1 //===----- X86CallFrameOptimization.cpp - Optimize x86 call sequences -----===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines a pass that optimizes call sequences on x86.
10 // Currently, it converts movs of function parameters onto the stack into
11 // pushes. This is beneficial for two main reasons:
12 // 1) The push instruction encoding is much smaller than a stack-ptr-based mov.
13 // 2) It is possible to push memory arguments directly. So, if the
14 // the transformation is performed pre-reg-alloc, it can help relieve
15 // register pressure.
16 //
17 //===----------------------------------------------------------------------===//
18
19 #include "MCTargetDesc/X86BaseInfo.h"
20 #include "X86.h"
21 #include "X86FrameLowering.h"
22 #include "X86InstrInfo.h"
23 #include "X86MachineFunctionInfo.h"
24 #include "X86RegisterInfo.h"
25 #include "X86Subtarget.h"
26 #include "llvm/ADT/DenseSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/StringRef.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineFunctionPass.h"
33 #include "llvm/CodeGen/MachineInstr.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineOperand.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetRegisterInfo.h"
39 #include "llvm/IR/DebugLoc.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/MC/MCDwarf.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include <cassert>
46 #include <cstddef>
47 #include <cstdint>
48 #include <iterator>
49
50 using namespace llvm;
51
52 #define DEBUG_TYPE "x86-cf-opt"
53
54 static cl::opt<bool>
55 NoX86CFOpt("no-x86-call-frame-opt",
56 cl::desc("Avoid optimizing x86 call frames for size"),
57 cl::init(false), cl::Hidden);
58
59 namespace {
60
61 class X86CallFrameOptimization : public MachineFunctionPass {
62 public:
X86CallFrameOptimization()63 X86CallFrameOptimization() : MachineFunctionPass(ID) { }
64
65 bool runOnMachineFunction(MachineFunction &MF) override;
66
67 static char ID;
68
69 private:
70 // Information we know about a particular call site
71 struct CallContext {
CallContext__anon1b6d6e6f0111::X86CallFrameOptimization::CallContext72 CallContext() : FrameSetup(nullptr), ArgStoreVector(4, nullptr) {}
73
74 // Iterator referring to the frame setup instruction
75 MachineBasicBlock::iterator FrameSetup;
76
77 // Actual call instruction
78 MachineInstr *Call = nullptr;
79
80 // A copy of the stack pointer
81 MachineInstr *SPCopy = nullptr;
82
83 // The total displacement of all passed parameters
84 int64_t ExpectedDist = 0;
85
86 // The sequence of storing instructions used to pass the parameters
87 SmallVector<MachineInstr *, 4> ArgStoreVector;
88
89 // True if this call site has no stack parameters
90 bool NoStackParams = false;
91
92 // True if this call site can use push instructions
93 bool UsePush = false;
94 };
95
96 typedef SmallVector<CallContext, 8> ContextVector;
97
98 bool isLegal(MachineFunction &MF);
99
100 bool isProfitable(MachineFunction &MF, ContextVector &CallSeqMap);
101
102 void collectCallInfo(MachineFunction &MF, MachineBasicBlock &MBB,
103 MachineBasicBlock::iterator I, CallContext &Context);
104
105 void adjustCallSequence(MachineFunction &MF, const CallContext &Context);
106
107 MachineInstr *canFoldIntoRegPush(MachineBasicBlock::iterator FrameSetup,
108 Register Reg);
109
110 enum InstClassification { Convert, Skip, Exit };
111
112 InstClassification classifyInstruction(MachineBasicBlock &MBB,
113 MachineBasicBlock::iterator MI,
114 const X86RegisterInfo &RegInfo,
115 DenseSet<unsigned int> &UsedRegs);
116
getPassName() const117 StringRef getPassName() const override { return "X86 Optimize Call Frame"; }
118
119 const X86InstrInfo *TII = nullptr;
120 const X86FrameLowering *TFL = nullptr;
121 const X86Subtarget *STI = nullptr;
122 MachineRegisterInfo *MRI = nullptr;
123 unsigned SlotSize = 0;
124 unsigned Log2SlotSize = 0;
125 };
126
127 } // end anonymous namespace
128 char X86CallFrameOptimization::ID = 0;
129 INITIALIZE_PASS(X86CallFrameOptimization, DEBUG_TYPE,
130 "X86 Call Frame Optimization", false, false)
131
132 // This checks whether the transformation is legal.
133 // Also returns false in cases where it's potentially legal, but
134 // we don't even want to try.
isLegal(MachineFunction & MF)135 bool X86CallFrameOptimization::isLegal(MachineFunction &MF) {
136 if (NoX86CFOpt.getValue())
137 return false;
138
139 // We can't encode multiple DW_CFA_GNU_args_size or DW_CFA_def_cfa_offset
140 // in the compact unwind encoding that Darwin uses. So, bail if there
141 // is a danger of that being generated.
142 if (STI->isTargetDarwin() &&
143 (!MF.getLandingPads().empty() ||
144 (MF.getFunction().needsUnwindTableEntry() && !TFL->hasFP(MF))))
145 return false;
146
147 // It is not valid to change the stack pointer outside the prolog/epilog
148 // on 64-bit Windows.
149 if (STI->isTargetWin64())
150 return false;
151
152 // You would expect straight-line code between call-frame setup and
153 // call-frame destroy. You would be wrong. There are circumstances (e.g.
154 // CMOV_GR8 expansion of a select that feeds a function call!) where we can
155 // end up with the setup and the destroy in different basic blocks.
156 // This is bad, and breaks SP adjustment.
157 // So, check that all of the frames in the function are closed inside
158 // the same block, and, for good measure, that there are no nested frames.
159 //
160 // If any call allocates more argument stack memory than the stack
161 // probe size, don't do this optimization. Otherwise, this pass
162 // would need to synthesize additional stack probe calls to allocate
163 // memory for arguments.
164 unsigned FrameSetupOpcode = TII->getCallFrameSetupOpcode();
165 unsigned FrameDestroyOpcode = TII->getCallFrameDestroyOpcode();
166 bool EmitStackProbeCall = STI->getTargetLowering()->hasStackProbeSymbol(MF);
167 unsigned StackProbeSize = STI->getTargetLowering()->getStackProbeSize(MF);
168 for (MachineBasicBlock &BB : MF) {
169 bool InsideFrameSequence = false;
170 for (MachineInstr &MI : BB) {
171 if (MI.getOpcode() == FrameSetupOpcode) {
172 if (TII->getFrameSize(MI) >= StackProbeSize && EmitStackProbeCall)
173 return false;
174 if (InsideFrameSequence)
175 return false;
176 InsideFrameSequence = true;
177 } else if (MI.getOpcode() == FrameDestroyOpcode) {
178 if (!InsideFrameSequence)
179 return false;
180 InsideFrameSequence = false;
181 }
182 }
183
184 if (InsideFrameSequence)
185 return false;
186 }
187
188 return true;
189 }
190
191 // Check whether this transformation is profitable for a particular
192 // function - in terms of code size.
isProfitable(MachineFunction & MF,ContextVector & CallSeqVector)193 bool X86CallFrameOptimization::isProfitable(MachineFunction &MF,
194 ContextVector &CallSeqVector) {
195 // This transformation is always a win when we do not expect to have
196 // a reserved call frame. Under other circumstances, it may be either
197 // a win or a loss, and requires a heuristic.
198 bool CannotReserveFrame = MF.getFrameInfo().hasVarSizedObjects();
199 if (CannotReserveFrame)
200 return true;
201
202 Align StackAlign = TFL->getStackAlign();
203
204 int64_t Advantage = 0;
205 for (const auto &CC : CallSeqVector) {
206 // Call sites where no parameters are passed on the stack
207 // do not affect the cost, since there needs to be no
208 // stack adjustment.
209 if (CC.NoStackParams)
210 continue;
211
212 if (!CC.UsePush) {
213 // If we don't use pushes for a particular call site,
214 // we pay for not having a reserved call frame with an
215 // additional sub/add esp pair. The cost is ~3 bytes per instruction,
216 // depending on the size of the constant.
217 // TODO: Callee-pop functions should have a smaller penalty, because
218 // an add is needed even with a reserved call frame.
219 Advantage -= 6;
220 } else {
221 // We can use pushes. First, account for the fixed costs.
222 // We'll need a add after the call.
223 Advantage -= 3;
224 // If we have to realign the stack, we'll also need a sub before
225 if (!isAligned(StackAlign, CC.ExpectedDist))
226 Advantage -= 3;
227 // Now, for each push, we save ~3 bytes. For small constants, we actually,
228 // save more (up to 5 bytes), but 3 should be a good approximation.
229 Advantage += (CC.ExpectedDist >> Log2SlotSize) * 3;
230 }
231 }
232
233 return Advantage >= 0;
234 }
235
runOnMachineFunction(MachineFunction & MF)236 bool X86CallFrameOptimization::runOnMachineFunction(MachineFunction &MF) {
237 STI = &MF.getSubtarget<X86Subtarget>();
238 TII = STI->getInstrInfo();
239 TFL = STI->getFrameLowering();
240 MRI = &MF.getRegInfo();
241
242 const X86RegisterInfo &RegInfo =
243 *static_cast<const X86RegisterInfo *>(STI->getRegisterInfo());
244 SlotSize = RegInfo.getSlotSize();
245 assert(isPowerOf2_32(SlotSize) && "Expect power of 2 stack slot size");
246 Log2SlotSize = Log2_32(SlotSize);
247
248 if (skipFunction(MF.getFunction()) || !isLegal(MF))
249 return false;
250
251 unsigned FrameSetupOpcode = TII->getCallFrameSetupOpcode();
252
253 bool Changed = false;
254
255 ContextVector CallSeqVector;
256
257 for (auto &MBB : MF)
258 for (auto &MI : MBB)
259 if (MI.getOpcode() == FrameSetupOpcode) {
260 CallContext Context;
261 collectCallInfo(MF, MBB, MI, Context);
262 CallSeqVector.push_back(Context);
263 }
264
265 if (!isProfitable(MF, CallSeqVector))
266 return false;
267
268 for (const auto &CC : CallSeqVector) {
269 if (CC.UsePush) {
270 adjustCallSequence(MF, CC);
271 Changed = true;
272 }
273 }
274
275 return Changed;
276 }
277
278 X86CallFrameOptimization::InstClassification
classifyInstruction(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,const X86RegisterInfo & RegInfo,DenseSet<unsigned int> & UsedRegs)279 X86CallFrameOptimization::classifyInstruction(
280 MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
281 const X86RegisterInfo &RegInfo, DenseSet<unsigned int> &UsedRegs) {
282 if (MI == MBB.end())
283 return Exit;
284
285 // The instructions we actually care about are movs onto the stack or special
286 // cases of constant-stores to stack
287 switch (MI->getOpcode()) {
288 case X86::AND16mi:
289 case X86::AND32mi:
290 case X86::AND64mi32: {
291 const MachineOperand &ImmOp = MI->getOperand(X86::AddrNumOperands);
292 return ImmOp.getImm() == 0 ? Convert : Exit;
293 }
294 case X86::OR16mi:
295 case X86::OR32mi:
296 case X86::OR64mi32: {
297 const MachineOperand &ImmOp = MI->getOperand(X86::AddrNumOperands);
298 return ImmOp.getImm() == -1 ? Convert : Exit;
299 }
300 case X86::MOV32mi:
301 case X86::MOV32mr:
302 case X86::MOV64mi32:
303 case X86::MOV64mr:
304 return Convert;
305 }
306
307 // Not all calling conventions have only stack MOVs between the stack
308 // adjust and the call.
309
310 // We want to tolerate other instructions, to cover more cases.
311 // In particular:
312 // a) PCrel calls, where we expect an additional COPY of the basereg.
313 // b) Passing frame-index addresses.
314 // c) Calling conventions that have inreg parameters. These generate
315 // both copies and movs into registers.
316 // To avoid creating lots of special cases, allow any instruction
317 // that does not write into memory, does not def or use the stack
318 // pointer, and does not def any register that was used by a preceding
319 // push.
320 // (Reading from memory is allowed, even if referenced through a
321 // frame index, since these will get adjusted properly in PEI)
322
323 // The reason for the last condition is that the pushes can't replace
324 // the movs in place, because the order must be reversed.
325 // So if we have a MOV32mr that uses EDX, then an instruction that defs
326 // EDX, and then the call, after the transformation the push will use
327 // the modified version of EDX, and not the original one.
328 // Since we are still in SSA form at this point, we only need to
329 // make sure we don't clobber any *physical* registers that were
330 // used by an earlier mov that will become a push.
331
332 if (MI->isCall() || MI->mayStore())
333 return Exit;
334
335 for (const MachineOperand &MO : MI->operands()) {
336 if (!MO.isReg())
337 continue;
338 Register Reg = MO.getReg();
339 if (!Reg.isPhysical())
340 continue;
341 if (RegInfo.regsOverlap(Reg, RegInfo.getStackRegister()))
342 return Exit;
343 if (MO.isDef()) {
344 for (unsigned int U : UsedRegs)
345 if (RegInfo.regsOverlap(Reg, U))
346 return Exit;
347 }
348 }
349
350 return Skip;
351 }
352
collectCallInfo(MachineFunction & MF,MachineBasicBlock & MBB,MachineBasicBlock::iterator I,CallContext & Context)353 void X86CallFrameOptimization::collectCallInfo(MachineFunction &MF,
354 MachineBasicBlock &MBB,
355 MachineBasicBlock::iterator I,
356 CallContext &Context) {
357 // Check that this particular call sequence is amenable to the
358 // transformation.
359 const X86RegisterInfo &RegInfo =
360 *static_cast<const X86RegisterInfo *>(STI->getRegisterInfo());
361
362 // We expect to enter this at the beginning of a call sequence
363 assert(I->getOpcode() == TII->getCallFrameSetupOpcode());
364 MachineBasicBlock::iterator FrameSetup = I++;
365 Context.FrameSetup = FrameSetup;
366
367 // How much do we adjust the stack? This puts an upper bound on
368 // the number of parameters actually passed on it.
369 unsigned int MaxAdjust = TII->getFrameSize(*FrameSetup) >> Log2SlotSize;
370
371 // A zero adjustment means no stack parameters
372 if (!MaxAdjust) {
373 Context.NoStackParams = true;
374 return;
375 }
376
377 // Skip over DEBUG_VALUE.
378 // For globals in PIC mode, we can have some LEAs here. Skip them as well.
379 // TODO: Extend this to something that covers more cases.
380 while (I->getOpcode() == X86::LEA32r || I->isDebugInstr())
381 ++I;
382
383 Register StackPtr = RegInfo.getStackRegister();
384 auto StackPtrCopyInst = MBB.end();
385 // SelectionDAG (but not FastISel) inserts a copy of ESP into a virtual
386 // register. If it's there, use that virtual register as stack pointer
387 // instead. Also, we need to locate this instruction so that we can later
388 // safely ignore it while doing the conservative processing of the call chain.
389 // The COPY can be located anywhere between the call-frame setup
390 // instruction and its first use. We use the call instruction as a boundary
391 // because it is usually cheaper to check if an instruction is a call than
392 // checking if an instruction uses a register.
393 for (auto J = I; !J->isCall(); ++J)
394 if (J->isCopy() && J->getOperand(0).isReg() && J->getOperand(1).isReg() &&
395 J->getOperand(1).getReg() == StackPtr) {
396 StackPtrCopyInst = J;
397 Context.SPCopy = &*J++;
398 StackPtr = Context.SPCopy->getOperand(0).getReg();
399 break;
400 }
401
402 // Scan the call setup sequence for the pattern we're looking for.
403 // We only handle a simple case - a sequence of store instructions that
404 // push a sequence of stack-slot-aligned values onto the stack, with
405 // no gaps between them.
406 if (MaxAdjust > 4)
407 Context.ArgStoreVector.resize(MaxAdjust, nullptr);
408
409 DenseSet<unsigned int> UsedRegs;
410
411 for (InstClassification Classification = Skip; Classification != Exit; ++I) {
412 // If this is the COPY of the stack pointer, it's ok to ignore.
413 if (I == StackPtrCopyInst)
414 continue;
415 Classification = classifyInstruction(MBB, I, RegInfo, UsedRegs);
416 if (Classification != Convert)
417 continue;
418 // We know the instruction has a supported store opcode.
419 // We only want movs of the form:
420 // mov imm/reg, k(%StackPtr)
421 // If we run into something else, bail.
422 // Note that AddrBaseReg may, counter to its name, not be a register,
423 // but rather a frame index.
424 // TODO: Support the fi case. This should probably work now that we
425 // have the infrastructure to track the stack pointer within a call
426 // sequence.
427 if (!I->getOperand(X86::AddrBaseReg).isReg() ||
428 (I->getOperand(X86::AddrBaseReg).getReg() != StackPtr) ||
429 !I->getOperand(X86::AddrScaleAmt).isImm() ||
430 (I->getOperand(X86::AddrScaleAmt).getImm() != 1) ||
431 (I->getOperand(X86::AddrIndexReg).getReg() != X86::NoRegister) ||
432 (I->getOperand(X86::AddrSegmentReg).getReg() != X86::NoRegister) ||
433 !I->getOperand(X86::AddrDisp).isImm())
434 return;
435
436 int64_t StackDisp = I->getOperand(X86::AddrDisp).getImm();
437 assert(StackDisp >= 0 &&
438 "Negative stack displacement when passing parameters");
439
440 // We really don't want to consider the unaligned case.
441 if (StackDisp & (SlotSize - 1))
442 return;
443 StackDisp >>= Log2SlotSize;
444
445 assert((size_t)StackDisp < Context.ArgStoreVector.size() &&
446 "Function call has more parameters than the stack is adjusted for.");
447
448 // If the same stack slot is being filled twice, something's fishy.
449 if (Context.ArgStoreVector[StackDisp] != nullptr)
450 return;
451 Context.ArgStoreVector[StackDisp] = &*I;
452
453 for (const MachineOperand &MO : I->uses()) {
454 if (!MO.isReg())
455 continue;
456 Register Reg = MO.getReg();
457 if (Reg.isPhysical())
458 UsedRegs.insert(Reg);
459 }
460 }
461
462 --I;
463
464 // We now expect the end of the sequence. If we stopped early,
465 // or reached the end of the block without finding a call, bail.
466 if (I == MBB.end() || !I->isCall())
467 return;
468
469 Context.Call = &*I;
470 if ((++I)->getOpcode() != TII->getCallFrameDestroyOpcode())
471 return;
472
473 // Now, go through the vector, and see that we don't have any gaps,
474 // but only a series of storing instructions.
475 auto MMI = Context.ArgStoreVector.begin(), MME = Context.ArgStoreVector.end();
476 for (; MMI != MME; ++MMI, Context.ExpectedDist += SlotSize)
477 if (*MMI == nullptr)
478 break;
479
480 // If the call had no parameters, do nothing
481 if (MMI == Context.ArgStoreVector.begin())
482 return;
483
484 // We are either at the last parameter, or a gap.
485 // Make sure it's not a gap
486 for (; MMI != MME; ++MMI)
487 if (*MMI != nullptr)
488 return;
489
490 Context.UsePush = true;
491 }
492
adjustCallSequence(MachineFunction & MF,const CallContext & Context)493 void X86CallFrameOptimization::adjustCallSequence(MachineFunction &MF,
494 const CallContext &Context) {
495 // Ok, we can in fact do the transformation for this call.
496 // Do not remove the FrameSetup instruction, but adjust the parameters.
497 // PEI will end up finalizing the handling of this.
498 MachineBasicBlock::iterator FrameSetup = Context.FrameSetup;
499 MachineBasicBlock &MBB = *(FrameSetup->getParent());
500 TII->setFrameAdjustment(*FrameSetup, Context.ExpectedDist);
501
502 const DebugLoc &DL = FrameSetup->getDebugLoc();
503 bool Is64Bit = STI->is64Bit();
504 // Now, iterate through the vector in reverse order, and replace the store to
505 // stack with pushes. MOVmi/MOVmr doesn't have any defs, so no need to
506 // replace uses.
507 for (int Idx = (Context.ExpectedDist >> Log2SlotSize) - 1; Idx >= 0; --Idx) {
508 MachineBasicBlock::iterator Store = *Context.ArgStoreVector[Idx];
509 const MachineOperand &PushOp = Store->getOperand(X86::AddrNumOperands);
510 MachineBasicBlock::iterator Push = nullptr;
511 unsigned PushOpcode;
512 switch (Store->getOpcode()) {
513 default:
514 llvm_unreachable("Unexpected Opcode!");
515 case X86::AND16mi:
516 case X86::AND32mi:
517 case X86::AND64mi32:
518 case X86::OR16mi:
519 case X86::OR32mi:
520 case X86::OR64mi32:
521 case X86::MOV32mi:
522 case X86::MOV64mi32:
523 PushOpcode = Is64Bit ? X86::PUSH64i32 : X86::PUSH32i;
524 Push = BuildMI(MBB, Context.Call, DL, TII->get(PushOpcode)).add(PushOp);
525 Push->cloneMemRefs(MF, *Store);
526 break;
527 case X86::MOV32mr:
528 case X86::MOV64mr: {
529 Register Reg = PushOp.getReg();
530
531 // If storing a 32-bit vreg on 64-bit targets, extend to a 64-bit vreg
532 // in preparation for the PUSH64. The upper 32 bits can be undef.
533 if (Is64Bit && Store->getOpcode() == X86::MOV32mr) {
534 Register UndefReg = MRI->createVirtualRegister(&X86::GR64RegClass);
535 Reg = MRI->createVirtualRegister(&X86::GR64RegClass);
536 BuildMI(MBB, Context.Call, DL, TII->get(X86::IMPLICIT_DEF), UndefReg);
537 BuildMI(MBB, Context.Call, DL, TII->get(X86::INSERT_SUBREG), Reg)
538 .addReg(UndefReg)
539 .add(PushOp)
540 .addImm(X86::sub_32bit);
541 }
542
543 // If PUSHrmm is not slow on this target, try to fold the source of the
544 // push into the instruction.
545 bool SlowPUSHrmm = STI->slowTwoMemOps();
546
547 // Check that this is legal to fold. Right now, we're extremely
548 // conservative about that.
549 MachineInstr *DefMov = nullptr;
550 if (!SlowPUSHrmm && (DefMov = canFoldIntoRegPush(FrameSetup, Reg))) {
551 PushOpcode = Is64Bit ? X86::PUSH64rmm : X86::PUSH32rmm;
552 Push = BuildMI(MBB, Context.Call, DL, TII->get(PushOpcode));
553
554 unsigned NumOps = DefMov->getDesc().getNumOperands();
555 for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
556 Push->addOperand(DefMov->getOperand(i));
557 Push->cloneMergedMemRefs(MF, {DefMov, &*Store});
558 DefMov->eraseFromParent();
559 } else {
560 PushOpcode = Is64Bit ? X86::PUSH64r : X86::PUSH32r;
561 Push = BuildMI(MBB, Context.Call, DL, TII->get(PushOpcode))
562 .addReg(Reg)
563 .getInstr();
564 Push->cloneMemRefs(MF, *Store);
565 }
566 break;
567 }
568 }
569
570 // For debugging, when using SP-based CFA, we need to adjust the CFA
571 // offset after each push.
572 // TODO: This is needed only if we require precise CFA.
573 if (!TFL->hasFP(MF))
574 TFL->BuildCFI(
575 MBB, std::next(Push), DL,
576 MCCFIInstruction::createAdjustCfaOffset(nullptr, SlotSize));
577
578 MBB.erase(Store);
579 }
580
581 // The stack-pointer copy is no longer used in the call sequences.
582 // There should not be any other users, but we can't commit to that, so:
583 if (Context.SPCopy && MRI->use_empty(Context.SPCopy->getOperand(0).getReg()))
584 Context.SPCopy->eraseFromParent();
585
586 // Once we've done this, we need to make sure PEI doesn't assume a reserved
587 // frame.
588 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
589 FuncInfo->setHasPushSequences(true);
590 }
591
canFoldIntoRegPush(MachineBasicBlock::iterator FrameSetup,Register Reg)592 MachineInstr *X86CallFrameOptimization::canFoldIntoRegPush(
593 MachineBasicBlock::iterator FrameSetup, Register Reg) {
594 // Do an extremely restricted form of load folding.
595 // ISel will often create patterns like:
596 // movl 4(%edi), %eax
597 // movl 8(%edi), %ecx
598 // movl 12(%edi), %edx
599 // movl %edx, 8(%esp)
600 // movl %ecx, 4(%esp)
601 // movl %eax, (%esp)
602 // call
603 // Get rid of those with prejudice.
604 if (!Reg.isVirtual())
605 return nullptr;
606
607 // Make sure this is the only use of Reg.
608 if (!MRI->hasOneNonDBGUse(Reg))
609 return nullptr;
610
611 MachineInstr &DefMI = *MRI->getVRegDef(Reg);
612
613 // Make sure the def is a MOV from memory.
614 // If the def is in another block, give up.
615 if ((DefMI.getOpcode() != X86::MOV32rm &&
616 DefMI.getOpcode() != X86::MOV64rm) ||
617 DefMI.getParent() != FrameSetup->getParent())
618 return nullptr;
619
620 // Make sure we don't have any instructions between DefMI and the
621 // push that make folding the load illegal.
622 for (MachineBasicBlock::iterator I = DefMI; I != FrameSetup; ++I)
623 if (I->isLoadFoldBarrier())
624 return nullptr;
625
626 return &DefMI;
627 }
628
createX86CallFrameOptimization()629 FunctionPass *llvm::createX86CallFrameOptimization() {
630 return new X86CallFrameOptimization();
631 }
632