PrologEpilogInserter.cpp revision ce48c1de828688b34cf5c2038fde23368a0a45f4
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#include "PrologEpilogInserter.h" 23#include "llvm/CodeGen/MachineDominators.h" 24#include "llvm/CodeGen/MachineLoopInfo.h" 25#include "llvm/CodeGen/MachineInstr.h" 26#include "llvm/CodeGen/MachineFrameInfo.h" 27#include "llvm/CodeGen/MachineRegisterInfo.h" 28#include "llvm/CodeGen/RegisterScavenging.h" 29#include "llvm/Target/TargetMachine.h" 30#include "llvm/Target/TargetRegisterInfo.h" 31#include "llvm/Target/TargetFrameInfo.h" 32#include "llvm/Target/TargetInstrInfo.h" 33#include "llvm/Support/CommandLine.h" 34#include "llvm/Support/Compiler.h" 35#include "llvm/ADT/IndexedMap.h" 36#include "llvm/ADT/STLExtras.h" 37#include <climits> 38 39using namespace llvm; 40 41char PEI::ID = 0; 42 43static RegisterPass<PEI> 44X("prologepilog", "Prologue/Epilogue Insertion"); 45 46/// createPrologEpilogCodeInserter - This function returns a pass that inserts 47/// prolog and epilog code, and eliminates abstract frame references. 48/// 49FunctionPass *llvm::createPrologEpilogCodeInserter() { return new PEI(); } 50 51/// runOnMachineFunction - Insert prolog/epilog code and replace abstract 52/// frame indexes with appropriate references. 53/// 54bool PEI::runOnMachineFunction(MachineFunction &Fn) { 55 const Function* F = Fn.getFunction(); 56 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 57 RS = TRI->requiresRegisterScavenging(Fn) ? new RegScavenger() : NULL; 58 FrameIndexVirtualScavenging = TRI->requiresFrameIndexScavenging(Fn); 59 FrameConstantRegMap.clear(); 60 61 // Calculate the MaxCallFrameSize and AdjustsStack variables for the 62 // function's frame information. Also eliminates call frame pseudo 63 // instructions. 64 calculateCallsInformation(Fn); 65 66 // Allow the target machine to make some adjustments to the function 67 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters. 68 TRI->processFunctionBeforeCalleeSavedScan(Fn, RS); 69 70 // Scan the function for modified callee saved registers and insert spill code 71 // for any callee saved registers that are modified. 72 calculateCalleeSavedRegisters(Fn); 73 74 // Determine placement of CSR spill/restore code: 75 // - with shrink wrapping, place spills and restores to tightly 76 // enclose regions in the Machine CFG of the function where 77 // they are used. Without shrink wrapping 78 // - default (no shrink wrapping), place all spills in the 79 // entry block, all restores in return blocks. 80 placeCSRSpillsAndRestores(Fn); 81 82 // Add the code to save and restore the callee saved registers 83 if (!F->hasFnAttr(Attribute::Naked)) 84 insertCSRSpillsAndRestores(Fn); 85 86 // Allow the target machine to make final modifications to the function 87 // before the frame layout is finalized. 88 TRI->processFunctionBeforeFrameFinalized(Fn); 89 90 // Calculate actual frame offsets for all abstract stack objects... 91 calculateFrameObjectOffsets(Fn); 92 93 // Add prolog and epilog code to the function. This function is required 94 // to align the stack frame as necessary for any stack variables or 95 // called functions. Because of this, calculateCalleeSavedRegisters() 96 // must be called before this function in order to set the AdjustsStack 97 // and MaxCallFrameSize variables. 98 if (!F->hasFnAttr(Attribute::Naked)) 99 insertPrologEpilogCode(Fn); 100 101 // Replace all MO_FrameIndex operands with physical register references 102 // and actual offsets. 103 // 104 replaceFrameIndices(Fn); 105 106 // If register scavenging is needed, as we've enabled doing it as a 107 // post-pass, scavenge the virtual registers that frame index elimiation 108 // inserted. 109 if (TRI->requiresRegisterScavenging(Fn) && FrameIndexVirtualScavenging) 110 scavengeFrameVirtualRegs(Fn); 111 112 delete RS; 113 clearAllSets(); 114 return true; 115} 116 117#if 0 118void PEI::getAnalysisUsage(AnalysisUsage &AU) const { 119 AU.setPreservesCFG(); 120 if (ShrinkWrapping || ShrinkWrapFunc != "") { 121 AU.addRequired<MachineLoopInfo>(); 122 AU.addRequired<MachineDominatorTree>(); 123 } 124 AU.addPreserved<MachineLoopInfo>(); 125 AU.addPreserved<MachineDominatorTree>(); 126 MachineFunctionPass::getAnalysisUsage(AU); 127} 128#endif 129 130/// calculateCallsInformation - Calculate the MaxCallFrameSize and AdjustsStack 131/// variables for the function's frame information and eliminate call frame 132/// pseudo instructions. 133void PEI::calculateCallsInformation(MachineFunction &Fn) { 134 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 135 MachineFrameInfo *MFI = Fn.getFrameInfo(); 136 137 unsigned MaxCallFrameSize = 0; 138 bool AdjustsStack = MFI->adjustsStack(); 139 140 // Get the function call frame set-up and tear-down instruction opcode 141 int FrameSetupOpcode = RegInfo->getCallFrameSetupOpcode(); 142 int FrameDestroyOpcode = RegInfo->getCallFrameDestroyOpcode(); 143 144 // Early exit for targets which have no call frame setup/destroy pseudo 145 // instructions. 146 if (FrameSetupOpcode == -1 && FrameDestroyOpcode == -1) 147 return; 148 149 std::vector<MachineBasicBlock::iterator> FrameSDOps; 150 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) 151 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) 152 if (I->getOpcode() == FrameSetupOpcode || 153 I->getOpcode() == FrameDestroyOpcode) { 154 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo" 155 " instructions should have a single immediate argument!"); 156 unsigned Size = I->getOperand(0).getImm(); 157 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size; 158 AdjustsStack = true; 159 FrameSDOps.push_back(I); 160 } else if (I->isInlineAsm()) { 161 // An InlineAsm might be a call; assume it is to get the stack frame 162 // aligned correctly for calls. 163 AdjustsStack = true; 164 } 165 166 MFI->setAdjustsStack(AdjustsStack); 167 MFI->setMaxCallFrameSize(MaxCallFrameSize); 168 169 for (std::vector<MachineBasicBlock::iterator>::iterator 170 i = FrameSDOps.begin(), e = FrameSDOps.end(); i != e; ++i) { 171 MachineBasicBlock::iterator I = *i; 172 173 // If call frames are not being included as part of the stack frame, and 174 // the target doesn't indicate otherwise, remove the call frame pseudos 175 // here. The sub/add sp instruction pairs are still inserted, but we don't 176 // need to track the SP adjustment for frame index elimination. 177 if (RegInfo->canSimplifyCallFramePseudos(Fn)) 178 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I); 179 } 180} 181 182 183/// calculateCalleeSavedRegisters - Scan the function for modified callee saved 184/// registers. 185void PEI::calculateCalleeSavedRegisters(MachineFunction &Fn) { 186 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 187 const TargetFrameInfo *TFI = Fn.getTarget().getFrameInfo(); 188 MachineFrameInfo *MFI = Fn.getFrameInfo(); 189 190 // Get the callee saved register list... 191 const unsigned *CSRegs = RegInfo->getCalleeSavedRegs(&Fn); 192 193 // These are used to keep track the callee-save area. Initialize them. 194 MinCSFrameIndex = INT_MAX; 195 MaxCSFrameIndex = 0; 196 197 // Early exit for targets which have no callee saved registers. 198 if (CSRegs == 0 || CSRegs[0] == 0) 199 return; 200 201 // In Naked functions we aren't going to save any registers. 202 if (Fn.getFunction()->hasFnAttr(Attribute::Naked)) 203 return; 204 205 std::vector<CalleeSavedInfo> CSI; 206 for (unsigned i = 0; CSRegs[i]; ++i) { 207 unsigned Reg = CSRegs[i]; 208 const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg); 209 if (Fn.getRegInfo().isPhysRegUsed(Reg)) { 210 // If the reg is modified, save it! 211 CSI.push_back(CalleeSavedInfo(Reg, RC)); 212 } else { 213 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg); 214 *AliasSet; ++AliasSet) { // Check alias registers too. 215 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) { 216 CSI.push_back(CalleeSavedInfo(Reg, RC)); 217 break; 218 } 219 } 220 } 221 } 222 223 if (CSI.empty()) 224 return; // Early exit if no callee saved registers are modified! 225 226 unsigned NumFixedSpillSlots; 227 const TargetFrameInfo::SpillSlot *FixedSpillSlots = 228 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots); 229 230 // Now that we know which registers need to be saved and restored, allocate 231 // stack slots for them. 232 for (std::vector<CalleeSavedInfo>::iterator 233 I = CSI.begin(), E = CSI.end(); I != E; ++I) { 234 unsigned Reg = I->getReg(); 235 const TargetRegisterClass *RC = I->getRegClass(); 236 237 int FrameIdx; 238 if (RegInfo->hasReservedSpillSlot(Fn, Reg, FrameIdx)) { 239 I->setFrameIdx(FrameIdx); 240 continue; 241 } 242 243 // Check to see if this physreg must be spilled to a particular stack slot 244 // on this target. 245 const TargetFrameInfo::SpillSlot *FixedSlot = FixedSpillSlots; 246 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots && 247 FixedSlot->Reg != Reg) 248 ++FixedSlot; 249 250 if (FixedSlot == FixedSpillSlots + NumFixedSpillSlots) { 251 // Nope, just spill it anywhere convenient. 252 unsigned Align = RC->getAlignment(); 253 unsigned StackAlign = TFI->getStackAlignment(); 254 255 // We may not be able to satisfy the desired alignment specification of 256 // the TargetRegisterClass if the stack alignment is smaller. Use the 257 // min. 258 Align = std::min(Align, StackAlign); 259 FrameIdx = MFI->CreateStackObject(RC->getSize(), Align, true); 260 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx; 261 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx; 262 } else { 263 // Spill it to the stack where we must. 264 FrameIdx = MFI->CreateFixedObject(RC->getSize(), FixedSlot->Offset, 265 true, false); 266 } 267 268 I->setFrameIdx(FrameIdx); 269 } 270 271 MFI->setCalleeSavedInfo(CSI); 272} 273 274/// insertCSRSpillsAndRestores - Insert spill and restore code for 275/// callee saved registers used in the function, handling shrink wrapping. 276/// 277void PEI::insertCSRSpillsAndRestores(MachineFunction &Fn) { 278 // Get callee saved register information. 279 MachineFrameInfo *MFI = Fn.getFrameInfo(); 280 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 281 282 MFI->setCalleeSavedInfoValid(true); 283 284 // Early exit if no callee saved registers are modified! 285 if (CSI.empty()) 286 return; 287 288 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 289 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 290 MachineBasicBlock::iterator I; 291 292 if (! ShrinkWrapThisFunction) { 293 // Spill using target interface. 294 I = EntryBlock->begin(); 295 if (!TII.spillCalleeSavedRegisters(*EntryBlock, I, CSI, TRI)) { 296 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 297 // Add the callee-saved register as live-in. 298 // It's killed at the spill. 299 EntryBlock->addLiveIn(CSI[i].getReg()); 300 301 // Insert the spill to the stack frame. 302 TII.storeRegToStackSlot(*EntryBlock, I, CSI[i].getReg(), true, 303 CSI[i].getFrameIdx(), CSI[i].getRegClass(),TRI); 304 } 305 } 306 307 // Restore using target interface. 308 for (unsigned ri = 0, re = ReturnBlocks.size(); ri != re; ++ri) { 309 MachineBasicBlock* MBB = ReturnBlocks[ri]; 310 I = MBB->end(); --I; 311 312 // Skip over all terminator instructions, which are part of the return 313 // sequence. 314 MachineBasicBlock::iterator I2 = I; 315 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 316 I = I2; 317 318 bool AtStart = I == MBB->begin(); 319 MachineBasicBlock::iterator BeforeI = I; 320 if (!AtStart) 321 --BeforeI; 322 323 // Restore all registers immediately before the return and any 324 // terminators that preceed it. 325 if (!TII.restoreCalleeSavedRegisters(*MBB, I, CSI, TRI)) { 326 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 327 TII.loadRegFromStackSlot(*MBB, I, CSI[i].getReg(), 328 CSI[i].getFrameIdx(), 329 CSI[i].getRegClass(), TRI); 330 assert(I != MBB->begin() && 331 "loadRegFromStackSlot didn't insert any code!"); 332 // Insert in reverse order. loadRegFromStackSlot can insert 333 // multiple instructions. 334 if (AtStart) 335 I = MBB->begin(); 336 else { 337 I = BeforeI; 338 ++I; 339 } 340 } 341 } 342 } 343 return; 344 } 345 346 // Insert spills. 347 std::vector<CalleeSavedInfo> blockCSI; 348 for (CSRegBlockMap::iterator BI = CSRSave.begin(), 349 BE = CSRSave.end(); BI != BE; ++BI) { 350 MachineBasicBlock* MBB = BI->first; 351 CSRegSet save = BI->second; 352 353 if (save.empty()) 354 continue; 355 356 blockCSI.clear(); 357 for (CSRegSet::iterator RI = save.begin(), 358 RE = save.end(); RI != RE; ++RI) { 359 blockCSI.push_back(CSI[*RI]); 360 } 361 assert(blockCSI.size() > 0 && 362 "Could not collect callee saved register info"); 363 364 I = MBB->begin(); 365 366 // When shrink wrapping, use stack slot stores/loads. 367 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 368 // Add the callee-saved register as live-in. 369 // It's killed at the spill. 370 MBB->addLiveIn(blockCSI[i].getReg()); 371 372 // Insert the spill to the stack frame. 373 TII.storeRegToStackSlot(*MBB, I, blockCSI[i].getReg(), 374 true, 375 blockCSI[i].getFrameIdx(), 376 blockCSI[i].getRegClass(), TRI); 377 } 378 } 379 380 for (CSRegBlockMap::iterator BI = CSRRestore.begin(), 381 BE = CSRRestore.end(); BI != BE; ++BI) { 382 MachineBasicBlock* MBB = BI->first; 383 CSRegSet restore = BI->second; 384 385 if (restore.empty()) 386 continue; 387 388 blockCSI.clear(); 389 for (CSRegSet::iterator RI = restore.begin(), 390 RE = restore.end(); RI != RE; ++RI) { 391 blockCSI.push_back(CSI[*RI]); 392 } 393 assert(blockCSI.size() > 0 && 394 "Could not find callee saved register info"); 395 396 // If MBB is empty and needs restores, insert at the _beginning_. 397 if (MBB->empty()) { 398 I = MBB->begin(); 399 } else { 400 I = MBB->end(); 401 --I; 402 403 // Skip over all terminator instructions, which are part of the 404 // return sequence. 405 if (! I->getDesc().isTerminator()) { 406 ++I; 407 } else { 408 MachineBasicBlock::iterator I2 = I; 409 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 410 I = I2; 411 } 412 } 413 414 bool AtStart = I == MBB->begin(); 415 MachineBasicBlock::iterator BeforeI = I; 416 if (!AtStart) 417 --BeforeI; 418 419 // Restore all registers immediately before the return and any 420 // terminators that preceed it. 421 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 422 TII.loadRegFromStackSlot(*MBB, I, blockCSI[i].getReg(), 423 blockCSI[i].getFrameIdx(), 424 blockCSI[i].getRegClass(), TRI); 425 assert(I != MBB->begin() && 426 "loadRegFromStackSlot didn't insert any code!"); 427 // Insert in reverse order. loadRegFromStackSlot can insert 428 // multiple instructions. 429 if (AtStart) 430 I = MBB->begin(); 431 else { 432 I = BeforeI; 433 ++I; 434 } 435 } 436 } 437} 438 439/// AdjustStackOffset - Helper function used to adjust the stack frame offset. 440static inline void 441AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, 442 bool StackGrowsDown, int64_t &Offset, 443 unsigned &MaxAlign) { 444 // If the stack grows down, add the object size to find the lowest address. 445 if (StackGrowsDown) 446 Offset += MFI->getObjectSize(FrameIdx); 447 448 unsigned Align = MFI->getObjectAlignment(FrameIdx); 449 450 // If the alignment of this object is greater than that of the stack, then 451 // increase the stack alignment to match. 452 MaxAlign = std::max(MaxAlign, Align); 453 454 // Adjust to alignment boundary. 455 Offset = (Offset + Align - 1) / Align * Align; 456 457 if (StackGrowsDown) { 458 MFI->setObjectOffset(FrameIdx, -Offset); // Set the computed offset 459 } else { 460 MFI->setObjectOffset(FrameIdx, Offset); 461 Offset += MFI->getObjectSize(FrameIdx); 462 } 463} 464 465/// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 466/// abstract stack objects. 467/// 468void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { 469 const TargetFrameInfo &TFI = *Fn.getTarget().getFrameInfo(); 470 471 bool StackGrowsDown = 472 TFI.getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 473 474 // Loop over all of the stack objects, assigning sequential addresses... 475 MachineFrameInfo *MFI = Fn.getFrameInfo(); 476 477 // Start at the beginning of the local area. 478 // The Offset is the distance from the stack top in the direction 479 // of stack growth -- so it's always nonnegative. 480 int LocalAreaOffset = TFI.getOffsetOfLocalArea(); 481 if (StackGrowsDown) 482 LocalAreaOffset = -LocalAreaOffset; 483 assert(LocalAreaOffset >= 0 484 && "Local area offset should be in direction of stack growth"); 485 int64_t Offset = LocalAreaOffset; 486 487 // If there are fixed sized objects that are preallocated in the local area, 488 // non-fixed objects can't be allocated right at the start of local area. 489 // We currently don't support filling in holes in between fixed sized 490 // objects, so we adjust 'Offset' to point to the end of last fixed sized 491 // preallocated object. 492 for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) { 493 int64_t FixedOff; 494 if (StackGrowsDown) { 495 // The maximum distance from the stack pointer is at lower address of 496 // the object -- which is given by offset. For down growing stack 497 // the offset is negative, so we negate the offset to get the distance. 498 FixedOff = -MFI->getObjectOffset(i); 499 } else { 500 // The maximum distance from the start pointer is at the upper 501 // address of the object. 502 FixedOff = MFI->getObjectOffset(i) + MFI->getObjectSize(i); 503 } 504 if (FixedOff > Offset) Offset = FixedOff; 505 } 506 507 // First assign frame offsets to stack objects that are used to spill 508 // callee saved registers. 509 if (StackGrowsDown) { 510 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) { 511 // If the stack grows down, we need to add the size to find the lowest 512 // address of the object. 513 Offset += MFI->getObjectSize(i); 514 515 unsigned Align = MFI->getObjectAlignment(i); 516 // Adjust to alignment boundary 517 Offset = (Offset+Align-1)/Align*Align; 518 519 MFI->setObjectOffset(i, -Offset); // Set the computed offset 520 } 521 } else { 522 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; 523 for (int i = MaxCSFI; i >= MinCSFI ; --i) { 524 unsigned Align = MFI->getObjectAlignment(i); 525 // Adjust to alignment boundary 526 Offset = (Offset+Align-1)/Align*Align; 527 528 MFI->setObjectOffset(i, Offset); 529 Offset += MFI->getObjectSize(i); 530 } 531 } 532 533 unsigned MaxAlign = MFI->getMaxAlignment(); 534 535 // Make sure the special register scavenging spill slot is closest to the 536 // frame pointer if a frame pointer is required. 537 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 538 if (RS && RegInfo->hasFP(Fn) && !RegInfo->needsStackRealignment(Fn)) { 539 int SFI = RS->getScavengingFrameIndex(); 540 if (SFI >= 0) 541 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 542 } 543 544 // Make sure that the stack protector comes before the local variables on the 545 // stack. 546 if (MFI->getStackProtectorIndex() >= 0) 547 AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), StackGrowsDown, 548 Offset, MaxAlign); 549 550 // Then assign frame offsets to stack objects that are not used to spill 551 // callee saved registers. 552 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 553 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 554 continue; 555 if (RS && (int)i == RS->getScavengingFrameIndex()) 556 continue; 557 if (MFI->isDeadObjectIndex(i)) 558 continue; 559 if (MFI->getStackProtectorIndex() == (int)i) 560 continue; 561 562 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 563 } 564 565 // Make sure the special register scavenging spill slot is closest to the 566 // stack pointer. 567 if (RS && (!RegInfo->hasFP(Fn) || RegInfo->needsStackRealignment(Fn))) { 568 int SFI = RS->getScavengingFrameIndex(); 569 if (SFI >= 0) 570 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 571 } 572 573 if (!RegInfo->targetHandlesStackFrameRounding()) { 574 // If we have reserved argument space for call sites in the function 575 // immediately on entry to the current function, count it as part of the 576 // overall stack size. 577 if (MFI->adjustsStack() && RegInfo->hasReservedCallFrame(Fn)) 578 Offset += MFI->getMaxCallFrameSize(); 579 580 // Round up the size to a multiple of the alignment. If the function has 581 // any calls or alloca's, align to the target's StackAlignment value to 582 // ensure that the callee's frame or the alloca data is suitably aligned; 583 // otherwise, for leaf functions, align to the TransientStackAlignment 584 // value. 585 unsigned StackAlign; 586 if (MFI->adjustsStack() || MFI->hasVarSizedObjects() || 587 (RegInfo->needsStackRealignment(Fn) && MFI->getObjectIndexEnd() != 0)) 588 StackAlign = TFI.getStackAlignment(); 589 else 590 StackAlign = TFI.getTransientStackAlignment(); 591 592 // If the frame pointer is eliminated, all frame offsets will be relative to 593 // SP not FP. Align to MaxAlign so this works. 594 StackAlign = std::max(StackAlign, MaxAlign); 595 unsigned AlignMask = StackAlign - 1; 596 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask); 597 } 598 599 // Update frame info to pretend that this is part of the stack... 600 MFI->setStackSize(Offset - LocalAreaOffset); 601} 602 603/// insertPrologEpilogCode - Scan the function for modified callee saved 604/// registers, insert spill code for these callee saved registers, then add 605/// prolog and epilog code to the function. 606/// 607void PEI::insertPrologEpilogCode(MachineFunction &Fn) { 608 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 609 610 // Add prologue to the function... 611 TRI->emitPrologue(Fn); 612 613 // Add epilogue to restore the callee-save registers in each exiting block 614 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 615 // If last instruction is a return instruction, add an epilogue 616 if (!I->empty() && I->back().getDesc().isReturn()) 617 TRI->emitEpilogue(Fn, *I); 618 } 619} 620 621/// replaceFrameIndices - Replace all MO_FrameIndex operands with physical 622/// register references and actual offsets. 623/// 624void PEI::replaceFrameIndices(MachineFunction &Fn) { 625 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do? 626 627 const TargetMachine &TM = Fn.getTarget(); 628 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!"); 629 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 630 const TargetFrameInfo *TFI = TM.getFrameInfo(); 631 bool StackGrowsDown = 632 TFI->getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 633 int FrameSetupOpcode = TRI.getCallFrameSetupOpcode(); 634 int FrameDestroyOpcode = TRI.getCallFrameDestroyOpcode(); 635 636 for (MachineFunction::iterator BB = Fn.begin(), 637 E = Fn.end(); BB != E; ++BB) { 638 int SPAdj = 0; // SP offset due to call frame setup / destroy. 639 if (RS && !FrameIndexVirtualScavenging) RS->enterBasicBlock(BB); 640 641 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 642 643 if (I->getOpcode() == FrameSetupOpcode || 644 I->getOpcode() == FrameDestroyOpcode) { 645 // Remember how much SP has been adjusted to create the call 646 // frame. 647 int Size = I->getOperand(0).getImm(); 648 649 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) || 650 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode)) 651 Size = -Size; 652 653 SPAdj += Size; 654 655 MachineBasicBlock::iterator PrevI = BB->end(); 656 if (I != BB->begin()) PrevI = prior(I); 657 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I); 658 659 // Visit the instructions created by eliminateCallFramePseudoInstr(). 660 if (PrevI == BB->end()) 661 I = BB->begin(); // The replaced instr was the first in the block. 662 else 663 I = llvm::next(PrevI); 664 continue; 665 } 666 667 MachineInstr *MI = I; 668 bool DoIncr = true; 669 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 670 if (MI->getOperand(i).isFI()) { 671 // Some instructions (e.g. inline asm instructions) can have 672 // multiple frame indices and/or cause eliminateFrameIndex 673 // to insert more than one instruction. We need the register 674 // scavenger to go through all of these instructions so that 675 // it can update its register information. We keep the 676 // iterator at the point before insertion so that we can 677 // revisit them in full. 678 bool AtBeginning = (I == BB->begin()); 679 if (!AtBeginning) --I; 680 681 // If this instruction has a FrameIndex operand, we need to 682 // use that target machine register info object to eliminate 683 // it. 684 TargetRegisterInfo::FrameIndexValue Value; 685 unsigned VReg = 686 TRI.eliminateFrameIndex(MI, SPAdj, &Value, 687 FrameIndexVirtualScavenging ? NULL : RS); 688 if (VReg) { 689 assert (FrameIndexVirtualScavenging && 690 "Not scavenging, but virtual returned from " 691 "eliminateFrameIndex()!"); 692 FrameConstantRegMap[VReg] = FrameConstantEntry(Value, SPAdj); 693 } 694 695 // Reset the iterator if we were at the beginning of the BB. 696 if (AtBeginning) { 697 I = BB->begin(); 698 DoIncr = false; 699 } 700 701 MI = 0; 702 break; 703 } 704 705 if (DoIncr && I != BB->end()) ++I; 706 707 // Update register states. 708 if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI); 709 } 710 711 assert(SPAdj == 0 && "Unbalanced call frame setup / destroy pairs?"); 712 } 713} 714 715/// findLastUseReg - find the killing use of the specified register within 716/// the instruciton range. Return the operand number of the kill in Operand. 717static MachineBasicBlock::iterator 718findLastUseReg(MachineBasicBlock::iterator I, MachineBasicBlock::iterator ME, 719 unsigned Reg) { 720 // Scan forward to find the last use of this virtual register 721 for (++I; I != ME; ++I) { 722 MachineInstr *MI = I; 723 bool isDefInsn = false; 724 bool isKillInsn = false; 725 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 726 if (MI->getOperand(i).isReg()) { 727 unsigned OpReg = MI->getOperand(i).getReg(); 728 if (OpReg == 0 || !TargetRegisterInfo::isVirtualRegister(OpReg)) 729 continue; 730 assert (OpReg == Reg 731 && "overlapping use of scavenged index register!"); 732 // If this is the killing use, we have a candidate. 733 if (MI->getOperand(i).isKill()) 734 isKillInsn = true; 735 else if (MI->getOperand(i).isDef()) 736 isDefInsn = true; 737 } 738 if (isKillInsn && !isDefInsn) 739 return I; 740 } 741 // If we hit the end of the basic block, there was no kill of 742 // the virtual register, which is wrong. 743 assert (0 && "scavenged index register never killed!"); 744 return ME; 745} 746 747/// scavengeFrameVirtualRegs - Replace all frame index virtual registers 748/// with physical registers. Use the register scavenger to find an 749/// appropriate register to use. 750void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) { 751 // Run through the instructions and find any virtual registers. 752 for (MachineFunction::iterator BB = Fn.begin(), 753 E = Fn.end(); BB != E; ++BB) { 754 RS->enterBasicBlock(BB); 755 756 // FIXME: The logic flow in this function is still too convoluted. 757 // It needs a cleanup refactoring. Do that in preparation for tracking 758 // more than one scratch register value and using ranges to find 759 // available scratch registers. 760 unsigned CurrentVirtReg = 0; 761 unsigned CurrentScratchReg = 0; 762 bool havePrevValue = false; 763 TargetRegisterInfo::FrameIndexValue PrevValue(0,0); 764 TargetRegisterInfo::FrameIndexValue Value(0,0); 765 MachineInstr *PrevLastUseMI = NULL; 766 unsigned PrevLastUseOp = 0; 767 bool trackingCurrentValue = false; 768 int SPAdj = 0; 769 770 // The instruction stream may change in the loop, so check BB->end() 771 // directly. 772 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 773 MachineInstr *MI = I; 774 bool isDefInsn = false; 775 bool isKillInsn = false; 776 bool clobbersScratchReg = false; 777 bool DoIncr = true; 778 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 779 if (MI->getOperand(i).isReg()) { 780 MachineOperand &MO = MI->getOperand(i); 781 unsigned Reg = MO.getReg(); 782 if (Reg == 0) 783 continue; 784 if (!TargetRegisterInfo::isVirtualRegister(Reg)) { 785 // If we have a previous scratch reg, check and see if anything 786 // here kills whatever value is in there. 787 if (Reg == CurrentScratchReg) { 788 if (MO.isUse()) { 789 // Two-address operands implicitly kill 790 if (MO.isKill() || MI->isRegTiedToDefOperand(i)) 791 clobbersScratchReg = true; 792 } else { 793 assert (MO.isDef()); 794 clobbersScratchReg = true; 795 } 796 } 797 continue; 798 } 799 // If this is a def, remember that this insn defines the value. 800 // This lets us properly consider insns which re-use the scratch 801 // register, such as r2 = sub r2, #imm, in the middle of the 802 // scratch range. 803 if (MO.isDef()) 804 isDefInsn = true; 805 806 // Have we already allocated a scratch register for this virtual? 807 if (Reg != CurrentVirtReg) { 808 // When we first encounter a new virtual register, it 809 // must be a definition. 810 assert(MI->getOperand(i).isDef() && 811 "frame index virtual missing def!"); 812 // We can't have nested virtual register live ranges because 813 // there's only a guarantee of one scavenged register at a time. 814 assert (CurrentVirtReg == 0 && 815 "overlapping frame index virtual registers!"); 816 817 // If the target gave us information about what's in the register, 818 // we can use that to re-use scratch regs. 819 DenseMap<unsigned, FrameConstantEntry>::iterator Entry = 820 FrameConstantRegMap.find(Reg); 821 trackingCurrentValue = Entry != FrameConstantRegMap.end(); 822 if (trackingCurrentValue) { 823 SPAdj = (*Entry).second.second; 824 Value = (*Entry).second.first; 825 } else { 826 SPAdj = 0; 827 Value.first = 0; 828 Value.second = 0; 829 } 830 831 // If the scratch register from the last allocation is still 832 // available, see if the value matches. If it does, just re-use it. 833 if (trackingCurrentValue && havePrevValue && PrevValue == Value) { 834 // FIXME: This assumes that the instructions in the live range 835 // for the virtual register are exclusively for the purpose 836 // of populating the value in the register. That's reasonable 837 // for these frame index registers, but it's still a very, very 838 // strong assumption. rdar://7322732. Better would be to 839 // explicitly check each instruction in the range for references 840 // to the virtual register. Only delete those insns that 841 // touch the virtual register. 842 843 // Find the last use of the new virtual register. Remove all 844 // instruction between here and there, and update the current 845 // instruction to reference the last use insn instead. 846 MachineBasicBlock::iterator LastUseMI = 847 findLastUseReg(I, BB->end(), Reg); 848 849 // Remove all instructions up 'til the last use, since they're 850 // just calculating the value we already have. 851 BB->erase(I, LastUseMI); 852 I = LastUseMI; 853 854 // Extend the live range of the scratch register 855 PrevLastUseMI->getOperand(PrevLastUseOp).setIsKill(false); 856 RS->setUsed(CurrentScratchReg); 857 CurrentVirtReg = Reg; 858 859 // We deleted the instruction we were scanning the operands of. 860 // Jump back to the instruction iterator loop. Don't increment 861 // past this instruction since we updated the iterator already. 862 DoIncr = false; 863 break; 864 } 865 866 // Scavenge a new scratch register 867 CurrentVirtReg = Reg; 868 const TargetRegisterClass *RC = Fn.getRegInfo().getRegClass(Reg); 869 CurrentScratchReg = RS->FindUnusedReg(RC); 870 if (CurrentScratchReg == 0) 871 // No register is "free". Scavenge a register. 872 CurrentScratchReg = RS->scavengeRegister(RC, I, SPAdj); 873 874 PrevValue = Value; 875 } 876 // replace this reference to the virtual register with the 877 // scratch register. 878 assert (CurrentScratchReg && "Missing scratch register!"); 879 MI->getOperand(i).setReg(CurrentScratchReg); 880 881 if (MI->getOperand(i).isKill()) { 882 isKillInsn = true; 883 PrevLastUseOp = i; 884 PrevLastUseMI = MI; 885 } 886 } 887 } 888 // If this is the last use of the scratch, stop tracking it. The 889 // last use will be a kill operand in an instruction that does 890 // not also define the scratch register. 891 if (isKillInsn && !isDefInsn) { 892 CurrentVirtReg = 0; 893 havePrevValue = trackingCurrentValue; 894 } 895 // Similarly, notice if instruction clobbered the value in the 896 // register we're tracking for possible later reuse. This is noted 897 // above, but enforced here since the value is still live while we 898 // process the rest of the operands of the instruction. 899 if (clobbersScratchReg) { 900 havePrevValue = false; 901 CurrentScratchReg = 0; 902 } 903 if (DoIncr) { 904 RS->forward(I); 905 ++I; 906 } 907 } 908 } 909} 910