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((intptr_t)&ID) {}
40 const char *getPassName() const {
41 return "Prolog/Epilog Insertion & Frame Finalization";
44 /// runOnMachineFunction - Insert prolog/epilog code and replace abstract
45 /// frame indexes with appropriate references.
47 bool runOnMachineFunction(MachineFunction &Fn) {
48 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo();
49 RS = TRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL;
51 // Get MachineModuleInfo so that we can track the construction of the
53 if (MachineModuleInfo *MMI = getAnalysisToUpdate<MachineModuleInfo>())
54 Fn.getFrameInfo()->setMachineModuleInfo(MMI);
56 // Allow the target machine to make some adjustments to the function
57 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters.
58 TRI->processFunctionBeforeCalleeSavedScan(Fn, RS);
60 // Scan the function for modified callee saved registers and insert spill
61 // code for any callee saved registers that are modified. Also calculate
62 // the MaxCallFrameSize and HasCalls variables for the function's frame
63 // information and eliminates call frame pseudo instructions.
64 calculateCalleeSavedRegisters(Fn);
66 // Add the code to save and restore the callee saved registers
67 saveCalleeSavedRegisters(Fn);
69 // Allow the target machine to make final modifications to the function
70 // before the frame layout is finalized.
71 Fn.getTarget().getRegisterInfo()->processFunctionBeforeFrameFinalized(Fn);
73 // Calculate actual frame offsets for all of the abstract stack objects...
74 calculateFrameObjectOffsets(Fn);
76 // Add prolog and epilog code to the function. This function is required
77 // to align the stack frame as necessary for any stack variables or
78 // called functions. Because of this, calculateCalleeSavedRegisters
79 // must be called before this function in order to set the HasCalls
80 // and MaxCallFrameSize variables.
81 insertPrologEpilogCode(Fn);
83 // Replace all MO_FrameIndex operands with physical register references
84 // and actual offsets.
86 replaceFrameIndices(Fn);
95 // MinCSFrameIndex, MaxCSFrameIndex - Keeps the range of callee saved
96 // stack frame indexes.
97 unsigned MinCSFrameIndex, MaxCSFrameIndex;
99 void calculateCalleeSavedRegisters(MachineFunction &Fn);
100 void saveCalleeSavedRegisters(MachineFunction &Fn);
101 void calculateFrameObjectOffsets(MachineFunction &Fn);
102 void replaceFrameIndices(MachineFunction &Fn);
103 void insertPrologEpilogCode(MachineFunction &Fn);
109 /// createPrologEpilogCodeInserter - This function returns a pass that inserts
110 /// prolog and epilog code, and eliminates abstract frame references.
112 FunctionPass *llvm::createPrologEpilogCodeInserter() { return new PEI(); }
115 /// calculateCalleeSavedRegisters - Scan the function for modified callee saved
116 /// registers. Also calculate the MaxCallFrameSize and HasCalls variables for
117 /// the function's frame information and eliminates call frame pseudo
120 void PEI::calculateCalleeSavedRegisters(MachineFunction &Fn) {
121 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo();
122 const TargetFrameInfo *TFI = Fn.getTarget().getFrameInfo();
124 // Get the callee saved register list...
125 const unsigned *CSRegs = RegInfo->getCalleeSavedRegs(&Fn);
127 // Get the function call frame set-up and tear-down instruction opcode
128 int FrameSetupOpcode = RegInfo->getCallFrameSetupOpcode();
129 int FrameDestroyOpcode = RegInfo->getCallFrameDestroyOpcode();
131 // These are used to keep track the callee-save area. Initialize them.
132 MinCSFrameIndex = INT_MAX;
135 // Early exit for targets which have no callee saved registers and no call
136 // frame setup/destroy pseudo instructions.
137 if ((CSRegs == 0 || CSRegs[0] == 0) &&
138 FrameSetupOpcode == -1 && FrameDestroyOpcode == -1)
141 unsigned MaxCallFrameSize = 0;
142 bool HasCalls = false;
144 std::vector<MachineBasicBlock::iterator> FrameSDOps;
145 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB)
146 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
147 if (I->getOpcode() == FrameSetupOpcode ||
148 I->getOpcode() == FrameDestroyOpcode) {
149 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo"
150 " instructions should have a single immediate argument!");
151 unsigned Size = I->getOperand(0).getImm();
152 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size;
154 FrameSDOps.push_back(I);
157 MachineFrameInfo *FFI = Fn.getFrameInfo();
158 FFI->setHasCalls(HasCalls);
159 FFI->setMaxCallFrameSize(MaxCallFrameSize);
161 for (unsigned i = 0, e = FrameSDOps.size(); i != e; ++i) {
162 MachineBasicBlock::iterator I = FrameSDOps[i];
163 // If call frames are not being included as part of the stack frame,
164 // and there is no dynamic allocation (therefore referencing frame slots
165 // off sp), leave the pseudo ops alone. We'll eliminate them later.
166 if (RegInfo->hasReservedCallFrame(Fn) || RegInfo->hasFP(Fn))
167 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I);
170 // Now figure out which *callee saved* registers are modified by the current
171 // function, thus needing to be saved and restored in the prolog/epilog.
173 const TargetRegisterClass* const *CSRegClasses =
174 RegInfo->getCalleeSavedRegClasses(&Fn);
175 std::vector<CalleeSavedInfo> CSI;
176 for (unsigned i = 0; CSRegs[i]; ++i) {
177 unsigned Reg = CSRegs[i];
178 if (Fn.getRegInfo().isPhysRegUsed(Reg)) {
179 // If the reg is modified, save it!
180 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i]));
182 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg);
183 *AliasSet; ++AliasSet) { // Check alias registers too.
184 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) {
185 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i]));
193 return; // Early exit if no callee saved registers are modified!
195 unsigned NumFixedSpillSlots;
196 const std::pair<unsigned,int> *FixedSpillSlots =
197 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots);
199 // Now that we know which registers need to be saved and restored, allocate
200 // stack slots for them.
201 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
202 unsigned Reg = CSI[i].getReg();
203 const TargetRegisterClass *RC = CSI[i].getRegClass();
205 // Check to see if this physreg must be spilled to a particular stack slot
207 const std::pair<unsigned,int> *FixedSlot = FixedSpillSlots;
208 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots &&
209 FixedSlot->first != Reg)
213 if (FixedSlot == FixedSpillSlots+NumFixedSpillSlots) {
214 // Nope, just spill it anywhere convenient.
215 unsigned Align = RC->getAlignment();
216 unsigned StackAlign = TFI->getStackAlignment();
217 // We may not be able to sastify the desired alignment specification of
218 // the TargetRegisterClass if the stack alignment is smaller. Use the min.
219 Align = std::min(Align, StackAlign);
220 FrameIdx = FFI->CreateStackObject(RC->getSize(), Align);
221 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx;
222 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx;
224 // Spill it to the stack where we must.
225 FrameIdx = FFI->CreateFixedObject(RC->getSize(), FixedSlot->second);
227 CSI[i].setFrameIdx(FrameIdx);
230 FFI->setCalleeSavedInfo(CSI);
233 /// saveCalleeSavedRegisters - Insert spill code for any callee saved registers
234 /// that are modified in the function.
236 void PEI::saveCalleeSavedRegisters(MachineFunction &Fn) {
237 // Get callee saved register information.
238 MachineFrameInfo *FFI = Fn.getFrameInfo();
239 const std::vector<CalleeSavedInfo> &CSI = FFI->getCalleeSavedInfo();
241 // Early exit if no callee saved registers are modified!
245 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo();
247 // Now that we have a stack slot for each register to be saved, insert spill
248 // code into the entry block.
249 MachineBasicBlock *MBB = Fn.begin();
250 MachineBasicBlock::iterator I = MBB->begin();
252 if (!TII.spillCalleeSavedRegisters(*MBB, I, CSI)) {
253 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
254 // Add the callee-saved register as live-in. It's killed at the spill.
255 MBB->addLiveIn(CSI[i].getReg());
257 // Insert the spill to the stack frame.
258 TII.storeRegToStackSlot(*MBB, I, CSI[i].getReg(), true,
259 CSI[i].getFrameIdx(), CSI[i].getRegClass());
263 // Add code to restore the callee-save registers in each exiting block.
264 for (MachineFunction::iterator FI = Fn.begin(), E = Fn.end(); FI != E; ++FI)
265 // If last instruction is a return instruction, add an epilogue.
266 if (!FI->empty() && FI->back().getDesc().isReturn()) {
270 // Skip over all terminator instructions, which are part of the return
272 MachineBasicBlock::iterator I2 = I;
273 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator())
276 bool AtStart = I == MBB->begin();
277 MachineBasicBlock::iterator BeforeI = I;
281 // Restore all registers immediately before the return and any terminators
283 if (!TII.restoreCalleeSavedRegisters(*MBB, I, CSI)) {
284 for (unsigned i = 0, e = CSI.size(); i != e; ++i) {
285 TII.loadRegFromStackSlot(*MBB, I, CSI[i].getReg(),
286 CSI[i].getFrameIdx(),
287 CSI[i].getRegClass());
288 assert(I != MBB->begin() &&
289 "loadRegFromStackSlot didn't insert any code!");
290 // Insert in reverse order. loadRegFromStackSlot can insert multiple
304 /// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the
305 /// abstract stack objects.
307 void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) {
308 const TargetFrameInfo &TFI = *Fn.getTarget().getFrameInfo();
310 bool StackGrowsDown =
311 TFI.getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown;
313 // Loop over all of the stack objects, assigning sequential addresses...
314 MachineFrameInfo *FFI = Fn.getFrameInfo();
316 unsigned MaxAlign = 0;
318 // Start at the beginning of the local area.
319 // The Offset is the distance from the stack top in the direction
320 // of stack growth -- so it's always positive.
321 int64_t Offset = TFI.getOffsetOfLocalArea();
325 && "Local area offset should be in direction of stack growth");
327 // If there are fixed sized objects that are preallocated in the local area,
328 // non-fixed objects can't be allocated right at the start of local area.
329 // We currently don't support filling in holes in between fixed sized objects,
330 // so we adjust 'Offset' to point to the end of last fixed sized
331 // preallocated object.
332 for (int i = FFI->getObjectIndexBegin(); i != 0; ++i) {
334 if (StackGrowsDown) {
335 // The maximum distance from the stack pointer is at lower address of
336 // the object -- which is given by offset. For down growing stack
337 // the offset is negative, so we negate the offset to get the distance.
338 FixedOff = -FFI->getObjectOffset(i);
340 // The maximum distance from the start pointer is at the upper
341 // address of the object.
342 FixedOff = FFI->getObjectOffset(i) + FFI->getObjectSize(i);
344 if (FixedOff > Offset) Offset = FixedOff;
347 // First assign frame offsets to stack objects that are used to spill
348 // callee saved registers.
349 if (StackGrowsDown) {
350 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) {
351 // If stack grows down, we need to add size of find the lowest
352 // address of the object.
353 Offset += FFI->getObjectSize(i);
355 unsigned Align = FFI->getObjectAlignment(i);
356 // If the alignment of this object is greater than that of the stack, then
357 // increase the stack alignment to match.
358 MaxAlign = std::max(MaxAlign, Align);
359 // Adjust to alignment boundary
360 Offset = (Offset+Align-1)/Align*Align;
362 FFI->setObjectOffset(i, -Offset); // Set the computed offset
365 for (unsigned i = MaxCSFrameIndex; i >= MinCSFrameIndex; --i) {
366 unsigned Align = FFI->getObjectAlignment(i);
367 // If the alignment of this object is greater than that of the stack, then
368 // increase the stack alignment to match.
369 MaxAlign = std::max(MaxAlign, Align);
370 // Adjust to alignment boundary
371 Offset = (Offset+Align-1)/Align*Align;
373 FFI->setObjectOffset(i, Offset);
374 Offset += FFI->getObjectSize(i);
378 // Make sure the special register scavenging spill slot is closest to the
379 // frame pointer if a frame pointer is required.
380 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo();
381 if (RS && RegInfo->hasFP(Fn)) {
382 int SFI = RS->getScavengingFrameIndex();
384 // If stack grows down, we need to add size of the lowest
385 // address of the object.
387 Offset += FFI->getObjectSize(SFI);
389 unsigned Align = FFI->getObjectAlignment(SFI);
390 // Adjust to alignment boundary
391 Offset = (Offset+Align-1)/Align*Align;
393 if (StackGrowsDown) {
394 FFI->setObjectOffset(SFI, -Offset); // Set the computed offset
396 FFI->setObjectOffset(SFI, Offset);
397 Offset += FFI->getObjectSize(SFI);
402 // Then assign frame offsets to stack objects that are not used to spill
403 // callee saved registers.
404 for (unsigned i = 0, e = FFI->getObjectIndexEnd(); i != e; ++i) {
405 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex)
407 if (RS && (int)i == RS->getScavengingFrameIndex())
409 if (FFI->isDeadObjectIndex(i))
412 // If stack grows down, we need to add size of find the lowest
413 // address of the object.
415 Offset += FFI->getObjectSize(i);
417 unsigned Align = FFI->getObjectAlignment(i);
418 // If the alignment of this object is greater than that of the stack, then
419 // increase the stack alignment to match.
420 MaxAlign = std::max(MaxAlign, Align);
421 // Adjust to alignment boundary
422 Offset = (Offset+Align-1)/Align*Align;
424 if (StackGrowsDown) {
425 FFI->setObjectOffset(i, -Offset); // Set the computed offset
427 FFI->setObjectOffset(i, Offset);
428 Offset += FFI->getObjectSize(i);
432 // Make sure the special register scavenging spill slot is closest to the
434 if (RS && !RegInfo->hasFP(Fn)) {
435 int SFI = RS->getScavengingFrameIndex();
437 // If stack grows down, we need to add size of find the lowest
438 // address of the object.
440 Offset += FFI->getObjectSize(SFI);
442 unsigned Align = FFI->getObjectAlignment(SFI);
443 // If the alignment of this object is greater than that of the
444 // stack, then increase the stack alignment to match.
445 MaxAlign = std::max(MaxAlign, Align);
446 // Adjust to alignment boundary
447 Offset = (Offset+Align-1)/Align*Align;
449 if (StackGrowsDown) {
450 FFI->setObjectOffset(SFI, -Offset); // Set the computed offset
452 FFI->setObjectOffset(SFI, Offset);
453 Offset += FFI->getObjectSize(SFI);
458 // Round up the size to a multiple of the alignment, but only if there are
459 // calls or alloca's in the function. This ensures that any calls to
460 // subroutines have their stack frames suitable aligned.
461 if (!RegInfo->targetHandlesStackFrameRounding() &&
462 (FFI->hasCalls() || FFI->hasVarSizedObjects())) {
463 // If we have reserved argument space for call sites in the function
464 // immediately on entry to the current function, count it as part of the
465 // overall stack size.
466 if (RegInfo->hasReservedCallFrame(Fn))
467 Offset += FFI->getMaxCallFrameSize();
469 unsigned AlignMask = TFI.getStackAlignment() - 1;
470 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask);
473 // Update frame info to pretend that this is part of the stack...
474 FFI->setStackSize(Offset+TFI.getOffsetOfLocalArea());
476 // Remember the required stack alignment in case targets need it to perform
477 // dynamic stack alignment.
478 assert(FFI->getMaxAlignment() == MaxAlign &&
479 "Stack alignment calculation broken!");
483 /// insertPrologEpilogCode - Scan the function for modified callee saved
484 /// registers, insert spill code for these callee saved registers, then add
485 /// prolog and epilog code to the function.
487 void PEI::insertPrologEpilogCode(MachineFunction &Fn) {
488 // Add prologue to the function...
489 Fn.getTarget().getRegisterInfo()->emitPrologue(Fn);
491 // Add epilogue to restore the callee-save registers in each exiting block
492 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) {
493 // If last instruction is a return instruction, add an epilogue
494 if (!I->empty() && I->back().getDesc().isReturn())
495 Fn.getTarget().getRegisterInfo()->emitEpilogue(Fn, *I);
500 /// replaceFrameIndices - Replace all MO_FrameIndex operands with physical
501 /// register references and actual offsets.
503 void PEI::replaceFrameIndices(MachineFunction &Fn) {
504 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do?
506 const TargetMachine &TM = Fn.getTarget();
507 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!");
508 const TargetRegisterInfo &TRI = *TM.getRegisterInfo();
509 const TargetFrameInfo *TFI = TM.getFrameInfo();
510 bool StackGrowsDown =
511 TFI->getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown;
512 int FrameSetupOpcode = TRI.getCallFrameSetupOpcode();
513 int FrameDestroyOpcode = TRI.getCallFrameDestroyOpcode();
515 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
516 int SPAdj = 0; // SP offset due to call frame setup / destroy.
517 if (RS) RS->enterBasicBlock(BB);
518 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) {
519 MachineInstr *MI = I;
521 if (I->getOpcode() == FrameSetupOpcode ||
522 I->getOpcode() == FrameDestroyOpcode) {
523 // Remember how much SP has been adjustment to create the call frame.
524 int Size = I->getOperand(0).getImm();
525 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) ||
526 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode))
529 MachineBasicBlock::iterator PrevI = prior(I);
530 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I);
531 // Visit the instructions created by eliminateCallFramePseudoInstr().
534 } else if (I->getOpcode() == TargetInstrInfo::DECLARE)
539 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
540 if (MI->getOperand(i).isFrameIndex()) {
541 // If this instruction has a FrameIndex operand, we need to use that
542 // target machine register info object to eliminate it.
543 TRI.eliminateFrameIndex(MI, SPAdj, RS);
545 // Revisit the instruction in full. Some instructions (e.g. inline
546 // asm instructions) can have multiple frame indices.
552 // Update register states.
553 if (RS && MI) RS->forward(MI);
555 assert(SPAdj == 0 && "Unbalanced call frame setup / destroy pairs?");