PrologEpilogInserter.cpp revision 108fb3202af6f500073cdbb7be32c25d7a273a2e
1//===-- PrologEpilogInserter.cpp - Insert Prolog/Epilog code in function --===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass is responsible for finalizing the functions frame layout, saving 11// callee saved registers, and for emitting prolog & epilog code for the 12// function. 13// 14// This pass must be run after register allocation. After this pass is 15// executed, it is illegal to construct MO_FrameIndex operands. 16// 17// This pass provides an optional shrink wrapping variant of prolog/epilog 18// insertion, enabled via --shrink-wrap. See ShrinkWrapping.cpp. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "pei" 23#include "PrologEpilogInserter.h" 24#include "llvm/ADT/IndexedMap.h" 25#include "llvm/ADT/STLExtras.h" 26#include "llvm/ADT/SmallSet.h" 27#include "llvm/ADT/Statistic.h" 28#include "llvm/CodeGen/MachineDominators.h" 29#include "llvm/CodeGen/MachineFrameInfo.h" 30#include "llvm/CodeGen/MachineInstr.h" 31#include "llvm/CodeGen/MachineLoopInfo.h" 32#include "llvm/CodeGen/MachineRegisterInfo.h" 33#include "llvm/CodeGen/RegisterScavenging.h" 34#include "llvm/IR/InlineAsm.h" 35#include "llvm/Support/CommandLine.h" 36#include "llvm/Support/Compiler.h" 37#include "llvm/Support/Debug.h" 38#include "llvm/Target/TargetFrameLowering.h" 39#include "llvm/Target/TargetInstrInfo.h" 40#include "llvm/Target/TargetMachine.h" 41#include "llvm/Target/TargetOptions.h" 42#include "llvm/Target/TargetRegisterInfo.h" 43#include <climits> 44 45using namespace llvm; 46 47char PEI::ID = 0; 48char &llvm::PrologEpilogCodeInserterID = PEI::ID; 49 50INITIALIZE_PASS_BEGIN(PEI, "prologepilog", 51 "Prologue/Epilogue Insertion", false, false) 52INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 53INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree) 54INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 55INITIALIZE_PASS_END(PEI, "prologepilog", 56 "Prologue/Epilogue Insertion & Frame Finalization", 57 false, false) 58 59STATISTIC(NumVirtualFrameRegs, "Number of virtual frame regs encountered"); 60STATISTIC(NumScavengedRegs, "Number of frame index regs scavenged"); 61STATISTIC(NumBytesStackSpace, 62 "Number of bytes used for stack in all functions"); 63 64/// runOnMachineFunction - Insert prolog/epilog code and replace abstract 65/// frame indexes with appropriate references. 66/// 67bool PEI::runOnMachineFunction(MachineFunction &Fn) { 68 const Function* F = Fn.getFunction(); 69 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 70 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 71 72 assert(!Fn.getRegInfo().getNumVirtRegs() && "Regalloc must assign all vregs"); 73 74 RS = TRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL; 75 FrameIndexVirtualScavenging = TRI->requiresFrameIndexScavenging(Fn); 76 77 // Calculate the MaxCallFrameSize and AdjustsStack variables for the 78 // function's frame information. Also eliminates call frame pseudo 79 // instructions. 80 calculateCallsInformation(Fn); 81 82 // Allow the target machine to make some adjustments to the function 83 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters. 84 TFI->processFunctionBeforeCalleeSavedScan(Fn, RS); 85 86 // Scan the function for modified callee saved registers and insert spill code 87 // for any callee saved registers that are modified. 88 calculateCalleeSavedRegisters(Fn); 89 90 // Determine placement of CSR spill/restore code: 91 // - With shrink wrapping, place spills and restores to tightly 92 // enclose regions in the Machine CFG of the function where 93 // they are used. 94 // - Without shink wrapping (default), place all spills in the 95 // entry block, all restores in return blocks. 96 placeCSRSpillsAndRestores(Fn); 97 98 // Add the code to save and restore the callee saved registers 99 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 100 Attribute::Naked)) 101 insertCSRSpillsAndRestores(Fn); 102 103 // Allow the target machine to make final modifications to the function 104 // before the frame layout is finalized. 105 TFI->processFunctionBeforeFrameFinalized(Fn); 106 107 // Calculate actual frame offsets for all abstract stack objects... 108 calculateFrameObjectOffsets(Fn); 109 110 // Add prolog and epilog code to the function. This function is required 111 // to align the stack frame as necessary for any stack variables or 112 // called functions. Because of this, calculateCalleeSavedRegisters() 113 // must be called before this function in order to set the AdjustsStack 114 // and MaxCallFrameSize variables. 115 if (!F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 116 Attribute::Naked)) 117 insertPrologEpilogCode(Fn); 118 119 // Replace all MO_FrameIndex operands with physical register references 120 // and actual offsets. 121 // 122 replaceFrameIndices(Fn); 123 124 // If register scavenging is needed, as we've enabled doing it as a 125 // post-pass, scavenge the virtual registers that frame index elimiation 126 // inserted. 127 if (TRI->requiresRegisterScavenging(Fn) && FrameIndexVirtualScavenging) 128 scavengeFrameVirtualRegs(Fn); 129 130 // Clear any vregs created by virtual scavenging. 131 Fn.getRegInfo().clearVirtRegs(); 132 133 delete RS; 134 clearAllSets(); 135 return true; 136} 137 138/// calculateCallsInformation - Calculate the MaxCallFrameSize and AdjustsStack 139/// variables for the function's frame information and eliminate call frame 140/// pseudo instructions. 141void PEI::calculateCallsInformation(MachineFunction &Fn) { 142 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 143 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 144 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 145 MachineFrameInfo *MFI = Fn.getFrameInfo(); 146 147 unsigned MaxCallFrameSize = 0; 148 bool AdjustsStack = MFI->adjustsStack(); 149 150 // Get the function call frame set-up and tear-down instruction opcode 151 int FrameSetupOpcode = TII.getCallFrameSetupOpcode(); 152 int FrameDestroyOpcode = TII.getCallFrameDestroyOpcode(); 153 154 // Early exit for targets which have no call frame setup/destroy pseudo 155 // instructions. 156 if (FrameSetupOpcode == -1 && FrameDestroyOpcode == -1) 157 return; 158 159 std::vector<MachineBasicBlock::iterator> FrameSDOps; 160 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) 161 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) 162 if (I->getOpcode() == FrameSetupOpcode || 163 I->getOpcode() == FrameDestroyOpcode) { 164 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo" 165 " instructions should have a single immediate argument!"); 166 unsigned Size = I->getOperand(0).getImm(); 167 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size; 168 AdjustsStack = true; 169 FrameSDOps.push_back(I); 170 } else if (I->isInlineAsm()) { 171 // Some inline asm's need a stack frame, as indicated by operand 1. 172 unsigned ExtraInfo = I->getOperand(InlineAsm::MIOp_ExtraInfo).getImm(); 173 if (ExtraInfo & InlineAsm::Extra_IsAlignStack) 174 AdjustsStack = true; 175 } 176 177 MFI->setAdjustsStack(AdjustsStack); 178 MFI->setMaxCallFrameSize(MaxCallFrameSize); 179 180 for (std::vector<MachineBasicBlock::iterator>::iterator 181 i = FrameSDOps.begin(), e = FrameSDOps.end(); i != e; ++i) { 182 MachineBasicBlock::iterator I = *i; 183 184 // If call frames are not being included as part of the stack frame, and 185 // the target doesn't indicate otherwise, remove the call frame pseudos 186 // here. The sub/add sp instruction pairs are still inserted, but we don't 187 // need to track the SP adjustment for frame index elimination. 188 if (TFI->canSimplifyCallFramePseudos(Fn)) 189 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I); 190 } 191} 192 193 194/// calculateCalleeSavedRegisters - Scan the function for modified callee saved 195/// registers. 196void PEI::calculateCalleeSavedRegisters(MachineFunction &F) { 197 const TargetRegisterInfo *RegInfo = F.getTarget().getRegisterInfo(); 198 const TargetFrameLowering *TFI = F.getTarget().getFrameLowering(); 199 MachineFrameInfo *MFI = F.getFrameInfo(); 200 201 // Get the callee saved register list... 202 const uint16_t *CSRegs = RegInfo->getCalleeSavedRegs(&F); 203 204 // These are used to keep track the callee-save area. Initialize them. 205 MinCSFrameIndex = INT_MAX; 206 MaxCSFrameIndex = 0; 207 208 // Early exit for targets which have no callee saved registers. 209 if (CSRegs == 0 || CSRegs[0] == 0) 210 return; 211 212 // In Naked functions we aren't going to save any registers. 213 if (F.getFunction()->getAttributes().hasAttribute(AttributeSet::FunctionIndex, 214 Attribute::Naked)) 215 return; 216 217 std::vector<CalleeSavedInfo> CSI; 218 for (unsigned i = 0; CSRegs[i]; ++i) { 219 unsigned Reg = CSRegs[i]; 220 if (F.getRegInfo().isPhysRegUsed(Reg)) { 221 // If the reg is modified, save it! 222 CSI.push_back(CalleeSavedInfo(Reg)); 223 } 224 } 225 226 if (CSI.empty()) 227 return; // Early exit if no callee saved registers are modified! 228 229 unsigned NumFixedSpillSlots; 230 const TargetFrameLowering::SpillSlot *FixedSpillSlots = 231 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots); 232 233 // Now that we know which registers need to be saved and restored, allocate 234 // stack slots for them. 235 for (std::vector<CalleeSavedInfo>::iterator 236 I = CSI.begin(), E = CSI.end(); I != E; ++I) { 237 unsigned Reg = I->getReg(); 238 const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg); 239 240 int FrameIdx; 241 if (RegInfo->hasReservedSpillSlot(F, Reg, FrameIdx)) { 242 I->setFrameIdx(FrameIdx); 243 continue; 244 } 245 246 // Check to see if this physreg must be spilled to a particular stack slot 247 // on this target. 248 const TargetFrameLowering::SpillSlot *FixedSlot = FixedSpillSlots; 249 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots && 250 FixedSlot->Reg != Reg) 251 ++FixedSlot; 252 253 if (FixedSlot == FixedSpillSlots + NumFixedSpillSlots) { 254 // Nope, just spill it anywhere convenient. 255 unsigned Align = RC->getAlignment(); 256 unsigned StackAlign = TFI->getStackAlignment(); 257 258 // We may not be able to satisfy the desired alignment specification of 259 // the TargetRegisterClass if the stack alignment is smaller. Use the 260 // min. 261 Align = std::min(Align, StackAlign); 262 FrameIdx = MFI->CreateStackObject(RC->getSize(), Align, true); 263 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx; 264 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx; 265 } else { 266 // Spill it to the stack where we must. 267 FrameIdx = MFI->CreateFixedObject(RC->getSize(), FixedSlot->Offset, true); 268 } 269 270 I->setFrameIdx(FrameIdx); 271 } 272 273 MFI->setCalleeSavedInfo(CSI); 274} 275 276/// insertCSRSpillsAndRestores - Insert spill and restore code for 277/// callee saved registers used in the function, handling shrink wrapping. 278/// 279void PEI::insertCSRSpillsAndRestores(MachineFunction &Fn) { 280 // Get callee saved register information. 281 MachineFrameInfo *MFI = Fn.getFrameInfo(); 282 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 283 284 MFI->setCalleeSavedInfoValid(true); 285 286 // Early exit if no callee saved registers are modified! 287 if (CSI.empty()) 288 return; 289 290 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 291 const TargetFrameLowering *TFI = Fn.getTarget().getFrameLowering(); 292 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 293 MachineBasicBlock::iterator I; 294 295 if (!ShrinkWrapThisFunction) { 296 // Spill using target interface. 297 I = EntryBlock->begin(); 298 if (!TFI->spillCalleeSavedRegisters(*EntryBlock, I, CSI, TRI)) { 299 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 300 // Add the callee-saved register as live-in. 301 // It's killed at the spill. 302 EntryBlock->addLiveIn(CSI[i].getReg()); 303 304 // Insert the spill to the stack frame. 305 unsigned Reg = CSI[i].getReg(); 306 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 307 TII.storeRegToStackSlot(*EntryBlock, I, Reg, true, 308 CSI[i].getFrameIdx(), RC, TRI); 309 } 310 } 311 312 // Restore using target interface. 313 for (unsigned ri = 0, re = ReturnBlocks.size(); ri != re; ++ri) { 314 MachineBasicBlock* MBB = ReturnBlocks[ri]; 315 I = MBB->end(); --I; 316 317 // Skip over all terminator instructions, which are part of the return 318 // sequence. 319 MachineBasicBlock::iterator I2 = I; 320 while (I2 != MBB->begin() && (--I2)->isTerminator()) 321 I = I2; 322 323 bool AtStart = I == MBB->begin(); 324 MachineBasicBlock::iterator BeforeI = I; 325 if (!AtStart) 326 --BeforeI; 327 328 // Restore all registers immediately before the return and any 329 // terminators that precede it. 330 if (!TFI->restoreCalleeSavedRegisters(*MBB, I, CSI, TRI)) { 331 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 332 unsigned Reg = CSI[i].getReg(); 333 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 334 TII.loadRegFromStackSlot(*MBB, I, Reg, 335 CSI[i].getFrameIdx(), 336 RC, TRI); 337 assert(I != MBB->begin() && 338 "loadRegFromStackSlot didn't insert any code!"); 339 // Insert in reverse order. loadRegFromStackSlot can insert 340 // multiple instructions. 341 if (AtStart) 342 I = MBB->begin(); 343 else { 344 I = BeforeI; 345 ++I; 346 } 347 } 348 } 349 } 350 return; 351 } 352 353 // Insert spills. 354 std::vector<CalleeSavedInfo> blockCSI; 355 for (CSRegBlockMap::iterator BI = CSRSave.begin(), 356 BE = CSRSave.end(); BI != BE; ++BI) { 357 MachineBasicBlock* MBB = BI->first; 358 CSRegSet save = BI->second; 359 360 if (save.empty()) 361 continue; 362 363 blockCSI.clear(); 364 for (CSRegSet::iterator RI = save.begin(), 365 RE = save.end(); RI != RE; ++RI) { 366 blockCSI.push_back(CSI[*RI]); 367 } 368 assert(blockCSI.size() > 0 && 369 "Could not collect callee saved register info"); 370 371 I = MBB->begin(); 372 373 // When shrink wrapping, use stack slot stores/loads. 374 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 375 // Add the callee-saved register as live-in. 376 // It's killed at the spill. 377 MBB->addLiveIn(blockCSI[i].getReg()); 378 379 // Insert the spill to the stack frame. 380 unsigned Reg = blockCSI[i].getReg(); 381 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 382 TII.storeRegToStackSlot(*MBB, I, Reg, 383 true, 384 blockCSI[i].getFrameIdx(), 385 RC, TRI); 386 } 387 } 388 389 for (CSRegBlockMap::iterator BI = CSRRestore.begin(), 390 BE = CSRRestore.end(); BI != BE; ++BI) { 391 MachineBasicBlock* MBB = BI->first; 392 CSRegSet restore = BI->second; 393 394 if (restore.empty()) 395 continue; 396 397 blockCSI.clear(); 398 for (CSRegSet::iterator RI = restore.begin(), 399 RE = restore.end(); RI != RE; ++RI) { 400 blockCSI.push_back(CSI[*RI]); 401 } 402 assert(blockCSI.size() > 0 && 403 "Could not find callee saved register info"); 404 405 // If MBB is empty and needs restores, insert at the _beginning_. 406 if (MBB->empty()) { 407 I = MBB->begin(); 408 } else { 409 I = MBB->end(); 410 --I; 411 412 // Skip over all terminator instructions, which are part of the 413 // return sequence. 414 if (! I->isTerminator()) { 415 ++I; 416 } else { 417 MachineBasicBlock::iterator I2 = I; 418 while (I2 != MBB->begin() && (--I2)->isTerminator()) 419 I = I2; 420 } 421 } 422 423 bool AtStart = I == MBB->begin(); 424 MachineBasicBlock::iterator BeforeI = I; 425 if (!AtStart) 426 --BeforeI; 427 428 // Restore all registers immediately before the return and any 429 // terminators that precede it. 430 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 431 unsigned Reg = blockCSI[i].getReg(); 432 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 433 TII.loadRegFromStackSlot(*MBB, I, Reg, 434 blockCSI[i].getFrameIdx(), 435 RC, TRI); 436 assert(I != MBB->begin() && 437 "loadRegFromStackSlot didn't insert any code!"); 438 // Insert in reverse order. loadRegFromStackSlot can insert 439 // multiple instructions. 440 if (AtStart) 441 I = MBB->begin(); 442 else { 443 I = BeforeI; 444 ++I; 445 } 446 } 447 } 448} 449 450/// AdjustStackOffset - Helper function used to adjust the stack frame offset. 451static inline void 452AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, 453 bool StackGrowsDown, int64_t &Offset, 454 unsigned &MaxAlign) { 455 // If the stack grows down, add the object size to find the lowest address. 456 if (StackGrowsDown) 457 Offset += MFI->getObjectSize(FrameIdx); 458 459 unsigned Align = MFI->getObjectAlignment(FrameIdx); 460 461 // If the alignment of this object is greater than that of the stack, then 462 // increase the stack alignment to match. 463 MaxAlign = std::max(MaxAlign, Align); 464 465 // Adjust to alignment boundary. 466 Offset = (Offset + Align - 1) / Align * Align; 467 468 if (StackGrowsDown) { 469 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << -Offset << "]\n"); 470 MFI->setObjectOffset(FrameIdx, -Offset); // Set the computed offset 471 } else { 472 DEBUG(dbgs() << "alloc FI(" << FrameIdx << ") at SP[" << Offset << "]\n"); 473 MFI->setObjectOffset(FrameIdx, Offset); 474 Offset += MFI->getObjectSize(FrameIdx); 475 } 476} 477 478/// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 479/// abstract stack objects. 480/// 481void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { 482 const TargetFrameLowering &TFI = *Fn.getTarget().getFrameLowering(); 483 484 bool StackGrowsDown = 485 TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 486 487 // Loop over all of the stack objects, assigning sequential addresses... 488 MachineFrameInfo *MFI = Fn.getFrameInfo(); 489 490 // Start at the beginning of the local area. 491 // The Offset is the distance from the stack top in the direction 492 // of stack growth -- so it's always nonnegative. 493 int LocalAreaOffset = TFI.getOffsetOfLocalArea(); 494 if (StackGrowsDown) 495 LocalAreaOffset = -LocalAreaOffset; 496 assert(LocalAreaOffset >= 0 497 && "Local area offset should be in direction of stack growth"); 498 int64_t Offset = LocalAreaOffset; 499 500 // If there are fixed sized objects that are preallocated in the local area, 501 // non-fixed objects can't be allocated right at the start of local area. 502 // We currently don't support filling in holes in between fixed sized 503 // objects, so we adjust 'Offset' to point to the end of last fixed sized 504 // preallocated object. 505 for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) { 506 int64_t FixedOff; 507 if (StackGrowsDown) { 508 // The maximum distance from the stack pointer is at lower address of 509 // the object -- which is given by offset. For down growing stack 510 // the offset is negative, so we negate the offset to get the distance. 511 FixedOff = -MFI->getObjectOffset(i); 512 } else { 513 // The maximum distance from the start pointer is at the upper 514 // address of the object. 515 FixedOff = MFI->getObjectOffset(i) + MFI->getObjectSize(i); 516 } 517 if (FixedOff > Offset) Offset = FixedOff; 518 } 519 520 // First assign frame offsets to stack objects that are used to spill 521 // callee saved registers. 522 if (StackGrowsDown) { 523 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) { 524 // If the stack grows down, we need to add the size to find the lowest 525 // address of the object. 526 Offset += MFI->getObjectSize(i); 527 528 unsigned Align = MFI->getObjectAlignment(i); 529 // Adjust to alignment boundary 530 Offset = (Offset+Align-1)/Align*Align; 531 532 MFI->setObjectOffset(i, -Offset); // Set the computed offset 533 } 534 } else { 535 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; 536 for (int i = MaxCSFI; i >= MinCSFI ; --i) { 537 unsigned Align = MFI->getObjectAlignment(i); 538 // Adjust to alignment boundary 539 Offset = (Offset+Align-1)/Align*Align; 540 541 MFI->setObjectOffset(i, Offset); 542 Offset += MFI->getObjectSize(i); 543 } 544 } 545 546 unsigned MaxAlign = MFI->getMaxAlignment(); 547 548 // Make sure the special register scavenging spill slot is closest to the 549 // frame pointer if a frame pointer is required. 550 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 551 if (RS && TFI.hasFP(Fn) && RegInfo->useFPForScavengingIndex(Fn) && 552 !RegInfo->needsStackRealignment(Fn)) { 553 int SFI = RS->getScavengingFrameIndex(); 554 if (SFI >= 0) 555 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 556 } 557 558 // FIXME: Once this is working, then enable flag will change to a target 559 // check for whether the frame is large enough to want to use virtual 560 // frame index registers. Functions which don't want/need this optimization 561 // will continue to use the existing code path. 562 if (MFI->getUseLocalStackAllocationBlock()) { 563 unsigned Align = MFI->getLocalFrameMaxAlign(); 564 565 // Adjust to alignment boundary. 566 Offset = (Offset + Align - 1) / Align * Align; 567 568 DEBUG(dbgs() << "Local frame base offset: " << Offset << "\n"); 569 570 // Resolve offsets for objects in the local block. 571 for (unsigned i = 0, e = MFI->getLocalFrameObjectCount(); i != e; ++i) { 572 std::pair<int, int64_t> Entry = MFI->getLocalFrameObjectMap(i); 573 int64_t FIOffset = (StackGrowsDown ? -Offset : Offset) + Entry.second; 574 DEBUG(dbgs() << "alloc FI(" << Entry.first << ") at SP[" << 575 FIOffset << "]\n"); 576 MFI->setObjectOffset(Entry.first, FIOffset); 577 } 578 // Allocate the local block 579 Offset += MFI->getLocalFrameSize(); 580 581 MaxAlign = std::max(Align, MaxAlign); 582 } 583 584 // Make sure that the stack protector comes before the local variables on the 585 // stack. 586 SmallSet<int, 16> LargeStackObjs; 587 if (MFI->getStackProtectorIndex() >= 0) { 588 AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), StackGrowsDown, 589 Offset, MaxAlign); 590 591 // Assign large stack objects first. 592 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 593 if (MFI->isObjectPreAllocated(i) && 594 MFI->getUseLocalStackAllocationBlock()) 595 continue; 596 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 597 continue; 598 if (RS && (int)i == RS->getScavengingFrameIndex()) 599 continue; 600 if (MFI->isDeadObjectIndex(i)) 601 continue; 602 if (MFI->getStackProtectorIndex() == (int)i) 603 continue; 604 if (!MFI->MayNeedStackProtector(i)) 605 continue; 606 607 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 608 LargeStackObjs.insert(i); 609 } 610 } 611 612 // Then assign frame offsets to stack objects that are not used to spill 613 // callee saved registers. 614 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 615 if (MFI->isObjectPreAllocated(i) && 616 MFI->getUseLocalStackAllocationBlock()) 617 continue; 618 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 619 continue; 620 if (RS && (int)i == RS->getScavengingFrameIndex()) 621 continue; 622 if (MFI->isDeadObjectIndex(i)) 623 continue; 624 if (MFI->getStackProtectorIndex() == (int)i) 625 continue; 626 if (LargeStackObjs.count(i)) 627 continue; 628 629 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 630 } 631 632 // Make sure the special register scavenging spill slot is closest to the 633 // stack pointer. 634 if (RS && (!TFI.hasFP(Fn) || RegInfo->needsStackRealignment(Fn) || 635 !RegInfo->useFPForScavengingIndex(Fn))) { 636 int SFI = RS->getScavengingFrameIndex(); 637 if (SFI >= 0) 638 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 639 } 640 641 if (!TFI.targetHandlesStackFrameRounding()) { 642 // If we have reserved argument space for call sites in the function 643 // immediately on entry to the current function, count it as part of the 644 // overall stack size. 645 if (MFI->adjustsStack() && TFI.hasReservedCallFrame(Fn)) 646 Offset += MFI->getMaxCallFrameSize(); 647 648 // Round up the size to a multiple of the alignment. If the function has 649 // any calls or alloca's, align to the target's StackAlignment value to 650 // ensure that the callee's frame or the alloca data is suitably aligned; 651 // otherwise, for leaf functions, align to the TransientStackAlignment 652 // value. 653 unsigned StackAlign; 654 if (MFI->adjustsStack() || MFI->hasVarSizedObjects() || 655 (RegInfo->needsStackRealignment(Fn) && MFI->getObjectIndexEnd() != 0)) 656 StackAlign = TFI.getStackAlignment(); 657 else 658 StackAlign = TFI.getTransientStackAlignment(); 659 660 // If the frame pointer is eliminated, all frame offsets will be relative to 661 // SP not FP. Align to MaxAlign so this works. 662 StackAlign = std::max(StackAlign, MaxAlign); 663 unsigned AlignMask = StackAlign - 1; 664 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask); 665 } 666 667 // Update frame info to pretend that this is part of the stack... 668 int64_t StackSize = Offset - LocalAreaOffset; 669 MFI->setStackSize(StackSize); 670 NumBytesStackSpace += StackSize; 671} 672 673/// insertPrologEpilogCode - Scan the function for modified callee saved 674/// registers, insert spill code for these callee saved registers, then add 675/// prolog and epilog code to the function. 676/// 677void PEI::insertPrologEpilogCode(MachineFunction &Fn) { 678 const TargetFrameLowering &TFI = *Fn.getTarget().getFrameLowering(); 679 680 // Add prologue to the function... 681 TFI.emitPrologue(Fn); 682 683 // Add epilogue to restore the callee-save registers in each exiting block 684 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 685 // If last instruction is a return instruction, add an epilogue 686 if (!I->empty() && I->back().isReturn()) 687 TFI.emitEpilogue(Fn, *I); 688 } 689 690 // Emit additional code that is required to support segmented stacks, if 691 // we've been asked for it. This, when linked with a runtime with support 692 // for segmented stacks (libgcc is one), will result in allocating stack 693 // space in small chunks instead of one large contiguous block. 694 if (Fn.getTarget().Options.EnableSegmentedStacks) 695 TFI.adjustForSegmentedStacks(Fn); 696} 697 698/// replaceFrameIndices - Replace all MO_FrameIndex operands with physical 699/// register references and actual offsets. 700/// 701void PEI::replaceFrameIndices(MachineFunction &Fn) { 702 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do? 703 704 const TargetMachine &TM = Fn.getTarget(); 705 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!"); 706 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 707 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 708 const TargetFrameLowering *TFI = TM.getFrameLowering(); 709 bool StackGrowsDown = 710 TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; 711 int FrameSetupOpcode = TII.getCallFrameSetupOpcode(); 712 int FrameDestroyOpcode = TII.getCallFrameDestroyOpcode(); 713 714 for (MachineFunction::iterator BB = Fn.begin(), 715 E = Fn.end(); BB != E; ++BB) { 716#ifndef NDEBUG 717 int SPAdjCount = 0; // frame setup / destroy count. 718#endif 719 int SPAdj = 0; // SP offset due to call frame setup / destroy. 720 if (RS && !FrameIndexVirtualScavenging) RS->enterBasicBlock(BB); 721 722 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 723 724 if (I->getOpcode() == FrameSetupOpcode || 725 I->getOpcode() == FrameDestroyOpcode) { 726#ifndef NDEBUG 727 // Track whether we see even pairs of them 728 SPAdjCount += I->getOpcode() == FrameSetupOpcode ? 1 : -1; 729#endif 730 // Remember how much SP has been adjusted to create the call 731 // frame. 732 int Size = I->getOperand(0).getImm(); 733 734 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) || 735 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode)) 736 Size = -Size; 737 738 SPAdj += Size; 739 740 MachineBasicBlock::iterator PrevI = BB->end(); 741 if (I != BB->begin()) PrevI = prior(I); 742 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I); 743 744 // Visit the instructions created by eliminateCallFramePseudoInstr(). 745 if (PrevI == BB->end()) 746 I = BB->begin(); // The replaced instr was the first in the block. 747 else 748 I = llvm::next(PrevI); 749 continue; 750 } 751 752 MachineInstr *MI = I; 753 bool DoIncr = true; 754 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 755 if (MI->getOperand(i).isFI()) { 756 // Some instructions (e.g. inline asm instructions) can have 757 // multiple frame indices and/or cause eliminateFrameIndex 758 // to insert more than one instruction. We need the register 759 // scavenger to go through all of these instructions so that 760 // it can update its register information. We keep the 761 // iterator at the point before insertion so that we can 762 // revisit them in full. 763 bool AtBeginning = (I == BB->begin()); 764 if (!AtBeginning) --I; 765 766 // If this instruction has a FrameIndex operand, we need to 767 // use that target machine register info object to eliminate 768 // it. 769 TRI.eliminateFrameIndex(MI, SPAdj, i, 770 FrameIndexVirtualScavenging ? NULL : RS); 771 772 // Reset the iterator if we were at the beginning of the BB. 773 if (AtBeginning) { 774 I = BB->begin(); 775 DoIncr = false; 776 } 777 778 MI = 0; 779 break; 780 } 781 782 if (DoIncr && I != BB->end()) ++I; 783 784 // Update register states. 785 if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI); 786 } 787 788 // If we have evenly matched pairs of frame setup / destroy instructions, 789 // make sure the adjustments come out to zero. If we don't have matched 790 // pairs, we can't be sure the missing bit isn't in another basic block 791 // due to a custom inserter playing tricks, so just asserting SPAdj==0 792 // isn't sufficient. See tMOVCC on Thumb1, for example. 793 assert((SPAdjCount || SPAdj == 0) && 794 "Unbalanced call frame setup / destroy pairs?"); 795 } 796} 797 798/// scavengeFrameVirtualRegs - Replace all frame index virtual registers 799/// with physical registers. Use the register scavenger to find an 800/// appropriate register to use. 801/// 802/// FIXME: Iterating over the instruction stream is unnecessary. We can simply 803/// iterate over the vreg use list, which at this point only contains machine 804/// operands for which eliminateFrameIndex need a new scratch reg. 805void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) { 806 // Run through the instructions and find any virtual registers. 807 for (MachineFunction::iterator BB = Fn.begin(), 808 E = Fn.end(); BB != E; ++BB) { 809 RS->enterBasicBlock(BB); 810 811 unsigned VirtReg = 0; 812 unsigned ScratchReg = 0; 813 int SPAdj = 0; 814 815 // The instruction stream may change in the loop, so check BB->end() 816 // directly. 817 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 818 MachineInstr *MI = I; 819 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 820 if (MI->getOperand(i).isReg()) { 821 MachineOperand &MO = MI->getOperand(i); 822 unsigned Reg = MO.getReg(); 823 if (Reg == 0) 824 continue; 825 if (!TargetRegisterInfo::isVirtualRegister(Reg)) 826 continue; 827 828 ++NumVirtualFrameRegs; 829 830 // Have we already allocated a scratch register for this virtual? 831 if (Reg != VirtReg) { 832 // When we first encounter a new virtual register, it 833 // must be a definition. 834 assert(MI->getOperand(i).isDef() && 835 "frame index virtual missing def!"); 836 // Scavenge a new scratch register 837 VirtReg = Reg; 838 const TargetRegisterClass *RC = Fn.getRegInfo().getRegClass(Reg); 839 ScratchReg = RS->scavengeRegister(RC, I, SPAdj); 840 ++NumScavengedRegs; 841 } 842 // Replace this reference to the virtual register with the 843 // scratch register. 844 assert (ScratchReg && "Missing scratch register!"); 845 MI->getOperand(i).setReg(ScratchReg); 846 847 } 848 } 849 RS->forward(I); 850 ++I; 851 } 852 } 853} 854