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 selectDebugInstr(MachineInstr &I, MachineRegisterInfo &MRI) const;
98 bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const;
99 bool selectUnmergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
100 MachineFunction &MF);
101 bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
102 MachineFunction &MF);
103 bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI,
104 MachineFunction &MF) const;
105 bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI,
106 MachineFunction &MF) const;
107 bool selectCondBranch(MachineInstr &I, MachineRegisterInfo &MRI,
108 MachineFunction &MF) const;
109 bool selectTurnIntoCOPY(MachineInstr &I, MachineRegisterInfo &MRI,
110 const unsigned DstReg,
111 const TargetRegisterClass *DstRC,
112 const unsigned SrcReg,
113 const TargetRegisterClass *SrcRC) const;
114 bool materializeFP(MachineInstr &I, MachineRegisterInfo &MRI,
115 MachineFunction &MF) const;
116 bool selectImplicitDefOrPHI(MachineInstr &I, MachineRegisterInfo &MRI) const;
117 bool selectDivRem(MachineInstr &I, MachineRegisterInfo &MRI,
118 MachineFunction &MF) const;
119 bool selectIntrinsicWSideEffects(MachineInstr &I, MachineRegisterInfo &MRI,
120 MachineFunction &MF) const;
121
122 // emit insert subreg instruction and insert it before MachineInstr &I
123 bool emitInsertSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
124 MachineRegisterInfo &MRI, MachineFunction &MF) const;
125 // emit extract subreg instruction and insert it before MachineInstr &I
126 bool emitExtractSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
127 MachineRegisterInfo &MRI, MachineFunction &MF) const;
128
129 const TargetRegisterClass *getRegClass(LLT Ty, const RegisterBank &RB) const;
130 const TargetRegisterClass *getRegClass(LLT Ty, unsigned Reg,
131 MachineRegisterInfo &MRI) const;
132
133 const X86TargetMachine &TM;
134 const X86Subtarget &STI;
135 const X86InstrInfo &TII;
136 const X86RegisterInfo &TRI;
137 const X86RegisterBankInfo &RBI;
138
139 #define GET_GLOBALISEL_PREDICATES_DECL
140 #include "X86GenGlobalISel.inc"
141 #undef GET_GLOBALISEL_PREDICATES_DECL
142
143 #define GET_GLOBALISEL_TEMPORARIES_DECL
144 #include "X86GenGlobalISel.inc"
145 #undef GET_GLOBALISEL_TEMPORARIES_DECL
146 };
147
148 } // end anonymous namespace
149
150 #define GET_GLOBALISEL_IMPL
151 #include "X86GenGlobalISel.inc"
152 #undef GET_GLOBALISEL_IMPL
153
X86InstructionSelector(const X86TargetMachine & TM,const X86Subtarget & STI,const X86RegisterBankInfo & RBI)154 X86InstructionSelector::X86InstructionSelector(const X86TargetMachine &TM,
155 const X86Subtarget &STI,
156 const X86RegisterBankInfo &RBI)
157 : TM(TM), STI(STI), TII(*STI.getInstrInfo()), TRI(*STI.getRegisterInfo()),
158 RBI(RBI),
159 #define GET_GLOBALISEL_PREDICATES_INIT
160 #include "X86GenGlobalISel.inc"
161 #undef GET_GLOBALISEL_PREDICATES_INIT
162 #define GET_GLOBALISEL_TEMPORARIES_INIT
163 #include "X86GenGlobalISel.inc"
164 #undef GET_GLOBALISEL_TEMPORARIES_INIT
165 {
166 }
167
168 // FIXME: This should be target-independent, inferred from the types declared
169 // for each class in the bank.
170 const TargetRegisterClass *
getRegClass(LLT Ty,const RegisterBank & RB) const171 X86InstructionSelector::getRegClass(LLT Ty, const RegisterBank &RB) const {
172 if (RB.getID() == X86::GPRRegBankID) {
173 if (Ty.getSizeInBits() <= 8)
174 return &X86::GR8RegClass;
175 if (Ty.getSizeInBits() == 16)
176 return &X86::GR16RegClass;
177 if (Ty.getSizeInBits() == 32)
178 return &X86::GR32RegClass;
179 if (Ty.getSizeInBits() == 64)
180 return &X86::GR64RegClass;
181 }
182 if (RB.getID() == X86::VECRRegBankID) {
183 if (Ty.getSizeInBits() == 16)
184 return STI.hasAVX512() ? &X86::FR16XRegClass : &X86::FR16RegClass;
185 if (Ty.getSizeInBits() == 32)
186 return STI.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass;
187 if (Ty.getSizeInBits() == 64)
188 return STI.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass;
189 if (Ty.getSizeInBits() == 128)
190 return STI.hasAVX512() ? &X86::VR128XRegClass : &X86::VR128RegClass;
191 if (Ty.getSizeInBits() == 256)
192 return STI.hasAVX512() ? &X86::VR256XRegClass : &X86::VR256RegClass;
193 if (Ty.getSizeInBits() == 512)
194 return &X86::VR512RegClass;
195 }
196
197 llvm_unreachable("Unknown RegBank!");
198 }
199
200 const TargetRegisterClass *
getRegClass(LLT Ty,unsigned Reg,MachineRegisterInfo & MRI) const201 X86InstructionSelector::getRegClass(LLT Ty, unsigned Reg,
202 MachineRegisterInfo &MRI) const {
203 const RegisterBank &RegBank = *RBI.getRegBank(Reg, MRI, TRI);
204 return getRegClass(Ty, RegBank);
205 }
206
getSubRegIndex(const TargetRegisterClass * RC)207 static unsigned getSubRegIndex(const TargetRegisterClass *RC) {
208 unsigned SubIdx = X86::NoSubRegister;
209 if (RC == &X86::GR32RegClass) {
210 SubIdx = X86::sub_32bit;
211 } else if (RC == &X86::GR16RegClass) {
212 SubIdx = X86::sub_16bit;
213 } else if (RC == &X86::GR8RegClass) {
214 SubIdx = X86::sub_8bit;
215 }
216
217 return SubIdx;
218 }
219
getRegClassFromGRPhysReg(Register Reg)220 static const TargetRegisterClass *getRegClassFromGRPhysReg(Register Reg) {
221 assert(Reg.isPhysical());
222 if (X86::GR64RegClass.contains(Reg))
223 return &X86::GR64RegClass;
224 if (X86::GR32RegClass.contains(Reg))
225 return &X86::GR32RegClass;
226 if (X86::GR16RegClass.contains(Reg))
227 return &X86::GR16RegClass;
228 if (X86::GR8RegClass.contains(Reg))
229 return &X86::GR8RegClass;
230
231 llvm_unreachable("Unknown RegClass for PhysReg!");
232 }
233
234 // FIXME: We need some sort of API in RBI/TRI to allow generic code to
235 // constrain operands of simple instructions given a TargetRegisterClass
236 // and LLT
selectDebugInstr(MachineInstr & I,MachineRegisterInfo & MRI) const237 bool X86InstructionSelector::selectDebugInstr(MachineInstr &I,
238 MachineRegisterInfo &MRI) const {
239 for (MachineOperand &MO : I.operands()) {
240 if (!MO.isReg())
241 continue;
242 Register Reg = MO.getReg();
243 if (!Reg)
244 continue;
245 if (Reg.isPhysical())
246 continue;
247 LLT Ty = MRI.getType(Reg);
248 const RegClassOrRegBank &RegClassOrBank = MRI.getRegClassOrRegBank(Reg);
249 const TargetRegisterClass *RC =
250 RegClassOrBank.dyn_cast<const TargetRegisterClass *>();
251 if (!RC) {
252 const RegisterBank &RB = *RegClassOrBank.get<const RegisterBank *>();
253 RC = getRegClass(Ty, RB);
254 if (!RC) {
255 LLVM_DEBUG(
256 dbgs() << "Warning: DBG_VALUE operand has unexpected size/bank\n");
257 break;
258 }
259 }
260 RBI.constrainGenericRegister(Reg, *RC, MRI);
261 }
262
263 return true;
264 }
265
266 // Set X86 Opcode and constrain DestReg.
selectCopy(MachineInstr & I,MachineRegisterInfo & MRI) const267 bool X86InstructionSelector::selectCopy(MachineInstr &I,
268 MachineRegisterInfo &MRI) const {
269 Register DstReg = I.getOperand(0).getReg();
270 const unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI);
271 const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);
272
273 Register SrcReg = I.getOperand(1).getReg();
274 const unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI);
275 const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);
276
277 if (DstReg.isPhysical()) {
278 assert(I.isCopy() && "Generic operators do not allow physical registers");
279
280 if (DstSize > SrcSize && SrcRegBank.getID() == X86::GPRRegBankID &&
281 DstRegBank.getID() == X86::GPRRegBankID) {
282
283 const TargetRegisterClass *SrcRC =
284 getRegClass(MRI.getType(SrcReg), SrcRegBank);
285 const TargetRegisterClass *DstRC = getRegClassFromGRPhysReg(DstReg);
286
287 if (SrcRC != DstRC) {
288 // This case can be generated by ABI lowering, performe anyext
289 Register ExtSrc = MRI.createVirtualRegister(DstRC);
290 BuildMI(*I.getParent(), I, I.getDebugLoc(),
291 TII.get(TargetOpcode::SUBREG_TO_REG))
292 .addDef(ExtSrc)
293 .addImm(0)
294 .addReg(SrcReg)
295 .addImm(getSubRegIndex(SrcRC));
296
297 I.getOperand(1).setReg(ExtSrc);
298 }
299 }
300
301 return true;
302 }
303
304 assert((!SrcReg.isPhysical() || I.isCopy()) &&
305 "No phys reg on generic operators");
306 assert((DstSize == SrcSize ||
307 // Copies are a mean to setup initial types, the number of
308 // bits may not exactly match.
309 (SrcReg.isPhysical() &&
310 DstSize <= RBI.getSizeInBits(SrcReg, MRI, TRI))) &&
311 "Copy with different width?!");
312
313 const TargetRegisterClass *DstRC =
314 getRegClass(MRI.getType(DstReg), DstRegBank);
315
316 if (SrcRegBank.getID() == X86::GPRRegBankID &&
317 DstRegBank.getID() == X86::GPRRegBankID && SrcSize > DstSize &&
318 SrcReg.isPhysical()) {
319 // Change the physical register to performe truncate.
320
321 const TargetRegisterClass *SrcRC = getRegClassFromGRPhysReg(SrcReg);
322
323 if (DstRC != SrcRC) {
324 I.getOperand(1).setSubReg(getSubRegIndex(DstRC));
325 I.getOperand(1).substPhysReg(SrcReg, TRI);
326 }
327 }
328
329 // No need to constrain SrcReg. It will get constrained when
330 // we hit another of its use or its defs.
331 // Copies do not have constraints.
332 const TargetRegisterClass *OldRC = MRI.getRegClassOrNull(DstReg);
333 if (!OldRC || !DstRC->hasSubClassEq(OldRC)) {
334 if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
335 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
336 << " operand\n");
337 return false;
338 }
339 }
340 I.setDesc(TII.get(X86::COPY));
341 return true;
342 }
343
select(MachineInstr & I)344 bool X86InstructionSelector::select(MachineInstr &I) {
345 assert(I.getParent() && "Instruction should be in a basic block!");
346 assert(I.getParent()->getParent() && "Instruction should be in a function!");
347
348 MachineBasicBlock &MBB = *I.getParent();
349 MachineFunction &MF = *MBB.getParent();
350 MachineRegisterInfo &MRI = MF.getRegInfo();
351
352 unsigned Opcode = I.getOpcode();
353 if (!isPreISelGenericOpcode(Opcode)) {
354 // Certain non-generic instructions also need some special handling.
355
356 if (Opcode == TargetOpcode::LOAD_STACK_GUARD)
357 return false;
358
359 if (I.isCopy())
360 return selectCopy(I, MRI);
361
362 if (I.isDebugInstr())
363 return selectDebugInstr(I, MRI);
364
365 return true;
366 }
367
368 assert(I.getNumOperands() == I.getNumExplicitOperands() &&
369 "Generic instruction has unexpected implicit operands\n");
370
371 if (selectImpl(I, *CoverageInfo))
372 return true;
373
374 LLVM_DEBUG(dbgs() << " C++ instruction selection: "; I.print(dbgs()));
375
376 // TODO: This should be implemented by tblgen.
377 switch (I.getOpcode()) {
378 default:
379 return false;
380 case TargetOpcode::G_STORE:
381 case TargetOpcode::G_LOAD:
382 return selectLoadStoreOp(I, MRI, MF);
383 case TargetOpcode::G_PTR_ADD:
384 case TargetOpcode::G_FRAME_INDEX:
385 return selectFrameIndexOrGep(I, MRI, MF);
386 case TargetOpcode::G_GLOBAL_VALUE:
387 return selectGlobalValue(I, MRI, MF);
388 case TargetOpcode::G_CONSTANT:
389 return selectConstant(I, MRI, MF);
390 case TargetOpcode::G_FCONSTANT:
391 return materializeFP(I, MRI, MF);
392 case TargetOpcode::G_PTRTOINT:
393 case TargetOpcode::G_TRUNC:
394 return selectTruncOrPtrToInt(I, MRI, MF);
395 case TargetOpcode::G_INTTOPTR:
396 return selectCopy(I, MRI);
397 case TargetOpcode::G_ZEXT:
398 return selectZext(I, MRI, MF);
399 case TargetOpcode::G_ANYEXT:
400 return selectAnyext(I, MRI, MF);
401 case TargetOpcode::G_ICMP:
402 return selectCmp(I, MRI, MF);
403 case TargetOpcode::G_FCMP:
404 return selectFCmp(I, MRI, MF);
405 case TargetOpcode::G_UADDE:
406 return selectUadde(I, MRI, MF);
407 case TargetOpcode::G_UNMERGE_VALUES:
408 return selectUnmergeValues(I, MRI, MF);
409 case TargetOpcode::G_MERGE_VALUES:
410 case TargetOpcode::G_CONCAT_VECTORS:
411 return selectMergeValues(I, MRI, MF);
412 case TargetOpcode::G_EXTRACT:
413 return selectExtract(I, MRI, MF);
414 case TargetOpcode::G_INSERT:
415 return selectInsert(I, MRI, MF);
416 case TargetOpcode::G_BRCOND:
417 return selectCondBranch(I, MRI, MF);
418 case TargetOpcode::G_IMPLICIT_DEF:
419 case TargetOpcode::G_PHI:
420 return selectImplicitDefOrPHI(I, MRI);
421 case TargetOpcode::G_SDIV:
422 case TargetOpcode::G_UDIV:
423 case TargetOpcode::G_SREM:
424 case TargetOpcode::G_UREM:
425 return selectDivRem(I, MRI, MF);
426 case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
427 return selectIntrinsicWSideEffects(I, MRI, MF);
428 }
429
430 return false;
431 }
432
getLoadStoreOp(const LLT & Ty,const RegisterBank & RB,unsigned Opc,Align Alignment) const433 unsigned X86InstructionSelector::getLoadStoreOp(const LLT &Ty,
434 const RegisterBank &RB,
435 unsigned Opc,
436 Align Alignment) const {
437 bool Isload = (Opc == TargetOpcode::G_LOAD);
438 bool HasAVX = STI.hasAVX();
439 bool HasAVX512 = STI.hasAVX512();
440 bool HasVLX = STI.hasVLX();
441
442 if (Ty == LLT::scalar(8)) {
443 if (X86::GPRRegBankID == RB.getID())
444 return Isload ? X86::MOV8rm : X86::MOV8mr;
445 } else if (Ty == LLT::scalar(16)) {
446 if (X86::GPRRegBankID == RB.getID())
447 return Isload ? X86::MOV16rm : X86::MOV16mr;
448 } else if (Ty == LLT::scalar(32) || Ty == LLT::pointer(0, 32)) {
449 if (X86::GPRRegBankID == RB.getID())
450 return Isload ? X86::MOV32rm : X86::MOV32mr;
451 if (X86::VECRRegBankID == RB.getID())
452 return Isload ? (HasAVX512 ? X86::VMOVSSZrm_alt :
453 HasAVX ? X86::VMOVSSrm_alt :
454 X86::MOVSSrm_alt)
455 : (HasAVX512 ? X86::VMOVSSZmr :
456 HasAVX ? X86::VMOVSSmr :
457 X86::MOVSSmr);
458 } else if (Ty == LLT::scalar(64) || Ty == LLT::pointer(0, 64)) {
459 if (X86::GPRRegBankID == RB.getID())
460 return Isload ? X86::MOV64rm : X86::MOV64mr;
461 if (X86::VECRRegBankID == RB.getID())
462 return Isload ? (HasAVX512 ? X86::VMOVSDZrm_alt :
463 HasAVX ? X86::VMOVSDrm_alt :
464 X86::MOVSDrm_alt)
465 : (HasAVX512 ? X86::VMOVSDZmr :
466 HasAVX ? X86::VMOVSDmr :
467 X86::MOVSDmr);
468 } else if (Ty.isVector() && Ty.getSizeInBits() == 128) {
469 if (Alignment >= Align(16))
470 return Isload ? (HasVLX ? X86::VMOVAPSZ128rm
471 : HasAVX512
472 ? X86::VMOVAPSZ128rm_NOVLX
473 : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm)
474 : (HasVLX ? X86::VMOVAPSZ128mr
475 : HasAVX512
476 ? X86::VMOVAPSZ128mr_NOVLX
477 : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
478 else
479 return Isload ? (HasVLX ? X86::VMOVUPSZ128rm
480 : HasAVX512
481 ? X86::VMOVUPSZ128rm_NOVLX
482 : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm)
483 : (HasVLX ? X86::VMOVUPSZ128mr
484 : HasAVX512
485 ? X86::VMOVUPSZ128mr_NOVLX
486 : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
487 } else if (Ty.isVector() && Ty.getSizeInBits() == 256) {
488 if (Alignment >= Align(32))
489 return Isload ? (HasVLX ? X86::VMOVAPSZ256rm
490 : HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX
491 : X86::VMOVAPSYrm)
492 : (HasVLX ? X86::VMOVAPSZ256mr
493 : HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX
494 : X86::VMOVAPSYmr);
495 else
496 return Isload ? (HasVLX ? X86::VMOVUPSZ256rm
497 : HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX
498 : X86::VMOVUPSYrm)
499 : (HasVLX ? X86::VMOVUPSZ256mr
500 : HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX
501 : X86::VMOVUPSYmr);
502 } else if (Ty.isVector() && Ty.getSizeInBits() == 512) {
503 if (Alignment >= Align(64))
504 return Isload ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
505 else
506 return Isload ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
507 }
508 return Opc;
509 }
510
511 // Fill in an address from the given instruction.
X86SelectAddress(const MachineInstr & I,const MachineRegisterInfo & MRI,X86AddressMode & AM)512 static void X86SelectAddress(const MachineInstr &I,
513 const MachineRegisterInfo &MRI,
514 X86AddressMode &AM) {
515 assert(I.getOperand(0).isReg() && "unsupported opperand.");
516 assert(MRI.getType(I.getOperand(0).getReg()).isPointer() &&
517 "unsupported type.");
518
519 if (I.getOpcode() == TargetOpcode::G_PTR_ADD) {
520 if (auto COff = getIConstantVRegSExtVal(I.getOperand(2).getReg(), MRI)) {
521 int64_t Imm = *COff;
522 if (isInt<32>(Imm)) { // Check for displacement overflow.
523 AM.Disp = static_cast<int32_t>(Imm);
524 AM.Base.Reg = I.getOperand(1).getReg();
525 return;
526 }
527 }
528 } else if (I.getOpcode() == TargetOpcode::G_FRAME_INDEX) {
529 AM.Base.FrameIndex = I.getOperand(1).getIndex();
530 AM.BaseType = X86AddressMode::FrameIndexBase;
531 return;
532 }
533
534 // Default behavior.
535 AM.Base.Reg = I.getOperand(0).getReg();
536 }
537
selectLoadStoreOp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const538 bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I,
539 MachineRegisterInfo &MRI,
540 MachineFunction &MF) const {
541 unsigned Opc = I.getOpcode();
542
543 assert((Opc == TargetOpcode::G_STORE || Opc == TargetOpcode::G_LOAD) &&
544 "unexpected instruction");
545
546 const Register DefReg = I.getOperand(0).getReg();
547 LLT Ty = MRI.getType(DefReg);
548 const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI);
549
550 assert(I.hasOneMemOperand());
551 auto &MemOp = **I.memoperands_begin();
552 if (MemOp.isAtomic()) {
553 // Note: for unordered operations, we rely on the fact the appropriate MMO
554 // is already on the instruction we're mutating, and thus we don't need to
555 // make any changes. So long as we select an opcode which is capable of
556 // loading or storing the appropriate size atomically, the rest of the
557 // backend is required to respect the MMO state.
558 if (!MemOp.isUnordered()) {
559 LLVM_DEBUG(dbgs() << "Atomic ordering not supported yet\n");
560 return false;
561 }
562 if (MemOp.getAlign() < Ty.getSizeInBits() / 8) {
563 LLVM_DEBUG(dbgs() << "Unaligned atomics not supported yet\n");
564 return false;
565 }
566 }
567
568 unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlign());
569 if (NewOpc == Opc)
570 return false;
571
572 X86AddressMode AM;
573 X86SelectAddress(*MRI.getVRegDef(I.getOperand(1).getReg()), MRI, AM);
574
575 I.setDesc(TII.get(NewOpc));
576 MachineInstrBuilder MIB(MF, I);
577 if (Opc == TargetOpcode::G_LOAD) {
578 I.removeOperand(1);
579 addFullAddress(MIB, AM);
580 } else {
581 // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL)
582 I.removeOperand(1);
583 I.removeOperand(0);
584 addFullAddress(MIB, AM).addUse(DefReg);
585 }
586 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
587 }
588
getLeaOP(LLT Ty,const X86Subtarget & STI)589 static unsigned getLeaOP(LLT Ty, const X86Subtarget &STI) {
590 if (Ty == LLT::pointer(0, 64))
591 return X86::LEA64r;
592 else if (Ty == LLT::pointer(0, 32))
593 return STI.isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r;
594 else
595 llvm_unreachable("Can't get LEA opcode. Unsupported type.");
596 }
597
selectFrameIndexOrGep(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const598 bool X86InstructionSelector::selectFrameIndexOrGep(MachineInstr &I,
599 MachineRegisterInfo &MRI,
600 MachineFunction &MF) const {
601 unsigned Opc = I.getOpcode();
602
603 assert((Opc == TargetOpcode::G_FRAME_INDEX || Opc == TargetOpcode::G_PTR_ADD) &&
604 "unexpected instruction");
605
606 const Register DefReg = I.getOperand(0).getReg();
607 LLT Ty = MRI.getType(DefReg);
608
609 // Use LEA to calculate frame index and GEP
610 unsigned NewOpc = getLeaOP(Ty, STI);
611 I.setDesc(TII.get(NewOpc));
612 MachineInstrBuilder MIB(MF, I);
613
614 if (Opc == TargetOpcode::G_FRAME_INDEX) {
615 addOffset(MIB, 0);
616 } else {
617 MachineOperand &InxOp = I.getOperand(2);
618 I.addOperand(InxOp); // set IndexReg
619 InxOp.ChangeToImmediate(1); // set Scale
620 MIB.addImm(0).addReg(0);
621 }
622
623 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
624 }
625
selectGlobalValue(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const626 bool X86InstructionSelector::selectGlobalValue(MachineInstr &I,
627 MachineRegisterInfo &MRI,
628 MachineFunction &MF) const {
629 assert((I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE) &&
630 "unexpected instruction");
631
632 auto GV = I.getOperand(1).getGlobal();
633 if (GV->isThreadLocal()) {
634 return false; // TODO: we don't support TLS yet.
635 }
636
637 // Can't handle alternate code models yet.
638 if (TM.getCodeModel() != CodeModel::Small)
639 return false;
640
641 X86AddressMode AM;
642 AM.GV = GV;
643 AM.GVOpFlags = STI.classifyGlobalReference(GV);
644
645 // TODO: The ABI requires an extra load. not supported yet.
646 if (isGlobalStubReference(AM.GVOpFlags))
647 return false;
648
649 // TODO: This reference is relative to the pic base. not supported yet.
650 if (isGlobalRelativeToPICBase(AM.GVOpFlags))
651 return false;
652
653 if (STI.isPICStyleRIPRel()) {
654 // Use rip-relative addressing.
655 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
656 AM.Base.Reg = X86::RIP;
657 }
658
659 const Register DefReg = I.getOperand(0).getReg();
660 LLT Ty = MRI.getType(DefReg);
661 unsigned NewOpc = getLeaOP(Ty, STI);
662
663 I.setDesc(TII.get(NewOpc));
664 MachineInstrBuilder MIB(MF, I);
665
666 I.removeOperand(1);
667 addFullAddress(MIB, AM);
668
669 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
670 }
671
selectConstant(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const672 bool X86InstructionSelector::selectConstant(MachineInstr &I,
673 MachineRegisterInfo &MRI,
674 MachineFunction &MF) const {
675 assert((I.getOpcode() == TargetOpcode::G_CONSTANT) &&
676 "unexpected instruction");
677
678 const Register DefReg = I.getOperand(0).getReg();
679 LLT Ty = MRI.getType(DefReg);
680
681 if (RBI.getRegBank(DefReg, MRI, TRI)->getID() != X86::GPRRegBankID)
682 return false;
683
684 uint64_t Val = 0;
685 if (I.getOperand(1).isCImm()) {
686 Val = I.getOperand(1).getCImm()->getZExtValue();
687 I.getOperand(1).ChangeToImmediate(Val);
688 } else if (I.getOperand(1).isImm()) {
689 Val = I.getOperand(1).getImm();
690 } else
691 llvm_unreachable("Unsupported operand type.");
692
693 unsigned NewOpc;
694 switch (Ty.getSizeInBits()) {
695 case 8:
696 NewOpc = X86::MOV8ri;
697 break;
698 case 16:
699 NewOpc = X86::MOV16ri;
700 break;
701 case 32:
702 NewOpc = X86::MOV32ri;
703 break;
704 case 64:
705 // TODO: in case isUInt<32>(Val), X86::MOV32ri can be used
706 if (isInt<32>(Val))
707 NewOpc = X86::MOV64ri32;
708 else
709 NewOpc = X86::MOV64ri;
710 break;
711 default:
712 llvm_unreachable("Can't select G_CONSTANT, unsupported type.");
713 }
714
715 I.setDesc(TII.get(NewOpc));
716 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
717 }
718
719 // Helper function for selectTruncOrPtrToInt and selectAnyext.
720 // Returns true if DstRC lives on a floating register class and
721 // SrcRC lives on a 128-bit vector class.
canTurnIntoCOPY(const TargetRegisterClass * DstRC,const TargetRegisterClass * SrcRC)722 static bool canTurnIntoCOPY(const TargetRegisterClass *DstRC,
723 const TargetRegisterClass *SrcRC) {
724 return (DstRC == &X86::FR32RegClass || DstRC == &X86::FR32XRegClass ||
725 DstRC == &X86::FR64RegClass || DstRC == &X86::FR64XRegClass) &&
726 (SrcRC == &X86::VR128RegClass || SrcRC == &X86::VR128XRegClass);
727 }
728
selectTurnIntoCOPY(MachineInstr & I,MachineRegisterInfo & MRI,const unsigned DstReg,const TargetRegisterClass * DstRC,const unsigned SrcReg,const TargetRegisterClass * SrcRC) const729 bool X86InstructionSelector::selectTurnIntoCOPY(
730 MachineInstr &I, MachineRegisterInfo &MRI, const unsigned DstReg,
731 const TargetRegisterClass *DstRC, const unsigned SrcReg,
732 const TargetRegisterClass *SrcRC) const {
733
734 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
735 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
736 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
737 << " operand\n");
738 return false;
739 }
740 I.setDesc(TII.get(X86::COPY));
741 return true;
742 }
743
selectTruncOrPtrToInt(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const744 bool X86InstructionSelector::selectTruncOrPtrToInt(MachineInstr &I,
745 MachineRegisterInfo &MRI,
746 MachineFunction &MF) const {
747 assert((I.getOpcode() == TargetOpcode::G_TRUNC ||
748 I.getOpcode() == TargetOpcode::G_PTRTOINT) &&
749 "unexpected instruction");
750
751 const Register DstReg = I.getOperand(0).getReg();
752 const Register SrcReg = I.getOperand(1).getReg();
753
754 const LLT DstTy = MRI.getType(DstReg);
755 const LLT SrcTy = MRI.getType(SrcReg);
756
757 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
758 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);
759
760 if (DstRB.getID() != SrcRB.getID()) {
761 LLVM_DEBUG(dbgs() << TII.getName(I.getOpcode())
762 << " input/output on different banks\n");
763 return false;
764 }
765
766 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
767 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);
768
769 if (!DstRC || !SrcRC)
770 return false;
771
772 // If that's truncation of the value that lives on the vector class and goes
773 // into the floating class, just replace it with copy, as we are able to
774 // select it as a regular move.
775 if (canTurnIntoCOPY(DstRC, SrcRC))
776 return selectTurnIntoCOPY(I, MRI, DstReg, DstRC, SrcReg, SrcRC);
777
778 if (DstRB.getID() != X86::GPRRegBankID)
779 return false;
780
781 unsigned SubIdx;
782 if (DstRC == SrcRC) {
783 // Nothing to be done
784 SubIdx = X86::NoSubRegister;
785 } else if (DstRC == &X86::GR32RegClass) {
786 SubIdx = X86::sub_32bit;
787 } else if (DstRC == &X86::GR16RegClass) {
788 SubIdx = X86::sub_16bit;
789 } else if (DstRC == &X86::GR8RegClass) {
790 SubIdx = X86::sub_8bit;
791 } else {
792 return false;
793 }
794
795 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);
796
797 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
798 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
799 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
800 << "\n");
801 return false;
802 }
803
804 I.getOperand(1).setSubReg(SubIdx);
805
806 I.setDesc(TII.get(X86::COPY));
807 return true;
808 }
809
selectZext(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const810 bool X86InstructionSelector::selectZext(MachineInstr &I,
811 MachineRegisterInfo &MRI,
812 MachineFunction &MF) const {
813 assert((I.getOpcode() == TargetOpcode::G_ZEXT) && "unexpected instruction");
814
815 const Register DstReg = I.getOperand(0).getReg();
816 const Register SrcReg = I.getOperand(1).getReg();
817
818 const LLT DstTy = MRI.getType(DstReg);
819 const LLT SrcTy = MRI.getType(SrcReg);
820
821 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(16)) &&
822 "8=>16 Zext is handled by tablegen");
823 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(32)) &&
824 "8=>32 Zext is handled by tablegen");
825 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(32)) &&
826 "16=>32 Zext is handled by tablegen");
827 assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(64)) &&
828 "8=>64 Zext is handled by tablegen");
829 assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(64)) &&
830 "16=>64 Zext is handled by tablegen");
831 assert(!(SrcTy == LLT::scalar(32) && DstTy == LLT::scalar(64)) &&
832 "32=>64 Zext is handled by tablegen");
833
834 if (SrcTy != LLT::scalar(1))
835 return false;
836
837 unsigned AndOpc;
838 if (DstTy == LLT::scalar(8))
839 AndOpc = X86::AND8ri;
840 else if (DstTy == LLT::scalar(16))
841 AndOpc = X86::AND16ri8;
842 else if (DstTy == LLT::scalar(32))
843 AndOpc = X86::AND32ri8;
844 else if (DstTy == LLT::scalar(64))
845 AndOpc = X86::AND64ri8;
846 else
847 return false;
848
849 Register DefReg = SrcReg;
850 if (DstTy != LLT::scalar(8)) {
851 Register ImpDefReg =
852 MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
853 BuildMI(*I.getParent(), I, I.getDebugLoc(),
854 TII.get(TargetOpcode::IMPLICIT_DEF), ImpDefReg);
855
856 DefReg = MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
857 BuildMI(*I.getParent(), I, I.getDebugLoc(),
858 TII.get(TargetOpcode::INSERT_SUBREG), DefReg)
859 .addReg(ImpDefReg)
860 .addReg(SrcReg)
861 .addImm(X86::sub_8bit);
862 }
863
864 MachineInstr &AndInst =
865 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(AndOpc), DstReg)
866 .addReg(DefReg)
867 .addImm(1);
868
869 constrainSelectedInstRegOperands(AndInst, TII, TRI, RBI);
870
871 I.eraseFromParent();
872 return true;
873 }
874
selectAnyext(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const875 bool X86InstructionSelector::selectAnyext(MachineInstr &I,
876 MachineRegisterInfo &MRI,
877 MachineFunction &MF) const {
878 assert((I.getOpcode() == TargetOpcode::G_ANYEXT) && "unexpected instruction");
879
880 const Register DstReg = I.getOperand(0).getReg();
881 const Register SrcReg = I.getOperand(1).getReg();
882
883 const LLT DstTy = MRI.getType(DstReg);
884 const LLT SrcTy = MRI.getType(SrcReg);
885
886 const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
887 const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);
888
889 assert(DstRB.getID() == SrcRB.getID() &&
890 "G_ANYEXT input/output on different banks\n");
891
892 assert(DstTy.getSizeInBits() > SrcTy.getSizeInBits() &&
893 "G_ANYEXT incorrect operand size");
894
895 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
896 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);
897
898 // If that's ANY_EXT of the value that lives on the floating class and goes
899 // into the vector class, just replace it with copy, as we are able to select
900 // it as a regular move.
901 if (canTurnIntoCOPY(SrcRC, DstRC))
902 return selectTurnIntoCOPY(I, MRI, SrcReg, SrcRC, DstReg, DstRC);
903
904 if (DstRB.getID() != X86::GPRRegBankID)
905 return false;
906
907 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
908 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
909 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
910 << " operand\n");
911 return false;
912 }
913
914 if (SrcRC == DstRC) {
915 I.setDesc(TII.get(X86::COPY));
916 return true;
917 }
918
919 BuildMI(*I.getParent(), I, I.getDebugLoc(),
920 TII.get(TargetOpcode::SUBREG_TO_REG))
921 .addDef(DstReg)
922 .addImm(0)
923 .addReg(SrcReg)
924 .addImm(getSubRegIndex(SrcRC));
925
926 I.eraseFromParent();
927 return true;
928 }
929
selectCmp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const930 bool X86InstructionSelector::selectCmp(MachineInstr &I,
931 MachineRegisterInfo &MRI,
932 MachineFunction &MF) const {
933 assert((I.getOpcode() == TargetOpcode::G_ICMP) && "unexpected instruction");
934
935 X86::CondCode CC;
936 bool SwapArgs;
937 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(
938 (CmpInst::Predicate)I.getOperand(1).getPredicate());
939
940 Register LHS = I.getOperand(2).getReg();
941 Register RHS = I.getOperand(3).getReg();
942
943 if (SwapArgs)
944 std::swap(LHS, RHS);
945
946 unsigned OpCmp;
947 LLT Ty = MRI.getType(LHS);
948
949 switch (Ty.getSizeInBits()) {
950 default:
951 return false;
952 case 8:
953 OpCmp = X86::CMP8rr;
954 break;
955 case 16:
956 OpCmp = X86::CMP16rr;
957 break;
958 case 32:
959 OpCmp = X86::CMP32rr;
960 break;
961 case 64:
962 OpCmp = X86::CMP64rr;
963 break;
964 }
965
966 MachineInstr &CmpInst =
967 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
968 .addReg(LHS)
969 .addReg(RHS);
970
971 MachineInstr &SetInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
972 TII.get(X86::SETCCr), I.getOperand(0).getReg()).addImm(CC);
973
974 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
975 constrainSelectedInstRegOperands(SetInst, TII, TRI, RBI);
976
977 I.eraseFromParent();
978 return true;
979 }
980
selectFCmp(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const981 bool X86InstructionSelector::selectFCmp(MachineInstr &I,
982 MachineRegisterInfo &MRI,
983 MachineFunction &MF) const {
984 assert((I.getOpcode() == TargetOpcode::G_FCMP) && "unexpected instruction");
985
986 Register LhsReg = I.getOperand(2).getReg();
987 Register RhsReg = I.getOperand(3).getReg();
988 CmpInst::Predicate Predicate =
989 (CmpInst::Predicate)I.getOperand(1).getPredicate();
990
991 // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
992 static const uint16_t SETFOpcTable[2][3] = {
993 {X86::COND_E, X86::COND_NP, X86::AND8rr},
994 {X86::COND_NE, X86::COND_P, X86::OR8rr}};
995 const uint16_t *SETFOpc = nullptr;
996 switch (Predicate) {
997 default:
998 break;
999 case CmpInst::FCMP_OEQ:
1000 SETFOpc = &SETFOpcTable[0][0];
1001 break;
1002 case CmpInst::FCMP_UNE:
1003 SETFOpc = &SETFOpcTable[1][0];
1004 break;
1005 }
1006
1007 // Compute the opcode for the CMP instruction.
1008 unsigned OpCmp;
1009 LLT Ty = MRI.getType(LhsReg);
1010 switch (Ty.getSizeInBits()) {
1011 default:
1012 return false;
1013 case 32:
1014 OpCmp = X86::UCOMISSrr;
1015 break;
1016 case 64:
1017 OpCmp = X86::UCOMISDrr;
1018 break;
1019 }
1020
1021 Register ResultReg = I.getOperand(0).getReg();
1022 RBI.constrainGenericRegister(
1023 ResultReg,
1024 *getRegClass(LLT::scalar(8), *RBI.getRegBank(ResultReg, MRI, TRI)), MRI);
1025 if (SETFOpc) {
1026 MachineInstr &CmpInst =
1027 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
1028 .addReg(LhsReg)
1029 .addReg(RhsReg);
1030
1031 Register FlagReg1 = MRI.createVirtualRegister(&X86::GR8RegClass);
1032 Register FlagReg2 = MRI.createVirtualRegister(&X86::GR8RegClass);
1033 MachineInstr &Set1 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1034 TII.get(X86::SETCCr), FlagReg1).addImm(SETFOpc[0]);
1035 MachineInstr &Set2 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1036 TII.get(X86::SETCCr), FlagReg2).addImm(SETFOpc[1]);
1037 MachineInstr &Set3 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1038 TII.get(SETFOpc[2]), ResultReg)
1039 .addReg(FlagReg1)
1040 .addReg(FlagReg2);
1041 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
1042 constrainSelectedInstRegOperands(Set1, TII, TRI, RBI);
1043 constrainSelectedInstRegOperands(Set2, TII, TRI, RBI);
1044 constrainSelectedInstRegOperands(Set3, TII, TRI, RBI);
1045
1046 I.eraseFromParent();
1047 return true;
1048 }
1049
1050 X86::CondCode CC;
1051 bool SwapArgs;
1052 std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate);
1053 assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");
1054
1055 if (SwapArgs)
1056 std::swap(LhsReg, RhsReg);
1057
1058 // Emit a compare of LHS/RHS.
1059 MachineInstr &CmpInst =
1060 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
1061 .addReg(LhsReg)
1062 .addReg(RhsReg);
1063
1064 MachineInstr &Set =
1065 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SETCCr), ResultReg).addImm(CC);
1066 constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
1067 constrainSelectedInstRegOperands(Set, TII, TRI, RBI);
1068 I.eraseFromParent();
1069 return true;
1070 }
1071
selectUadde(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1072 bool X86InstructionSelector::selectUadde(MachineInstr &I,
1073 MachineRegisterInfo &MRI,
1074 MachineFunction &MF) const {
1075 assert((I.getOpcode() == TargetOpcode::G_UADDE) && "unexpected instruction");
1076
1077 const Register DstReg = I.getOperand(0).getReg();
1078 const Register CarryOutReg = I.getOperand(1).getReg();
1079 const Register Op0Reg = I.getOperand(2).getReg();
1080 const Register Op1Reg = I.getOperand(3).getReg();
1081 Register CarryInReg = I.getOperand(4).getReg();
1082
1083 const LLT DstTy = MRI.getType(DstReg);
1084
1085 if (DstTy != LLT::scalar(32))
1086 return false;
1087
1088 // find CarryIn def instruction.
1089 MachineInstr *Def = MRI.getVRegDef(CarryInReg);
1090 while (Def->getOpcode() == TargetOpcode::G_TRUNC) {
1091 CarryInReg = Def->getOperand(1).getReg();
1092 Def = MRI.getVRegDef(CarryInReg);
1093 }
1094
1095 unsigned Opcode;
1096 if (Def->getOpcode() == TargetOpcode::G_UADDE) {
1097 // carry set by prev ADD.
1098
1099 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), X86::EFLAGS)
1100 .addReg(CarryInReg);
1101
1102 if (!RBI.constrainGenericRegister(CarryInReg, X86::GR32RegClass, MRI))
1103 return false;
1104
1105 Opcode = X86::ADC32rr;
1106 } else if (auto val = getIConstantVRegVal(CarryInReg, MRI)) {
1107 // carry is constant, support only 0.
1108 if (*val != 0)
1109 return false;
1110
1111 Opcode = X86::ADD32rr;
1112 } else
1113 return false;
1114
1115 MachineInstr &AddInst =
1116 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg)
1117 .addReg(Op0Reg)
1118 .addReg(Op1Reg);
1119
1120 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), CarryOutReg)
1121 .addReg(X86::EFLAGS);
1122
1123 if (!constrainSelectedInstRegOperands(AddInst, TII, TRI, RBI) ||
1124 !RBI.constrainGenericRegister(CarryOutReg, X86::GR32RegClass, MRI))
1125 return false;
1126
1127 I.eraseFromParent();
1128 return true;
1129 }
1130
selectExtract(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1131 bool X86InstructionSelector::selectExtract(MachineInstr &I,
1132 MachineRegisterInfo &MRI,
1133 MachineFunction &MF) const {
1134 assert((I.getOpcode() == TargetOpcode::G_EXTRACT) &&
1135 "unexpected instruction");
1136
1137 const Register DstReg = I.getOperand(0).getReg();
1138 const Register SrcReg = I.getOperand(1).getReg();
1139 int64_t Index = I.getOperand(2).getImm();
1140
1141 const LLT DstTy = MRI.getType(DstReg);
1142 const LLT SrcTy = MRI.getType(SrcReg);
1143
1144 // Meanwile handle vector type only.
1145 if (!DstTy.isVector())
1146 return false;
1147
1148 if (Index % DstTy.getSizeInBits() != 0)
1149 return false; // Not extract subvector.
1150
1151 if (Index == 0) {
1152 // Replace by extract subreg copy.
1153 if (!emitExtractSubreg(DstReg, SrcReg, I, MRI, MF))
1154 return false;
1155
1156 I.eraseFromParent();
1157 return true;
1158 }
1159
1160 bool HasAVX = STI.hasAVX();
1161 bool HasAVX512 = STI.hasAVX512();
1162 bool HasVLX = STI.hasVLX();
1163
1164 if (SrcTy.getSizeInBits() == 256 && DstTy.getSizeInBits() == 128) {
1165 if (HasVLX)
1166 I.setDesc(TII.get(X86::VEXTRACTF32x4Z256rr));
1167 else if (HasAVX)
1168 I.setDesc(TII.get(X86::VEXTRACTF128rr));
1169 else
1170 return false;
1171 } else if (SrcTy.getSizeInBits() == 512 && HasAVX512) {
1172 if (DstTy.getSizeInBits() == 128)
1173 I.setDesc(TII.get(X86::VEXTRACTF32x4Zrr));
1174 else if (DstTy.getSizeInBits() == 256)
1175 I.setDesc(TII.get(X86::VEXTRACTF64x4Zrr));
1176 else
1177 return false;
1178 } else
1179 return false;
1180
1181 // Convert to X86 VEXTRACT immediate.
1182 Index = Index / DstTy.getSizeInBits();
1183 I.getOperand(2).setImm(Index);
1184
1185 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
1186 }
1187
emitExtractSubreg(unsigned DstReg,unsigned SrcReg,MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1188 bool X86InstructionSelector::emitExtractSubreg(unsigned DstReg, unsigned SrcReg,
1189 MachineInstr &I,
1190 MachineRegisterInfo &MRI,
1191 MachineFunction &MF) const {
1192 const LLT DstTy = MRI.getType(DstReg);
1193 const LLT SrcTy = MRI.getType(SrcReg);
1194 unsigned SubIdx = X86::NoSubRegister;
1195
1196 if (!DstTy.isVector() || !SrcTy.isVector())
1197 return false;
1198
1199 assert(SrcTy.getSizeInBits() > DstTy.getSizeInBits() &&
1200 "Incorrect Src/Dst register size");
1201
1202 if (DstTy.getSizeInBits() == 128)
1203 SubIdx = X86::sub_xmm;
1204 else if (DstTy.getSizeInBits() == 256)
1205 SubIdx = X86::sub_ymm;
1206 else
1207 return false;
1208
1209 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);
1210 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);
1211
1212 SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);
1213
1214 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
1215 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
1216 LLVM_DEBUG(dbgs() << "Failed to constrain EXTRACT_SUBREG\n");
1217 return false;
1218 }
1219
1220 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), DstReg)
1221 .addReg(SrcReg, 0, SubIdx);
1222
1223 return true;
1224 }
1225
emitInsertSubreg(unsigned DstReg,unsigned SrcReg,MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1226 bool X86InstructionSelector::emitInsertSubreg(unsigned DstReg, unsigned SrcReg,
1227 MachineInstr &I,
1228 MachineRegisterInfo &MRI,
1229 MachineFunction &MF) const {
1230 const LLT DstTy = MRI.getType(DstReg);
1231 const LLT SrcTy = MRI.getType(SrcReg);
1232 unsigned SubIdx = X86::NoSubRegister;
1233
1234 // TODO: support scalar types
1235 if (!DstTy.isVector() || !SrcTy.isVector())
1236 return false;
1237
1238 assert(SrcTy.getSizeInBits() < DstTy.getSizeInBits() &&
1239 "Incorrect Src/Dst register size");
1240
1241 if (SrcTy.getSizeInBits() == 128)
1242 SubIdx = X86::sub_xmm;
1243 else if (SrcTy.getSizeInBits() == 256)
1244 SubIdx = X86::sub_ymm;
1245 else
1246 return false;
1247
1248 const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);
1249 const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);
1250
1251 if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
1252 !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
1253 LLVM_DEBUG(dbgs() << "Failed to constrain INSERT_SUBREG\n");
1254 return false;
1255 }
1256
1257 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY))
1258 .addReg(DstReg, RegState::DefineNoRead, SubIdx)
1259 .addReg(SrcReg);
1260
1261 return true;
1262 }
1263
selectInsert(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1264 bool X86InstructionSelector::selectInsert(MachineInstr &I,
1265 MachineRegisterInfo &MRI,
1266 MachineFunction &MF) const {
1267 assert((I.getOpcode() == TargetOpcode::G_INSERT) && "unexpected instruction");
1268
1269 const Register DstReg = I.getOperand(0).getReg();
1270 const Register SrcReg = I.getOperand(1).getReg();
1271 const Register InsertReg = I.getOperand(2).getReg();
1272 int64_t Index = I.getOperand(3).getImm();
1273
1274 const LLT DstTy = MRI.getType(DstReg);
1275 const LLT InsertRegTy = MRI.getType(InsertReg);
1276
1277 // Meanwile handle vector type only.
1278 if (!DstTy.isVector())
1279 return false;
1280
1281 if (Index % InsertRegTy.getSizeInBits() != 0)
1282 return false; // Not insert subvector.
1283
1284 if (Index == 0 && MRI.getVRegDef(SrcReg)->isImplicitDef()) {
1285 // Replace by subreg copy.
1286 if (!emitInsertSubreg(DstReg, InsertReg, I, MRI, MF))
1287 return false;
1288
1289 I.eraseFromParent();
1290 return true;
1291 }
1292
1293 bool HasAVX = STI.hasAVX();
1294 bool HasAVX512 = STI.hasAVX512();
1295 bool HasVLX = STI.hasVLX();
1296
1297 if (DstTy.getSizeInBits() == 256 && InsertRegTy.getSizeInBits() == 128) {
1298 if (HasVLX)
1299 I.setDesc(TII.get(X86::VINSERTF32x4Z256rr));
1300 else if (HasAVX)
1301 I.setDesc(TII.get(X86::VINSERTF128rr));
1302 else
1303 return false;
1304 } else if (DstTy.getSizeInBits() == 512 && HasAVX512) {
1305 if (InsertRegTy.getSizeInBits() == 128)
1306 I.setDesc(TII.get(X86::VINSERTF32x4Zrr));
1307 else if (InsertRegTy.getSizeInBits() == 256)
1308 I.setDesc(TII.get(X86::VINSERTF64x4Zrr));
1309 else
1310 return false;
1311 } else
1312 return false;
1313
1314 // Convert to X86 VINSERT immediate.
1315 Index = Index / InsertRegTy.getSizeInBits();
1316
1317 I.getOperand(3).setImm(Index);
1318
1319 return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
1320 }
1321
selectUnmergeValues(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF)1322 bool X86InstructionSelector::selectUnmergeValues(
1323 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
1324 assert((I.getOpcode() == TargetOpcode::G_UNMERGE_VALUES) &&
1325 "unexpected instruction");
1326
1327 // Split to extracts.
1328 unsigned NumDefs = I.getNumOperands() - 1;
1329 Register SrcReg = I.getOperand(NumDefs).getReg();
1330 unsigned DefSize = MRI.getType(I.getOperand(0).getReg()).getSizeInBits();
1331
1332 for (unsigned Idx = 0; Idx < NumDefs; ++Idx) {
1333 MachineInstr &ExtrInst =
1334 *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1335 TII.get(TargetOpcode::G_EXTRACT), I.getOperand(Idx).getReg())
1336 .addReg(SrcReg)
1337 .addImm(Idx * DefSize);
1338
1339 if (!select(ExtrInst))
1340 return false;
1341 }
1342
1343 I.eraseFromParent();
1344 return true;
1345 }
1346
selectMergeValues(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF)1347 bool X86InstructionSelector::selectMergeValues(
1348 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
1349 assert((I.getOpcode() == TargetOpcode::G_MERGE_VALUES ||
1350 I.getOpcode() == TargetOpcode::G_CONCAT_VECTORS) &&
1351 "unexpected instruction");
1352
1353 // Split to inserts.
1354 Register DstReg = I.getOperand(0).getReg();
1355 Register SrcReg0 = I.getOperand(1).getReg();
1356
1357 const LLT DstTy = MRI.getType(DstReg);
1358 const LLT SrcTy = MRI.getType(SrcReg0);
1359 unsigned SrcSize = SrcTy.getSizeInBits();
1360
1361 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
1362
1363 // For the first src use insertSubReg.
1364 Register DefReg = MRI.createGenericVirtualRegister(DstTy);
1365 MRI.setRegBank(DefReg, RegBank);
1366 if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF))
1367 return false;
1368
1369 for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) {
1370 Register Tmp = MRI.createGenericVirtualRegister(DstTy);
1371 MRI.setRegBank(Tmp, RegBank);
1372
1373 MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1374 TII.get(TargetOpcode::G_INSERT), Tmp)
1375 .addReg(DefReg)
1376 .addReg(I.getOperand(Idx).getReg())
1377 .addImm((Idx - 1) * SrcSize);
1378
1379 DefReg = Tmp;
1380
1381 if (!select(InsertInst))
1382 return false;
1383 }
1384
1385 MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
1386 TII.get(TargetOpcode::COPY), DstReg)
1387 .addReg(DefReg);
1388
1389 if (!select(CopyInst))
1390 return false;
1391
1392 I.eraseFromParent();
1393 return true;
1394 }
1395
selectCondBranch(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1396 bool X86InstructionSelector::selectCondBranch(MachineInstr &I,
1397 MachineRegisterInfo &MRI,
1398 MachineFunction &MF) const {
1399 assert((I.getOpcode() == TargetOpcode::G_BRCOND) && "unexpected instruction");
1400
1401 const Register CondReg = I.getOperand(0).getReg();
1402 MachineBasicBlock *DestMBB = I.getOperand(1).getMBB();
1403
1404 MachineInstr &TestInst =
1405 *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TEST8ri))
1406 .addReg(CondReg)
1407 .addImm(1);
1408 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::JCC_1))
1409 .addMBB(DestMBB).addImm(X86::COND_NE);
1410
1411 constrainSelectedInstRegOperands(TestInst, TII, TRI, RBI);
1412
1413 I.eraseFromParent();
1414 return true;
1415 }
1416
materializeFP(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1417 bool X86InstructionSelector::materializeFP(MachineInstr &I,
1418 MachineRegisterInfo &MRI,
1419 MachineFunction &MF) const {
1420 assert((I.getOpcode() == TargetOpcode::G_FCONSTANT) &&
1421 "unexpected instruction");
1422
1423 // Can't handle alternate code models yet.
1424 CodeModel::Model CM = TM.getCodeModel();
1425 if (CM != CodeModel::Small && CM != CodeModel::Large)
1426 return false;
1427
1428 const Register DstReg = I.getOperand(0).getReg();
1429 const LLT DstTy = MRI.getType(DstReg);
1430 const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
1431 Align Alignment = Align(DstTy.getSizeInBytes());
1432 const DebugLoc &DbgLoc = I.getDebugLoc();
1433
1434 unsigned Opc =
1435 getLoadStoreOp(DstTy, RegBank, TargetOpcode::G_LOAD, Alignment);
1436
1437 // Create the load from the constant pool.
1438 const ConstantFP *CFP = I.getOperand(1).getFPImm();
1439 unsigned CPI = MF.getConstantPool()->getConstantPoolIndex(CFP, Alignment);
1440 MachineInstr *LoadInst = nullptr;
1441 unsigned char OpFlag = STI.classifyLocalReference(nullptr);
1442
1443 if (CM == CodeModel::Large && STI.is64Bit()) {
1444 // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
1445 // they cannot be folded into immediate fields.
1446
1447 Register AddrReg = MRI.createVirtualRegister(&X86::GR64RegClass);
1448 BuildMI(*I.getParent(), I, DbgLoc, TII.get(X86::MOV64ri), AddrReg)
1449 .addConstantPoolIndex(CPI, 0, OpFlag);
1450
1451 MachineMemOperand *MMO = MF.getMachineMemOperand(
1452 MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
1453 LLT::pointer(0, MF.getDataLayout().getPointerSizeInBits()), Alignment);
1454
1455 LoadInst =
1456 addDirectMem(BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg),
1457 AddrReg)
1458 .addMemOperand(MMO);
1459
1460 } else if (CM == CodeModel::Small || !STI.is64Bit()) {
1461 // Handle the case when globals fit in our immediate field.
1462 // This is true for X86-32 always and X86-64 when in -mcmodel=small mode.
1463
1464 // x86-32 PIC requires a PIC base register for constant pools.
1465 unsigned PICBase = 0;
1466 if (OpFlag == X86II::MO_PIC_BASE_OFFSET || OpFlag == X86II::MO_GOTOFF) {
1467 // PICBase can be allocated by TII.getGlobalBaseReg(&MF).
1468 // In DAGISEL the code that initialize it generated by the CGBR pass.
1469 return false; // TODO support the mode.
1470 } else if (STI.is64Bit() && TM.getCodeModel() == CodeModel::Small)
1471 PICBase = X86::RIP;
1472
1473 LoadInst = addConstantPoolReference(
1474 BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), CPI, PICBase,
1475 OpFlag);
1476 } else
1477 return false;
1478
1479 constrainSelectedInstRegOperands(*LoadInst, TII, TRI, RBI);
1480 I.eraseFromParent();
1481 return true;
1482 }
1483
selectImplicitDefOrPHI(MachineInstr & I,MachineRegisterInfo & MRI) const1484 bool X86InstructionSelector::selectImplicitDefOrPHI(
1485 MachineInstr &I, MachineRegisterInfo &MRI) const {
1486 assert((I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF ||
1487 I.getOpcode() == TargetOpcode::G_PHI) &&
1488 "unexpected instruction");
1489
1490 Register DstReg = I.getOperand(0).getReg();
1491
1492 if (!MRI.getRegClassOrNull(DstReg)) {
1493 const LLT DstTy = MRI.getType(DstReg);
1494 const TargetRegisterClass *RC = getRegClass(DstTy, DstReg, MRI);
1495
1496 if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
1497 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
1498 << " operand\n");
1499 return false;
1500 }
1501 }
1502
1503 if (I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
1504 I.setDesc(TII.get(X86::IMPLICIT_DEF));
1505 else
1506 I.setDesc(TII.get(X86::PHI));
1507
1508 return true;
1509 }
1510
selectDivRem(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1511 bool X86InstructionSelector::selectDivRem(MachineInstr &I,
1512 MachineRegisterInfo &MRI,
1513 MachineFunction &MF) const {
1514 // The implementation of this function is taken from X86FastISel.
1515 assert((I.getOpcode() == TargetOpcode::G_SDIV ||
1516 I.getOpcode() == TargetOpcode::G_SREM ||
1517 I.getOpcode() == TargetOpcode::G_UDIV ||
1518 I.getOpcode() == TargetOpcode::G_UREM) &&
1519 "unexpected instruction");
1520
1521 const Register DstReg = I.getOperand(0).getReg();
1522 const Register Op1Reg = I.getOperand(1).getReg();
1523 const Register Op2Reg = I.getOperand(2).getReg();
1524
1525 const LLT RegTy = MRI.getType(DstReg);
1526 assert(RegTy == MRI.getType(Op1Reg) && RegTy == MRI.getType(Op2Reg) &&
1527 "Arguments and return value types must match");
1528
1529 const RegisterBank *RegRB = RBI.getRegBank(DstReg, MRI, TRI);
1530 if (!RegRB || RegRB->getID() != X86::GPRRegBankID)
1531 return false;
1532
1533 const static unsigned NumTypes = 4; // i8, i16, i32, i64
1534 const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem
1535 const static bool S = true; // IsSigned
1536 const static bool U = false; // !IsSigned
1537 const static unsigned Copy = TargetOpcode::COPY;
1538 // For the X86 IDIV instruction, in most cases the dividend
1539 // (numerator) must be in a specific register pair highreg:lowreg,
1540 // producing the quotient in lowreg and the remainder in highreg.
1541 // For most data types, to set up the instruction, the dividend is
1542 // copied into lowreg, and lowreg is sign-extended into highreg. The
1543 // exception is i8, where the dividend is defined as a single register rather
1544 // than a register pair, and we therefore directly sign-extend the dividend
1545 // into lowreg, instead of copying, and ignore the highreg.
1546 const static struct DivRemEntry {
1547 // The following portion depends only on the data type.
1548 unsigned SizeInBits;
1549 unsigned LowInReg; // low part of the register pair
1550 unsigned HighInReg; // high part of the register pair
1551 // The following portion depends on both the data type and the operation.
1552 struct DivRemResult {
1553 unsigned OpDivRem; // The specific DIV/IDIV opcode to use.
1554 unsigned OpSignExtend; // Opcode for sign-extending lowreg into
1555 // highreg, or copying a zero into highreg.
1556 unsigned OpCopy; // Opcode for copying dividend into lowreg, or
1557 // zero/sign-extending into lowreg for i8.
1558 unsigned DivRemResultReg; // Register containing the desired result.
1559 bool IsOpSigned; // Whether to use signed or unsigned form.
1560 } ResultTable[NumOps];
1561 } OpTable[NumTypes] = {
1562 {8,
1563 X86::AX,
1564 0,
1565 {
1566 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S}, // SDiv
1567 {X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S}, // SRem
1568 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U}, // UDiv
1569 {X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U}, // URem
1570 }}, // i8
1571 {16,
1572 X86::AX,
1573 X86::DX,
1574 {
1575 {X86::IDIV16r, X86::CWD, Copy, X86::AX, S}, // SDiv
1576 {X86::IDIV16r, X86::CWD, Copy, X86::DX, S}, // SRem
1577 {X86::DIV16r, X86::MOV32r0, Copy, X86::AX, U}, // UDiv
1578 {X86::DIV16r, X86::MOV32r0, Copy, X86::DX, U}, // URem
1579 }}, // i16
1580 {32,
1581 X86::EAX,
1582 X86::EDX,
1583 {
1584 {X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S}, // SDiv
1585 {X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S}, // SRem
1586 {X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U}, // UDiv
1587 {X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U}, // URem
1588 }}, // i32
1589 {64,
1590 X86::RAX,
1591 X86::RDX,
1592 {
1593 {X86::IDIV64r, X86::CQO, Copy, X86::RAX, S}, // SDiv
1594 {X86::IDIV64r, X86::CQO, Copy, X86::RDX, S}, // SRem
1595 {X86::DIV64r, X86::MOV32r0, Copy, X86::RAX, U}, // UDiv
1596 {X86::DIV64r, X86::MOV32r0, Copy, X86::RDX, U}, // URem
1597 }}, // i64
1598 };
1599
1600 auto OpEntryIt = llvm::find_if(OpTable, [RegTy](const DivRemEntry &El) {
1601 return El.SizeInBits == RegTy.getSizeInBits();
1602 });
1603 if (OpEntryIt == std::end(OpTable))
1604 return false;
1605
1606 unsigned OpIndex;
1607 switch (I.getOpcode()) {
1608 default:
1609 llvm_unreachable("Unexpected div/rem opcode");
1610 case TargetOpcode::G_SDIV:
1611 OpIndex = 0;
1612 break;
1613 case TargetOpcode::G_SREM:
1614 OpIndex = 1;
1615 break;
1616 case TargetOpcode::G_UDIV:
1617 OpIndex = 2;
1618 break;
1619 case TargetOpcode::G_UREM:
1620 OpIndex = 3;
1621 break;
1622 }
1623
1624 const DivRemEntry &TypeEntry = *OpEntryIt;
1625 const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
1626
1627 const TargetRegisterClass *RegRC = getRegClass(RegTy, *RegRB);
1628 if (!RBI.constrainGenericRegister(Op1Reg, *RegRC, MRI) ||
1629 !RBI.constrainGenericRegister(Op2Reg, *RegRC, MRI) ||
1630 !RBI.constrainGenericRegister(DstReg, *RegRC, MRI)) {
1631 LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
1632 << " operand\n");
1633 return false;
1634 }
1635
1636 // Move op1 into low-order input register.
1637 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpCopy),
1638 TypeEntry.LowInReg)
1639 .addReg(Op1Reg);
1640 // Zero-extend or sign-extend into high-order input register.
1641 if (OpEntry.OpSignExtend) {
1642 if (OpEntry.IsOpSigned)
1643 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1644 TII.get(OpEntry.OpSignExtend));
1645 else {
1646 Register Zero32 = MRI.createVirtualRegister(&X86::GR32RegClass);
1647 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::MOV32r0),
1648 Zero32);
1649
1650 // Copy the zero into the appropriate sub/super/identical physical
1651 // register. Unfortunately the operations needed are not uniform enough
1652 // to fit neatly into the table above.
1653 if (RegTy.getSizeInBits() == 16) {
1654 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
1655 TypeEntry.HighInReg)
1656 .addReg(Zero32, 0, X86::sub_16bit);
1657 } else if (RegTy.getSizeInBits() == 32) {
1658 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
1659 TypeEntry.HighInReg)
1660 .addReg(Zero32);
1661 } else if (RegTy.getSizeInBits() == 64) {
1662 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1663 TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
1664 .addImm(0)
1665 .addReg(Zero32)
1666 .addImm(X86::sub_32bit);
1667 }
1668 }
1669 }
1670 // Generate the DIV/IDIV instruction.
1671 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpDivRem))
1672 .addReg(Op2Reg);
1673 // For i8 remainder, we can't reference ah directly, as we'll end
1674 // up with bogus copies like %r9b = COPY %ah. Reference ax
1675 // instead to prevent ah references in a rex instruction.
1676 //
1677 // The current assumption of the fast register allocator is that isel
1678 // won't generate explicit references to the GR8_NOREX registers. If
1679 // the allocator and/or the backend get enhanced to be more robust in
1680 // that regard, this can be, and should be, removed.
1681 if ((I.getOpcode() == Instruction::SRem ||
1682 I.getOpcode() == Instruction::URem) &&
1683 OpEntry.DivRemResultReg == X86::AH && STI.is64Bit()) {
1684 Register SourceSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
1685 Register ResultSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
1686 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), SourceSuperReg)
1687 .addReg(X86::AX);
1688
1689 // Shift AX right by 8 bits instead of using AH.
1690 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SHR16ri),
1691 ResultSuperReg)
1692 .addReg(SourceSuperReg)
1693 .addImm(8);
1694
1695 // Now reference the 8-bit subreg of the result.
1696 BuildMI(*I.getParent(), I, I.getDebugLoc(),
1697 TII.get(TargetOpcode::SUBREG_TO_REG))
1698 .addDef(DstReg)
1699 .addImm(0)
1700 .addReg(ResultSuperReg)
1701 .addImm(X86::sub_8bit);
1702 } else {
1703 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY),
1704 DstReg)
1705 .addReg(OpEntry.DivRemResultReg);
1706 }
1707 I.eraseFromParent();
1708 return true;
1709 }
1710
selectIntrinsicWSideEffects(MachineInstr & I,MachineRegisterInfo & MRI,MachineFunction & MF) const1711 bool X86InstructionSelector::selectIntrinsicWSideEffects(
1712 MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) const {
1713
1714 assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS &&
1715 "unexpected instruction");
1716
1717 if (I.getOperand(0).getIntrinsicID() != Intrinsic::trap)
1718 return false;
1719
1720 BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TRAP));
1721
1722 I.eraseFromParent();
1723 return true;
1724 }
1725
1726 InstructionSelector *
createX86InstructionSelector(const X86TargetMachine & TM,X86Subtarget & Subtarget,X86RegisterBankInfo & RBI)1727 llvm::createX86InstructionSelector(const X86TargetMachine &TM,
1728 X86Subtarget &Subtarget,
1729 X86RegisterBankInfo &RBI) {
1730 return new X86InstructionSelector(TM, Subtarget, RBI);
1731 }
1732