1 //===- PrologEpilogInserter.cpp - Insert Prolog/Epilog code in function ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass is responsible for finalizing the functions frame layout, saving
11 // callee saved registers, and for emitting prolog & epilog code for the
14 // This pass must be run after register allocation. After this pass is
15 // executed, it is illegal to construct MO_FrameIndex operands.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/BitVector.h"
21 #include "llvm/ADT/DepthFirstIterator.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/SetVector.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/SmallVector.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineDominators.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineInstr.h"
35 #include "llvm/CodeGen/MachineLoopInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineOperand.h"
38 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
39 #include "llvm/CodeGen/MachineRegisterInfo.h"
40 #include "llvm/CodeGen/RegisterScavenging.h"
41 #include "llvm/CodeGen/TargetFrameLowering.h"
42 #include "llvm/CodeGen/TargetInstrInfo.h"
43 #include "llvm/CodeGen/TargetOpcodes.h"
44 #include "llvm/CodeGen/TargetRegisterInfo.h"
45 #include "llvm/CodeGen/TargetSubtargetInfo.h"
46 #include "llvm/CodeGen/WinEHFuncInfo.h"
47 #include "llvm/IR/Attributes.h"
48 #include "llvm/IR/CallingConv.h"
49 #include "llvm/IR/DebugInfoMetadata.h"
50 #include "llvm/IR/DiagnosticInfo.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/InlineAsm.h"
53 #include "llvm/IR/LLVMContext.h"
54 #include "llvm/MC/MCRegisterInfo.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/CodeGen.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Debug.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Support/MathExtras.h"
61 #include "llvm/Support/raw_ostream.h"
62 #include "llvm/Target/TargetMachine.h"
63 #include "llvm/Target/TargetOptions.h"
74 #define DEBUG_TYPE "prologepilog"
76 using MBBVector = SmallVector<MachineBasicBlock *, 4>;
78 STATISTIC(NumLeafFuncWithSpills, "Number of leaf functions with CSRs");
79 STATISTIC(NumFuncSeen, "Number of functions seen in PEI");
84 class PEI : public MachineFunctionPass {
88 PEI() : MachineFunctionPass(ID) {
89 initializePEIPass(*PassRegistry::getPassRegistry());
92 void getAnalysisUsage(AnalysisUsage &AU) const override;
94 /// runOnMachineFunction - Insert prolog/epilog code and replace abstract
95 /// frame indexes with appropriate references.
96 bool runOnMachineFunction(MachineFunction &MF) override;
101 // MinCSFrameIndex, MaxCSFrameIndex - Keeps the range of callee saved
102 // stack frame indexes.
103 unsigned MinCSFrameIndex = std::numeric_limits<unsigned>::max();
104 unsigned MaxCSFrameIndex = 0;
106 // Save and Restore blocks of the current function. Typically there is a
107 // single save block, unless Windows EH funclets are involved.
108 MBBVector SaveBlocks;
109 MBBVector RestoreBlocks;
111 // Flag to control whether to use the register scavenger to resolve
112 // frame index materialization registers. Set according to
113 // TRI->requiresFrameIndexScavenging() for the current function.
114 bool FrameIndexVirtualScavenging;
116 // Flag to control whether the scavenger should be passed even though
117 // FrameIndexVirtualScavenging is used.
118 bool FrameIndexEliminationScavenging;
121 MachineOptimizationRemarkEmitter *ORE = nullptr;
123 void calculateCallFrameInfo(MachineFunction &MF);
124 void calculateSaveRestoreBlocks(MachineFunction &MF);
125 void spillCalleeSavedRegs(MachineFunction &MF);
127 void calculateFrameObjectOffsets(MachineFunction &MF);
128 void replaceFrameIndices(MachineFunction &MF);
129 void replaceFrameIndices(MachineBasicBlock *BB, MachineFunction &MF,
131 void insertPrologEpilogCode(MachineFunction &MF);
134 } // end anonymous namespace
138 char &llvm::PrologEpilogCodeInserterID = PEI::ID;
140 static cl::opt<unsigned>
141 WarnStackSize("warn-stack-size", cl::Hidden, cl::init((unsigned)-1),
142 cl::desc("Warn for stack size bigger than the given"
145 INITIALIZE_PASS_BEGIN(PEI, DEBUG_TYPE, "Prologue/Epilogue Insertion", false,
147 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
148 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
149 INITIALIZE_PASS_DEPENDENCY(MachineOptimizationRemarkEmitterPass)
150 INITIALIZE_PASS_END(PEI, DEBUG_TYPE,
151 "Prologue/Epilogue Insertion & Frame Finalization", false,
154 MachineFunctionPass *llvm::createPrologEpilogInserterPass() {
158 STATISTIC(NumBytesStackSpace,
159 "Number of bytes used for stack in all functions");
161 void PEI::getAnalysisUsage(AnalysisUsage &AU) const {
162 AU.setPreservesCFG();
163 AU.addPreserved<MachineLoopInfo>();
164 AU.addPreserved<MachineDominatorTree>();
165 AU.addRequired<MachineOptimizationRemarkEmitterPass>();
166 MachineFunctionPass::getAnalysisUsage(AU);
169 /// StackObjSet - A set of stack object indexes
170 using StackObjSet = SmallSetVector<int, 8>;
172 /// runOnMachineFunction - Insert prolog/epilog code and replace abstract
173 /// frame indexes with appropriate references.
174 bool PEI::runOnMachineFunction(MachineFunction &MF) {
176 const Function &F = MF.getFunction();
177 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
178 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
180 RS = TRI->requiresRegisterScavenging(MF) ? new RegScavenger() : nullptr;
181 FrameIndexVirtualScavenging = TRI->requiresFrameIndexScavenging(MF);
182 FrameIndexEliminationScavenging = (RS && !FrameIndexVirtualScavenging) ||
183 TRI->requiresFrameIndexReplacementScavenging(MF);
184 ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE();
186 // Calculate the MaxCallFrameSize and AdjustsStack variables for the
187 // function's frame information. Also eliminates call frame pseudo
189 calculateCallFrameInfo(MF);
191 // Determine placement of CSR spill/restore code and prolog/epilog code:
192 // place all spills in the entry block, all restores in return blocks.
193 calculateSaveRestoreBlocks(MF);
195 // Handle CSR spilling and restoring, for targets that need it.
196 if (MF.getTarget().usesPhysRegsForPEI())
197 spillCalleeSavedRegs(MF);
199 // Allow the target machine to make final modifications to the function
200 // before the frame layout is finalized.
201 TFI->processFunctionBeforeFrameFinalized(MF, RS);
203 // Calculate actual frame offsets for all abstract stack objects...
204 calculateFrameObjectOffsets(MF);
206 // Add prolog and epilog code to the function. This function is required
207 // to align the stack frame as necessary for any stack variables or
208 // called functions. Because of this, calculateCalleeSavedRegisters()
209 // must be called before this function in order to set the AdjustsStack
210 // and MaxCallFrameSize variables.
211 if (!F.hasFnAttribute(Attribute::Naked))
212 insertPrologEpilogCode(MF);
214 // Replace all MO_FrameIndex operands with physical register references
215 // and actual offsets.
217 replaceFrameIndices(MF);
219 // If register scavenging is needed, as we've enabled doing it as a
220 // post-pass, scavenge the virtual registers that frame index elimination
222 if (TRI->requiresRegisterScavenging(MF) && FrameIndexVirtualScavenging)
223 scavengeFrameVirtualRegs(MF, *RS);
225 // Warn on stack size when we exceeds the given limit.
226 MachineFrameInfo &MFI = MF.getFrameInfo();
227 uint64_t StackSize = MFI.getStackSize();
228 if (WarnStackSize.getNumOccurrences() > 0 && WarnStackSize < StackSize) {
229 DiagnosticInfoStackSize DiagStackSize(F, StackSize);
230 F.getContext().diagnose(DiagStackSize);
233 return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "StackSize",
234 MF.getFunction().getSubprogram(),
236 << ore::NV("NumStackBytes", StackSize) << " stack bytes in function";
241 RestoreBlocks.clear();
242 MFI.setSavePoint(nullptr);
243 MFI.setRestorePoint(nullptr);
247 /// Calculate the MaxCallFrameSize and AdjustsStack
248 /// variables for the function's frame information and eliminate call frame
249 /// pseudo instructions.
250 void PEI::calculateCallFrameInfo(MachineFunction &MF) {
251 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
252 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
253 MachineFrameInfo &MFI = MF.getFrameInfo();
255 unsigned MaxCallFrameSize = 0;
256 bool AdjustsStack = MFI.adjustsStack();
258 // Get the function call frame set-up and tear-down instruction opcode
259 unsigned FrameSetupOpcode = TII.getCallFrameSetupOpcode();
260 unsigned FrameDestroyOpcode = TII.getCallFrameDestroyOpcode();
262 // Early exit for targets which have no call frame setup/destroy pseudo
264 if (FrameSetupOpcode == ~0u && FrameDestroyOpcode == ~0u)
267 std::vector<MachineBasicBlock::iterator> FrameSDOps;
268 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
269 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
270 if (TII.isFrameInstr(*I)) {
271 unsigned Size = TII.getFrameSize(*I);
272 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size;
274 FrameSDOps.push_back(I);
275 } else if (I->isInlineAsm()) {
276 // Some inline asm's need a stack frame, as indicated by operand 1.
277 unsigned ExtraInfo = I->getOperand(InlineAsm::MIOp_ExtraInfo).getImm();
278 if (ExtraInfo & InlineAsm::Extra_IsAlignStack)
282 assert(!MFI.isMaxCallFrameSizeComputed() ||
283 (MFI.getMaxCallFrameSize() == MaxCallFrameSize &&
284 MFI.adjustsStack() == AdjustsStack));
285 MFI.setAdjustsStack(AdjustsStack);
286 MFI.setMaxCallFrameSize(MaxCallFrameSize);
288 for (std::vector<MachineBasicBlock::iterator>::iterator
289 i = FrameSDOps.begin(), e = FrameSDOps.end(); i != e; ++i) {
290 MachineBasicBlock::iterator I = *i;
292 // If call frames are not being included as part of the stack frame, and
293 // the target doesn't indicate otherwise, remove the call frame pseudos
294 // here. The sub/add sp instruction pairs are still inserted, but we don't
295 // need to track the SP adjustment for frame index elimination.
296 if (TFI->canSimplifyCallFramePseudos(MF))
297 TFI->eliminateCallFramePseudoInstr(MF, *I->getParent(), I);
301 /// Compute the sets of entry and return blocks for saving and restoring
302 /// callee-saved registers, and placing prolog and epilog code.
303 void PEI::calculateSaveRestoreBlocks(MachineFunction &MF) {
304 const MachineFrameInfo &MFI = MF.getFrameInfo();
306 // Even when we do not change any CSR, we still want to insert the
307 // prologue and epilogue of the function.
308 // So set the save points for those.
310 // Use the points found by shrink-wrapping, if any.
311 if (MFI.getSavePoint()) {
312 SaveBlocks.push_back(MFI.getSavePoint());
313 assert(MFI.getRestorePoint() && "Both restore and save must be set");
314 MachineBasicBlock *RestoreBlock = MFI.getRestorePoint();
315 // If RestoreBlock does not have any successor and is not a return block
316 // then the end point is unreachable and we do not need to insert any
318 if (!RestoreBlock->succ_empty() || RestoreBlock->isReturnBlock())
319 RestoreBlocks.push_back(RestoreBlock);
323 // Save refs to entry and return blocks.
324 SaveBlocks.push_back(&MF.front());
325 for (MachineBasicBlock &MBB : MF) {
326 if (MBB.isEHFuncletEntry())
327 SaveBlocks.push_back(&MBB);
328 if (MBB.isReturnBlock())
329 RestoreBlocks.push_back(&MBB);
333 static void assignCalleeSavedSpillSlots(MachineFunction &F,
334 const BitVector &SavedRegs,
335 unsigned &MinCSFrameIndex,
336 unsigned &MaxCSFrameIndex) {
337 if (SavedRegs.empty())
340 const TargetRegisterInfo *RegInfo = F.getSubtarget().getRegisterInfo();
341 const MCPhysReg *CSRegs = F.getRegInfo().getCalleeSavedRegs();
343 std::vector<CalleeSavedInfo> CSI;
344 for (unsigned i = 0; CSRegs[i]; ++i) {
345 unsigned Reg = CSRegs[i];
346 if (SavedRegs.test(Reg))
347 CSI.push_back(CalleeSavedInfo(Reg));
350 const TargetFrameLowering *TFI = F.getSubtarget().getFrameLowering();
351 MachineFrameInfo &MFI = F.getFrameInfo();
352 if (!TFI->assignCalleeSavedSpillSlots(F, RegInfo, CSI)) {
353 // If target doesn't implement this, use generic code.
356 return; // Early exit if no callee saved registers are modified!
358 unsigned NumFixedSpillSlots;
359 const TargetFrameLowering::SpillSlot *FixedSpillSlots =
360 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots);
362 // Now that we know which registers need to be saved and restored, allocate
363 // stack slots for them.
364 for (auto &CS : CSI) {
365 // If the target has spilled this register to another register, we don't
366 // need to allocate a stack slot.
367 if (CS.isSpilledToReg())
370 unsigned Reg = CS.getReg();
371 const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg);
374 if (RegInfo->hasReservedSpillSlot(F, Reg, FrameIdx)) {
375 CS.setFrameIdx(FrameIdx);
379 // Check to see if this physreg must be spilled to a particular stack slot
381 const TargetFrameLowering::SpillSlot *FixedSlot = FixedSpillSlots;
382 while (FixedSlot != FixedSpillSlots + NumFixedSpillSlots &&
383 FixedSlot->Reg != Reg)
386 unsigned Size = RegInfo->getSpillSize(*RC);
387 if (FixedSlot == FixedSpillSlots + NumFixedSpillSlots) {
388 // Nope, just spill it anywhere convenient.
389 unsigned Align = RegInfo->getSpillAlignment(*RC);
390 unsigned StackAlign = TFI->getStackAlignment();
392 // We may not be able to satisfy the desired alignment specification of
393 // the TargetRegisterClass if the stack alignment is smaller. Use the
395 Align = std::min(Align, StackAlign);
396 FrameIdx = MFI.CreateStackObject(Size, Align, true);
397 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx;
398 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx;
400 // Spill it to the stack where we must.
401 FrameIdx = MFI.CreateFixedSpillStackObject(Size, FixedSlot->Offset);
404 CS.setFrameIdx(FrameIdx);
408 MFI.setCalleeSavedInfo(CSI);
411 /// Helper function to update the liveness information for the callee-saved
413 static void updateLiveness(MachineFunction &MF) {
414 MachineFrameInfo &MFI = MF.getFrameInfo();
415 // Visited will contain all the basic blocks that are in the region
416 // where the callee saved registers are alive:
417 // - Anything that is not Save or Restore -> LiveThrough.
419 // - Restore -> LiveOut.
420 // The live-out is not attached to the block, so no need to keep
421 // Restore in this set.
422 SmallPtrSet<MachineBasicBlock *, 8> Visited;
423 SmallVector<MachineBasicBlock *, 8> WorkList;
424 MachineBasicBlock *Entry = &MF.front();
425 MachineBasicBlock *Save = MFI.getSavePoint();
431 WorkList.push_back(Entry);
432 Visited.insert(Entry);
434 Visited.insert(Save);
436 MachineBasicBlock *Restore = MFI.getRestorePoint();
438 // By construction Restore cannot be visited, otherwise it
439 // means there exists a path to Restore that does not go
441 WorkList.push_back(Restore);
443 while (!WorkList.empty()) {
444 const MachineBasicBlock *CurBB = WorkList.pop_back_val();
445 // By construction, the region that is after the save point is
446 // dominated by the Save and post-dominated by the Restore.
447 if (CurBB == Save && Save != Restore)
449 // Enqueue all the successors not already visited.
450 // Those are by construction either before Save or after Restore.
451 for (MachineBasicBlock *SuccBB : CurBB->successors())
452 if (Visited.insert(SuccBB).second)
453 WorkList.push_back(SuccBB);
456 const std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
458 MachineRegisterInfo &MRI = MF.getRegInfo();
459 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
460 for (MachineBasicBlock *MBB : Visited) {
461 MCPhysReg Reg = CSI[i].getReg();
462 // Add the callee-saved register as live-in.
463 // It's killed at the spill.
464 if (!MRI.isReserved(Reg) && !MBB->isLiveIn(Reg))
467 // If callee-saved register is spilled to another register rather than
468 // spilling to stack, the destination register has to be marked as live for
469 // each MBB between the prologue and epilogue so that it is not clobbered
470 // before it is reloaded in the epilogue. The Visited set contains all
471 // blocks outside of the region delimited by prologue/epilogue.
472 if (CSI[i].isSpilledToReg()) {
473 for (MachineBasicBlock &MBB : MF) {
474 if (Visited.count(&MBB))
476 MCPhysReg DstReg = CSI[i].getDstReg();
477 if (!MBB.isLiveIn(DstReg))
478 MBB.addLiveIn(DstReg);
485 /// Insert restore code for the callee-saved registers used in the function.
486 static void insertCSRSaves(MachineBasicBlock &SaveBlock,
487 ArrayRef<CalleeSavedInfo> CSI) {
488 MachineFunction &MF = *SaveBlock.getParent();
489 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
490 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
491 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
493 MachineBasicBlock::iterator I = SaveBlock.begin();
494 if (!TFI->spillCalleeSavedRegisters(SaveBlock, I, CSI, TRI)) {
495 for (const CalleeSavedInfo &CS : CSI) {
496 // Insert the spill to the stack frame.
497 unsigned Reg = CS.getReg();
498 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
499 TII.storeRegToStackSlot(SaveBlock, I, Reg, true, CS.getFrameIdx(), RC,
505 /// Insert restore code for the callee-saved registers used in the function.
506 static void insertCSRRestores(MachineBasicBlock &RestoreBlock,
507 std::vector<CalleeSavedInfo> &CSI) {
508 MachineFunction &MF = *RestoreBlock.getParent();
509 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
510 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
511 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
513 // Restore all registers immediately before the return and any
514 // terminators that precede it.
515 MachineBasicBlock::iterator I = RestoreBlock.getFirstTerminator();
517 if (!TFI->restoreCalleeSavedRegisters(RestoreBlock, I, CSI, TRI)) {
518 for (const CalleeSavedInfo &CI : reverse(CSI)) {
519 unsigned Reg = CI.getReg();
520 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg);
521 TII.loadRegFromStackSlot(RestoreBlock, I, Reg, CI.getFrameIdx(), RC, TRI);
522 assert(I != RestoreBlock.begin() &&
523 "loadRegFromStackSlot didn't insert any code!");
524 // Insert in reverse order. loadRegFromStackSlot can insert
525 // multiple instructions.
530 void PEI::spillCalleeSavedRegs(MachineFunction &MF) {
531 // We can't list this requirement in getRequiredProperties because some
532 // targets (WebAssembly) use virtual registers past this point, and the pass
533 // pipeline is set up without giving the passes a chance to look at the
535 // FIXME: Find a way to express this in getRequiredProperties.
536 assert(MF.getProperties().hasProperty(
537 MachineFunctionProperties::Property::NoVRegs));
539 const Function &F = MF.getFunction();
540 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
541 MachineFrameInfo &MFI = MF.getFrameInfo();
542 MinCSFrameIndex = std::numeric_limits<unsigned>::max();
545 // Determine which of the registers in the callee save list should be saved.
547 TFI->determineCalleeSaves(MF, SavedRegs, RS);
549 // Assign stack slots for any callee-saved registers that must be spilled.
550 assignCalleeSavedSpillSlots(MF, SavedRegs, MinCSFrameIndex, MaxCSFrameIndex);
552 // Add the code to save and restore the callee saved registers.
553 if (!F.hasFnAttribute(Attribute::Naked)) {
554 MFI.setCalleeSavedInfoValid(true);
556 std::vector<CalleeSavedInfo> &CSI = MFI.getCalleeSavedInfo();
559 NumLeafFuncWithSpills++;
561 for (MachineBasicBlock *SaveBlock : SaveBlocks) {
562 insertCSRSaves(*SaveBlock, CSI);
563 // Update the live-in information of all the blocks up to the save
567 for (MachineBasicBlock *RestoreBlock : RestoreBlocks)
568 insertCSRRestores(*RestoreBlock, CSI);
573 /// AdjustStackOffset - Helper function used to adjust the stack frame offset.
575 AdjustStackOffset(MachineFrameInfo &MFI, int FrameIdx,
576 bool StackGrowsDown, int64_t &Offset,
577 unsigned &MaxAlign, unsigned Skew) {
578 // If the stack grows down, add the object size to find the lowest address.
580 Offset += MFI.getObjectSize(FrameIdx);
582 unsigned Align = MFI.getObjectAlignment(FrameIdx);
584 // If the alignment of this object is greater than that of the stack, then
585 // increase the stack alignment to match.
586 MaxAlign = std::max(MaxAlign, Align);
588 // Adjust to alignment boundary.
589 Offset = alignTo(Offset, Align, Skew);
591 if (StackGrowsDown) {
592 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << -Offset
594 MFI.setObjectOffset(FrameIdx, -Offset); // Set the computed offset
596 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << Offset
598 MFI.setObjectOffset(FrameIdx, Offset);
599 Offset += MFI.getObjectSize(FrameIdx);
603 /// Compute which bytes of fixed and callee-save stack area are unused and keep
604 /// track of them in StackBytesFree.
606 computeFreeStackSlots(MachineFrameInfo &MFI, bool StackGrowsDown,
607 unsigned MinCSFrameIndex, unsigned MaxCSFrameIndex,
608 int64_t FixedCSEnd, BitVector &StackBytesFree) {
609 // Avoid undefined int64_t -> int conversion below in extreme case.
610 if (FixedCSEnd > std::numeric_limits<int>::max())
613 StackBytesFree.resize(FixedCSEnd, true);
615 SmallVector<int, 16> AllocatedFrameSlots;
616 // Add fixed objects.
617 for (int i = MFI.getObjectIndexBegin(); i != 0; ++i)
618 AllocatedFrameSlots.push_back(i);
619 // Add callee-save objects.
620 for (int i = MinCSFrameIndex; i <= (int)MaxCSFrameIndex; ++i)
621 AllocatedFrameSlots.push_back(i);
623 for (int i : AllocatedFrameSlots) {
624 // These are converted from int64_t, but they should always fit in int
625 // because of the FixedCSEnd check above.
626 int ObjOffset = MFI.getObjectOffset(i);
627 int ObjSize = MFI.getObjectSize(i);
628 int ObjStart, ObjEnd;
629 if (StackGrowsDown) {
630 // ObjOffset is negative when StackGrowsDown is true.
631 ObjStart = -ObjOffset - ObjSize;
634 ObjStart = ObjOffset;
635 ObjEnd = ObjOffset + ObjSize;
637 // Ignore fixed holes that are in the previous stack frame.
639 StackBytesFree.reset(ObjStart, ObjEnd);
643 /// Assign frame object to an unused portion of the stack in the fixed stack
644 /// object range. Return true if the allocation was successful.
645 static inline bool scavengeStackSlot(MachineFrameInfo &MFI, int FrameIdx,
646 bool StackGrowsDown, unsigned MaxAlign,
647 BitVector &StackBytesFree) {
648 if (MFI.isVariableSizedObjectIndex(FrameIdx))
651 if (StackBytesFree.none()) {
652 // clear it to speed up later scavengeStackSlot calls to
653 // StackBytesFree.none()
654 StackBytesFree.clear();
658 unsigned ObjAlign = MFI.getObjectAlignment(FrameIdx);
659 if (ObjAlign > MaxAlign)
662 int64_t ObjSize = MFI.getObjectSize(FrameIdx);
664 for (FreeStart = StackBytesFree.find_first(); FreeStart != -1;
665 FreeStart = StackBytesFree.find_next(FreeStart)) {
667 // Check that free space has suitable alignment.
668 unsigned ObjStart = StackGrowsDown ? FreeStart + ObjSize : FreeStart;
669 if (alignTo(ObjStart, ObjAlign) != ObjStart)
672 if (FreeStart + ObjSize > StackBytesFree.size())
675 bool AllBytesFree = true;
676 for (unsigned Byte = 0; Byte < ObjSize; ++Byte)
677 if (!StackBytesFree.test(FreeStart + Byte)) {
678 AllBytesFree = false;
688 if (StackGrowsDown) {
689 int ObjStart = -(FreeStart + ObjSize);
690 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") scavenged at SP["
691 << ObjStart << "]\n");
692 MFI.setObjectOffset(FrameIdx, ObjStart);
694 LLVM_DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") scavenged at SP["
695 << FreeStart << "]\n");
696 MFI.setObjectOffset(FrameIdx, FreeStart);
699 StackBytesFree.reset(FreeStart, FreeStart + ObjSize);
703 /// AssignProtectedObjSet - Helper function to assign large stack objects (i.e.,
704 /// those required to be close to the Stack Protector) to stack offsets.
706 AssignProtectedObjSet(const StackObjSet &UnassignedObjs,
707 SmallSet<int, 16> &ProtectedObjs,
708 MachineFrameInfo &MFI, bool StackGrowsDown,
709 int64_t &Offset, unsigned &MaxAlign, unsigned Skew) {
711 for (StackObjSet::const_iterator I = UnassignedObjs.begin(),
712 E = UnassignedObjs.end(); I != E; ++I) {
714 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign, Skew);
715 ProtectedObjs.insert(i);
719 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the
720 /// abstract stack objects.
721 void PEI::calculateFrameObjectOffsets(MachineFunction &MF) {
722 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
724 bool StackGrowsDown =
725 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown;
727 // Loop over all of the stack objects, assigning sequential addresses...
728 MachineFrameInfo &MFI = MF.getFrameInfo();
730 // Start at the beginning of the local area.
731 // The Offset is the distance from the stack top in the direction
732 // of stack growth -- so it's always nonnegative.
733 int LocalAreaOffset = TFI.getOffsetOfLocalArea();
735 LocalAreaOffset = -LocalAreaOffset;
736 assert(LocalAreaOffset >= 0
737 && "Local area offset should be in direction of stack growth");
738 int64_t Offset = LocalAreaOffset;
740 // Skew to be applied to alignment.
741 unsigned Skew = TFI.getStackAlignmentSkew(MF);
743 // If there are fixed sized objects that are preallocated in the local area,
744 // non-fixed objects can't be allocated right at the start of local area.
745 // Adjust 'Offset' to point to the end of last fixed sized preallocated
747 for (int i = MFI.getObjectIndexBegin(); i != 0; ++i) {
749 if (StackGrowsDown) {
750 // The maximum distance from the stack pointer is at lower address of
751 // the object -- which is given by offset. For down growing stack
752 // the offset is negative, so we negate the offset to get the distance.
753 FixedOff = -MFI.getObjectOffset(i);
755 // The maximum distance from the start pointer is at the upper
756 // address of the object.
757 FixedOff = MFI.getObjectOffset(i) + MFI.getObjectSize(i);
759 if (FixedOff > Offset) Offset = FixedOff;
762 // First assign frame offsets to stack objects that are used to spill
763 // callee saved registers.
764 if (StackGrowsDown) {
765 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) {
766 // If the stack grows down, we need to add the size to find the lowest
767 // address of the object.
768 Offset += MFI.getObjectSize(i);
770 unsigned Align = MFI.getObjectAlignment(i);
771 // Adjust to alignment boundary
772 Offset = alignTo(Offset, Align, Skew);
774 LLVM_DEBUG(dbgs() << "alloc FI(" << i << ") at SP[" << -Offset << "]\n");
775 MFI.setObjectOffset(i, -Offset); // Set the computed offset
777 } else if (MaxCSFrameIndex >= MinCSFrameIndex) {
778 // Be careful about underflow in comparisons agains MinCSFrameIndex.
779 for (unsigned i = MaxCSFrameIndex; i != MinCSFrameIndex - 1; --i) {
780 if (MFI.isDeadObjectIndex(i))
783 unsigned Align = MFI.getObjectAlignment(i);
784 // Adjust to alignment boundary
785 Offset = alignTo(Offset, Align, Skew);
787 LLVM_DEBUG(dbgs() << "alloc FI(" << i << ") at SP[" << Offset << "]\n");
788 MFI.setObjectOffset(i, Offset);
789 Offset += MFI.getObjectSize(i);
793 // FixedCSEnd is the stack offset to the end of the fixed and callee-save
795 int64_t FixedCSEnd = Offset;
796 unsigned MaxAlign = MFI.getMaxAlignment();
798 // Make sure the special register scavenging spill slot is closest to the
799 // incoming stack pointer if a frame pointer is required and is closer
800 // to the incoming rather than the final stack pointer.
801 const TargetRegisterInfo *RegInfo = MF.getSubtarget().getRegisterInfo();
802 bool EarlyScavengingSlots = (TFI.hasFP(MF) &&
803 TFI.isFPCloseToIncomingSP() &&
804 RegInfo->useFPForScavengingIndex(MF) &&
805 !RegInfo->needsStackRealignment(MF));
806 if (RS && EarlyScavengingSlots) {
807 SmallVector<int, 2> SFIs;
808 RS->getScavengingFrameIndices(SFIs);
809 for (SmallVectorImpl<int>::iterator I = SFIs.begin(),
810 IE = SFIs.end(); I != IE; ++I)
811 AdjustStackOffset(MFI, *I, StackGrowsDown, Offset, MaxAlign, Skew);
814 // FIXME: Once this is working, then enable flag will change to a target
815 // check for whether the frame is large enough to want to use virtual
816 // frame index registers. Functions which don't want/need this optimization
817 // will continue to use the existing code path.
818 if (MFI.getUseLocalStackAllocationBlock()) {
819 unsigned Align = MFI.getLocalFrameMaxAlign();
821 // Adjust to alignment boundary.
822 Offset = alignTo(Offset, Align, Skew);
824 LLVM_DEBUG(dbgs() << "Local frame base offset: " << Offset << "\n");
826 // Resolve offsets for objects in the local block.
827 for (unsigned i = 0, e = MFI.getLocalFrameObjectCount(); i != e; ++i) {
828 std::pair<int, int64_t> Entry = MFI.getLocalFrameObjectMap(i);
829 int64_t FIOffset = (StackGrowsDown ? -Offset : Offset) + Entry.second;
830 LLVM_DEBUG(dbgs() << "alloc FI(" << Entry.first << ") at SP[" << FIOffset
832 MFI.setObjectOffset(Entry.first, FIOffset);
834 // Allocate the local block
835 Offset += MFI.getLocalFrameSize();
837 MaxAlign = std::max(Align, MaxAlign);
840 // Retrieve the Exception Handler registration node.
841 int EHRegNodeFrameIndex = std::numeric_limits<int>::max();
842 if (const WinEHFuncInfo *FuncInfo = MF.getWinEHFuncInfo())
843 EHRegNodeFrameIndex = FuncInfo->EHRegNodeFrameIndex;
845 // Make sure that the stack protector comes before the local variables on the
847 SmallSet<int, 16> ProtectedObjs;
848 if (MFI.getStackProtectorIndex() >= 0) {
849 StackObjSet LargeArrayObjs;
850 StackObjSet SmallArrayObjs;
851 StackObjSet AddrOfObjs;
853 AdjustStackOffset(MFI, MFI.getStackProtectorIndex(), StackGrowsDown,
854 Offset, MaxAlign, Skew);
856 // Assign large stack objects first.
857 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
858 if (MFI.isObjectPreAllocated(i) &&
859 MFI.getUseLocalStackAllocationBlock())
861 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex)
863 if (RS && RS->isScavengingFrameIndex((int)i))
865 if (MFI.isDeadObjectIndex(i))
867 if (MFI.getStackProtectorIndex() == (int)i ||
868 EHRegNodeFrameIndex == (int)i)
871 switch (MFI.getObjectSSPLayout(i)) {
872 case MachineFrameInfo::SSPLK_None:
874 case MachineFrameInfo::SSPLK_SmallArray:
875 SmallArrayObjs.insert(i);
877 case MachineFrameInfo::SSPLK_AddrOf:
878 AddrOfObjs.insert(i);
880 case MachineFrameInfo::SSPLK_LargeArray:
881 LargeArrayObjs.insert(i);
884 llvm_unreachable("Unexpected SSPLayoutKind.");
887 AssignProtectedObjSet(LargeArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
888 Offset, MaxAlign, Skew);
889 AssignProtectedObjSet(SmallArrayObjs, ProtectedObjs, MFI, StackGrowsDown,
890 Offset, MaxAlign, Skew);
891 AssignProtectedObjSet(AddrOfObjs, ProtectedObjs, MFI, StackGrowsDown,
892 Offset, MaxAlign, Skew);
895 SmallVector<int, 8> ObjectsToAllocate;
897 // Then prepare to assign frame offsets to stack objects that are not used to
898 // spill callee saved registers.
899 for (unsigned i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
900 if (MFI.isObjectPreAllocated(i) && MFI.getUseLocalStackAllocationBlock())
902 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex)
904 if (RS && RS->isScavengingFrameIndex((int)i))
906 if (MFI.isDeadObjectIndex(i))
908 if (MFI.getStackProtectorIndex() == (int)i ||
909 EHRegNodeFrameIndex == (int)i)
911 if (ProtectedObjs.count(i))
914 // Add the objects that we need to allocate to our working set.
915 ObjectsToAllocate.push_back(i);
918 // Allocate the EH registration node first if one is present.
919 if (EHRegNodeFrameIndex != std::numeric_limits<int>::max())
920 AdjustStackOffset(MFI, EHRegNodeFrameIndex, StackGrowsDown, Offset,
923 // Give the targets a chance to order the objects the way they like it.
924 if (MF.getTarget().getOptLevel() != CodeGenOpt::None &&
925 MF.getTarget().Options.StackSymbolOrdering)
926 TFI.orderFrameObjects(MF, ObjectsToAllocate);
928 // Keep track of which bytes in the fixed and callee-save range are used so we
929 // can use the holes when allocating later stack objects. Only do this if
930 // stack protector isn't being used and the target requests it and we're
932 BitVector StackBytesFree;
933 if (!ObjectsToAllocate.empty() &&
934 MF.getTarget().getOptLevel() != CodeGenOpt::None &&
935 MFI.getStackProtectorIndex() < 0 && TFI.enableStackSlotScavenging(MF))
936 computeFreeStackSlots(MFI, StackGrowsDown, MinCSFrameIndex, MaxCSFrameIndex,
937 FixedCSEnd, StackBytesFree);
939 // Now walk the objects and actually assign base offsets to them.
940 for (auto &Object : ObjectsToAllocate)
941 if (!scavengeStackSlot(MFI, Object, StackGrowsDown, MaxAlign,
943 AdjustStackOffset(MFI, Object, StackGrowsDown, Offset, MaxAlign, Skew);
945 // Make sure the special register scavenging spill slot is closest to the
947 if (RS && !EarlyScavengingSlots) {
948 SmallVector<int, 2> SFIs;
949 RS->getScavengingFrameIndices(SFIs);
950 for (SmallVectorImpl<int>::iterator I = SFIs.begin(),
951 IE = SFIs.end(); I != IE; ++I)
952 AdjustStackOffset(MFI, *I, StackGrowsDown, Offset, MaxAlign, Skew);
955 if (!TFI.targetHandlesStackFrameRounding()) {
956 // If we have reserved argument space for call sites in the function
957 // immediately on entry to the current function, count it as part of the
958 // overall stack size.
959 if (MFI.adjustsStack() && TFI.hasReservedCallFrame(MF))
960 Offset += MFI.getMaxCallFrameSize();
962 // Round up the size to a multiple of the alignment. If the function has
963 // any calls or alloca's, align to the target's StackAlignment value to
964 // ensure that the callee's frame or the alloca data is suitably aligned;
965 // otherwise, for leaf functions, align to the TransientStackAlignment
968 if (MFI.adjustsStack() || MFI.hasVarSizedObjects() ||
969 (RegInfo->needsStackRealignment(MF) && MFI.getObjectIndexEnd() != 0))
970 StackAlign = TFI.getStackAlignment();
972 StackAlign = TFI.getTransientStackAlignment();
974 // If the frame pointer is eliminated, all frame offsets will be relative to
975 // SP not FP. Align to MaxAlign so this works.
976 StackAlign = std::max(StackAlign, MaxAlign);
977 Offset = alignTo(Offset, StackAlign, Skew);
980 // Update frame info to pretend that this is part of the stack...
981 int64_t StackSize = Offset - LocalAreaOffset;
982 MFI.setStackSize(StackSize);
983 NumBytesStackSpace += StackSize;
986 /// insertPrologEpilogCode - Scan the function for modified callee saved
987 /// registers, insert spill code for these callee saved registers, then add
988 /// prolog and epilog code to the function.
989 void PEI::insertPrologEpilogCode(MachineFunction &MF) {
990 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
992 // Add prologue to the function...
993 for (MachineBasicBlock *SaveBlock : SaveBlocks)
994 TFI.emitPrologue(MF, *SaveBlock);
996 // Add epilogue to restore the callee-save registers in each exiting block.
997 for (MachineBasicBlock *RestoreBlock : RestoreBlocks)
998 TFI.emitEpilogue(MF, *RestoreBlock);
1000 for (MachineBasicBlock *SaveBlock : SaveBlocks)
1001 TFI.inlineStackProbe(MF, *SaveBlock);
1003 // Emit additional code that is required to support segmented stacks, if
1004 // we've been asked for it. This, when linked with a runtime with support
1005 // for segmented stacks (libgcc is one), will result in allocating stack
1006 // space in small chunks instead of one large contiguous block.
1007 if (MF.shouldSplitStack()) {
1008 for (MachineBasicBlock *SaveBlock : SaveBlocks)
1009 TFI.adjustForSegmentedStacks(MF, *SaveBlock);
1010 // Record that there are split-stack functions, so we will emit a
1011 // special section to tell the linker.
1012 MF.getMMI().setHasSplitStack(true);
1014 MF.getMMI().setHasNosplitStack(true);
1016 // Emit additional code that is required to explicitly handle the stack in
1017 // HiPE native code (if needed) when loaded in the Erlang/OTP runtime. The
1018 // approach is rather similar to that of Segmented Stacks, but it uses a
1019 // different conditional check and another BIF for allocating more stack
1021 if (MF.getFunction().getCallingConv() == CallingConv::HiPE)
1022 for (MachineBasicBlock *SaveBlock : SaveBlocks)
1023 TFI.adjustForHiPEPrologue(MF, *SaveBlock);
1026 /// replaceFrameIndices - Replace all MO_FrameIndex operands with physical
1027 /// register references and actual offsets.
1028 void PEI::replaceFrameIndices(MachineFunction &MF) {
1029 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
1030 if (!TFI.needsFrameIndexResolution(MF)) return;
1032 // Store SPAdj at exit of a basic block.
1033 SmallVector<int, 8> SPState;
1034 SPState.resize(MF.getNumBlockIDs());
1035 df_iterator_default_set<MachineBasicBlock*> Reachable;
1037 // Iterate over the reachable blocks in DFS order.
1038 for (auto DFI = df_ext_begin(&MF, Reachable), DFE = df_ext_end(&MF, Reachable);
1039 DFI != DFE; ++DFI) {
1041 // Check the exit state of the DFS stack predecessor.
1042 if (DFI.getPathLength() >= 2) {
1043 MachineBasicBlock *StackPred = DFI.getPath(DFI.getPathLength() - 2);
1044 assert(Reachable.count(StackPred) &&
1045 "DFS stack predecessor is already visited.\n");
1046 SPAdj = SPState[StackPred->getNumber()];
1048 MachineBasicBlock *BB = *DFI;
1049 replaceFrameIndices(BB, MF, SPAdj);
1050 SPState[BB->getNumber()] = SPAdj;
1053 // Handle the unreachable blocks.
1054 for (auto &BB : MF) {
1055 if (Reachable.count(&BB))
1056 // Already handled in DFS traversal.
1059 replaceFrameIndices(&BB, MF, SPAdj);
1063 void PEI::replaceFrameIndices(MachineBasicBlock *BB, MachineFunction &MF,
1065 assert(MF.getSubtarget().getRegisterInfo() &&
1066 "getRegisterInfo() must be implemented!");
1067 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1068 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
1069 const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
1071 if (RS && FrameIndexEliminationScavenging)
1072 RS->enterBasicBlock(*BB);
1074 bool InsideCallSequence = false;
1076 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) {
1077 if (TII.isFrameInstr(*I)) {
1078 InsideCallSequence = TII.isFrameSetup(*I);
1079 SPAdj += TII.getSPAdjust(*I);
1080 I = TFI->eliminateCallFramePseudoInstr(MF, *BB, I);
1084 MachineInstr &MI = *I;
1086 bool DidFinishLoop = true;
1087 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1088 if (!MI.getOperand(i).isFI())
1091 // Frame indices in debug values are encoded in a target independent
1092 // way with simply the frame index and offset rather than any
1093 // target-specific addressing mode.
1094 if (MI.isDebugValue()) {
1095 assert(i == 0 && "Frame indices can only appear as the first "
1096 "operand of a DBG_VALUE machine instruction");
1099 TFI->getFrameIndexReference(MF, MI.getOperand(0).getIndex(), Reg);
1100 MI.getOperand(0).ChangeToRegister(Reg, false /*isDef*/);
1101 MI.getOperand(0).setIsDebug();
1102 auto *DIExpr = DIExpression::prepend(MI.getDebugExpression(),
1103 DIExpression::NoDeref, Offset);
1104 MI.getOperand(3).setMetadata(DIExpr);
1108 // TODO: This code should be commoned with the code for
1109 // PATCHPOINT. There's no good reason for the difference in
1110 // implementation other than historical accident. The only
1111 // remaining difference is the unconditional use of the stack
1112 // pointer as the base register.
1113 if (MI.getOpcode() == TargetOpcode::STATEPOINT) {
1114 assert((!MI.isDebugValue() || i == 0) &&
1115 "Frame indicies can only appear as the first operand of a "
1116 "DBG_VALUE machine instruction");
1118 MachineOperand &Offset = MI.getOperand(i + 1);
1119 int refOffset = TFI->getFrameIndexReferencePreferSP(
1120 MF, MI.getOperand(i).getIndex(), Reg, /*IgnoreSPUpdates*/ false);
1121 Offset.setImm(Offset.getImm() + refOffset + SPAdj);
1122 MI.getOperand(i).ChangeToRegister(Reg, false /*isDef*/);
1126 // Some instructions (e.g. inline asm instructions) can have
1127 // multiple frame indices and/or cause eliminateFrameIndex
1128 // to insert more than one instruction. We need the register
1129 // scavenger to go through all of these instructions so that
1130 // it can update its register information. We keep the
1131 // iterator at the point before insertion so that we can
1132 // revisit them in full.
1133 bool AtBeginning = (I == BB->begin());
1134 if (!AtBeginning) --I;
1136 // If this instruction has a FrameIndex operand, we need to
1137 // use that target machine register info object to eliminate
1139 TRI.eliminateFrameIndex(MI, SPAdj, i,
1140 FrameIndexEliminationScavenging ? RS : nullptr);
1142 // Reset the iterator if we were at the beginning of the BB.
1148 DidFinishLoop = false;
1152 // If we are looking at a call sequence, we need to keep track of
1153 // the SP adjustment made by each instruction in the sequence.
1154 // This includes both the frame setup/destroy pseudos (handled above),
1155 // as well as other instructions that have side effects w.r.t the SP.
1156 // Note that this must come after eliminateFrameIndex, because
1157 // if I itself referred to a frame index, we shouldn't count its own
1159 if (DidFinishLoop && InsideCallSequence)
1160 SPAdj += TII.getSPAdjust(MI);
1162 if (DoIncr && I != BB->end()) ++I;
1164 // Update register states.
1165 if (RS && FrameIndexEliminationScavenging && DidFinishLoop)