PrologEpilogInserter.cpp revision 8c5358c93675b009ba2d57c3a5980f6bc58ba536
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 HasCalls variables for the function's 62 // frame information. Also eliminates call frame pseudo instructions. 63 calculateCallsInformation(Fn); 64 65 // Allow the target machine to make some adjustments to the function 66 // e.g. UsedPhysRegs before calculateCalleeSavedRegisters. 67 TRI->processFunctionBeforeCalleeSavedScan(Fn, RS); 68 69 // Scan the function for modified callee saved registers and insert spill code 70 // for any callee saved registers that are modified. 71 calculateCalleeSavedRegisters(Fn); 72 73 // Determine placement of CSR spill/restore code: 74 // - with shrink wrapping, place spills and restores to tightly 75 // enclose regions in the Machine CFG of the function where 76 // they are used. Without shrink wrapping 77 // - default (no shrink wrapping), place all spills in the 78 // entry block, all restores in return blocks. 79 placeCSRSpillsAndRestores(Fn); 80 81 // Add the code to save and restore the callee saved registers 82 if (!F->hasFnAttr(Attribute::Naked)) 83 insertCSRSpillsAndRestores(Fn); 84 85 // Allow the target machine to make final modifications to the function 86 // before the frame layout is finalized. 87 TRI->processFunctionBeforeFrameFinalized(Fn); 88 89 // Calculate actual frame offsets for all abstract stack objects... 90 calculateFrameObjectOffsets(Fn); 91 92 // Add prolog and epilog code to the function. This function is required 93 // to align the stack frame as necessary for any stack variables or 94 // called functions. Because of this, calculateCalleeSavedRegisters 95 // must be called before this function in order to set the HasCalls 96 // and MaxCallFrameSize variables. 97 if (!F->hasFnAttr(Attribute::Naked)) 98 insertPrologEpilogCode(Fn); 99 100 // Replace all MO_FrameIndex operands with physical register references 101 // and actual offsets. 102 // 103 replaceFrameIndices(Fn); 104 105 // If register scavenging is needed, as we've enabled doing it as a 106 // post-pass, scavenge the virtual registers that frame index elimiation 107 // inserted. 108 if (TRI->requiresRegisterScavenging(Fn) && FrameIndexVirtualScavenging) 109 scavengeFrameVirtualRegs(Fn); 110 111 delete RS; 112 clearAllSets(); 113 return true; 114} 115 116#if 0 117void PEI::getAnalysisUsage(AnalysisUsage &AU) const { 118 AU.setPreservesCFG(); 119 if (ShrinkWrapping || ShrinkWrapFunc != "") { 120 AU.addRequired<MachineLoopInfo>(); 121 AU.addRequired<MachineDominatorTree>(); 122 } 123 AU.addPreserved<MachineLoopInfo>(); 124 AU.addPreserved<MachineDominatorTree>(); 125 MachineFunctionPass::getAnalysisUsage(AU); 126} 127#endif 128 129/// calculateCallsInformation - Calculate the MaxCallFrameSize and HasCalls 130/// variables for the function's frame information and eliminate call frame 131/// pseudo instructions. 132void PEI::calculateCallsInformation(MachineFunction &Fn) { 133 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 134 MachineFrameInfo *MFI = Fn.getFrameInfo(); 135 136 unsigned MaxCallFrameSize = 0; 137 bool HasCalls = MFI->hasCalls(); 138 139 // Get the function call frame set-up and tear-down instruction opcode 140 int FrameSetupOpcode = RegInfo->getCallFrameSetupOpcode(); 141 int FrameDestroyOpcode = RegInfo->getCallFrameDestroyOpcode(); 142 143 // Early exit for targets which have no call frame setup/destroy pseudo 144 // instructions. 145 if (FrameSetupOpcode == -1 && FrameDestroyOpcode == -1) 146 return; 147 148 std::vector<MachineBasicBlock::iterator> FrameSDOps; 149 for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) 150 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) 151 if (I->getOpcode() == FrameSetupOpcode || 152 I->getOpcode() == FrameDestroyOpcode) { 153 assert(I->getNumOperands() >= 1 && "Call Frame Setup/Destroy Pseudo" 154 " instructions should have a single immediate argument!"); 155 unsigned Size = I->getOperand(0).getImm(); 156 if (Size > MaxCallFrameSize) MaxCallFrameSize = Size; 157 HasCalls = true; 158 FrameSDOps.push_back(I); 159 } else if (I->isInlineAsm()) { 160 // An InlineAsm might be a call; assume it is to get the stack frame 161 // aligned correctly for calls. 162 HasCalls = true; 163 } 164 165 MFI->setHasCalls(HasCalls); 166 MFI->setMaxCallFrameSize(MaxCallFrameSize); 167 168 for (std::vector<MachineBasicBlock::iterator>::iterator 169 i = FrameSDOps.begin(), e = FrameSDOps.end(); i != e; ++i) { 170 MachineBasicBlock::iterator I = *i; 171 172 // If call frames are not being included as part of the stack frame, and 173 // the target doesn't indicate otherwise, remove the call frame pseudos 174 // here. The sub/add sp instruction pairs are still inserted, but we don't 175 // need to track the SP adjustment for frame index elimination. 176 if (RegInfo->canSimplifyCallFramePseudos(Fn)) 177 RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I); 178 } 179} 180 181 182/// calculateCalleeSavedRegisters - Scan the function for modified callee saved 183/// registers. 184void PEI::calculateCalleeSavedRegisters(MachineFunction &Fn) { 185 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 186 const TargetFrameInfo *TFI = Fn.getTarget().getFrameInfo(); 187 MachineFrameInfo *MFI = Fn.getFrameInfo(); 188 189 // Get the callee saved register list... 190 const unsigned *CSRegs = RegInfo->getCalleeSavedRegs(&Fn); 191 192 // These are used to keep track the callee-save area. Initialize them. 193 MinCSFrameIndex = INT_MAX; 194 MaxCSFrameIndex = 0; 195 196 // Early exit for targets which have no callee saved registers. 197 if (CSRegs == 0 || CSRegs[0] == 0) 198 return; 199 200 // In Naked functions we aren't going to save any registers. 201 if (Fn.getFunction()->hasFnAttr(Attribute::Naked)) 202 return; 203 204 // Figure out which *callee saved* registers are modified by the current 205 // function, thus needing to be saved and restored in the prolog/epilog. 206 const TargetRegisterClass * const *CSRegClasses = 207 RegInfo->getCalleeSavedRegClasses(&Fn); 208 209 std::vector<CalleeSavedInfo> CSI; 210 for (unsigned i = 0; CSRegs[i]; ++i) { 211 unsigned Reg = CSRegs[i]; 212 if (Fn.getRegInfo().isPhysRegUsed(Reg)) { 213 // If the reg is modified, save it! 214 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i])); 215 } else { 216 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg); 217 *AliasSet; ++AliasSet) { // Check alias registers too. 218 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) { 219 CSI.push_back(CalleeSavedInfo(Reg, CSRegClasses[i])); 220 break; 221 } 222 } 223 } 224 } 225 226 if (CSI.empty()) 227 return; // Early exit if no callee saved registers are modified! 228 229 unsigned NumFixedSpillSlots; 230 const TargetFrameInfo::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 = I->getRegClass(); 239 240 int FrameIdx; 241 if (RegInfo->hasReservedSpillSlot(Fn, 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 TargetFrameInfo::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, 268 true, false); 269 } 270 271 I->setFrameIdx(FrameIdx); 272 } 273 274 MFI->setCalleeSavedInfo(CSI); 275} 276 277/// insertCSRSpillsAndRestores - Insert spill and restore code for 278/// callee saved registers used in the function, handling shrink wrapping. 279/// 280void PEI::insertCSRSpillsAndRestores(MachineFunction &Fn) { 281 // Get callee saved register information. 282 MachineFrameInfo *MFI = Fn.getFrameInfo(); 283 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 284 285 MFI->setCalleeSavedInfoValid(true); 286 287 // Early exit if no callee saved registers are modified! 288 if (CSI.empty()) 289 return; 290 291 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 292 MachineBasicBlock::iterator I; 293 294 if (! ShrinkWrapThisFunction) { 295 // Spill using target interface. 296 I = EntryBlock->begin(); 297 if (!TII.spillCalleeSavedRegisters(*EntryBlock, I, CSI)) { 298 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 299 // Add the callee-saved register as live-in. 300 // It's killed at the spill. 301 EntryBlock->addLiveIn(CSI[i].getReg()); 302 303 // Insert the spill to the stack frame. 304 TII.storeRegToStackSlot(*EntryBlock, I, CSI[i].getReg(), true, 305 CSI[i].getFrameIdx(), CSI[i].getRegClass()); 306 } 307 } 308 309 // Restore using target interface. 310 for (unsigned ri = 0, re = ReturnBlocks.size(); ri != re; ++ri) { 311 MachineBasicBlock* MBB = ReturnBlocks[ri]; 312 I = MBB->end(); --I; 313 314 // Skip over all terminator instructions, which are part of the return 315 // sequence. 316 MachineBasicBlock::iterator I2 = I; 317 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 318 I = I2; 319 320 bool AtStart = I == MBB->begin(); 321 MachineBasicBlock::iterator BeforeI = I; 322 if (!AtStart) 323 --BeforeI; 324 325 // Restore all registers immediately before the return and any 326 // terminators that preceed it. 327 if (!TII.restoreCalleeSavedRegisters(*MBB, I, CSI)) { 328 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 329 TII.loadRegFromStackSlot(*MBB, I, CSI[i].getReg(), 330 CSI[i].getFrameIdx(), 331 CSI[i].getRegClass()); 332 assert(I != MBB->begin() && 333 "loadRegFromStackSlot didn't insert any code!"); 334 // Insert in reverse order. loadRegFromStackSlot can insert 335 // multiple instructions. 336 if (AtStart) 337 I = MBB->begin(); 338 else { 339 I = BeforeI; 340 ++I; 341 } 342 } 343 } 344 } 345 return; 346 } 347 348 // Insert spills. 349 std::vector<CalleeSavedInfo> blockCSI; 350 for (CSRegBlockMap::iterator BI = CSRSave.begin(), 351 BE = CSRSave.end(); BI != BE; ++BI) { 352 MachineBasicBlock* MBB = BI->first; 353 CSRegSet save = BI->second; 354 355 if (save.empty()) 356 continue; 357 358 blockCSI.clear(); 359 for (CSRegSet::iterator RI = save.begin(), 360 RE = save.end(); RI != RE; ++RI) { 361 blockCSI.push_back(CSI[*RI]); 362 } 363 assert(blockCSI.size() > 0 && 364 "Could not collect callee saved register info"); 365 366 I = MBB->begin(); 367 368 // When shrink wrapping, use stack slot stores/loads. 369 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 370 // Add the callee-saved register as live-in. 371 // It's killed at the spill. 372 MBB->addLiveIn(blockCSI[i].getReg()); 373 374 // Insert the spill to the stack frame. 375 TII.storeRegToStackSlot(*MBB, I, blockCSI[i].getReg(), 376 true, 377 blockCSI[i].getFrameIdx(), 378 blockCSI[i].getRegClass()); 379 } 380 } 381 382 for (CSRegBlockMap::iterator BI = CSRRestore.begin(), 383 BE = CSRRestore.end(); BI != BE; ++BI) { 384 MachineBasicBlock* MBB = BI->first; 385 CSRegSet restore = BI->second; 386 387 if (restore.empty()) 388 continue; 389 390 blockCSI.clear(); 391 for (CSRegSet::iterator RI = restore.begin(), 392 RE = restore.end(); RI != RE; ++RI) { 393 blockCSI.push_back(CSI[*RI]); 394 } 395 assert(blockCSI.size() > 0 && 396 "Could not find callee saved register info"); 397 398 // If MBB is empty and needs restores, insert at the _beginning_. 399 if (MBB->empty()) { 400 I = MBB->begin(); 401 } else { 402 I = MBB->end(); 403 --I; 404 405 // Skip over all terminator instructions, which are part of the 406 // return sequence. 407 if (! I->getDesc().isTerminator()) { 408 ++I; 409 } else { 410 MachineBasicBlock::iterator I2 = I; 411 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 412 I = I2; 413 } 414 } 415 416 bool AtStart = I == MBB->begin(); 417 MachineBasicBlock::iterator BeforeI = I; 418 if (!AtStart) 419 --BeforeI; 420 421 // Restore all registers immediately before the return and any 422 // terminators that preceed it. 423 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 424 TII.loadRegFromStackSlot(*MBB, I, blockCSI[i].getReg(), 425 blockCSI[i].getFrameIdx(), 426 blockCSI[i].getRegClass()); 427 assert(I != MBB->begin() && 428 "loadRegFromStackSlot didn't insert any code!"); 429 // Insert in reverse order. loadRegFromStackSlot can insert 430 // multiple instructions. 431 if (AtStart) 432 I = MBB->begin(); 433 else { 434 I = BeforeI; 435 ++I; 436 } 437 } 438 } 439} 440 441/// AdjustStackOffset - Helper function used to adjust the stack frame offset. 442static inline void 443AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, 444 bool StackGrowsDown, int64_t &Offset, 445 unsigned &MaxAlign) { 446 // If the stack grows down, add the object size to find the lowest address. 447 if (StackGrowsDown) 448 Offset += MFI->getObjectSize(FrameIdx); 449 450 unsigned Align = MFI->getObjectAlignment(FrameIdx); 451 452 // 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); 455 456 // Adjust to alignment boundary. 457 Offset = (Offset + Align - 1) / Align * Align; 458 459 if (StackGrowsDown) { 460 MFI->setObjectOffset(FrameIdx, -Offset); // Set the computed offset 461 } else { 462 MFI->setObjectOffset(FrameIdx, Offset); 463 Offset += MFI->getObjectSize(FrameIdx); 464 } 465} 466 467/// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 468/// abstract stack objects. 469/// 470void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { 471 const TargetFrameInfo &TFI = *Fn.getTarget().getFrameInfo(); 472 473 bool StackGrowsDown = 474 TFI.getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 475 476 // Loop over all of the stack objects, assigning sequential addresses... 477 MachineFrameInfo *MFI = Fn.getFrameInfo(); 478 479 // Start at the beginning of the local area. 480 // The Offset is the distance from the stack top in the direction 481 // of stack growth -- so it's always nonnegative. 482 int LocalAreaOffset = TFI.getOffsetOfLocalArea(); 483 if (StackGrowsDown) 484 LocalAreaOffset = -LocalAreaOffset; 485 assert(LocalAreaOffset >= 0 486 && "Local area offset should be in direction of stack growth"); 487 int64_t Offset = LocalAreaOffset; 488 489 // If there are fixed sized objects that are preallocated in the local area, 490 // non-fixed objects can't be allocated right at the start of local area. 491 // We currently don't support filling in holes in between fixed sized 492 // objects, so we adjust 'Offset' to point to the end of last fixed sized 493 // preallocated object. 494 for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) { 495 int64_t FixedOff; 496 if (StackGrowsDown) { 497 // The maximum distance from the stack pointer is at lower address of 498 // the object -- which is given by offset. For down growing stack 499 // the offset is negative, so we negate the offset to get the distance. 500 FixedOff = -MFI->getObjectOffset(i); 501 } else { 502 // The maximum distance from the start pointer is at the upper 503 // address of the object. 504 FixedOff = MFI->getObjectOffset(i) + MFI->getObjectSize(i); 505 } 506 if (FixedOff > Offset) Offset = FixedOff; 507 } 508 509 // First assign frame offsets to stack objects that are used to spill 510 // callee saved registers. 511 if (StackGrowsDown) { 512 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) { 513 // If stack grows down, we need to add size of find the lowest 514 // address of the object. 515 Offset += MFI->getObjectSize(i); 516 517 unsigned Align = MFI->getObjectAlignment(i); 518 // Adjust to alignment boundary 519 Offset = (Offset+Align-1)/Align*Align; 520 521 MFI->setObjectOffset(i, -Offset); // Set the computed offset 522 } 523 } else { 524 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; 525 for (int i = MaxCSFI; i >= MinCSFI ; --i) { 526 unsigned Align = MFI->getObjectAlignment(i); 527 // Adjust to alignment boundary 528 Offset = (Offset+Align-1)/Align*Align; 529 530 MFI->setObjectOffset(i, Offset); 531 Offset += MFI->getObjectSize(i); 532 } 533 } 534 535 unsigned MaxAlign = MFI->getMaxAlignment(); 536 537 // Make sure the special register scavenging spill slot is closest to the 538 // frame pointer if a frame pointer is required. 539 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 540 if (RS && RegInfo->hasFP(Fn) && !RegInfo->needsStackRealignment(Fn)) { 541 int SFI = RS->getScavengingFrameIndex(); 542 if (SFI >= 0) 543 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 544 } 545 546 // Make sure that the stack protector comes before the local variables on the 547 // stack. 548 if (MFI->getStackProtectorIndex() >= 0) 549 AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), StackGrowsDown, 550 Offset, MaxAlign); 551 552 // Then assign frame offsets to stack objects that are not used to spill 553 // callee saved registers. 554 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 555 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 556 continue; 557 if (RS && (int)i == RS->getScavengingFrameIndex()) 558 continue; 559 if (MFI->isDeadObjectIndex(i)) 560 continue; 561 if (MFI->getStackProtectorIndex() == (int)i) 562 continue; 563 564 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 565 } 566 567 // Make sure the special register scavenging spill slot is closest to the 568 // stack pointer. 569 if (RS && (!RegInfo->hasFP(Fn) || RegInfo->needsStackRealignment(Fn))) { 570 int SFI = RS->getScavengingFrameIndex(); 571 if (SFI >= 0) 572 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 573 } 574 575 if (!RegInfo->targetHandlesStackFrameRounding()) { 576 // If we have reserved argument space for call sites in the function 577 // immediately on entry to the current function, count it as part of the 578 // overall stack size. 579 if (MFI->hasCalls() && RegInfo->hasReservedCallFrame(Fn)) 580 Offset += MFI->getMaxCallFrameSize(); 581 582 // Round up the size to a multiple of the alignment. If the function has 583 // any calls or alloca's, align to the target's StackAlignment value to 584 // ensure that the callee's frame or the alloca data is suitably aligned; 585 // otherwise, for leaf functions, align to the TransientStackAlignment 586 // value. 587 unsigned StackAlign; 588 if (MFI->hasCalls() || MFI->hasVarSizedObjects() || 589 (RegInfo->needsStackRealignment(Fn) && MFI->getObjectIndexEnd() != 0)) 590 StackAlign = TFI.getStackAlignment(); 591 else 592 StackAlign = TFI.getTransientStackAlignment(); 593 // If the frame pointer is eliminated, all frame offsets will be relative 594 // to SP not FP; align to MaxAlign so this works. 595 StackAlign = std::max(StackAlign, MaxAlign); 596 unsigned AlignMask = StackAlign - 1; 597 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask); 598 } 599 600 // Update frame info to pretend that this is part of the stack... 601 MFI->setStackSize(Offset - LocalAreaOffset); 602} 603 604 605/// insertPrologEpilogCode - Scan the function for modified callee saved 606/// registers, insert spill code for these callee saved registers, then add 607/// prolog and epilog code to the function. 608/// 609void PEI::insertPrologEpilogCode(MachineFunction &Fn) { 610 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 611 612 // Add prologue to the function... 613 TRI->emitPrologue(Fn); 614 615 // Add epilogue to restore the callee-save registers in each exiting block 616 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 617 // If last instruction is a return instruction, add an epilogue 618 if (!I->empty() && I->back().getDesc().isReturn()) 619 TRI->emitEpilogue(Fn, *I); 620 } 621} 622 623 624/// replaceFrameIndices - Replace all MO_FrameIndex operands with physical 625/// register references and actual offsets. 626/// 627void PEI::replaceFrameIndices(MachineFunction &Fn) { 628 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do? 629 630 const TargetMachine &TM = Fn.getTarget(); 631 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!"); 632 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 633 const TargetFrameInfo *TFI = TM.getFrameInfo(); 634 bool StackGrowsDown = 635 TFI->getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 636 int FrameSetupOpcode = TRI.getCallFrameSetupOpcode(); 637 int FrameDestroyOpcode = TRI.getCallFrameDestroyOpcode(); 638 639 for (MachineFunction::iterator BB = Fn.begin(), 640 E = Fn.end(); BB != E; ++BB) { 641 int SPAdj = 0; // SP offset due to call frame setup / destroy. 642 if (RS && !FrameIndexVirtualScavenging) RS->enterBasicBlock(BB); 643 644 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 645 646 if (I->getOpcode() == FrameSetupOpcode || 647 I->getOpcode() == FrameDestroyOpcode) { 648 // Remember how much SP has been adjusted to create the call 649 // frame. 650 int Size = I->getOperand(0).getImm(); 651 652 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) || 653 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode)) 654 Size = -Size; 655 656 SPAdj += Size; 657 658 MachineBasicBlock::iterator PrevI = BB->end(); 659 if (I != BB->begin()) PrevI = prior(I); 660 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I); 661 662 // Visit the instructions created by eliminateCallFramePseudoInstr(). 663 if (PrevI == BB->end()) 664 I = BB->begin(); // The replaced instr was the first in the block. 665 else 666 I = llvm::next(PrevI); 667 continue; 668 } 669 670 MachineInstr *MI = I; 671 bool DoIncr = true; 672 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 673 if (MI->getOperand(i).isFI()) { 674 // Some instructions (e.g. inline asm instructions) can have 675 // multiple frame indices and/or cause eliminateFrameIndex 676 // to insert more than one instruction. We need the register 677 // scavenger to go through all of these instructions so that 678 // it can update its register information. We keep the 679 // iterator at the point before insertion so that we can 680 // revisit them in full. 681 bool AtBeginning = (I == BB->begin()); 682 if (!AtBeginning) --I; 683 684 // If this instruction has a FrameIndex operand, we need to 685 // use that target machine register info object to eliminate 686 // it. 687 TargetRegisterInfo::FrameIndexValue Value; 688 unsigned VReg = 689 TRI.eliminateFrameIndex(MI, SPAdj, &Value, 690 FrameIndexVirtualScavenging ? NULL : RS); 691 if (VReg) { 692 assert (FrameIndexVirtualScavenging && 693 "Not scavenging, but virtual returned from " 694 "eliminateFrameIndex()!"); 695 FrameConstantRegMap[VReg] = FrameConstantEntry(Value, SPAdj); 696 } 697 698 // Reset the iterator if we were at the beginning of the BB. 699 if (AtBeginning) { 700 I = BB->begin(); 701 DoIncr = false; 702 } 703 704 MI = 0; 705 break; 706 } 707 708 if (DoIncr && I != BB->end()) ++I; 709 710 // Update register states. 711 if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI); 712 } 713 714 assert(SPAdj == 0 && "Unbalanced call frame setup / destroy pairs?"); 715 } 716} 717 718/// findLastUseReg - find the killing use of the specified register within 719/// the instruciton range. Return the operand number of the kill in Operand. 720static MachineBasicBlock::iterator 721findLastUseReg(MachineBasicBlock::iterator I, MachineBasicBlock::iterator ME, 722 unsigned Reg) { 723 // Scan forward to find the last use of this virtual register 724 for (++I; I != ME; ++I) { 725 MachineInstr *MI = I; 726 bool isDefInsn = false; 727 bool isKillInsn = false; 728 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 729 if (MI->getOperand(i).isReg()) { 730 unsigned OpReg = MI->getOperand(i).getReg(); 731 if (OpReg == 0 || !TargetRegisterInfo::isVirtualRegister(OpReg)) 732 continue; 733 assert (OpReg == Reg 734 && "overlapping use of scavenged index register!"); 735 // If this is the killing use, we have a candidate. 736 if (MI->getOperand(i).isKill()) 737 isKillInsn = true; 738 else if (MI->getOperand(i).isDef()) 739 isDefInsn = true; 740 } 741 if (isKillInsn && !isDefInsn) 742 return I; 743 } 744 // If we hit the end of the basic block, there was no kill of 745 // the virtual register, which is wrong. 746 assert (0 && "scavenged index register never killed!"); 747 return ME; 748} 749 750/// scavengeFrameVirtualRegs - Replace all frame index virtual registers 751/// with physical registers. Use the register scavenger to find an 752/// appropriate register to use. 753void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) { 754 // Run through the instructions and find any virtual registers. 755 for (MachineFunction::iterator BB = Fn.begin(), 756 E = Fn.end(); BB != E; ++BB) { 757 RS->enterBasicBlock(BB); 758 759 // FIXME: The logic flow in this function is still too convoluted. 760 // It needs a cleanup refactoring. Do that in preparation for tracking 761 // more than one scratch register value and using ranges to find 762 // available scratch registers. 763 unsigned CurrentVirtReg = 0; 764 unsigned CurrentScratchReg = 0; 765 bool havePrevValue = false; 766 TargetRegisterInfo::FrameIndexValue PrevValue(0,0); 767 TargetRegisterInfo::FrameIndexValue Value(0,0); 768 MachineInstr *PrevLastUseMI = NULL; 769 unsigned PrevLastUseOp = 0; 770 bool trackingCurrentValue = false; 771 int SPAdj = 0; 772 773 // The instruction stream may change in the loop, so check BB->end() 774 // directly. 775 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 776 MachineInstr *MI = I; 777 bool isDefInsn = false; 778 bool isKillInsn = false; 779 bool clobbersScratchReg = false; 780 bool DoIncr = true; 781 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 782 if (MI->getOperand(i).isReg()) { 783 MachineOperand &MO = MI->getOperand(i); 784 unsigned Reg = MO.getReg(); 785 if (Reg == 0) 786 continue; 787 if (!TargetRegisterInfo::isVirtualRegister(Reg)) { 788 // If we have a previous scratch reg, check and see if anything 789 // here kills whatever value is in there. 790 if (Reg == CurrentScratchReg) { 791 if (MO.isUse()) { 792 // Two-address operands implicitly kill 793 if (MO.isKill() || MI->isRegTiedToDefOperand(i)) 794 clobbersScratchReg = true; 795 } else { 796 assert (MO.isDef()); 797 clobbersScratchReg = true; 798 } 799 } 800 continue; 801 } 802 // If this is a def, remember that this insn defines the value. 803 // This lets us properly consider insns which re-use the scratch 804 // register, such as r2 = sub r2, #imm, in the middle of the 805 // scratch range. 806 if (MO.isDef()) 807 isDefInsn = true; 808 809 // Have we already allocated a scratch register for this virtual? 810 if (Reg != CurrentVirtReg) { 811 // When we first encounter a new virtual register, it 812 // must be a definition. 813 assert(MI->getOperand(i).isDef() && 814 "frame index virtual missing def!"); 815 // We can't have nested virtual register live ranges because 816 // there's only a guarantee of one scavenged register at a time. 817 assert (CurrentVirtReg == 0 && 818 "overlapping frame index virtual registers!"); 819 820 // If the target gave us information about what's in the register, 821 // we can use that to re-use scratch regs. 822 DenseMap<unsigned, FrameConstantEntry>::iterator Entry = 823 FrameConstantRegMap.find(Reg); 824 trackingCurrentValue = Entry != FrameConstantRegMap.end(); 825 if (trackingCurrentValue) { 826 SPAdj = (*Entry).second.second; 827 Value = (*Entry).second.first; 828 } else { 829 SPAdj = 0; 830 Value.first = 0; 831 Value.second = 0; 832 } 833 834 // If the scratch register from the last allocation is still 835 // available, see if the value matches. If it does, just re-use it. 836 if (trackingCurrentValue && havePrevValue && PrevValue == Value) { 837 // FIXME: This assumes that the instructions in the live range 838 // for the virtual register are exclusively for the purpose 839 // of populating the value in the register. That's reasonable 840 // for these frame index registers, but it's still a very, very 841 // strong assumption. rdar://7322732. Better would be to 842 // explicitly check each instruction in the range for references 843 // to the virtual register. Only delete those insns that 844 // touch the virtual register. 845 846 // Find the last use of the new virtual register. Remove all 847 // instruction between here and there, and update the current 848 // instruction to reference the last use insn instead. 849 MachineBasicBlock::iterator LastUseMI = 850 findLastUseReg(I, BB->end(), Reg); 851 852 // Remove all instructions up 'til the last use, since they're 853 // just calculating the value we already have. 854 BB->erase(I, LastUseMI); 855 I = LastUseMI; 856 857 // Extend the live range of the scratch register 858 PrevLastUseMI->getOperand(PrevLastUseOp).setIsKill(false); 859 RS->setUsed(CurrentScratchReg); 860 CurrentVirtReg = Reg; 861 862 // We deleted the instruction we were scanning the operands of. 863 // Jump back to the instruction iterator loop. Don't increment 864 // past this instruction since we updated the iterator already. 865 DoIncr = false; 866 break; 867 } 868 869 // Scavenge a new scratch register 870 CurrentVirtReg = Reg; 871 const TargetRegisterClass *RC = Fn.getRegInfo().getRegClass(Reg); 872 CurrentScratchReg = RS->FindUnusedReg(RC); 873 if (CurrentScratchReg == 0) 874 // No register is "free". Scavenge a register. 875 CurrentScratchReg = RS->scavengeRegister(RC, I, SPAdj); 876 877 PrevValue = Value; 878 } 879 // replace this reference to the virtual register with the 880 // scratch register. 881 assert (CurrentScratchReg && "Missing scratch register!"); 882 MI->getOperand(i).setReg(CurrentScratchReg); 883 884 if (MI->getOperand(i).isKill()) { 885 isKillInsn = true; 886 PrevLastUseOp = i; 887 PrevLastUseMI = MI; 888 } 889 } 890 } 891 // If this is the last use of the scratch, stop tracking it. The 892 // last use will be a kill operand in an instruction that does 893 // not also define the scratch register. 894 if (isKillInsn && !isDefInsn) { 895 CurrentVirtReg = 0; 896 havePrevValue = trackingCurrentValue; 897 } 898 // Similarly, notice if instruction clobbered the value in the 899 // register we're tracking for possible later reuse. This is noted 900 // above, but enforced here since the value is still live while we 901 // process the rest of the operands of the instruction. 902 if (clobbersScratchReg) { 903 havePrevValue = false; 904 CurrentScratchReg = 0; 905 } 906 if (DoIncr) { 907 RS->forward(I); 908 ++I; 909 } 910 } 911 } 912} 913