PrologEpilogInserter.cpp revision 063fc74fa18287afc12477192f73d059e9947f6a
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 // Some inline asm's need a stack frame, as indicated by operand 1. 162 if (I->getOperand(1).getImm()) 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 if (Fn.getRegInfo().isPhysRegUsed(Reg)) { 209 // If the reg is modified, save it! 210 CSI.push_back(CalleeSavedInfo(Reg)); 211 } else { 212 for (const unsigned *AliasSet = RegInfo->getAliasSet(Reg); 213 *AliasSet; ++AliasSet) { // Check alias registers too. 214 if (Fn.getRegInfo().isPhysRegUsed(*AliasSet)) { 215 CSI.push_back(CalleeSavedInfo(Reg)); 216 break; 217 } 218 } 219 } 220 } 221 222 if (CSI.empty()) 223 return; // Early exit if no callee saved registers are modified! 224 225 unsigned NumFixedSpillSlots; 226 const TargetFrameInfo::SpillSlot *FixedSpillSlots = 227 TFI->getCalleeSavedSpillSlots(NumFixedSpillSlots); 228 229 // Now that we know which registers need to be saved and restored, allocate 230 // stack slots for them. 231 for (std::vector<CalleeSavedInfo>::iterator 232 I = CSI.begin(), E = CSI.end(); I != E; ++I) { 233 unsigned Reg = I->getReg(); 234 const TargetRegisterClass *RC = RegInfo->getMinimalPhysRegClass(Reg); 235 236 int FrameIdx; 237 if (RegInfo->hasReservedSpillSlot(Fn, Reg, FrameIdx)) { 238 I->setFrameIdx(FrameIdx); 239 continue; 240 } 241 242 // Check to see if this physreg must be spilled to a particular stack slot 243 // on this target. 244 const TargetFrameInfo::SpillSlot *FixedSlot = FixedSpillSlots; 245 while (FixedSlot != FixedSpillSlots+NumFixedSpillSlots && 246 FixedSlot->Reg != Reg) 247 ++FixedSlot; 248 249 if (FixedSlot == FixedSpillSlots + NumFixedSpillSlots) { 250 // Nope, just spill it anywhere convenient. 251 unsigned Align = RC->getAlignment(); 252 unsigned StackAlign = TFI->getStackAlignment(); 253 254 // We may not be able to satisfy the desired alignment specification of 255 // the TargetRegisterClass if the stack alignment is smaller. Use the 256 // min. 257 Align = std::min(Align, StackAlign); 258 FrameIdx = MFI->CreateStackObject(RC->getSize(), Align, true); 259 if ((unsigned)FrameIdx < MinCSFrameIndex) MinCSFrameIndex = FrameIdx; 260 if ((unsigned)FrameIdx > MaxCSFrameIndex) MaxCSFrameIndex = FrameIdx; 261 } else { 262 // Spill it to the stack where we must. 263 FrameIdx = MFI->CreateFixedObject(RC->getSize(), FixedSlot->Offset, true); 264 } 265 266 I->setFrameIdx(FrameIdx); 267 } 268 269 MFI->setCalleeSavedInfo(CSI); 270} 271 272/// insertCSRSpillsAndRestores - Insert spill and restore code for 273/// callee saved registers used in the function, handling shrink wrapping. 274/// 275void PEI::insertCSRSpillsAndRestores(MachineFunction &Fn) { 276 // Get callee saved register information. 277 MachineFrameInfo *MFI = Fn.getFrameInfo(); 278 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 279 280 MFI->setCalleeSavedInfoValid(true); 281 282 // Early exit if no callee saved registers are modified! 283 if (CSI.empty()) 284 return; 285 286 const TargetInstrInfo &TII = *Fn.getTarget().getInstrInfo(); 287 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 288 MachineBasicBlock::iterator I; 289 290 if (! ShrinkWrapThisFunction) { 291 // Spill using target interface. 292 I = EntryBlock->begin(); 293 if (!TII.spillCalleeSavedRegisters(*EntryBlock, I, CSI, TRI)) { 294 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 295 // Add the callee-saved register as live-in. 296 // It's killed at the spill. 297 EntryBlock->addLiveIn(CSI[i].getReg()); 298 299 // Insert the spill to the stack frame. 300 unsigned Reg = CSI[i].getReg(); 301 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 302 TII.storeRegToStackSlot(*EntryBlock, I, Reg, true, 303 CSI[i].getFrameIdx(), RC, 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 unsigned Reg = CSI[i].getReg(); 328 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 329 TII.loadRegFromStackSlot(*MBB, I, Reg, 330 CSI[i].getFrameIdx(), 331 RC, TRI); 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 unsigned Reg = blockCSI[i].getReg(); 376 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 377 TII.storeRegToStackSlot(*MBB, I, Reg, 378 true, 379 blockCSI[i].getFrameIdx(), 380 RC, TRI); 381 } 382 } 383 384 for (CSRegBlockMap::iterator BI = CSRRestore.begin(), 385 BE = CSRRestore.end(); BI != BE; ++BI) { 386 MachineBasicBlock* MBB = BI->first; 387 CSRegSet restore = BI->second; 388 389 if (restore.empty()) 390 continue; 391 392 blockCSI.clear(); 393 for (CSRegSet::iterator RI = restore.begin(), 394 RE = restore.end(); RI != RE; ++RI) { 395 blockCSI.push_back(CSI[*RI]); 396 } 397 assert(blockCSI.size() > 0 && 398 "Could not find callee saved register info"); 399 400 // If MBB is empty and needs restores, insert at the _beginning_. 401 if (MBB->empty()) { 402 I = MBB->begin(); 403 } else { 404 I = MBB->end(); 405 --I; 406 407 // Skip over all terminator instructions, which are part of the 408 // return sequence. 409 if (! I->getDesc().isTerminator()) { 410 ++I; 411 } else { 412 MachineBasicBlock::iterator I2 = I; 413 while (I2 != MBB->begin() && (--I2)->getDesc().isTerminator()) 414 I = I2; 415 } 416 } 417 418 bool AtStart = I == MBB->begin(); 419 MachineBasicBlock::iterator BeforeI = I; 420 if (!AtStart) 421 --BeforeI; 422 423 // Restore all registers immediately before the return and any 424 // terminators that preceed it. 425 for (unsigned i = 0, e = blockCSI.size(); i != e; ++i) { 426 unsigned Reg = blockCSI[i].getReg(); 427 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 428 TII.loadRegFromStackSlot(*MBB, I, Reg, 429 blockCSI[i].getFrameIdx(), 430 RC, TRI); 431 assert(I != MBB->begin() && 432 "loadRegFromStackSlot didn't insert any code!"); 433 // Insert in reverse order. loadRegFromStackSlot can insert 434 // multiple instructions. 435 if (AtStart) 436 I = MBB->begin(); 437 else { 438 I = BeforeI; 439 ++I; 440 } 441 } 442 } 443} 444 445/// AdjustStackOffset - Helper function used to adjust the stack frame offset. 446static inline void 447AdjustStackOffset(MachineFrameInfo *MFI, int FrameIdx, 448 bool StackGrowsDown, int64_t &Offset, 449 unsigned &MaxAlign) { 450 // If the stack grows down, add the object size to find the lowest address. 451 if (StackGrowsDown) 452 Offset += MFI->getObjectSize(FrameIdx); 453 454 unsigned Align = MFI->getObjectAlignment(FrameIdx); 455 456 // If the alignment of this object is greater than that of the stack, then 457 // increase the stack alignment to match. 458 MaxAlign = std::max(MaxAlign, Align); 459 460 // Adjust to alignment boundary. 461 Offset = (Offset + Align - 1) / Align * Align; 462 463 if (StackGrowsDown) { 464 MFI->setObjectOffset(FrameIdx, -Offset); // Set the computed offset 465 } else { 466 MFI->setObjectOffset(FrameIdx, Offset); 467 Offset += MFI->getObjectSize(FrameIdx); 468 } 469} 470 471/// calculateFrameObjectOffsets - Calculate actual frame offsets for all of the 472/// abstract stack objects. 473/// 474void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { 475 const TargetFrameInfo &TFI = *Fn.getTarget().getFrameInfo(); 476 477 bool StackGrowsDown = 478 TFI.getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 479 480 // Loop over all of the stack objects, assigning sequential addresses... 481 MachineFrameInfo *MFI = Fn.getFrameInfo(); 482 483 // Start at the beginning of the local area. 484 // The Offset is the distance from the stack top in the direction 485 // of stack growth -- so it's always nonnegative. 486 int LocalAreaOffset = TFI.getOffsetOfLocalArea(); 487 if (StackGrowsDown) 488 LocalAreaOffset = -LocalAreaOffset; 489 assert(LocalAreaOffset >= 0 490 && "Local area offset should be in direction of stack growth"); 491 int64_t Offset = LocalAreaOffset; 492 493 // If there are fixed sized objects that are preallocated in the local area, 494 // non-fixed objects can't be allocated right at the start of local area. 495 // We currently don't support filling in holes in between fixed sized 496 // objects, so we adjust 'Offset' to point to the end of last fixed sized 497 // preallocated object. 498 for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) { 499 int64_t FixedOff; 500 if (StackGrowsDown) { 501 // The maximum distance from the stack pointer is at lower address of 502 // the object -- which is given by offset. For down growing stack 503 // the offset is negative, so we negate the offset to get the distance. 504 FixedOff = -MFI->getObjectOffset(i); 505 } else { 506 // The maximum distance from the start pointer is at the upper 507 // address of the object. 508 FixedOff = MFI->getObjectOffset(i) + MFI->getObjectSize(i); 509 } 510 if (FixedOff > Offset) Offset = FixedOff; 511 } 512 513 // First assign frame offsets to stack objects that are used to spill 514 // callee saved registers. 515 if (StackGrowsDown) { 516 for (unsigned i = MinCSFrameIndex; i <= MaxCSFrameIndex; ++i) { 517 // If the stack grows down, we need to add the size to find the lowest 518 // address of the object. 519 Offset += MFI->getObjectSize(i); 520 521 unsigned Align = MFI->getObjectAlignment(i); 522 // Adjust to alignment boundary 523 Offset = (Offset+Align-1)/Align*Align; 524 525 MFI->setObjectOffset(i, -Offset); // Set the computed offset 526 } 527 } else { 528 int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; 529 for (int i = MaxCSFI; i >= MinCSFI ; --i) { 530 unsigned Align = MFI->getObjectAlignment(i); 531 // Adjust to alignment boundary 532 Offset = (Offset+Align-1)/Align*Align; 533 534 MFI->setObjectOffset(i, Offset); 535 Offset += MFI->getObjectSize(i); 536 } 537 } 538 539 unsigned MaxAlign = MFI->getMaxAlignment(); 540 541 // Make sure the special register scavenging spill slot is closest to the 542 // frame pointer if a frame pointer is required. 543 const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); 544 if (RS && RegInfo->hasFP(Fn) && !RegInfo->needsStackRealignment(Fn)) { 545 int SFI = RS->getScavengingFrameIndex(); 546 if (SFI >= 0) 547 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 548 } 549 550 // Make sure that the stack protector comes before the local variables on the 551 // stack. 552 if (MFI->getStackProtectorIndex() >= 0) 553 AdjustStackOffset(MFI, MFI->getStackProtectorIndex(), StackGrowsDown, 554 Offset, MaxAlign); 555 556 // Then assign frame offsets to stack objects that are not used to spill 557 // callee saved registers. 558 for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 559 if (i >= MinCSFrameIndex && i <= MaxCSFrameIndex) 560 continue; 561 if (RS && (int)i == RS->getScavengingFrameIndex()) 562 continue; 563 if (MFI->isDeadObjectIndex(i)) 564 continue; 565 if (MFI->getStackProtectorIndex() == (int)i) 566 continue; 567 568 AdjustStackOffset(MFI, i, StackGrowsDown, Offset, MaxAlign); 569 } 570 571 // Make sure the special register scavenging spill slot is closest to the 572 // stack pointer. 573 if (RS && (!RegInfo->hasFP(Fn) || RegInfo->needsStackRealignment(Fn))) { 574 int SFI = RS->getScavengingFrameIndex(); 575 if (SFI >= 0) 576 AdjustStackOffset(MFI, SFI, StackGrowsDown, Offset, MaxAlign); 577 } 578 579 if (!RegInfo->targetHandlesStackFrameRounding()) { 580 // If we have reserved argument space for call sites in the function 581 // immediately on entry to the current function, count it as part of the 582 // overall stack size. 583 if (MFI->adjustsStack() && RegInfo->hasReservedCallFrame(Fn)) 584 Offset += MFI->getMaxCallFrameSize(); 585 586 // Round up the size to a multiple of the alignment. If the function has 587 // any calls or alloca's, align to the target's StackAlignment value to 588 // ensure that the callee's frame or the alloca data is suitably aligned; 589 // otherwise, for leaf functions, align to the TransientStackAlignment 590 // value. 591 unsigned StackAlign; 592 if (MFI->adjustsStack() || MFI->hasVarSizedObjects() || 593 (RegInfo->needsStackRealignment(Fn) && MFI->getObjectIndexEnd() != 0)) 594 StackAlign = TFI.getStackAlignment(); 595 else 596 StackAlign = TFI.getTransientStackAlignment(); 597 598 // If the frame pointer is eliminated, all frame offsets will be relative to 599 // SP not FP. Align to MaxAlign so this works. 600 StackAlign = std::max(StackAlign, MaxAlign); 601 unsigned AlignMask = StackAlign - 1; 602 Offset = (Offset + AlignMask) & ~uint64_t(AlignMask); 603 } 604 605 // Update frame info to pretend that this is part of the stack... 606 MFI->setStackSize(Offset - LocalAreaOffset); 607} 608 609/// insertPrologEpilogCode - Scan the function for modified callee saved 610/// registers, insert spill code for these callee saved registers, then add 611/// prolog and epilog code to the function. 612/// 613void PEI::insertPrologEpilogCode(MachineFunction &Fn) { 614 const TargetRegisterInfo *TRI = Fn.getTarget().getRegisterInfo(); 615 616 // Add prologue to the function... 617 TRI->emitPrologue(Fn); 618 619 // Add epilogue to restore the callee-save registers in each exiting block 620 for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 621 // If last instruction is a return instruction, add an epilogue 622 if (!I->empty() && I->back().getDesc().isReturn()) 623 TRI->emitEpilogue(Fn, *I); 624 } 625} 626 627/// replaceFrameIndices - Replace all MO_FrameIndex operands with physical 628/// register references and actual offsets. 629/// 630void PEI::replaceFrameIndices(MachineFunction &Fn) { 631 if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do? 632 633 const TargetMachine &TM = Fn.getTarget(); 634 assert(TM.getRegisterInfo() && "TM::getRegisterInfo() must be implemented!"); 635 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 636 const TargetFrameInfo *TFI = TM.getFrameInfo(); 637 bool StackGrowsDown = 638 TFI->getStackGrowthDirection() == TargetFrameInfo::StackGrowsDown; 639 int FrameSetupOpcode = TRI.getCallFrameSetupOpcode(); 640 int FrameDestroyOpcode = TRI.getCallFrameDestroyOpcode(); 641 642 for (MachineFunction::iterator BB = Fn.begin(), 643 E = Fn.end(); BB != E; ++BB) { 644#ifndef NDEBUG 645 int SPAdjCount = 0; // frame setup / destroy count. 646#endif 647 int SPAdj = 0; // SP offset due to call frame setup / destroy. 648 if (RS && !FrameIndexVirtualScavenging) RS->enterBasicBlock(BB); 649 650 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 651 652 if (I->getOpcode() == FrameSetupOpcode || 653 I->getOpcode() == FrameDestroyOpcode) { 654#ifndef NDEBUG 655 // Track whether we see even pairs of them 656 SPAdjCount += I->getOpcode() == FrameSetupOpcode ? 1 : -1; 657#endif 658 // Remember how much SP has been adjusted to create the call 659 // frame. 660 int Size = I->getOperand(0).getImm(); 661 662 if ((!StackGrowsDown && I->getOpcode() == FrameSetupOpcode) || 663 (StackGrowsDown && I->getOpcode() == FrameDestroyOpcode)) 664 Size = -Size; 665 666 SPAdj += Size; 667 668 MachineBasicBlock::iterator PrevI = BB->end(); 669 if (I != BB->begin()) PrevI = prior(I); 670 TRI.eliminateCallFramePseudoInstr(Fn, *BB, I); 671 672 // Visit the instructions created by eliminateCallFramePseudoInstr(). 673 if (PrevI == BB->end()) 674 I = BB->begin(); // The replaced instr was the first in the block. 675 else 676 I = llvm::next(PrevI); 677 continue; 678 } 679 680 MachineInstr *MI = I; 681 bool DoIncr = true; 682 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 683 if (MI->getOperand(i).isFI()) { 684 // Some instructions (e.g. inline asm instructions) can have 685 // multiple frame indices and/or cause eliminateFrameIndex 686 // to insert more than one instruction. We need the register 687 // scavenger to go through all of these instructions so that 688 // it can update its register information. We keep the 689 // iterator at the point before insertion so that we can 690 // revisit them in full. 691 bool AtBeginning = (I == BB->begin()); 692 if (!AtBeginning) --I; 693 694 // If this instruction has a FrameIndex operand, we need to 695 // use that target machine register info object to eliminate 696 // it. 697 TargetRegisterInfo::FrameIndexValue Value; 698 unsigned VReg = 699 TRI.eliminateFrameIndex(MI, SPAdj, &Value, 700 FrameIndexVirtualScavenging ? NULL : RS); 701 if (VReg) { 702 assert (FrameIndexVirtualScavenging && 703 "Not scavenging, but virtual returned from " 704 "eliminateFrameIndex()!"); 705 FrameConstantRegMap[VReg] = FrameConstantEntry(Value, SPAdj); 706 } 707 708 // Reset the iterator if we were at the beginning of the BB. 709 if (AtBeginning) { 710 I = BB->begin(); 711 DoIncr = false; 712 } 713 714 MI = 0; 715 break; 716 } 717 718 if (DoIncr && I != BB->end()) ++I; 719 720 // Update register states. 721 if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI); 722 } 723 724 // If we have evenly matched pairs of frame setup / destroy instructions, 725 // make sure the adjustments come out to zero. If we don't have matched 726 // pairs, we can't be sure the missing bit isn't in another basic block 727 // due to a custom inserter playing tricks, so just asserting SPAdj==0 728 // isn't sufficient. See tMOVCC on Thumb1, for example. 729 assert((SPAdjCount || SPAdj == 0) && 730 "Unbalanced call frame setup / destroy pairs?"); 731 } 732} 733 734/// findLastUseReg - find the killing use of the specified register within 735/// the instruciton range. Return the operand number of the kill in Operand. 736static MachineBasicBlock::iterator 737findLastUseReg(MachineBasicBlock::iterator I, MachineBasicBlock::iterator ME, 738 unsigned Reg) { 739 // Scan forward to find the last use of this virtual register 740 for (++I; I != ME; ++I) { 741 MachineInstr *MI = I; 742 bool isDefInsn = false; 743 bool isKillInsn = false; 744 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) 745 if (MI->getOperand(i).isReg()) { 746 unsigned OpReg = MI->getOperand(i).getReg(); 747 if (OpReg == 0 || !TargetRegisterInfo::isVirtualRegister(OpReg)) 748 continue; 749 assert (OpReg == Reg 750 && "overlapping use of scavenged index register!"); 751 // If this is the killing use, we have a candidate. 752 if (MI->getOperand(i).isKill()) 753 isKillInsn = true; 754 else if (MI->getOperand(i).isDef()) 755 isDefInsn = true; 756 } 757 if (isKillInsn && !isDefInsn) 758 return I; 759 } 760 // If we hit the end of the basic block, there was no kill of 761 // the virtual register, which is wrong. 762 assert (0 && "scavenged index register never killed!"); 763 return ME; 764} 765 766/// scavengeFrameVirtualRegs - Replace all frame index virtual registers 767/// with physical registers. Use the register scavenger to find an 768/// appropriate register to use. 769void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) { 770 // Run through the instructions and find any virtual registers. 771 for (MachineFunction::iterator BB = Fn.begin(), 772 E = Fn.end(); BB != E; ++BB) { 773 RS->enterBasicBlock(BB); 774 775 // FIXME: The logic flow in this function is still too convoluted. 776 // It needs a cleanup refactoring. Do that in preparation for tracking 777 // more than one scratch register value and using ranges to find 778 // available scratch registers. 779 unsigned CurrentVirtReg = 0; 780 unsigned CurrentScratchReg = 0; 781 bool havePrevValue = false; 782 TargetRegisterInfo::FrameIndexValue PrevValue(0,0); 783 TargetRegisterInfo::FrameIndexValue Value(0,0); 784 MachineInstr *PrevLastUseMI = NULL; 785 unsigned PrevLastUseOp = 0; 786 bool trackingCurrentValue = false; 787 int SPAdj = 0; 788 789 // The instruction stream may change in the loop, so check BB->end() 790 // directly. 791 for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) { 792 MachineInstr *MI = I; 793 bool isDefInsn = false; 794 bool isKillInsn = false; 795 bool clobbersScratchReg = false; 796 bool DoIncr = true; 797 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 798 if (MI->getOperand(i).isReg()) { 799 MachineOperand &MO = MI->getOperand(i); 800 unsigned Reg = MO.getReg(); 801 if (Reg == 0) 802 continue; 803 if (!TargetRegisterInfo::isVirtualRegister(Reg)) { 804 // If we have a previous scratch reg, check and see if anything 805 // here kills whatever value is in there. 806 if (Reg == CurrentScratchReg) { 807 if (MO.isUse()) { 808 // Two-address operands implicitly kill 809 if (MO.isKill() || MI->isRegTiedToDefOperand(i)) 810 clobbersScratchReg = true; 811 } else { 812 assert (MO.isDef()); 813 clobbersScratchReg = true; 814 } 815 } 816 continue; 817 } 818 // If this is a def, remember that this insn defines the value. 819 // This lets us properly consider insns which re-use the scratch 820 // register, such as r2 = sub r2, #imm, in the middle of the 821 // scratch range. 822 if (MO.isDef()) 823 isDefInsn = true; 824 825 // Have we already allocated a scratch register for this virtual? 826 if (Reg != CurrentVirtReg) { 827 // When we first encounter a new virtual register, it 828 // must be a definition. 829 assert(MI->getOperand(i).isDef() && 830 "frame index virtual missing def!"); 831 // We can't have nested virtual register live ranges because 832 // there's only a guarantee of one scavenged register at a time. 833 assert (CurrentVirtReg == 0 && 834 "overlapping frame index virtual registers!"); 835 836 // If the target gave us information about what's in the register, 837 // we can use that to re-use scratch regs. 838 DenseMap<unsigned, FrameConstantEntry>::iterator Entry = 839 FrameConstantRegMap.find(Reg); 840 trackingCurrentValue = Entry != FrameConstantRegMap.end(); 841 if (trackingCurrentValue) { 842 SPAdj = (*Entry).second.second; 843 Value = (*Entry).second.first; 844 } else { 845 SPAdj = 0; 846 Value.first = 0; 847 Value.second = 0; 848 } 849 850 // If the scratch register from the last allocation is still 851 // available, see if the value matches. If it does, just re-use it. 852 if (trackingCurrentValue && havePrevValue && PrevValue == Value) { 853 // FIXME: This assumes that the instructions in the live range 854 // for the virtual register are exclusively for the purpose 855 // of populating the value in the register. That's reasonable 856 // for these frame index registers, but it's still a very, very 857 // strong assumption. rdar://7322732. Better would be to 858 // explicitly check each instruction in the range for references 859 // to the virtual register. Only delete those insns that 860 // touch the virtual register. 861 862 // Find the last use of the new virtual register. Remove all 863 // instruction between here and there, and update the current 864 // instruction to reference the last use insn instead. 865 MachineBasicBlock::iterator LastUseMI = 866 findLastUseReg(I, BB->end(), Reg); 867 868 // Remove all instructions up 'til the last use, since they're 869 // just calculating the value we already have. 870 BB->erase(I, LastUseMI); 871 I = LastUseMI; 872 873 // Extend the live range of the scratch register 874 PrevLastUseMI->getOperand(PrevLastUseOp).setIsKill(false); 875 RS->setUsed(CurrentScratchReg); 876 CurrentVirtReg = Reg; 877 878 // We deleted the instruction we were scanning the operands of. 879 // Jump back to the instruction iterator loop. Don't increment 880 // past this instruction since we updated the iterator already. 881 DoIncr = false; 882 break; 883 } 884 885 // Scavenge a new scratch register 886 CurrentVirtReg = Reg; 887 const TargetRegisterClass *RC = Fn.getRegInfo().getRegClass(Reg); 888 CurrentScratchReg = RS->scavengeRegister(RC, I, SPAdj); 889 PrevValue = Value; 890 } 891 // replace this reference to the virtual register with the 892 // scratch register. 893 assert (CurrentScratchReg && "Missing scratch register!"); 894 MI->getOperand(i).setReg(CurrentScratchReg); 895 896 if (MI->getOperand(i).isKill()) { 897 isKillInsn = true; 898 PrevLastUseOp = i; 899 PrevLastUseMI = MI; 900 } 901 } 902 } 903 // If this is the last use of the scratch, stop tracking it. The 904 // last use will be a kill operand in an instruction that does 905 // not also define the scratch register. 906 if (isKillInsn && !isDefInsn) { 907 CurrentVirtReg = 0; 908 havePrevValue = trackingCurrentValue; 909 } 910 // Similarly, notice if instruction clobbered the value in the 911 // register we're tracking for possible later reuse. This is noted 912 // above, but enforced here since the value is still live while we 913 // process the rest of the operands of the instruction. 914 if (clobbersScratchReg) { 915 havePrevValue = false; 916 CurrentScratchReg = 0; 917 } 918 if (DoIncr) { 919 RS->forward(I); 920 ++I; 921 } 922 } 923 } 924} 925