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/CodeGen/Passes.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineInstr.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineModuleInfo.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/RegisterScavenging.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Target/TargetRegisterInfo.h"
28 #include "llvm/Target/TargetFrameInfo.h"
29 #include "llvm/Target/TargetInstrInfo.h"
30 #include "llvm/Support/Compiler.h"
31 #include "llvm/ADT/STLExtras.h"
36 struct VISIBILITY_HIDDEN PEI : public MachineFunctionPass {
38 PEI() : MachineFunctionPass(&ID) {}
40 const char *getPassName() const {
41 return "Prolog/Epilog Insertion & Frame Finalization";
44 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
45 AU.addPreservedID(MachineLoopInfoID);
46 AU.addPreservedID(MachineDominatorsID);
47 MachineFunctionPass::getAnalysisUsage(AU);
50 /// runOnMachineFunction - Insert prolog/epilog code and replace abstract
51 /// frame indexes with appropriate references.
53 bool runOnMachineFunction(MachineFunction &Fn) {
54 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo();
55 RS = TRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL;
57 // Get MachineModuleInfo so that we can track the construction of the
59 if (MachineModuleInfo *MMI = getAnalysisToUpdate<MachineModuleInfo>())
60 Fn.getFrameInfo()->setMachineModuleInfo(MMI);
62 // Allow the target machine to make some adjustments to the function
63 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters.
64 TRI->processFunctionBeforeCalleeSavedScan(Fn, RS);
66 // Scan the function for modified callee saved registers and insert spill
67 // code for any callee saved registers that are modified. Also calculate
68 // the MaxCallFrameSize and HasCalls variables for the function's frame
69 // information and eliminates call frame pseudo instructions.
70 calculateCalleeSavedRegisters(Fn);
72 // Add the code to save and restore the callee saved registers
73 saveCalleeSavedRegisters(Fn);
75 // Allow the target machine to make final modifications to the function
76 // before the frame layout is finalized.
77 TRI->processFunctionBeforeFrameFinalized(Fn);
79 // Calculate actual frame offsets for all of the abstract stack objects...
80 calculateFrameObjectOffsets(Fn);
82 // Add prolog and epilog code to the function. This function is required
83 // to align the stack frame as necessary for any stack variables or
84 // called functions. Because of this, calculateCalleeSavedRegisters
85 // must be called before this function in order to set the HasCalls
86 // and MaxCallFrameSize variables.
87 insertPrologEpilogCode(Fn);
89 // Replace all MO_FrameIndex operands with physical register references
90 // and actual offsets.
92 replaceFrameIndices(Fn);
101 // MinCSFrameIndex, MaxCSFrameIndex - Keeps the range of callee saved
102 // stack frame indexes.
103 unsigned MinCSFrameIndex, MaxCSFrameIndex;
105 void calculateCalleeSavedRegisters(MachineFunction &Fn);
106 void saveCalleeSavedRegisters(MachineFunction &Fn);
107 void calculateFrameObjectOffsets(MachineFunction &Fn);
108 void replaceFrameIndices(MachineFunction &Fn);
109 void insertPrologEpilogCode(MachineFunction &Fn);
115 /// createPrologEpilogCodeInserter - This function returns a pass that inserts
116 /// prolog and epilog code, and eliminates abstract frame references.
118 FunctionPass *llvm::createPrologEpilogCodeInserter() { return new PEI(); }
121 /// calculateCalleeSavedRegisters - Scan the function for modified callee saved
122 /// registers. Also calculate the MaxCallFrameSize and HasCalls variables for
123 /// the function's frame information and eliminates call frame pseudo
126 void PEI::calculateCalleeSavedRegisters(MachineFunction &Fn) {
127 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo();
128 const TargetFrameInfo *TFI = Fn.getTarget().getFrameInfo();
130 // Get the callee saved register list...
131 const unsigned *CSRegs = RegInfo->getCalleeSavedRegs(&Fn);
133 // Get the function call frame set-up and tear-down instruction opcode
134 int FrameSetupOpcode = RegInfo->getCallFrameSetupOpcode();
135 int FrameDestroyOpcode = RegInfo->getCallFrameDestroyOpcode();
137 // These are used to keep track the callee-save area. Initialize them.
138 MinCSFrameIndex = INT_MAX;
141 // Early exit for targets which have no callee saved registers and no call
142 // frame setup/destroy pseudo instructions.
143 if ((CSRegs == 0 || CSRegs[0] == 0) &&
144 FrameSetupOpcode == -1 && FrameDestroyOpcode == -1)
147 unsigned MaxCallFrameSize = 0;
148 bool HasCalls = false;
150 std::vector<MachineBasicBlock::iterator> FrameSDOps;
151 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB)
152 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
153 if (I->getOpcode() == FrameSetupOpcode ||
154 I->getOpcode() == FrameDestroyOpcode) {
155 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo"
156 " instructions should have a single immediate argument!");
157 unsigned Size = I->getOperand(0).getImm();
158 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size;
160 FrameSDOps.push_back(I);
163 MachineFrameInfo *FFI = Fn.getFrameInfo();
164 FFI->setHasCalls(HasCalls);
165 FFI->setMaxCallFrameSize(MaxCallFrameSize);
167 for (unsigned i = 0, e = FrameSDOps.size(); i != e; ++i) {
168 MachineBasicBlock::iterator I = FrameSDOps[i];
169 // If call frames are not being included as part of the stack frame,
170 // and there is no dynamic allocation (therefore referencing frame slots
171 // off sp), leave the pseudo ops alone. We'll eliminate them later.
172 if (RegInfo->hasReservedCallFrame(Fn) || RegInfo->hasFP(Fn))
173 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I);
176 // Now figure out which *callee saved* registers are modified by the current
177 // function, thus needing to be saved and restored in the prolog/epilog.
179 const TargetRegisterClass* const *CSRegClasses =
180 RegInfo->getCalleeSavedRegClasses(&Fn);
181 std::vector<CalleeSavedInfo> CSI;
182 for (unsigned i = 0; CSRegs[i]; ++i) {
183 unsigned Reg = CSRegs[i];
184 if (Fn.getRegInfo().isPhysRegUsed(Reg)) {
185 // If the reg is modified, save it!
186 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i]));
188 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg);
189 *AliasSet; ++AliasSet) { // Check alias registers too.
190 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) {
191 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i]));
199 return; // Early exit if no callee saved registers are modified!
201 unsigned NumFixedSpillSlots;
202 const std::pair<unsigned,int> *FixedSpillSlots =
203 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots);
205 // Now that we know which registers need to be saved and restored, allocate
206 // stack slots for them.
207 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
208 unsigned Reg = CSI[i].getReg();
209 const TargetRegisterClass *RC = CSI[i].getRegClass();
211 // Check to see if this physreg must be spilled to a particular stack slot
213 const std::pair<unsigned,int> *FixedSlot = FixedSpillSlots;
214 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots &&
215 FixedSlot->first != Reg)
219 if (FixedSlot == FixedSpillSlots+NumFixedSpillSlots) {
220 // Nope, just spill it anywhere convenient.
221 unsigned Align = RC->getAlignment();
222 unsigned StackAlign = TFI->getStackAlignment();
223 // We may not be able to sastify the desired alignment specification of
224 // the TargetRegisterClass if the stack alignment is smaller. Use the min.
225 Align = std::min(Align, StackAlign);
226 FrameIdx = FFI->CreateStackObject(RC->getSize(), Align);
227 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx;
228 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx;
230 // Spill it to the stack where we must.
231 FrameIdx = FFI->CreateFixedObject(RC->getSize(), FixedSlot->second);
233 CSI[i].setFrameIdx(FrameIdx);
236 FFI->setCalleeSavedInfo(CSI);
239 /// saveCalleeSavedRegisters - Insert spill code for any callee saved registers
240 /// that are modified in the function.
242 void PEI::saveCalleeSavedRegisters(MachineFunction &Fn) {
243 // Get callee saved register information.
244 MachineFrameInfo *FFI = Fn.getFrameInfo();
245 const std::vector<CalleeSavedInfo> &CSI = FFI->getCalleeSavedInfo();
247 // Early exit if no callee saved registers are modified!
251 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo();
253 // Now that we have a stack slot for each register to be saved, insert spill
254 // code into the entry block.
255 MachineBasicBlock *MBB = Fn.begin();
256 MachineBasicBlock::iterator I = MBB->begin();
258 if (!TII.spillCalleeSavedRegisters(*MBB, I, CSI)) {
259 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
260 // Add the callee-saved register as live-in. It's killed at the spill.
261 MBB->addLiveIn(CSI[i].getReg());
263 // Insert the spill to the stack frame.
264 TII.storeRegToStackSlot(*MBB, I, CSI[i].getReg(), true,
265 CSI[i].getFrameIdx(), CSI[i].getRegClass());
269 // Add code to restore the callee-save registers in each exiting block.
270 for (MachineFunction::iterator FI = Fn.begin(), E = Fn.end(); FI != E; ++FI)
271 // If last instruction is a return instruction, add an epilogue.
272 if (!FI->empty() && FI->back().getDesc().isReturn()) {
276 // Skip over all terminator instructions, which are part of the return
278 MachineBasicBlock::iterator I2 = I;
279 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator())
282 bool AtStart = I == MBB->begin();
283 MachineBasicBlock::iterator BeforeI = I;
287 // Restore all registers immediately before the return and any terminators
289 if (!TII.restoreCalleeSavedRegisters(*MBB, I, CSI)) {
290 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
291 TII.loadRegFromStackSlot(*MBB, I, CSI[i].getReg(),
292 CSI[i].getFrameIdx(),
293 CSI[i].getRegClass());
294 assert(I != MBB->begin() &&
295 "loadRegFromStackSlot didn't insert any code!");
296 // Insert in reverse order. loadRegFromStackSlot can insert multiple
310 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the
311 /// abstract stack objects.
313 void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) {
314 const TargetFrameInfo &TFI = *Fn.getTarget().getFrameInfo();
316 bool StackGrowsDown =
317 TFI.getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown;
319 // Loop over all of the stack objects, assigning sequential addresses...
320 MachineFrameInfo *FFI = Fn.getFrameInfo();
322 unsigned MaxAlign = FFI->getMaxAlignment();
324 // Start at the beginning of the local area.
325 // The Offset is the distance from the stack top in the direction
326 // of stack growth -- so it's always nonnegative.
327 int64_t Offset = TFI.getOffsetOfLocalArea();
331 && "Local area offset should be in direction of stack growth");
333 // If there are fixed sized objects that are preallocated in the local area,
334 // non-fixed objects can't be allocated right at the start of local area.
335 // We currently don't support filling in holes in between fixed sized objects,
336 // so we adjust 'Offset' to point to the end of last fixed sized
337 // preallocated object.
338 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
340 if (StackGrowsDown) {
341 // The maximum distance from the stack pointer is at lower address of
342 // the object -- which is given by offset. For down growing stack
343 // the offset is negative, so we negate the offset to get the distance.
344 FixedOff = -FFI->getObjectOffset(i);
346 // The maximum distance from the start pointer is at the upper
347 // address of the object.
348 FixedOff = FFI->getObjectOffset(i) + FFI->getObjectSize(i);
350 if (FixedOff > Offset) Offset = FixedOff;
353 // First assign frame offsets to stack objects that are used to spill
354 // callee saved registers.
355 if (StackGrowsDown) {
356 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) {
357 // If stack grows down, we need to add size of find the lowest
358 // address of the object.
359 Offset += FFI->getObjectSize(i);
361 unsigned Align = FFI->getObjectAlignment(i);
362 // If the alignment of this object is greater than that of the stack, then
363 // increase the stack alignment to match.
364 MaxAlign = std::max(MaxAlign, Align);
365 // Adjust to alignment boundary
366 Offset = (Offset+Align-1)/Align*Align;
368 FFI->setObjectOffset(i, -Offset); // Set the computed offset
371 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex;
372 for (int i = MaxCSFI; i >= MinCSFI ; --i) {
373 unsigned Align = FFI->getObjectAlignment(i);
374 // If the alignment of this object is greater than that of the stack, then
375 // increase the stack alignment to match.
376 MaxAlign = std::max(MaxAlign, Align);
377 // Adjust to alignment boundary
378 Offset = (Offset+Align-1)/Align*Align;
380 FFI->setObjectOffset(i, Offset);
381 Offset += FFI->getObjectSize(i);
385 // Make sure the special register scavenging spill slot is closest to the
386 // frame pointer if a frame pointer is required.
387 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo();
388 if (RS && RegInfo->hasFP(Fn)) {
389 int SFI = RS->getScavengingFrameIndex();
391 // If stack grows down, we need to add size of the lowest
392 // address of the object.
394 Offset += FFI->getObjectSize(SFI);
396 unsigned Align = FFI->getObjectAlignment(SFI);
397 // Adjust to alignment boundary
398 Offset = (Offset+Align-1)/Align*Align;
400 if (StackGrowsDown) {
401 FFI->setObjectOffset(SFI, -Offset); // Set the computed offset
403 FFI->setObjectOffset(SFI, Offset);
404 Offset += FFI->getObjectSize(SFI);
409 // Make sure that the stack protector comes before the local variables on the
411 if (FFI->getStackProtectorIndex() >= 0) {
412 int FI = FFI->getStackProtectorIndex();
414 // If stack grows down, we need to add size of find the lowest
415 // address of the object.
417 Offset += FFI->getObjectSize(FI);
419 unsigned Align = FFI->getObjectAlignment(FI);
421 // If the alignment of this object is greater than that of the stack, then
422 // increase the stack alignment to match.
423 MaxAlign = std::max(MaxAlign, Align);
425 // Adjust to alignment boundary.
426 Offset = (Offset + Align - 1) / Align * Align;
428 if (StackGrowsDown) {
429 FFI->setObjectOffset(FI, -Offset); // Set the computed offset
431 FFI->setObjectOffset(FI, Offset);
432 Offset += FFI->getObjectSize(FI);
436 // Then assign frame offsets to stack objects that are not used to spill
437 // callee saved registers.
438 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
439 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex)
441 if (RS && (int)i == RS->getScavengingFrameIndex())
443 if (FFI->isDeadObjectIndex(i))
446 // If stack grows down, we need to add size of find the lowest
447 // address of the object.
449 Offset += FFI->getObjectSize(i);
451 unsigned Align = FFI->getObjectAlignment(i);
452 // If the alignment of this object is greater than that of the stack, then
453 // increase the stack alignment to match.
454 MaxAlign = std::max(MaxAlign, Align);
455 // Adjust to alignment boundary
456 Offset = (Offset+Align-1)/Align*Align;
458 if (StackGrowsDown) {
459 FFI->setObjectOffset(i, -Offset); // Set the computed offset
461 FFI->setObjectOffset(i, Offset);
462 Offset += FFI->getObjectSize(i);
466 // Make sure the special register scavenging spill slot is closest to the
468 if (RS && !RegInfo->hasFP(Fn)) {
469 int SFI = RS->getScavengingFrameIndex();
471 // If stack grows down, we need to add size of find the lowest
472 // address of the object.
474 Offset += FFI->getObjectSize(SFI);
476 unsigned Align = FFI->getObjectAlignment(SFI);
477 // If the alignment of this object is greater than that of the
478 // stack, then increase the stack alignment to match.
479 MaxAlign = std::max(MaxAlign, Align);
480 // Adjust to alignment boundary
481 Offset = (Offset+Align-1)/Align*Align;
483 if (StackGrowsDown) {
484 FFI->setObjectOffset(SFI, -Offset); // Set the computed offset
486 FFI->setObjectOffset(SFI, Offset);
487 Offset += FFI->getObjectSize(SFI);
492 // Round up the size to a multiple of the alignment, but only if there are
493 // calls or alloca's in the function. This ensures that any calls to
494 // subroutines have their stack frames suitable aligned.
495 // Also do this if we need runtime alignment of the stack. In this case
496 // offsets will be relative to SP not FP; round up the stack size so this
498 if (!RegInfo->targetHandlesStackFrameRounding() &&
499 (FFI->hasCalls() || FFI->hasVarSizedObjects() ||
500 (RegInfo->needsStackRealignment(Fn) &&
501 FFI->getObjectIndexEnd() != 0))) {
502 // If we have reserved argument space for call sites in the function
503 // immediately on entry to the current function, count it as part of the
504 // overall stack size.
505 if (RegInfo->hasReservedCallFrame(Fn))
506 Offset += FFI->getMaxCallFrameSize();
508 unsigned AlignMask = std::max(TFI.getStackAlignment(),MaxAlign) - 1;
509 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask);
512 // Update frame info to pretend that this is part of the stack...
513 FFI->setStackSize(Offset+TFI.getOffsetOfLocalArea());
515 // Remember the required stack alignment in case targets need it to perform
516 // dynamic stack alignment.
517 FFI->setMaxAlignment(MaxAlign);
521 /// insertPrologEpilogCode - Scan the function for modified callee saved
522 /// registers, insert spill code for these callee saved registers, then add
523 /// prolog and epilog code to the function.
525 void PEI::insertPrologEpilogCode(MachineFunction &Fn) {
526 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo();
528 // Add prologue to the function...
529 TRI->emitPrologue(Fn);
531 // Add epilogue to restore the callee-save registers in each exiting block
532 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
533 // If last instruction is a return instruction, add an epilogue
534 if (!I->empty() && I->back().getDesc().isReturn())
535 TRI->emitEpilogue(Fn, *I);
540 /// replaceFrameIndices - Replace all MO_FrameIndex operands with physical
541 /// register references and actual offsets.
543 void PEI::replaceFrameIndices(MachineFunction &Fn) {
544 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do?
546 const TargetMachine &TM = Fn.getTarget();
547 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!");
548 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
549 const TargetFrameInfo *TFI = TM.getFrameInfo();
550 bool StackGrowsDown =
551 TFI->getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown;
552 int FrameSetupOpcode = TRI.getCallFrameSetupOpcode();
553 int FrameDestroyOpcode = TRI.getCallFrameDestroyOpcode();
555 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
556 int SPAdj = 0; // SP offset due to call frame setup / destroy.
557 if (RS) RS->enterBasicBlock(BB);
558 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) {
559 MachineInstr *MI = I;
561 if (I->getOpcode() == TargetInstrInfo::DECLARE) {
567 if (I->getOpcode() == FrameSetupOpcode ||
568 I->getOpcode() == FrameDestroyOpcode) {
569 // Remember how much SP has been adjusted to create the call
571 int Size = I->getOperand(0).getImm();
573 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) ||
574 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode))
579 MachineBasicBlock::iterator PrevI = prior(I);
580 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I);
582 // Visit the instructions created by eliminateCallFramePseudoInstr().
589 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
590 if (MI->getOperand(i).isFI()) {
591 // Some instructions (e.g. inline asm instructions) can have
592 // multiple frame indices and/or cause eliminateFrameIndex
593 // to insert more than one instruction. We need the register
594 // scavenger to go through all of these instructions so that
595 // it can update its register information. We keep the
596 // iterator at the point before insertion so that we can
597 // revisit them in full.
598 bool AtBeginning = (I == BB->begin());
599 if (!AtBeginning) --I;
601 // If this instruction has a FrameIndex operand, we need to
602 // use that target machine register info object to eliminate
604 TRI.eliminateFrameIndex(MI, SPAdj, RS);
606 // Reset the iterator if we were at the beginning of the BB.
618 // Update register states.
619 if (RS && MI) RS->forward(MI);
622 assert(SPAdj == 0 && "Unbalanced call frame setup / destroy pairs?");