_ZEXT:
362 return selectZext(I, MRI, MF);
363 case TargetOpcode::G_ANYEXT:
364 return selectAnyext(I, MRI, MF);
365 case TargetOpcode::G_ICMP:
366 return selectCmp(I, MRI, MF);
367 case TargetOpcode::G_FCMP:
368 return selectFCmp(I, MRI, MF);
369 case TargetOpcode::G_UADDE:
370 return selectUadde(I, MRI, MF);
371 case TargetOpcode::G_UNMERGE_VALUES:
372 return selectUnmergeValues(I, MRI, MF);
373 case TargetOpcode::G_MERGE_VALUES:
374 case TargetOpcode::G_CONCAT_VECTORS:
375 return selectMergeValues(I, MRI, MF);
376 case TargetOpcode::G_EXTRACT:
377 return selectExtract(I, MRI, MF);
378 case TargetOpcode::G_INSERT:
379 return selectInsert(I, MRI, MF);
380 case TargetOpcode::G_BRCOND:
381 return selectCondBranch(I, MRI, MF);
382 case TargetOpcode::G_IMPLICIT_DEF:
383 case TargetOpcode::G_PHI:
384 return selectImplicitDefOrPHI(I, MRI);
385 case TargetOpcode::G_SDIV:
386 case TargetOpcode::G_UDIV:
387 case TargetOpcode::G_SREM:
388 case TargetOpcode::G_UREM:
389 return selectDivRem(I, MRI, MF);
390 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
391 return selectIntrinsicWSideEffects(I, MRI, MF);
392 }
393
394 return false;
395 }
396
getLoadStoreOp(const LLT & Ty,const RegisterBank & RB,unsigned Opc,Align Alignment) const397 unsigned X86InstructionSelector::getLoadStoreOp(const LLT &Ty,
398 const RegisterBank &RB,
399 unsigned Opc,
400 Align Alignment) const {
401 bool Isload = (Opc == TargetOpcode::G_LOAD);
402 bool HasAVX = STI.hasAVX();
403 bool HasAVX512 = STI.hasAVX512();
404 bool HasVLX = STI.hasVLX();
405
406 if (Ty == LLT::scalar(8)) {
407 if (X86::GPRRegBankID == RB.getID())
408 return Isload ? X86::MOV8rm : X86::MOV8mr;
409 } else if (Ty == LLT::scalar(16)) {
410 if (X86::GPRRegBankID == RB.getID())
411 return Isload ? X86::MOV16rm : X86::MOV16mr;
412 } else if (Ty == LLT::scalar(32) || Ty == LLT::pointer(0, 32)) {
413 if (X86::GPRRegBankID == RB.getID())
414 return Isload ? X86::MOV32rm : X86::MOV32mr;
415 if (X86::VECRRegBankID == RB.getID())
416 return Isload ? (HasAVX512 ? X86::VMOVSSZrm_alt :
417 HasAVX ? X86::VMOVSSrm_alt :
418 X86::MOVSSrm_alt)
419 : (HasAVX512 ? X86::VMOVSSZmr :
420 HasAVX ? X86::VMOVSSmr :
421 X86::MOVSSmr);
422 } else if (Ty == LLT::scalar(64) || Ty == LLT::pointer(0, 64)) {
423 if (X86::GPRRegBankID == RB.getID())
424 return Isload ? X86::MOV64rm : X86::MOV64mr;
425 if (X86::VECRRegBankID == RB.getID())
426 return Isload ? (HasAVX512 ? X86::VMOVSDZrm_alt :
427 HasAVX ? X86::VMOVSDrm_alt :
428 X86::MOVSDrm_alt)
429 : (HasAVX512 ? X86::VMOVSDZmr :
430 HasAVX ? X86::VMOVSDmr :
431 X86::MOVSDmr);
432 } else if (Ty.isVector() && Ty.getSizeInBits() == 128) {
433 if (Alignment >= Align(16))
434 return Isload ? (HasVLX ? X86::VMOVAPSZ128rm
435 : HasAVX512
436 ? X86::VMOVAPSZ128rm_NOVLX
437 : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm)
438 : (HasVLX ? X86::VMOVAPSZ128mr
439 : HasAVX512
440 ? X86::VMOVAPSZ128mr_NOVLX
441 : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
442 else
443 return Isload ? (HasVLX ? X86::VMOVUPSZ128rm
444 : HasAVX512
445 ? X86::VMOVUPSZ128rm_NOVLX
446 : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm)
447 : (HasVLX ? X86::VMOVUPSZ128mr
448 : HasAVX512
449 ? X86::VMOVUPSZ128mr_NOVLX
450 : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
451 } else if (Ty.isVector() && Ty.getSizeInBits() == 256) {
452 if (Alignment >= Align(32))
453 return Isload ? (HasVLX ? X86::VMOVAPSZ256rm
454 : HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX
455 : X86::VMOVAPSYrm)
456 : (HasVLX ? X86::VMOVAPSZ256mr
457 : HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX
458 : X86::VMOVAPSYmr);
459 else
460 return Isload ? (HasVLX ? X86::VMOVUPSZ256rm
461 : HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX
462 : X86::VMOVUPSYrm)
463 : (HasVLX ? X86::VMOVUPSZ256mr
464 : HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX
465 : X86::VMOVUPSYmr);
466 } else if (Ty.isVector() && Ty.getSizeInBits() == 512) {
467 if (Alignment >= Align(64))
468 return Isload ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
469 else
470 return Isload ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
471 }
472 return Opc;
473 }
474
475 // Fill in an address from the given instruction.
X86SelectAddress(const MachineInstr & I,const MachineRegisterInfo & MRI,X86AddressMode & AM)476 static void X86SelectAddress(const MachineInstr &I,
477 const MachineRegisterInfo &MRI,
478 X86AddressMode &AM) {
479 assert(I.getOperand(0).isReg() && "unsupported opperand.");
480 assert(MRI.getType(I.getOperand(0).getReg()).isPointer() &&
481 "unsupported type.");
482
483 if (I.getOpcode() == TargetOpcode::G_PTR_ADD) {
484 if (auto COff = getIConstantVRegSExtVal(I.getOperand(2).getReg(), MRI)) {
485 int64_t Imm = *COff;
486 if (isInt<32>(Imm)) { // Check for displacement overflow.
487 AM.Disp = static_cast<int32_t>(Imm);
488 AM.Base.Reg = I.getOperand(1).getReg();
489 return;
490 }
491 }
492 } else if (I.getOpcode() == TargetOpcode::G_FRAME_INDEX) {
493 AM.Base.FrameIndex = I.getOperand(1).getIndex();
494 AM.BaseType = X86AddressMode::FrameIndexBase;
495 return;
496 }
497
498 // Default behavior.
499 AM.Base.Reg = I.getOperand(0).getReg();
500 }
501
selectLoadStoreOp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const502 bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I,
503 MachineRegisterInfo &MRI,
504 MachineFunction &MF) const {
505 unsigned Opc = I.getOpcode();
506
507 assert((Opc == TargetOpcode::G_STORE || Opc == TargetOpcode::G_LOAD) &&
508 "unexpected instruction");
509
510 const Register DefReg = I.getOperand(0).getReg();
511 LLT Ty = MRI.getType(DefReg);
512 const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI);
513
514 assert(I.hasOneMemOperand());
515 auto &MemOp = **I.memoperands_begin();
516 if (MemOp.isAtomic()) {
517 // Note: for unordered operations, we rely on the fact the appropriate MMO
518 // is already on the instruction we're mutating, and thus we don't need to
519 // make any changes. So long as we select an opcode which is capable of
520 // loading or storing the appropriate size atomically, the rest of the
521 // backend is required to respect the MMO state.
522 if (!MemOp.isUnordered()) {
523 LLVM_DEBUG(dbgs() << "Atomic ordering not supported yet\n");
524 return false;
525 }
526 if (MemOp.getAlign() < Ty.getSizeInBits() / 8) {
527 LLVM_DEBUG(dbgs() << "Unaligned atomics not supported yet\n");
528 return false;
529 }
530 }
531
532 unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlign());
533 if (NewOpc == Opc)
534 return false;
535
536 X86AddressMode AM;
537 X86SelectAddress(*MRI.getVRegDef(I.getOperand(1).getReg()), MRI, AM);
538
539 I.setDesc(TII.get(NewOpc));
540 MachineInstrBuilder MIB(MF, I);
541 if (Opc == TargetOpcode::G_LOAD) {
542 I.removeOperand(1);
543 addFullAddress(MIB, AM);
544 } else {
545 // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL)
546 I.removeOperand(1);
547 I.removeOperand(0);
548 addFullAddress(MIB, AM).addUse(DefReg);
549 }
550 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
551 }
552
getLeaOP(LLT Ty,const X86Subtarget & STI)553 static unsigned getLeaOP(LLT Ty, const X86Subtarget &STI) {
554 if (Ty == LLT::pointer(0, 64))
555 return X86::LEA64r;
556 else if (Ty == LLT::pointer(0, 32))
557 return STI.isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r;
558 else
559 llvm_unreachable("Can't get LEA opcode. Unsupported type.");
560 }
561
selectFrameIndexOrGep(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const562 bool X86InstructionSelector::selectFrameIndexOrGep(MachineInstr &I,
563 MachineRegisterInfo &MRI,
564 MachineFunction &MF) const {
565 unsigned Opc = I.getOpcode();
566
567 assert((Opc == TargetOpcode::G_FRAME_INDEX || Opc == TargetOpcode::G_PTR_ADD) &&
568 "unexpected instruction");
569
570 const Register DefReg = I.getOperand(0).getReg();
571 LLT Ty = MRI.getType(DefReg);
572
573 // Use LEA to calculate frame index and GEP
574 unsigned NewOpc = getLeaOP(Ty, STI);
575 I.setDesc(TII.get(NewOpc));
576 MachineInstrBuilder MIB(MF, I);
577
578 if (Opc == TargetOpcode::G_FRAME_INDEX) {
579 addOffset(MIB, 0);
580 } else {
581 MachineOperand &InxOp = I.getOperand(2);
582 I.addOperand(InxOp); // set IndexReg
583 InxOp.ChangeToImmediate(1); // set Scale
584 MIB.addImm(0).addReg(0);
585 }
586
587 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
588 }
589
selectGlobalValue(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const590 bool X86InstructionSelector::selectGlobalValue(MachineInstr &I,
591 MachineRegisterInfo &MRI,
592 MachineFunction &MF) const {
593 assert((I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE) &&
594 "unexpected instruction");
595
596 auto GV = I.getOperand(1).getGlobal();
597 if (GV->isThreadLocal()) {
598 return false; // TODO: we don't support TLS yet.
599 }
600
601 // Can't handle alternate code models yet.
602 if (TM.getCodeModel() != CodeModel::Small)
603 return false;
604
605 X86AddressMode AM;
606 AM.GV = GV;
607 AM.GVOpFlags = STI.classifyGlobalReference(GV);
608
609 // TODO: The ABI requires an extra load. not supported yet.
610 if (isGlobalStubReference(AM.GVOpFlags))
611 return false;
612
613 // TODO: This reference is relative to the pic base. not supported yet.
614 if (isGlobalRelativeToPICBase(AM.GVOpFlags))
615 return false;
616
617 if (STI.isPICStyleRIPRel()) {
618 // Use rip-relative addressing.
619 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
620 AM.Base.Reg = X86::RIP;
621 }
622
623 const Register DefReg = I.getOperand(0).getReg();
624 LLT Ty = MRI.getType(DefReg);
625 unsigned NewOpc = getLeaOP(Ty, STI);
626
627 I.setDesc(TII.get(NewOpc));
628 MachineInstrBuilder MIB(MF, I);
629
630 I.removeOperand(1);
631 addFullAddress(MIB, AM);
632
633 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
634 }
635
selectConstant(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const636 bool X86InstructionSelector::selectConstant(MachineInstr &I,
637 MachineRegisterInfo &MRI,
638 MachineFunction &MF) const {
639 assert((I.getOpcode() == TargetOpcode::G_CONSTANT) &&
640 "unexpected instruction");
641
642 const Register DefReg = I.getOperand(0).getReg();
643 LLT Ty = MRI.getType(DefReg);
644
645 if (RBI.getRegBank(DefReg, MRI, TRI)->getID() != X86::GPRRegBankID)
646 return false;
647
648 uint64_t Val = 0;
649 if (I.getOperand(1).isCImm()) {
650 Val = I.getOperand(1).getCImm()->getZExtValue();
651 I.getOperand(1).ChangeToImmediate(Val);
652 } else if (I.getOperand(1).isImm()) {
653 Val = I.getOperand(1).getImm();
654 } else
655 llvm_unreachable("Unsupported operand type.");
656
657 unsigned NewOpc;
658 switch (Ty.getSizeInBits()) {
659 case 8:
660 NewOpc = X86::MOV8ri;
661 break;
662 case 16:
663 NewOpc = X86::MOV16ri;
664 break;
665 case 32:
666 NewOpc = X86::MOV32ri;
667 break;
668 case 64:
669 // TODO: in case isUInt<32>(Val), X86::MOV32ri can be used
670 if (isInt<32>(Val))
671 NewOpc = X86::MOV64ri32;
672 else
673 NewOpc = X86::MOV64ri;
674 break;
675 default:
676 llvm_unreachable("Can't select G_CONSTANT, unsupported type.");
677 }
678
679 I.setDesc(TII.get(NewOpc));
680 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
681 }
682
683 // Helper function for selectTruncOrPtrToInt and selectAnyext.
684 // Returns true if DstRC lives on a floating register class and
685 // SrcRC lives on a 128-bit vector class.
canTurnIntoCOPY(const TargetRegisterClass * DstRC,const TargetRegisterClass * SrcRC)686 static bool canTurnIntoCOPY(const TargetRegisterClass *DstRC,
687 const TargetRegisterClass *SrcRC) {
688 return (DstRC == &X86::FR32RegClass || DstRC == &X86::FR32XRegClass ||
689 DstRC == &X86::FR64RegClass || DstRC == &X86::FR64XRegClass) &&
690 (SrcRC == &X86::VR128RegClass || SrcRC == &X86::VR128XRegClass);
691 }
692
selectTurnIntoCOPY(MachineInstr & I,MachineRegisterInfo & MRI,const unsigned DstReg,const TargetRegisterClass * DstRC,const unsigned SrcReg,const TargetRegisterClass * SrcRC) const693 bool X86InstructionSelector::selectTurnIntoCOPY(
694 MachineInstr &I, MachineRegisterInfo &MRI, const unsigned DstReg,
695 const TargetRegisterClass *DstRC, const unsigned SrcReg,
696 const TargetRegisterClass *SrcRC) const {
697
698 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
699 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
700 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
701 << " operand\n");
702 return false;
703 }
704 I.setDesc(TII.get(X86::COPY));
705 return true;
706 }
707
selectTruncOrPtrToInt(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const708 bool X86InstructionSelector::selectTruncOrPtrToInt(MachineInstr &I,
709 MachineRegisterInfo &MRI,
710 MachineFunction &MF) const {
711 assert((I.getOpcode() == TargetOpcode::G_TRUNC ||
712 I.getOpcode() == TargetOpcode::G_PTRTOINT) &&
713 "unexpected instruction");
714
715 const Register DstReg = I.getOperand(0).getReg();
716 const Register SrcReg = I.getOperand(1).getReg();
717
718 const LLT DstTy = MRI.getType(DstReg);
719 const LLT SrcTy = MRI.getType(SrcReg);
720
721 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
722 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);
723
724 if (DstRB.getID() != SrcRB.getID()) {
725 LLVM_DEBUG(dbgs() << TII.getName(I.getOpcode())
726 << " input/output on different banks\n");
727 return false;
728 }
729
730 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
731 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);
732
733 if (!DstRC || !SrcRC)
734 return false;
735
736 // If that's truncation of the value that lives on the vector class and goes
737 // into the floating class, just replace it with copy, as we are able to
738 // select it as a regular move.
739 if (canTurnIntoCOPY(DstRC, SrcRC))
740 return selectTurnIntoCOPY(I, MRI, DstReg, DstRC, SrcReg, SrcRC);
741
742 if (DstRB.getID() != X86::GPRRegBankID)
743 return false;
744
745 unsigned SubIdx;
746 if (DstRC == SrcRC) {
747 // Nothing to be done
748 SubIdx = X86::NoSubRegister;
749 } else if (DstRC == &X86::GR32RegClass) {
750 SubIdx = X86::sub_32bit;
751 } else if (DstRC == &X86::GR16RegClass) {
752 SubIdx = X86::sub_16bit;
753 } else if (DstRC == &X86::GR8RegClass) {
754 SubIdx = X86::sub_8bit;
755 } else {
756 return false;
757 }
758
759 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);
760
761 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
762 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
763 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
764 << "\n");
765 return false;
766 }
767
768 I.getOperand(1).setSubReg(SubIdx);
769
770 I.setDesc(TII.get(X86::COPY));
771 return true;
772 }
773
selectZext(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const774 bool X86InstructionSelector::selectZext(MachineInstr &I,
775 MachineRegisterInfo &MRI,
776 MachineFunction &MF) const {
777 assert((I.getOpcode() == TargetOpcode::G_ZEXT) && "unexpected instruction");
778
779 const Register DstReg = I.getOperand(0).getReg();
780 const Register SrcReg = I.getOperand(1).getReg();
781
782 const LLT DstTy = MRI.getType(DstReg);
783 const LLT SrcTy = MRI.getType(SrcReg);
784
785 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(16)) &&
786 "8=>16 Zext is handled by tablegen");
787 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(32)) &&
788 "8=>32 Zext is handled by tablegen");
789 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(32)) &&
790 "16=>32 Zext is handled by tablegen");
791 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(64)) &&
792 "8=>64 Zext is handled by tablegen");
793 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(64)) &&
794 "16=>64 Zext is handled by tablegen");
795 assert(!(SrcTy == LLT::scalar(32) && DstTy == LLT::scalar(64)) &&
796 "32=>64 Zext is handled by tablegen");
797
798 if (SrcTy != LLT::scalar(1))
799 return false;
800
801 unsigned AndOpc;
802 if (DstTy == LLT::scalar(8))
803 AndOpc = X86::AND8ri;
804 else if (DstTy == LLT::scalar(16))
805 AndOpc = X86::AND16ri8;
806 else if (DstTy == LLT::scalar(32))
807 AndOpc = X86::AND32ri8;
808 else if (DstTy == LLT::scalar(64))
809 AndOpc = X86::AND64ri8;
810 else
811 return false;
812
813 Register DefReg = SrcReg;
814 if (DstTy != LLT::scalar(8)) {
815 Register ImpDefReg =
816 MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
817 BuildMI(*I.getParent(), I, I.getDebugLoc(),
818 TII.get(TargetOpcode::IMPLICIT_DEF), ImpDefReg);
819
820 DefReg = MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
821 BuildMI(*I.getParent(), I, I.getDebugLoc(),
822 TII.get(TargetOpcode::INSERT_SUBREG), DefReg)
823 .addReg(ImpDefReg)
824 .addReg(SrcReg)
825 .addImm(X86::sub_8bit);
826 }
827
828 MachineInstr &AndInst =
829 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(AndOpc), DstReg)
830 .addReg(DefReg)
831 .addImm(1);
832
833 constrainSelectedInstRegOperands(AndInst, TII, TRI, RBI);
834
835 I.eraseFromParent();
836 return true;
837 }
838
selectAnyext(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const839 bool X86InstructionSelector::selectAnyext(MachineInstr &I,
840 MachineRegisterInfo &MRI,
841 MachineFunction &MF) const {
842 assert((I.getOpcode() == TargetOpcode::G_ANYEXT) && "unexpected instruction");
843
844 const Register DstReg = I.getOperand(0).getReg();
845 const Register SrcReg = I.getOperand(1).getReg();
846
847 const LLT DstTy = MRI.getType(DstReg);
848 const LLT SrcTy = MRI.getType(SrcReg);
849
850 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
851 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);
852
853 assert(DstRB.getID() == SrcRB.getID() &&
854 "G_ANYEXT input/output on different banks\n");
855
856 assert(DstTy.getSizeInBits() > SrcTy.getSizeInBits() &&
857 "G_ANYEXT incorrect operand size");
858
859 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
860 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);
861
862 // If that's ANY_EXT of the value that lives on the floating class and goes
863 // into the vector class, just replace it with copy, as we are able to select
864 // it as a regular move.
865 if (canTurnIntoCOPY(SrcRC, DstRC))
866 return selectTurnIntoCOPY(I, MRI, SrcReg, SrcRC, DstReg, DstRC);
867
868 if (DstRB.getID() != X86::GPRRegBankID)
869 return false;
870
871 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
872 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
873 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
874 << " operand\n");
875 return false;
876 }
877
878 if (SrcRC == DstRC) {
879 I.setDesc(TII.get(X86::COPY));
880 return true;
881 }
882
883 BuildMI(*I.getParent(), I, I.getDebugLoc(),
884 TII.get(TargetOpcode::SUBREG_TO_REG))
885 .addDef(DstReg)
886 .addImm(0)
887 .addReg(SrcReg)
888 .addImm(getSubRegIndex(SrcRC));
889
890 I.eraseFromParent();
891 return true;
892 }
893
selectCmp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const894 bool X86InstructionSelector::selectCmp(MachineInstr &I,
895 MachineRegisterInfo &MRI,
896 MachineFunction &MF) const {
897 assert((I.getOpcode() == TargetOpcode::G_ICMP) && "unexpected instruction");
898
899 X86::CondCode CC;
900 bool SwapArgs;
901 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(
902 (CmpInst::Predicate)I.getOperand(1).getPredicate());
903
904 Register LHS = I.getOperand(2).getReg();
905 Register RHS = I.getOperand(3).getReg();
906
907 if (SwapArgs)
908 std::swap(LHS, RHS);
909
910 unsigned OpCmp;
911 LLT Ty = MRI.getType(LHS);
912
913 switch (Ty.getSizeInBits()) {
914 default:
915 return false;
916 case 8:
917 OpCmp = X86::CMP8rr;
918 break;
919 case 16:
920 OpCmp = X86::CMP16rr;
921 break;
922 case 32:
923 OpCmp = X86::CMP32rr;
924 break;
925 case 64:
926 OpCmp = X86::CMP64rr;
927 break;
928 }
929
930 MachineInstr &CmpInst =
931 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
932 .addReg(LHS)
933 .addReg(RHS);
934
935 MachineInstr &SetInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
936 TII.get(X86::SETCCr), I.getOperand(0).getReg()).addImm(CC);
937
938 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
939 constrainSelectedInstRegOperands(SetInst, TII, TRI, RBI);
940
941 I.eraseFromParent();
942 return true;
943 }
944
selectFCmp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const945 bool X86InstructionSelector::selectFCmp(MachineInstr &I,
946 MachineRegisterInfo &MRI,
947 MachineFunction &MF) const {
948 assert((I.getOpcode() == TargetOpcode::G_FCMP) && "unexpected instruction");
949
950 Register LhsReg = I.getOperand(2).getReg();
951 Register RhsReg = I.getOperand(3).getReg();
952 CmpInst::Predicate Predicate =
953 (CmpInst::Predicate)I.getOperand(1).getPredicate();
954
955 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
956 static const uint16_t SETFOpcTable[2][3] = {
957 {X86::COND_E, X86::COND_NP, X86::AND8rr},
958 {X86::COND_NE, X86::COND_P, X86::OR8rr}};
959 const uint16_t *SETFOpc = nullptr;
960 switch (Predicate) {
961 default:
962 break;
963 case CmpInst::FCMP_OEQ:
964 SETFOpc = &SETFOpcTable[0][0];
965 break;
966 case CmpInst::FCMP_UNE:
967 SETFOpc = &SETFOpcTable[1][0];
968 break;
969 }
970
971 // Compute the opcode for the CMP instruction.
972 unsigned OpCmp;
973 LLT Ty = MRI.getType(LhsReg);
974 switch (Ty.getSizeInBits()) {
975 default:
976 return false;
977 case 32:
978 OpCmp = X86::UCOMISSrr;
979 break;
980 case 64:
981 OpCmp = X86::UCOMISDrr;
982 break;
983 }
984
985 Register ResultReg = I.getOperand(0).getReg();
986 RBI.constrainGenericRegister(
987 ResultReg,
988 *getRegClass(LLT::scalar(8), *RBI.getRegBank(ResultReg, MRI, TRI)), MRI);
989 if (SETFOpc) {
990 MachineInstr &CmpInst =
991 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
992 .addReg(LhsReg)
993 .addReg(RhsReg);
994
995 Register FlagReg1 = MRI.createVirtualRegister(&X86::GR8RegClass);
996 Register FlagReg2 = MRI.createVirtualRegister(&X86::GR8RegClass);
997 MachineInstr &Set1 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
998 TII.get(X86::SETCCr), FlagReg1).addImm(SETFOpc[0]);
999 MachineInstr &Set2 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1000 TII.get(X86::SETCCr), FlagReg2).addImm(SETFOpc[1]);
1001 MachineInstr &Set3 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1002 TII.get(SETFOpc[2]), ResultReg)
1003 .addReg(FlagReg1)
1004 .addReg(FlagReg2);
1005 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
1006 constrainSelectedInstRegOperands(Set1, TII, TRI, RBI);
1007 constrainSelectedInstRegOperands(Set2, TII, TRI, RBI);
1008 constrainSelectedInstRegOperands(Set3, TII, TRI, RBI);
1009
1010 I.eraseFromParent();
1011 return true;
1012 }
1013
1014 X86::CondCode CC;
1015 bool SwapArgs;
1016 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate);
1017 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");
1018
1019 if (SwapArgs)
1020 std::swap(LhsReg, RhsReg);
1021
1022 // Emit a compare of LHS/RHS.
1023 MachineInstr &CmpInst =
1024 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
1025 .addReg(LhsReg)
1026 .addReg(RhsReg);
1027
1028 MachineInstr &Set =
1029 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SETCCr), ResultReg).addImm(CC);
1030 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
1031 constrainSelectedInstRegOperands(Set, TII, TRI, RBI);
1032 I.eraseFromParent();
1033 return true;
1034 }
1035
selectUadde(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1036 bool X86InstructionSelector::selectUadde(MachineInstr &I,
1037 MachineRegisterInfo &MRI,
1038 MachineFunction &MF) const {
1039 assert((I.getOpcode() == TargetOpcode::G_UADDE) && "unexpected instruction");
1040
1041 const Register DstReg = I.getOperand(0).getReg();
1042 const Register CarryOutReg = I.getOperand(1).getReg();
1043 const Register Op0Reg = I.getOperand(2).getReg();
1044 const Register Op1Reg = I.getOperand(3).getReg();
1045 Register CarryInReg = I.getOperand(4).getReg();
1046
1047 const LLT DstTy = MRI.getType(DstReg);
1048
1049 if (DstTy != LLT::scalar(32))
1050 return false;
1051
1052 // find CarryIn def instruction.
1053 MachineInstr *Def = MRI.getVRegDef(CarryInReg);
1054 while (Def->getOpcode() == TargetOpcode::G_TRUNC) {
1055 CarryInReg = Def->getOperand(1).getReg();
1056 Def = MRI.getVRegDef(CarryInReg);
1057 }
1058
1059 unsigned Opcode;
1060 if (Def->getOpcode() == TargetOpcode::G_UADDE) {
1061 // carry set by prev ADD.
1062
1063 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), X86::EFLAGS)
1064 .addReg(CarryInReg);
1065
1066 if (!RBI.constrainGenericRegister(CarryInReg, X86::GR32RegClass, MRI))
1067 return false;
1068
1069 Opcode = X86::ADC32rr;
1070 } else if (auto val = getIConstantVRegVal(CarryInReg, MRI)) {
1071 // carry is constant, support only 0.
1072 if (*val != 0)
1073 return false;
1074
1075 Opcode = X86::ADD32rr;
1076 } else
1077 return false;
1078
1079 MachineInstr &AddInst =
1080 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg)
1081 .addReg(Op0Reg)
1082 .addReg(Op1Reg);
1083
1084 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), CarryOutReg)
1085 .addReg(X86::EFLAGS);
1086
1087 if (!constrainSelectedInstRegOperands(AddInst, TII, TRI, RBI) ||
1088 !RBI.constrainGenericRegister(CarryOutReg, X86::GR32RegClass, MRI))
1089 return false;
1090
1091 I.eraseFromParent();
1092 return true;
1093 }
1094
selectExtract(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1095 bool X86InstructionSelector::selectExtract(MachineInstr &I,
1096 MachineRegisterInfo &MRI,
1097 MachineFunction &MF) const {
1098 assert((I.getOpcode() == TargetOpcode::G_EXTRACT) &&
1099 "unexpected instruction");
1100
1101 const Register DstReg = I.getOperand(0).getReg();
1102 const Register SrcReg = I.getOperand(1).getReg();
1103 int64_t Index = I.getOperand(2).getImm();
1104
1105 const LLT DstTy = MRI.getType(DstReg);
1106 const LLT SrcTy = MRI.getType(SrcReg);
1107
1108 // Meanwile handle vector type only.
1109 if (!DstTy.isVector())
1110 return false;
1111
1112 if (Index % DstTy.getSizeInBits() != 0)
1113 return false; // Not extract subvector.
1114
1115 if (Index == 0) {
1116 // Replace by extract subreg copy.
1117 if (!emitExtractSubreg(DstReg, SrcReg, I, MRI, MF))
1118 return false;
1119
1120 I.eraseFromParent();
1121 return true;
1122 }
1123
1124 bool HasAVX = STI.hasAVX();
1125 bool HasAVX512 = STI.hasAVX512();
1126 bool HasVLX = STI.hasVLX();
1127
1128 if (SrcTy.getSizeInBits() == 256 && DstTy.getSizeInBits() == 128) {
1129 if (HasVLX)
1130 I.setDesc(TII.get(X86::VEXTRACTF32x4Z256rr));
1131 else if (HasAVX)
1132 I.setDesc(TII.get(X86::VEXTRACTF128rr));
1133 else
1134 return false;
1135 } else if (SrcTy.getSizeInBits() == 512 && HasAVX512) {
1136 if (DstTy.getSizeInBits() == 128)
1137 I.setDesc(TII.get(X86::VEXTRACTF32x4Zrr));
1138 else if (DstTy.getSizeInBits() == 256)
1139 I.setDesc(TII.get(X86::VEXTRACTF64x4Zrr));
1140 else
1141 return false;
1142 } else
1143 return false;
1144
1145 // Convert to X86 VEXTRACT immediate.
1146 Index = Index / DstTy.getSizeInBits();
1147 I.getOperand(2).setImm(Index);
1148
1149 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
1150 }
1151
emitExtractSubreg(unsigned DstReg,unsigned SrcReg,MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1152 bool X86InstructionSelector::emitExtractSubreg(unsigned DstReg, unsigned SrcReg,
1153 MachineInstr &I,
1154 MachineRegisterInfo &MRI,
1155 MachineFunction &MF) const {
1156 const LLT DstTy = MRI.getType(DstReg);
1157 const LLT SrcTy = MRI.getType(SrcReg);
1158 unsigned SubIdx = X86::NoSubRegister;
1159
1160 if (!DstTy.isVector() || !SrcTy.isVector())
1161 return false;
1162
1163 assert(SrcTy.getSizeInBits() > DstTy.getSizeInBits() &&
1164 "Incorrect Src/Dst register size");
1165
1166 if (DstTy.getSizeInBits() == 128)
1167 SubIdx = X86::sub_xmm;
1168 else if (DstTy.getSizeInBits() == 256)
1169 SubIdx = X86::sub_ymm;
1170 else
1171 return false;
1172
1173 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);
1174 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);
1175
1176 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);
1177
1178 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
1179 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
1180 LLVM_DEBUG(dbgs() << "Failed to constrain EXTRACT_SUBREG\n");
1181 return false;
1182 }
1183
1184 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), DstReg)
1185 .addReg(SrcReg, 0, SubIdx);
1186
1187 return true;
1188 }
1189
emitInsertSubreg(unsigned DstReg,unsigned SrcReg,MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1190 bool X86InstructionSelector::emitInsertSubreg(unsigned DstReg, unsigned SrcReg,
1191 MachineInstr &I,
1192 MachineRegisterInfo &MRI,
1193 MachineFunction &MF) const {
1194 const LLT DstTy = MRI.getType(DstReg);
1195 const LLT SrcTy = MRI.getType(SrcReg);
1196 unsigned SubIdx = X86::NoSubRegister;
1197
1198 // TODO: support scalar types
1199 if (!DstTy.isVector() || !SrcTy.isVector())
1200 return false;
1201
1202 assert(SrcTy.getSizeInBits() < DstTy.getSizeInBits() &&
1203 "Incorrect Src/Dst register size");
1204
1205 if (SrcTy.getSizeInBits() == 128)
1206 SubIdx = X86::sub_xmm;
1207 else if (SrcTy.getSizeInBits() == 256)
1208 SubIdx = X86::sub_ymm;
1209 else
1210 return false;
1211
1212 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);
1213 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);
1214
1215 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
1216 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
1217 LLVM_DEBUG(dbgs() << "Failed to constrain INSERT_SUBREG\n");
1218 return false;
1219 }
1220
1221 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY))
1222 .addReg(DstReg, RegState::DefineNoRead, SubIdx)
1223 .addReg(SrcReg);
1224
1225 return true;
1226 }
1227
selectInsert(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1228 bool X86InstructionSelector::selectInsert(MachineInstr &I,
1229 MachineRegisterInfo &MRI,
1230 MachineFunction &MF) const {
1231 assert((I.getOpcode() == TargetOpcode::G_INSERT) && "unexpected instruction");
1232
1233 const Register DstReg = I.getOperand(0).getReg();
1234 const Register SrcReg = I.getOperand(1).getReg();
1235 const Register InsertReg = I.getOperand(2).getReg();
1236 int64_t Index = I.getOperand(3).getImm();
1237
1238 const LLT DstTy = MRI.getType(DstReg);
1239 const LLT InsertRegTy = MRI.getType(InsertReg);
1240
1241 // Meanwile handle vector type only.
1242 if (!DstTy.isVector())
1243 return false;
1244
1245 if (Index % InsertRegTy.getSizeInBits() != 0)
1246 return false; // Not insert subvector.
1247
1248 if (Index == 0 && MRI.getVRegDef(SrcReg)->isImplicitDef()) {
1249 // Replace by subreg copy.
1250 if (!emitInsertSubreg(DstReg, InsertReg, I, MRI, MF))
1251 return false;
1252
1253 I.eraseFromParent();
1254 return true;
1255 }
1256
1257 bool HasAVX = STI.hasAVX();
1258 bool HasAVX512 = STI.hasAVX512();
1259 bool HasVLX = STI.hasVLX();
1260
1261 if (DstTy.getSizeInBits() == 256 && InsertRegTy.getSizeInBits() == 128) {
1262 if (HasVLX)
1263 I.setDesc(TII.get(X86::VINSERTF32x4Z256rr));
1264 else if (HasAVX)
1265 I.setDesc(TII.get(X86::VINSERTF128rr));
1266 else
1267 return false;
1268 } else if (DstTy.getSizeInBits() == 512 && HasAVX512) {
1269 if (InsertRegTy.getSizeInBits() == 128)
1270 I.setDesc(TII.get(X86::VINSERTF32x4Zrr));
1271 else if (InsertRegTy.getSizeInBits() == 256)
1272 I.setDesc(TII.get(X86::VINSERTF64x4Zrr));
1273 else
1274 return false;
1275 } else
1276 return false;
1277
1278 // Convert to X86 VINSERT immediate.
1279 Index = Index / InsertRegTy.getSizeInBits();
1280
1281 I.getOperand(3).setImm(Index);
1282
1283 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
1284 }
1285
selectUnmergeValues(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF)1286 bool X86InstructionSelector::selectUnmergeValues(
1287 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
1288 assert((I.getOpcode() == TargetOpcode::G_UNMERGE_VALUES) &&
1289 "unexpected instruction");
1290
1291 // Split to extracts.
1292 unsigned NumDefs = I.getNumOperands() - 1;
1293 Register SrcReg = I.getOperand(NumDefs).getReg();
1294 unsigned DefSize = MRI.getType(I.getOperand(0).getReg()).getSizeInBits();
1295
1296 for (unsigned Idx = 0; Idx < NumDefs; ++Idx) {
1297 MachineInstr &ExtrInst =
1298 *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1299 TII.get(TargetOpcode::G_EXTRACT), I.getOperand(Idx).getReg())
1300 .addReg(SrcReg)
1301 .addImm(Idx * DefSize);
1302
1303 if (!select(ExtrInst))
1304 return false;
1305 }
1306
1307 I.eraseFromParent();
1308 return true;
1309 }
1310
selectMergeValues(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF)1311 bool X86InstructionSelector::selectMergeValues(
1312 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
1313 assert((I.getOpcode() == TargetOpcode::G_MERGE_VALUES ||
1314 I.getOpcode() == TargetOpcode::G_CONCAT_VECTORS) &&
1315 "unexpected instruction");
1316
1317 // Split to inserts.
1318 Register DstReg = I.getOperand(0).getReg();
1319 Register SrcReg0 = I.getOperand(1).getReg();
1320
1321 const LLT DstTy = MRI.getType(DstReg);
1322 const LLT SrcTy = MRI.getType(SrcReg0);
1323 unsigned SrcSize = SrcTy.getSizeInBits();
1324
1325 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
1326
1327 // For the first src use insertSubReg.
1328 Register DefReg = MRI.createGenericVirtualRegister(DstTy);
1329 MRI.setRegBank(DefReg, RegBank);
1330 if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF))
1331 return false;
1332
1333 for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) {
1334 Register Tmp = MRI.createGenericVirtualRegister(DstTy);
1335 MRI.setRegBank(Tmp, RegBank);
1336
1337 MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1338 TII.get(TargetOpcode::G_INSERT), Tmp)
1339 .addReg(DefReg)
1340 .addReg(I.getOperand(Idx).getReg())
1341 .addImm((Idx - 1) * SrcSize);
1342
1343 DefReg = Tmp;
1344
1345 if (!select(InsertInst))
1346 return false;
1347 }
1348
1349 MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1350 TII.get(TargetOpcode::COPY), DstReg)
1351 .addReg(DefReg);
1352
1353 if (!select(CopyInst))
1354 return false;
1355
1356 I.eraseFromParent();
1357 return true;
1358 }
1359
selectCondBranch(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1360 bool X86InstructionSelector::selectCondBranch(MachineInstr &I,
1361 MachineRegisterInfo &MRI,
1362 MachineFunction &MF) const {
1363 assert((I.getOpcode() == TargetOpcode::G_BRCOND) && "unexpected instruction");
1364
1365 const Register CondReg = I.getOperand(0).getReg();
1366 MachineBasicBlock *DestMBB = I.getOperand(1).getMBB();
1367
1368 MachineInstr &TestInst =
1369 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TEST8ri))
1370 .addReg(CondReg)
1371 .addImm(1);
1372 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::JCC_1))
1373 .addMBB(DestMBB).addImm(X86::COND_NE);
1374
1375 constrainSelectedInstRegOperands(TestInst, TII, TRI, RBI);
1376
1377 I.eraseFromParent();
1378 return true;
1379 }
1380
materializeFP(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1381 bool X86InstructionSelector::materializeFP(MachineInstr &I,
1382 MachineRegisterInfo &MRI,
1383 MachineFunction &MF) const {
1384 assert((I.getOpcode() == TargetOpcode::G_FCONSTANT) &&
1385 "unexpected instruction");
1386
1387 // Can't handle alternate code models yet.
1388 CodeModel::Model CM = TM.getCodeModel();
1389 if (CM != CodeModel::Small && CM != CodeModel::Large)
1390 return false;
1391
1392 const Register DstReg = I.getOperand(0).getReg();
1393 const LLT DstTy = MRI.getType(DstReg);
1394 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
1395 Align Alignment = Align(DstTy.getSizeInBytes());
1396 const DebugLoc &DbgLoc = I.getDebugLoc();
1397
1398 unsigned Opc =
1399 getLoadStoreOp(DstTy, RegBank, TargetOpcode::G_LOAD, Alignment);
1400
1401 // Create the load from the constant pool.
1402 const ConstantFP *CFP = I.getOperand(1).getFPImm();
1403 unsigned CPI = MF.getConstantPool()->getConstantPoolIndex(CFP, Alignment);
1404 MachineInstr *LoadInst = nullptr;
1405 unsigned char OpFlag = STI.classifyLocalReference(nullptr);
1406
1407 if (CM == CodeModel::Large && STI.is64Bit()) {
1408 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
1409 // they cannot be folded into immediate fields.
1410
1411 Register AddrReg = MRI.createVirtualRegister(&X86::GR64RegClass);
1412 BuildMI(*I.getParent(), I, DbgLoc, TII.get(X86::MOV64ri), AddrReg)
1413 .addConstantPoolIndex(CPI, 0, OpFlag);
1414
1415 MachineMemOperand *MMO = MF.getMachineMemOperand(
1416 MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
1417 LLT::pointer(0, MF.getDataLayout().getPointerSizeInBits()), Alignment);
1418
1419 LoadInst =
1420 addDirectMem(BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg),
1421 AddrReg)
1422 .addMemOperand(MMO);
1423
1424 } else if (CM == CodeModel::Small || !STI.is64Bit()) {
1425 // Handle the case when globals fit in our immediate field.
1426 // This is true for X86-32 always and X86-64 when in -mcmodel=small mode.
1427
1428 // x86-32 PIC requires a PIC base register for constant pools.
1429 unsigned PICBase = 0;
1430 if (OpFlag == X86II::MO_PIC_BASE_OFFSET || OpFlag == X86II::MO_GOTOFF) {
1431 // PICBase can be allocated by TII.getGlobalBaseReg(&MF).
1432 // In DAGISEL the code that initialize it generated by the CGBR pass.
1433 return false; // TODO support the mode.
1434 } else if (STI.is64Bit() && TM.getCodeModel() == CodeModel::Small)
1435 PICBase = X86::RIP;
1436
1437 LoadInst = addConstantPoolReference(
1438 BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), CPI, PICBase,
1439 OpFlag);
1440 } else
1441 return false;
1442
1443 constrainSelectedInstRegOperands(*LoadInst, TII, TRI, RBI);
1444 I.eraseFromParent();
1445 return true;
1446 }
1447
selectImplicitDefOrPHI(MachineInstr & I,MachineRegisterInfo & MRI) const1448 bool X86InstructionSelector::selectImplicitDefOrPHI(
1449 MachineInstr &I, MachineRegisterInfo &MRI) const {
1450 assert((I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF ||
1451 I.getOpcode() == TargetOpcode::G_PHI) &&
1452 "unexpected instruction");
1453
1454 Register DstReg = I.getOperand(0).getReg();
1455
1456 if (!MRI.getRegClassOrNull(DstReg)) {
1457 const LLT DstTy = MRI.getType(DstReg);
1458 const TargetRegisterClass *RC = getRegClass(DstTy, DstReg, MRI);
1459
1460 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
1461 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
1462 << " operand\n");
1463 return false;
1464 }
1465 }
1466
1467 if (I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
1468 I.setDesc(TII.get(X86::IMPLICIT_DEF));
1469 else
1470 I.setDesc(TII.get(X86::PHI));
1471
1472 return true;
1473 }
1474
selectDivRem(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1475 bool X86InstructionSelector::selectDivRem(MachineInstr &I,
1476 MachineRegisterInfo &MRI,
1477 MachineFunction &MF) const {
1478 // The implementation of this function is taken from X86FastISel.
1479 assert((I.getOpcode() == TargetOpcode::G_SDIV ||
1480 I.getOpcode() == TargetOpcode::G_SREM ||
1481 I.getOpcode() == TargetOpcode::G_UDIV ||
1482 I.getOpcode() == TargetOpcode::G_UREM) &&
1483 "unexpected instruction");
1484
1485 const Register DstReg = I.getOperand(0).getReg();
1486 const Register Op1Reg = I.getOperand(1).getReg();
1487 const Register Op2Reg = I.getOperand(2).getReg();
1488
1489 const LLT RegTy = MRI.getType(DstReg);
1490 assert(RegTy == MRI.getType(Op1Reg) && RegTy == MRI.getType(Op2Reg) &&
1491 "Arguments and return value types must match");
1492
1493 const RegisterBank *RegRB = RBI.getRegBank(DstReg, MRI, TRI);
1494 if (!RegRB || RegRB->getID() != X86::GPRRegBankID)
1495 return false;
1496
1497 const static unsigned NumTypes = 4; // i8, i16, i32, i64
1498 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem
1499 const static bool S = true; // IsSigned
1500 const static bool U = false; // !IsSigned
1501 const static unsigned Copy = TargetOpcode::COPY;
1502 // For the X86 IDIV instruction, in most cases the dividend
1503 // (numerator) must be in a specific register pair highreg:lowreg,
1504 // producing the quotient in lowreg and the remainder in highreg.
1505 // For most data types, to set up the instruction, the dividend is
1506 // copied into lowreg, and lowreg is sign-extended into highreg. The
1507 // exception is i8, where the dividend is defined as a single register rather
1508 // than a register pair, and we therefore directly sign-extend the dividend
1509 // into lowreg, instead of copying, and ignore the highreg.
1510 const static struct DivRemEntry {
1511 // The following portion depends only on the data type.
1512 unsigned SizeInBits;
1513 unsigned LowInReg; // low part of the register pair
1514 unsigned HighInReg; // high part of the register pair
1515 // The following portion depends on both the data type and the operation.
1516 struct DivRemResult {
1517 unsigned OpDivRem; // The specific DIV/IDIV opcode to use.
1518 unsigned OpSignExtend; // Opcode for sign-extending lowreg into
1519 // highreg, or copying a zero into highreg.
1520 unsigned OpCopy; // Opcode for copying dividend into lowreg, or
1521 // zero/sign-extending into lowreg for i8.
1522 unsigned DivRemResultReg; // Register containing the desired result.
1523 bool IsOpSigned; // Whether to use signed or unsigned form.
1524 } ResultTable[NumOps];
1525 } OpTable[NumTypes] = {
1526 {8,
1527 X86::AX,
1528 0,
1529 {
1530 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S}, // SDiv
1531 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S}, // SRem
1532 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U}, // UDiv
1533 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U}, // URem
1534 }}, // i8
1535 {16,
1536 X86::AX,
1537 X86::DX,
1538 {
1539 {X86::IDIV16r, X86::CWD, Copy, X86::AX, S}, // SDiv
1540 {X86::IDIV16r, X86::CWD, Copy, X86::DX, S}, // SRem
1541 {X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U}, // UDiv
1542 {X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U}, // URem
1543 }}, // i16
1544 {32,
1545 X86::EAX,
1546 X86::EDX,
1547 {
1548 {X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S}, // SDiv
1549 {X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S}, // SRem
1550 {X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U}, // UDiv
1551 {X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U}, // URem
1552 }}, // i32
1553 {64,
1554 X86::RAX,
1555 X86::RDX,
1556 {
1557 {X86::IDIV64r, X86::CQO, Copy, X86::RAX, S}, // SDiv
1558 {X86::IDIV64r, X86::CQO, Copy, X86::RDX, S}, // SRem
1559 {X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U}, // UDiv
1560 {X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U}, // URem
1561 }}, // i64
1562 };
1563
1564 auto OpEntryIt = llvm::find_if(OpTable, [RegTy](const DivRemEntry &El) {
1565 return El.SizeInBits == RegTy.getSizeInBits();
1566 });
1567 if (OpEntryIt == std::end(OpTable))
1568 return false;
1569
1570 unsigned OpIndex;
1571 switch (I.getOpcode()) {
1572 default:
1573 llvm_unreachable("Unexpected div/rem opcode");
1574 case TargetOpcode::G_SDIV:
1575 OpIndex = 0;
1576 break;
1577 case TargetOpcode::G_SREM:
1578 OpIndex = 1;
1579 break;
1580 case TargetOpcode::G_UDIV:
1581 OpIndex = 2;
1582 break;
1583 case TargetOpcode::G_UREM:
1584 OpIndex = 3;
1585 break;
1586 }
1587
1588 const DivRemEntry &TypeEntry = *OpEntryIt;
1589 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
1590
1591 const TargetRegisterClass *RegRC = getRegClass(RegTy, *RegRB);
1592 if (!RBI.constrainGenericRegister(Op1Reg, *RegRC, MRI) ||
1593 !RBI.constrainGenericRegister(Op2Reg, *RegRC, MRI) ||
1594 !RBI.constrainGenericRegister(DstReg, *RegRC, MRI)) {
1595 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
1596 << " operand\n");
1597 return false;
1598 }
1599
1600 // Move op1 into low-order input register.
1601 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpCopy),
1602 TypeEntry.LowInReg)
1603 .addReg(Op1Reg);
1604 // Zero-extend or sign-extend into high-order input register.
1605 if (OpEntry.OpSignExtend) {
1606 if (OpEntry.IsOpSigned)
1607 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1608 TII.get(OpEntry.OpSignExtend));
1609 else {
1610 Register Zero32 = MRI.createVirtualRegister(&X86::GR32RegClass);
1611 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::MOV32r0),
1612 Zero32);
1613
1614 // Copy the zero into the appropriate sub/super/identical physical
1615 // register. Unfortunately the operations needed are not uniform enough
1616 // to fit neatly into the table above.
1617 if (RegTy.getSizeInBits() == 16) {
1618 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
1619 TypeEntry.HighInReg)
1620 .addReg(Zero32, 0, X86::sub_16bit);
1621 } else if (RegTy.getSizeInBits() == 32) {
1622 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
1623 TypeEntry.HighInReg)
1624 .addReg(Zero32);
1625 } else if (RegTy.getSizeInBits() == 64) {
1626 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1627 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
1628 .addImm(0)
1629 .addReg(Zero32)
1630 .addImm(X86::sub_32bit);
1631 }
1632 }
1633 }
1634 // Generate the DIV/IDIV instruction.
1635 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpDivRem))
1636 .addReg(Op2Reg);
1637 // For i8 remainder, we can't reference ah directly, as we'll end
1638 // up with bogus copies like %r9b = COPY %ah. Reference ax
1639 // instead to prevent ah references in a rex instruction.
1640 //
1641 // The current assumption of the fast register allocator is that isel
1642 // won't generate explicit references to the GR8_NOREX registers. If
1643 // the allocator and/or the backend get enhanced to be more robust in
1644 // that regard, this can be, and should be, removed.
1645 if ((I.getOpcode() == Instruction::SRem ||
1646 I.getOpcode() == Instruction::URem) &&
1647 OpEntry.DivRemResultReg == X86::AH && STI.is64Bit()) {
1648 Register SourceSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
1649 Register ResultSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
1650 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), SourceSuperReg)
1651 .addReg(X86::AX);
1652
1653 // Shift AX right by 8 bits instead of using AH.
1654 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SHR16ri),
1655 ResultSuperReg)
1656 .addReg(SourceSuperReg)
1657 .addImm(8);
1658
1659 // Now reference the 8-bit subreg of the result.
1660 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1661 TII.get(TargetOpcode::SUBREG_TO_REG))
1662 .addDef(DstReg)
1663 .addImm(0)
1664 .addReg(ResultSuperReg)
1665 .addImm(X86::sub_8bit);
1666 } else {
1667 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY),
1668 DstReg)
1669 .addReg(OpEntry.DivRemResultReg);
1670 }
1671 I.eraseFromParent();
1672 return true;
1673 }
1674
selectIntrinsicWSideEffects(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1675 bool X86InstructionSelector::selectIntrinsicWSideEffects(
1676 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) const {
1677
1678 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS &&
1679 "unexpected instruction");
1680
1681 if (I.getOperand(0).getIntrinsicID() != Intrinsic::trap)
1682 return false;
1683
1684 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TRAP));
1685
1686 I.eraseFromParent();
1687 return true;
1688 }
1689
1690 InstructionSelector *
createX86InstructionSelector(const X86TargetMachine & TM,X86Subtarget & Subtarget,X86RegisterBankInfo & RBI)1691 llvm::createX86InstructionSelector(const X86TargetMachine &TM,
1692 X86Subtarget &Subtarget,
1693 X86RegisterBankInfo &RBI) {
1694 return new X86InstructionSelector(TM, Subtarget, RBI);
1695 }
1696