InlineFunction.cpp revision 9580641d5facf436e549982a5b0a5b3b694f9dc0
1//===- InlineFunction.cpp - Code to perform function inlining -------------===// 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 file implements inlining of a function into a call site, resolving 11// parameters and the return value as appropriate. 12// 13// The code in this file for handling inlines through invoke 14// instructions preserves semantics only under some assumptions about 15// the behavior of unwinders which correspond to gcc-style libUnwind 16// exception personality functions. Eventually the IR will be 17// improved to make this unnecessary, but until then, this code is 18// marked [LIBUNWIND]. 19// 20//===----------------------------------------------------------------------===// 21 22#include "llvm/Transforms/Utils/Cloning.h" 23#include "llvm/Constants.h" 24#include "llvm/DerivedTypes.h" 25#include "llvm/Module.h" 26#include "llvm/Instructions.h" 27#include "llvm/IntrinsicInst.h" 28#include "llvm/Intrinsics.h" 29#include "llvm/Attributes.h" 30#include "llvm/Analysis/CallGraph.h" 31#include "llvm/Analysis/DebugInfo.h" 32#include "llvm/Analysis/InstructionSimplify.h" 33#include "llvm/Target/TargetData.h" 34#include "llvm/Transforms/Utils/Local.h" 35#include "llvm/ADT/SmallVector.h" 36#include "llvm/ADT/StringExtras.h" 37#include "llvm/Support/CallSite.h" 38#include "llvm/Support/IRBuilder.h" 39using namespace llvm; 40 41bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { 42 return InlineFunction(CallSite(CI), IFI); 43} 44bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { 45 return InlineFunction(CallSite(II), IFI); 46} 47 48/// [LIBUNWIND] Look for an llvm.eh.exception call in the given block. 49static EHExceptionInst *findExceptionInBlock(BasicBlock *bb) { 50 for (BasicBlock::iterator i = bb->begin(), e = bb->end(); i != e; i++) { 51 EHExceptionInst *exn = dyn_cast<EHExceptionInst>(i); 52 if (exn) return exn; 53 } 54 55 return 0; 56} 57 58/// [LIBUNWIND] Look for the 'best' llvm.eh.selector instruction for 59/// the given llvm.eh.exception call. 60static EHSelectorInst *findSelectorForException(EHExceptionInst *exn) { 61 BasicBlock *exnBlock = exn->getParent(); 62 63 EHSelectorInst *outOfBlockSelector = 0; 64 for (Instruction::use_iterator 65 ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) { 66 EHSelectorInst *sel = dyn_cast<EHSelectorInst>(*ui); 67 if (!sel) continue; 68 69 // Immediately accept an eh.selector in the same block as the 70 // excepton call. 71 if (sel->getParent() == exnBlock) return sel; 72 73 // Otherwise, use the first selector we see. 74 if (!outOfBlockSelector) outOfBlockSelector = sel; 75 } 76 77 return outOfBlockSelector; 78} 79 80/// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector 81/// in the given landing pad. In principle, llvm.eh.exception is 82/// required to be in the landing pad; in practice, SplitCriticalEdge 83/// can break that invariant, and then inlining can break it further. 84/// There's a real need for a reliable solution here, but until that 85/// happens, we have some fragile workarounds here. 86static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) { 87 // Look for an exception call in the actual landing pad. 88 EHExceptionInst *exn = findExceptionInBlock(lpad); 89 if (exn) return findSelectorForException(exn); 90 91 // Okay, if that failed, look for one in an obvious successor. If 92 // we find one, we'll fix the IR by moving things back to the 93 // landing pad. 94 95 bool dominates = true; // does the lpad dominate the exn call 96 BasicBlock *nonDominated = 0; // if not, the first non-dominated block 97 BasicBlock *lastDominated = 0; // and the block which branched to it 98 99 BasicBlock *exnBlock = lpad; 100 101 // We need to protect against lpads that lead into infinite loops. 102 SmallPtrSet<BasicBlock*,4> visited; 103 visited.insert(exnBlock); 104 105 do { 106 // We're not going to apply this hack to anything more complicated 107 // than a series of unconditional branches, so if the block 108 // doesn't terminate in an unconditional branch, just fail. More 109 // complicated cases can arise when, say, sinking a call into a 110 // split unwind edge and then inlining it; but that can do almost 111 // *anything* to the CFG, including leaving the selector 112 // completely unreachable. The only way to fix that properly is 113 // to (1) prohibit transforms which move the exception or selector 114 // values away from the landing pad, e.g. by producing them with 115 // instructions that are pinned to an edge like a phi, or 116 // producing them with not-really-instructions, and (2) making 117 // transforms which split edges deal with that. 118 BranchInst *branch = dyn_cast<BranchInst>(&exnBlock->back()); 119 if (!branch || branch->isConditional()) return 0; 120 121 BasicBlock *successor = branch->getSuccessor(0); 122 123 // Fail if we found an infinite loop. 124 if (!visited.insert(successor)) return 0; 125 126 // If the successor isn't dominated by exnBlock: 127 if (!successor->getSinglePredecessor()) { 128 // We don't want to have to deal with threading the exception 129 // through multiple levels of phi, so give up if we've already 130 // followed a non-dominating edge. 131 if (!dominates) return 0; 132 133 // Otherwise, remember this as a non-dominating edge. 134 dominates = false; 135 nonDominated = successor; 136 lastDominated = exnBlock; 137 } 138 139 exnBlock = successor; 140 141 // Can we stop here? 142 exn = findExceptionInBlock(exnBlock); 143 } while (!exn); 144 145 // Look for a selector call for the exception we found. 146 EHSelectorInst *selector = findSelectorForException(exn); 147 if (!selector) return 0; 148 149 // The easy case is when the landing pad still dominates the 150 // exception call, in which case we can just move both calls back to 151 // the landing pad. 152 if (dominates) { 153 selector->moveBefore(lpad->getFirstNonPHI()); 154 exn->moveBefore(selector); 155 return selector; 156 } 157 158 // Otherwise, we have to split at the first non-dominating block. 159 // The CFG looks basically like this: 160 // lpad: 161 // phis_0 162 // insnsAndBranches_1 163 // br label %nonDominated 164 // nonDominated: 165 // phis_2 166 // insns_3 167 // %exn = call i8* @llvm.eh.exception() 168 // insnsAndBranches_4 169 // %selector = call @llvm.eh.selector(i8* %exn, ... 170 // We need to turn this into: 171 // lpad: 172 // phis_0 173 // %exn0 = call i8* @llvm.eh.exception() 174 // %selector0 = call @llvm.eh.selector(i8* %exn0, ... 175 // insnsAndBranches_1 176 // br label %split // from lastDominated 177 // nonDominated: 178 // phis_2 (without edge from lastDominated) 179 // %exn1 = call i8* @llvm.eh.exception() 180 // %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ... 181 // br label %split 182 // split: 183 // phis_2 (edge from lastDominated, edge from split) 184 // %exn = phi ... 185 // %selector = phi ... 186 // insns_3 187 // insnsAndBranches_4 188 189 assert(nonDominated); 190 assert(lastDominated); 191 192 // First, make clones of the intrinsics to go in lpad. 193 EHExceptionInst *lpadExn = cast<EHExceptionInst>(exn->clone()); 194 EHSelectorInst *lpadSelector = cast<EHSelectorInst>(selector->clone()); 195 lpadSelector->setArgOperand(0, lpadExn); 196 lpadSelector->insertBefore(lpad->getFirstNonPHI()); 197 lpadExn->insertBefore(lpadSelector); 198 199 // Split the non-dominated block. 200 BasicBlock *split = 201 nonDominated->splitBasicBlock(nonDominated->getFirstNonPHI(), 202 nonDominated->getName() + ".lpad-fix"); 203 204 // Redirect the last dominated branch there. 205 cast<BranchInst>(lastDominated->back()).setSuccessor(0, split); 206 207 // Move the existing intrinsics to the end of the old block. 208 selector->moveBefore(&nonDominated->back()); 209 exn->moveBefore(selector); 210 211 Instruction *splitIP = &split->front(); 212 213 // For all the phis in nonDominated, make a new phi in split to join 214 // that phi with the edge from lastDominated. 215 for (BasicBlock::iterator 216 i = nonDominated->begin(), e = nonDominated->end(); i != e; ++i) { 217 PHINode *phi = dyn_cast<PHINode>(i); 218 if (!phi) break; 219 220 PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(), 221 splitIP); 222 phi->replaceAllUsesWith(splitPhi); 223 splitPhi->addIncoming(phi, nonDominated); 224 splitPhi->addIncoming(phi->removeIncomingValue(lastDominated), 225 lastDominated); 226 } 227 228 // Make new phis for the exception and selector. 229 PHINode *exnPhi = PHINode::Create(exn->getType(), 2, "", splitIP); 230 exn->replaceAllUsesWith(exnPhi); 231 selector->setArgOperand(0, exn); // except for this use 232 exnPhi->addIncoming(exn, nonDominated); 233 exnPhi->addIncoming(lpadExn, lastDominated); 234 235 PHINode *selectorPhi = PHINode::Create(selector->getType(), 2, "", splitIP); 236 selector->replaceAllUsesWith(selectorPhi); 237 selectorPhi->addIncoming(selector, nonDominated); 238 selectorPhi->addIncoming(lpadSelector, lastDominated); 239 240 return lpadSelector; 241} 242 243namespace { 244 /// A class for recording information about inlining through an invoke. 245 class InvokeInliningInfo { 246 BasicBlock *OuterUnwindDest; 247 EHSelectorInst *OuterSelector; 248 BasicBlock *InnerUnwindDest; 249 PHINode *InnerExceptionPHI; 250 PHINode *InnerSelectorPHI; 251 SmallVector<Value*, 8> UnwindDestPHIValues; 252 253 // New EH: 254 BasicBlock *OuterResumeDest; //< Destination of the invoke's unwind. 255 BasicBlock *InnerResumeDest; //< Destination for the callee's resume. 256 LandingPadInst *CallerLPad; //< LandingPadInst associated with the invoke. 257 PHINode *InnerEHValuesPHI; //< PHI for EH values from landingpad insts. 258 259 public: 260 InvokeInliningInfo(InvokeInst *II) 261 : OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), 262 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0), 263 264 OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0), 265 CallerLPad(0), InnerEHValuesPHI(0) { 266 // If there are PHI nodes in the unwind destination block, we need to keep 267 // track of which values came into them from the invoke before removing 268 // the edge from this block. 269 llvm::BasicBlock *InvokeBB = II->getParent(); 270 BasicBlock::iterator I = OuterUnwindDest->begin(); 271 for (; isa<PHINode>(I); ++I) { 272 // Save the value to use for this edge. 273 PHINode *PHI = cast<PHINode>(I); 274 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 275 } 276 277 // FIXME: With the new EH, this if/dyn_cast should be a 'cast'. 278 if (LandingPadInst *LPI = dyn_cast<LandingPadInst>(I)) { 279 CallerLPad = LPI; 280 } 281 } 282 283 /// The outer unwind destination is the target of unwind edges 284 /// introduced for calls within the inlined function. 285 BasicBlock *getOuterUnwindDest() const { 286 return OuterUnwindDest; 287 } 288 289 EHSelectorInst *getOuterSelector() { 290 if (!OuterSelector) 291 OuterSelector = findSelectorForLandingPad(OuterUnwindDest); 292 return OuterSelector; 293 } 294 295 BasicBlock *getInnerUnwindDest(); 296 BasicBlock *getInnerUnwindDest_new(); 297 298 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 299 300 bool forwardEHResume(CallInst *call, BasicBlock *src); 301 302 /// forwardResume - Forward the 'resume' instruction to the caller's landing 303 /// pad block. When the landing pad block has only one predecessor, this is 304 /// a simple branch. When there is more than one predecessor, we need to 305 /// split the landing pad block after the landingpad instruction and jump 306 /// to there. 307 void forwardResume(ResumeInst *RI); 308 309 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind 310 /// destination block for the given basic block, using the values for the 311 /// original invoke's source block. 312 void addIncomingPHIValuesFor(BasicBlock *BB) const { 313 addIncomingPHIValuesForInto(BB, OuterUnwindDest); 314 } 315 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 316 BasicBlock::iterator I = dest->begin(); 317 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 318 PHINode *PHI = cast<PHINode>(I); 319 PHI->addIncoming(UnwindDestPHIValues[i], src); 320 } 321 } 322 }; 323} 324 325/// Get or create a target for the branch out of rewritten calls to 326/// llvm.eh.resume. 327BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { 328 if (InnerUnwindDest) return InnerUnwindDest; 329 330 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls 331 // in the outer landing pad to immediately following the phis. 332 EHSelectorInst *selector = getOuterSelector(); 333 if (!selector) return 0; 334 335 // The call to llvm.eh.exception *must* be in the landing pad. 336 Instruction *exn = cast<Instruction>(selector->getArgOperand(0)); 337 assert(exn->getParent() == OuterUnwindDest); 338 339 // TODO: recognize when we've already done this, so that we don't 340 // get a linear number of these when inlining calls into lots of 341 // invokes with the same landing pad. 342 343 // Do the hoisting. 344 Instruction *splitPoint = exn->getParent()->getFirstNonPHI(); 345 assert(splitPoint != selector && "selector-on-exception dominance broken!"); 346 if (splitPoint == exn) { 347 selector->removeFromParent(); 348 selector->insertAfter(exn); 349 splitPoint = selector->getNextNode(); 350 } else { 351 exn->moveBefore(splitPoint); 352 selector->moveBefore(splitPoint); 353 } 354 355 // Split the landing pad. 356 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint, 357 OuterUnwindDest->getName() + ".body"); 358 359 // The number of incoming edges we expect to the inner landing pad. 360 const unsigned phiCapacity = 2; 361 362 // Create corresponding new phis for all the phis in the outer landing pad. 363 BasicBlock::iterator insertPoint = InnerUnwindDest->begin(); 364 BasicBlock::iterator I = OuterUnwindDest->begin(); 365 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 366 PHINode *outerPhi = cast<PHINode>(I); 367 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity, 368 outerPhi->getName() + ".lpad-body", 369 insertPoint); 370 outerPhi->replaceAllUsesWith(innerPhi); 371 innerPhi->addIncoming(outerPhi, OuterUnwindDest); 372 } 373 374 // Create a phi for the exception value... 375 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity, 376 "exn.lpad-body", insertPoint); 377 exn->replaceAllUsesWith(InnerExceptionPHI); 378 selector->setArgOperand(0, exn); // restore this use 379 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest); 380 381 // ...and the selector. 382 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity, 383 "selector.lpad-body", insertPoint); 384 selector->replaceAllUsesWith(InnerSelectorPHI); 385 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest); 386 387 // All done. 388 return InnerUnwindDest; 389} 390 391/// [LIBUNWIND] Try to forward the given call, which logically occurs 392/// at the end of the given block, as a branch to the inner unwind 393/// block. Returns true if the call was forwarded. 394bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) { 395 // First, check whether this is a call to the intrinsic. 396 Function *fn = dyn_cast<Function>(call->getCalledValue()); 397 if (!fn || fn->getName() != "llvm.eh.resume") 398 return false; 399 400 // At this point, we need to return true on all paths, because 401 // otherwise we'll construct an invoke of the intrinsic, which is 402 // not well-formed. 403 404 // Try to find or make an inner unwind dest, which will fail if we 405 // can't find a selector call for the outer unwind dest. 406 BasicBlock *dest = getInnerUnwindDest(); 407 bool hasSelector = (dest != 0); 408 409 // If we failed, just use the outer unwind dest, dropping the 410 // exception and selector on the floor. 411 if (!hasSelector) 412 dest = OuterUnwindDest; 413 414 // Make a branch. 415 BranchInst::Create(dest, src); 416 417 // Update the phis in the destination. They were inserted in an 418 // order which makes this work. 419 addIncomingPHIValuesForInto(src, dest); 420 421 if (hasSelector) { 422 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src); 423 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src); 424 } 425 426 return true; 427} 428 429/// Get or create a target for the branch from ResumeInsts. 430BasicBlock *InvokeInliningInfo::getInnerUnwindDest_new() { 431 if (InnerResumeDest) return InnerResumeDest; 432 433 // Split the landing pad. 434 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint; 435 InnerResumeDest = 436 OuterResumeDest->splitBasicBlock(SplitPoint, 437 OuterResumeDest->getName() + ".body"); 438 439 // The number of incoming edges we expect to the inner landing pad. 440 const unsigned PHICapacity = 2; 441 442 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 443 BasicBlock::iterator InsertPoint = InnerResumeDest->begin(); 444 BasicBlock::iterator I = OuterResumeDest->begin(); 445 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 446 PHINode *OuterPHI = cast<PHINode>(I); 447 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 448 OuterPHI->getName() + ".lpad-body", 449 InsertPoint); 450 OuterPHI->replaceAllUsesWith(InnerPHI); 451 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 452 } 453 454 // Create a PHI for the exception values. 455 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 456 "eh.lpad-body", InsertPoint); 457 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 458 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 459 460 // All done. 461 return InnerResumeDest; 462} 463 464/// forwardResume - Forward the 'resume' instruction to the caller's landing pad 465/// block. When the landing pad block has only one predecessor, this is a simple 466/// branch. When there is more than one predecessor, we need to split the 467/// landing pad block after the landingpad instruction and jump to there. 468void InvokeInliningInfo::forwardResume(ResumeInst *RI) { 469 BasicBlock *Dest = getInnerUnwindDest_new(); 470 BasicBlock *Src = RI->getParent(); 471 472 BranchInst::Create(Dest, Src); 473 474 // Update the PHIs in the destination. They were inserted in an order which 475 // makes this work. 476 addIncomingPHIValuesForInto(Src, Dest); 477 478 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 479 RI->eraseFromParent(); 480} 481 482/// [LIBUNWIND] Check whether this selector is "only cleanups": 483/// call i32 @llvm.eh.selector(blah, blah, i32 0) 484static bool isCleanupOnlySelector(EHSelectorInst *selector) { 485 if (selector->getNumArgOperands() != 3) return false; 486 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2)); 487 return (val && val->isZero()); 488} 489 490/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 491/// an invoke, we have to turn all of the calls that can throw into 492/// invokes. This function analyze BB to see if there are any calls, and if so, 493/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 494/// nodes in that block with the values specified in InvokeDestPHIValues. 495/// 496/// Returns true to indicate that the next block should be skipped. 497static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 498 InvokeInliningInfo &Invoke) { 499 LandingPadInst *LPI = Invoke.getLandingPadInst(); 500 501 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 502 Instruction *I = BBI++; 503 504 if (LPI) // FIXME: This won't be NULL in the new EH. 505 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) { 506 unsigned NumClauses = LPI->getNumClauses(); 507 L->reserveClauses(NumClauses); 508 for (unsigned i = 0; i != NumClauses; ++i) 509 L->addClause(LPI->getClauseType(i), LPI->getClauseValue(i)); 510 } 511 512 // We only need to check for function calls: inlined invoke 513 // instructions require no special handling. 514 CallInst *CI = dyn_cast<CallInst>(I); 515 if (CI == 0) continue; 516 517 // LIBUNWIND: merge selector instructions. 518 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) { 519 EHSelectorInst *Outer = Invoke.getOuterSelector(); 520 if (!Outer) continue; 521 522 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); 523 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); 524 525 // If both selectors contain only cleanups, we don't need to do 526 // anything. TODO: this is really just a very specific instance 527 // of a much more general optimization. 528 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; 529 530 // Otherwise, we just append the outer selector to the inner selector. 531 SmallVector<Value*, 16> NewSelector; 532 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) 533 NewSelector.push_back(Inner->getArgOperand(i)); 534 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) 535 NewSelector.push_back(Outer->getArgOperand(i)); 536 537 CallInst *NewInner = 538 IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector); 539 // No need to copy attributes, calling convention, etc. 540 NewInner->takeName(Inner); 541 Inner->replaceAllUsesWith(NewInner); 542 Inner->eraseFromParent(); 543 continue; 544 } 545 546 // If this call cannot unwind, don't convert it to an invoke. 547 if (CI->doesNotThrow()) 548 continue; 549 550 // Convert this function call into an invoke instruction. 551 // First, split the basic block. 552 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 553 554 // Delete the unconditional branch inserted by splitBasicBlock 555 BB->getInstList().pop_back(); 556 557 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch 558 // directly to the new landing pad. 559 if (Invoke.forwardEHResume(CI, BB)) { 560 // TODO: 'Split' is now unreachable; clean it up. 561 562 // We want to leave the original call intact so that the call 563 // graph and other structures won't get misled. We also have to 564 // avoid processing the next block, or we'll iterate here forever. 565 return true; 566 } 567 568 // Otherwise, create the new invoke instruction. 569 ImmutableCallSite CS(CI); 570 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 571 InvokeInst *II = 572 InvokeInst::Create(CI->getCalledValue(), Split, 573 Invoke.getOuterUnwindDest(), 574 InvokeArgs, CI->getName(), BB); 575 II->setCallingConv(CI->getCallingConv()); 576 II->setAttributes(CI->getAttributes()); 577 578 // Make sure that anything using the call now uses the invoke! This also 579 // updates the CallGraph if present, because it uses a WeakVH. 580 CI->replaceAllUsesWith(II); 581 582 Split->getInstList().pop_front(); // Delete the original call 583 584 // Update any PHI nodes in the exceptional block to indicate that 585 // there is now a new entry in them. 586 Invoke.addIncomingPHIValuesFor(BB); 587 return false; 588 } 589 590 return false; 591} 592 593 594/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 595/// in the body of the inlined function into invokes and turn unwind 596/// instructions into branches to the invoke unwind dest. 597/// 598/// II is the invoke instruction being inlined. FirstNewBlock is the first 599/// block of the inlined code (the last block is the end of the function), 600/// and InlineCodeInfo is information about the code that got inlined. 601static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 602 ClonedCodeInfo &InlinedCodeInfo) { 603 BasicBlock *InvokeDest = II->getUnwindDest(); 604 605 Function *Caller = FirstNewBlock->getParent(); 606 607 // The inlined code is currently at the end of the function, scan from the 608 // start of the inlined code to its end, checking for stuff we need to 609 // rewrite. If the code doesn't have calls or unwinds, we know there is 610 // nothing to rewrite. 611 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 612 // Now that everything is happy, we have one final detail. The PHI nodes in 613 // the exception destination block still have entries due to the original 614 // invoke instruction. Eliminate these entries (which might even delete the 615 // PHI node) now. 616 InvokeDest->removePredecessor(II->getParent()); 617 return; 618 } 619 620 InvokeInliningInfo Invoke(II); 621 622 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 623 if (InlinedCodeInfo.ContainsCalls) 624 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 625 // Honor a request to skip the next block. We don't need to 626 // consider UnwindInsts in this case either. 627 ++BB; 628 continue; 629 } 630 631 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 632 // An UnwindInst requires special handling when it gets inlined into an 633 // invoke site. Once this happens, we know that the unwind would cause 634 // a control transfer to the invoke exception destination, so we can 635 // transform it into a direct branch to the exception destination. 636 BranchInst::Create(InvokeDest, UI); 637 638 // Delete the unwind instruction! 639 UI->eraseFromParent(); 640 641 // Update any PHI nodes in the exceptional block to indicate that 642 // there is now a new entry in them. 643 Invoke.addIncomingPHIValuesFor(BB); 644 } 645 646 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 647 Invoke.forwardResume(RI); 648 } 649 } 650 651 // Now that everything is happy, we have one final detail. The PHI nodes in 652 // the exception destination block still have entries due to the original 653 // invoke instruction. Eliminate these entries (which might even delete the 654 // PHI node) now. 655 InvokeDest->removePredecessor(II->getParent()); 656} 657 658/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 659/// into the caller, update the specified callgraph to reflect the changes we 660/// made. Note that it's possible that not all code was copied over, so only 661/// some edges of the callgraph may remain. 662static void UpdateCallGraphAfterInlining(CallSite CS, 663 Function::iterator FirstNewBlock, 664 ValueToValueMapTy &VMap, 665 InlineFunctionInfo &IFI) { 666 CallGraph &CG = *IFI.CG; 667 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 668 const Function *Callee = CS.getCalledFunction(); 669 CallGraphNode *CalleeNode = CG[Callee]; 670 CallGraphNode *CallerNode = CG[Caller]; 671 672 // Since we inlined some uninlined call sites in the callee into the caller, 673 // add edges from the caller to all of the callees of the callee. 674 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 675 676 // Consider the case where CalleeNode == CallerNode. 677 CallGraphNode::CalledFunctionsVector CallCache; 678 if (CalleeNode == CallerNode) { 679 CallCache.assign(I, E); 680 I = CallCache.begin(); 681 E = CallCache.end(); 682 } 683 684 for (; I != E; ++I) { 685 const Value *OrigCall = I->first; 686 687 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 688 // Only copy the edge if the call was inlined! 689 if (VMI == VMap.end() || VMI->second == 0) 690 continue; 691 692 // If the call was inlined, but then constant folded, there is no edge to 693 // add. Check for this case. 694 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 695 if (NewCall == 0) continue; 696 697 // Remember that this call site got inlined for the client of 698 // InlineFunction. 699 IFI.InlinedCalls.push_back(NewCall); 700 701 // It's possible that inlining the callsite will cause it to go from an 702 // indirect to a direct call by resolving a function pointer. If this 703 // happens, set the callee of the new call site to a more precise 704 // destination. This can also happen if the call graph node of the caller 705 // was just unnecessarily imprecise. 706 if (I->second->getFunction() == 0) 707 if (Function *F = CallSite(NewCall).getCalledFunction()) { 708 // Indirect call site resolved to direct call. 709 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 710 711 continue; 712 } 713 714 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 715 } 716 717 // Update the call graph by deleting the edge from Callee to Caller. We must 718 // do this after the loop above in case Caller and Callee are the same. 719 CallerNode->removeCallEdgeFor(CS); 720} 721 722/// HandleByValArgument - When inlining a call site that has a byval argument, 723/// we have to make the implicit memcpy explicit by adding it. 724static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 725 const Function *CalledFunc, 726 InlineFunctionInfo &IFI, 727 unsigned ByValAlignment) { 728 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 729 730 // If the called function is readonly, then it could not mutate the caller's 731 // copy of the byval'd memory. In this case, it is safe to elide the copy and 732 // temporary. 733 if (CalledFunc->onlyReadsMemory()) { 734 // If the byval argument has a specified alignment that is greater than the 735 // passed in pointer, then we either have to round up the input pointer or 736 // give up on this transformation. 737 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 738 return Arg; 739 740 // If the pointer is already known to be sufficiently aligned, or if we can 741 // round it up to a larger alignment, then we don't need a temporary. 742 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 743 IFI.TD) >= ByValAlignment) 744 return Arg; 745 746 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 747 // for code quality, but rarely happens and is required for correctness. 748 } 749 750 LLVMContext &Context = Arg->getContext(); 751 752 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 753 754 // Create the alloca. If we have TargetData, use nice alignment. 755 unsigned Align = 1; 756 if (IFI.TD) 757 Align = IFI.TD->getPrefTypeAlignment(AggTy); 758 759 // If the byval had an alignment specified, we *must* use at least that 760 // alignment, as it is required by the byval argument (and uses of the 761 // pointer inside the callee). 762 Align = std::max(Align, ByValAlignment); 763 764 Function *Caller = TheCall->getParent()->getParent(); 765 766 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 767 &*Caller->begin()->begin()); 768 // Emit a memcpy. 769 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 770 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 771 Intrinsic::memcpy, 772 Tys); 773 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 774 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 775 776 Value *Size; 777 if (IFI.TD == 0) 778 Size = ConstantExpr::getSizeOf(AggTy); 779 else 780 Size = ConstantInt::get(Type::getInt64Ty(Context), 781 IFI.TD->getTypeStoreSize(AggTy)); 782 783 // Always generate a memcpy of alignment 1 here because we don't know 784 // the alignment of the src pointer. Other optimizations can infer 785 // better alignment. 786 Value *CallArgs[] = { 787 DestCast, SrcCast, Size, 788 ConstantInt::get(Type::getInt32Ty(Context), 1), 789 ConstantInt::getFalse(Context) // isVolatile 790 }; 791 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs); 792 793 // Uses of the argument in the function should use our new alloca 794 // instead. 795 return NewAlloca; 796} 797 798// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 799// intrinsic. 800static bool isUsedByLifetimeMarker(Value *V) { 801 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 802 ++UI) { 803 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 804 switch (II->getIntrinsicID()) { 805 default: break; 806 case Intrinsic::lifetime_start: 807 case Intrinsic::lifetime_end: 808 return true; 809 } 810 } 811 } 812 return false; 813} 814 815// hasLifetimeMarkers - Check whether the given alloca already has 816// lifetime.start or lifetime.end intrinsics. 817static bool hasLifetimeMarkers(AllocaInst *AI) { 818 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 819 if (AI->getType() == Int8PtrTy) 820 return isUsedByLifetimeMarker(AI); 821 822 // Do a scan to find all the casts to i8*. 823 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 824 ++I) { 825 if (I->getType() != Int8PtrTy) continue; 826 if (I->stripPointerCasts() != AI) continue; 827 if (isUsedByLifetimeMarker(*I)) 828 return true; 829 } 830 return false; 831} 832 833/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively 834/// update InlinedAtEntry of a DebugLoc. 835static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 836 const DebugLoc &InlinedAtDL, 837 LLVMContext &Ctx) { 838 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 839 DebugLoc NewInlinedAtDL 840 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 841 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 842 NewInlinedAtDL.getAsMDNode(Ctx)); 843 } 844 845 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 846 InlinedAtDL.getAsMDNode(Ctx)); 847} 848 849 850/// fixupLineNumbers - Update inlined instructions' line numbers to 851/// to encode location where these instructions are inlined. 852static void fixupLineNumbers(Function *Fn, Function::iterator FI, 853 Instruction *TheCall) { 854 DebugLoc TheCallDL = TheCall->getDebugLoc(); 855 if (TheCallDL.isUnknown()) 856 return; 857 858 for (; FI != Fn->end(); ++FI) { 859 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 860 BI != BE; ++BI) { 861 DebugLoc DL = BI->getDebugLoc(); 862 if (!DL.isUnknown()) { 863 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 864 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 865 LLVMContext &Ctx = BI->getContext(); 866 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 867 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 868 InlinedAt, Ctx)); 869 } 870 } 871 } 872 } 873} 874 875// InlineFunction - This function inlines the called function into the basic 876// block of the caller. This returns false if it is not possible to inline this 877// call. The program is still in a well defined state if this occurs though. 878// 879// Note that this only does one level of inlining. For example, if the 880// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 881// exists in the instruction stream. Similarly this will inline a recursive 882// function by one level. 883// 884bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 885 Instruction *TheCall = CS.getInstruction(); 886 LLVMContext &Context = TheCall->getContext(); 887 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 888 "Instruction not in function!"); 889 890 // If IFI has any state in it, zap it before we fill it in. 891 IFI.reset(); 892 893 const Function *CalledFunc = CS.getCalledFunction(); 894 if (CalledFunc == 0 || // Can't inline external function or indirect 895 CalledFunc->isDeclaration() || // call, or call to a vararg function! 896 CalledFunc->getFunctionType()->isVarArg()) return false; 897 898 // If the call to the callee is not a tail call, we must clear the 'tail' 899 // flags on any calls that we inline. 900 bool MustClearTailCallFlags = 901 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 902 903 // If the call to the callee cannot throw, set the 'nounwind' flag on any 904 // calls that we inline. 905 bool MarkNoUnwind = CS.doesNotThrow(); 906 907 BasicBlock *OrigBB = TheCall->getParent(); 908 Function *Caller = OrigBB->getParent(); 909 910 // GC poses two hazards to inlining, which only occur when the callee has GC: 911 // 1. If the caller has no GC, then the callee's GC must be propagated to the 912 // caller. 913 // 2. If the caller has a differing GC, it is invalid to inline. 914 if (CalledFunc->hasGC()) { 915 if (!Caller->hasGC()) 916 Caller->setGC(CalledFunc->getGC()); 917 else if (CalledFunc->getGC() != Caller->getGC()) 918 return false; 919 } 920 921 // Find the personality function used by the landing pads of the caller. If it 922 // exists, then check to see that it matches the personality function used in 923 // the callee. 924 for (Function::const_iterator 925 I = Caller->begin(), E = Caller->end(); I != E; ++I) 926 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 927 const BasicBlock *BB = II->getUnwindDest(); 928 // FIXME: This 'isa' here should become go away once the new EH system is 929 // in place. 930 if (!isa<LandingPadInst>(BB->getFirstNonPHI())) 931 continue; 932 const LandingPadInst *LP = cast<LandingPadInst>(BB->getFirstNonPHI()); 933 const Value *CallerPersFn = LP->getPersonalityFn(); 934 935 // If the personality functions match, then we can perform the 936 // inlining. Otherwise, we can't inline. 937 // TODO: This isn't 100% true. Some personality functions are proper 938 // supersets of others and can be used in place of the other. 939 for (Function::const_iterator 940 I = CalledFunc->begin(), E = CalledFunc->end(); I != E; ++I) 941 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 942 const BasicBlock *BB = II->getUnwindDest(); 943 // FIXME: This 'if/dyn_cast' here should become a normal 'cast' once 944 // the new EH system is in place. 945 if (const LandingPadInst *LP = 946 dyn_cast<LandingPadInst>(BB->getFirstNonPHI())) 947 if (CallerPersFn != LP->getPersonalityFn()) 948 return false; 949 break; 950 } 951 952 break; 953 } 954 955 // Get an iterator to the last basic block in the function, which will have 956 // the new function inlined after it. 957 // 958 Function::iterator LastBlock = &Caller->back(); 959 960 // Make sure to capture all of the return instructions from the cloned 961 // function. 962 SmallVector<ReturnInst*, 8> Returns; 963 ClonedCodeInfo InlinedFunctionInfo; 964 Function::iterator FirstNewBlock; 965 966 { // Scope to destroy VMap after cloning. 967 ValueToValueMapTy VMap; 968 969 assert(CalledFunc->arg_size() == CS.arg_size() && 970 "No varargs calls can be inlined!"); 971 972 // Calculate the vector of arguments to pass into the function cloner, which 973 // matches up the formal to the actual argument values. 974 CallSite::arg_iterator AI = CS.arg_begin(); 975 unsigned ArgNo = 0; 976 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 977 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 978 Value *ActualArg = *AI; 979 980 // When byval arguments actually inlined, we need to make the copy implied 981 // by them explicit. However, we don't do this if the callee is readonly 982 // or readnone, because the copy would be unneeded: the callee doesn't 983 // modify the struct. 984 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 985 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 986 CalledFunc->getParamAlignment(ArgNo+1)); 987 988 // Calls that we inline may use the new alloca, so we need to clear 989 // their 'tail' flags if HandleByValArgument introduced a new alloca and 990 // the callee has calls. 991 MustClearTailCallFlags |= ActualArg != *AI; 992 } 993 994 VMap[I] = ActualArg; 995 } 996 997 // We want the inliner to prune the code as it copies. We would LOVE to 998 // have no dead or constant instructions leftover after inlining occurs 999 // (which can happen, e.g., because an argument was constant), but we'll be 1000 // happy with whatever the cloner can do. 1001 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 1002 /*ModuleLevelChanges=*/false, Returns, ".i", 1003 &InlinedFunctionInfo, IFI.TD, TheCall); 1004 1005 // Remember the first block that is newly cloned over. 1006 FirstNewBlock = LastBlock; ++FirstNewBlock; 1007 1008 // Update the callgraph if requested. 1009 if (IFI.CG) 1010 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 1011 1012 // Update inlined instructions' line number information. 1013 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 1014 } 1015 1016 // If there are any alloca instructions in the block that used to be the entry 1017 // block for the callee, move them to the entry block of the caller. First 1018 // calculate which instruction they should be inserted before. We insert the 1019 // instructions at the end of the current alloca list. 1020 // 1021 { 1022 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 1023 for (BasicBlock::iterator I = FirstNewBlock->begin(), 1024 E = FirstNewBlock->end(); I != E; ) { 1025 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 1026 if (AI == 0) continue; 1027 1028 // If the alloca is now dead, remove it. This often occurs due to code 1029 // specialization. 1030 if (AI->use_empty()) { 1031 AI->eraseFromParent(); 1032 continue; 1033 } 1034 1035 if (!isa<Constant>(AI->getArraySize())) 1036 continue; 1037 1038 // Keep track of the static allocas that we inline into the caller. 1039 IFI.StaticAllocas.push_back(AI); 1040 1041 // Scan for the block of allocas that we can move over, and move them 1042 // all at once. 1043 while (isa<AllocaInst>(I) && 1044 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 1045 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 1046 ++I; 1047 } 1048 1049 // Transfer all of the allocas over in a block. Using splice means 1050 // that the instructions aren't removed from the symbol table, then 1051 // reinserted. 1052 Caller->getEntryBlock().getInstList().splice(InsertPoint, 1053 FirstNewBlock->getInstList(), 1054 AI, I); 1055 } 1056 } 1057 1058 // Leave lifetime markers for the static alloca's, scoping them to the 1059 // function we just inlined. 1060 if (!IFI.StaticAllocas.empty()) { 1061 IRBuilder<> builder(FirstNewBlock->begin()); 1062 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 1063 AllocaInst *AI = IFI.StaticAllocas[ai]; 1064 1065 // If the alloca is already scoped to something smaller than the whole 1066 // function then there's no need to add redundant, less accurate markers. 1067 if (hasLifetimeMarkers(AI)) 1068 continue; 1069 1070 builder.CreateLifetimeStart(AI); 1071 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 1072 IRBuilder<> builder(Returns[ri]); 1073 builder.CreateLifetimeEnd(AI); 1074 } 1075 } 1076 } 1077 1078 // If the inlined code contained dynamic alloca instructions, wrap the inlined 1079 // code with llvm.stacksave/llvm.stackrestore intrinsics. 1080 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 1081 Module *M = Caller->getParent(); 1082 // Get the two intrinsics we care about. 1083 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 1084 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 1085 1086 // Insert the llvm.stacksave. 1087 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 1088 .CreateCall(StackSave, "savedstack"); 1089 1090 // Insert a call to llvm.stackrestore before any return instructions in the 1091 // inlined function. 1092 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1093 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); 1094 } 1095 1096 // Count the number of StackRestore calls we insert. 1097 unsigned NumStackRestores = Returns.size(); 1098 1099 // If we are inlining an invoke instruction, insert restores before each 1100 // unwind. These unwinds will be rewritten into branches later. 1101 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 1102 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1103 BB != E; ++BB) 1104 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 1105 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr); 1106 ++NumStackRestores; 1107 } 1108 } 1109 } 1110 1111 // If we are inlining tail call instruction through a call site that isn't 1112 // marked 'tail', we must remove the tail marker for any calls in the inlined 1113 // code. Also, calls inlined through a 'nounwind' call site should be marked 1114 // 'nounwind'. 1115 if (InlinedFunctionInfo.ContainsCalls && 1116 (MustClearTailCallFlags || MarkNoUnwind)) { 1117 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1118 BB != E; ++BB) 1119 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 1120 if (CallInst *CI = dyn_cast<CallInst>(I)) { 1121 if (MustClearTailCallFlags) 1122 CI->setTailCall(false); 1123 if (MarkNoUnwind) 1124 CI->setDoesNotThrow(); 1125 } 1126 } 1127 1128 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 1129 // instructions are unreachable. 1130 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 1131 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1132 BB != E; ++BB) { 1133 TerminatorInst *Term = BB->getTerminator(); 1134 if (isa<UnwindInst>(Term)) { 1135 new UnreachableInst(Context, Term); 1136 BB->getInstList().erase(Term); 1137 } 1138 } 1139 1140 // If we are inlining for an invoke instruction, we must make sure to rewrite 1141 // any inlined 'unwind' instructions into branches to the invoke exception 1142 // destination, and call instructions into invoke instructions. 1143 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 1144 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 1145 1146 // If we cloned in _exactly one_ basic block, and if that block ends in a 1147 // return instruction, we splice the body of the inlined callee directly into 1148 // the calling basic block. 1149 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 1150 // Move all of the instructions right before the call. 1151 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 1152 FirstNewBlock->begin(), FirstNewBlock->end()); 1153 // Remove the cloned basic block. 1154 Caller->getBasicBlockList().pop_back(); 1155 1156 // If the call site was an invoke instruction, add a branch to the normal 1157 // destination. 1158 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 1159 BranchInst::Create(II->getNormalDest(), TheCall); 1160 1161 // If the return instruction returned a value, replace uses of the call with 1162 // uses of the returned value. 1163 if (!TheCall->use_empty()) { 1164 ReturnInst *R = Returns[0]; 1165 if (TheCall == R->getReturnValue()) 1166 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1167 else 1168 TheCall->replaceAllUsesWith(R->getReturnValue()); 1169 } 1170 // Since we are now done with the Call/Invoke, we can delete it. 1171 TheCall->eraseFromParent(); 1172 1173 // Since we are now done with the return instruction, delete it also. 1174 Returns[0]->eraseFromParent(); 1175 1176 // We are now done with the inlining. 1177 return true; 1178 } 1179 1180 // Otherwise, we have the normal case, of more than one block to inline or 1181 // multiple return sites. 1182 1183 // We want to clone the entire callee function into the hole between the 1184 // "starter" and "ender" blocks. How we accomplish this depends on whether 1185 // this is an invoke instruction or a call instruction. 1186 BasicBlock *AfterCallBB; 1187 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 1188 1189 // Add an unconditional branch to make this look like the CallInst case... 1190 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 1191 1192 // Split the basic block. This guarantees that no PHI nodes will have to be 1193 // updated due to new incoming edges, and make the invoke case more 1194 // symmetric to the call case. 1195 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 1196 CalledFunc->getName()+".exit"); 1197 1198 } else { // It's a call 1199 // If this is a call instruction, we need to split the basic block that 1200 // the call lives in. 1201 // 1202 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 1203 CalledFunc->getName()+".exit"); 1204 } 1205 1206 // Change the branch that used to go to AfterCallBB to branch to the first 1207 // basic block of the inlined function. 1208 // 1209 TerminatorInst *Br = OrigBB->getTerminator(); 1210 assert(Br && Br->getOpcode() == Instruction::Br && 1211 "splitBasicBlock broken!"); 1212 Br->setOperand(0, FirstNewBlock); 1213 1214 1215 // Now that the function is correct, make it a little bit nicer. In 1216 // particular, move the basic blocks inserted from the end of the function 1217 // into the space made by splitting the source basic block. 1218 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 1219 FirstNewBlock, Caller->end()); 1220 1221 // Handle all of the return instructions that we just cloned in, and eliminate 1222 // any users of the original call/invoke instruction. 1223 Type *RTy = CalledFunc->getReturnType(); 1224 1225 PHINode *PHI = 0; 1226 if (Returns.size() > 1) { 1227 // The PHI node should go at the front of the new basic block to merge all 1228 // possible incoming values. 1229 if (!TheCall->use_empty()) { 1230 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 1231 AfterCallBB->begin()); 1232 // Anything that used the result of the function call should now use the 1233 // PHI node as their operand. 1234 TheCall->replaceAllUsesWith(PHI); 1235 } 1236 1237 // Loop over all of the return instructions adding entries to the PHI node 1238 // as appropriate. 1239 if (PHI) { 1240 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1241 ReturnInst *RI = Returns[i]; 1242 assert(RI->getReturnValue()->getType() == PHI->getType() && 1243 "Ret value not consistent in function!"); 1244 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 1245 } 1246 } 1247 1248 1249 // Add a branch to the merge points and remove return instructions. 1250 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1251 ReturnInst *RI = Returns[i]; 1252 BranchInst::Create(AfterCallBB, RI); 1253 RI->eraseFromParent(); 1254 } 1255 } else if (!Returns.empty()) { 1256 // Otherwise, if there is exactly one return value, just replace anything 1257 // using the return value of the call with the computed value. 1258 if (!TheCall->use_empty()) { 1259 if (TheCall == Returns[0]->getReturnValue()) 1260 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1261 else 1262 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 1263 } 1264 1265 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 1266 BasicBlock *ReturnBB = Returns[0]->getParent(); 1267 ReturnBB->replaceAllUsesWith(AfterCallBB); 1268 1269 // Splice the code from the return block into the block that it will return 1270 // to, which contains the code that was after the call. 1271 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 1272 ReturnBB->getInstList()); 1273 1274 // Delete the return instruction now and empty ReturnBB now. 1275 Returns[0]->eraseFromParent(); 1276 ReturnBB->eraseFromParent(); 1277 } else if (!TheCall->use_empty()) { 1278 // No returns, but something is using the return value of the call. Just 1279 // nuke the result. 1280 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1281 } 1282 1283 // Since we are now done with the Call/Invoke, we can delete it. 1284 TheCall->eraseFromParent(); 1285 1286 // We should always be able to fold the entry block of the function into the 1287 // single predecessor of the block... 1288 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 1289 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 1290 1291 // Splice the code entry block into calling block, right before the 1292 // unconditional branch. 1293 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 1294 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 1295 1296 // Remove the unconditional branch. 1297 OrigBB->getInstList().erase(Br); 1298 1299 // Now we can remove the CalleeEntry block, which is now empty. 1300 Caller->getBasicBlockList().erase(CalleeEntry); 1301 1302 // If we inserted a phi node, check to see if it has a single value (e.g. all 1303 // the entries are the same or undef). If so, remove the PHI so it doesn't 1304 // block other optimizations. 1305 if (PHI) 1306 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 1307 PHI->replaceAllUsesWith(V); 1308 PHI->eraseFromParent(); 1309 } 1310 1311 return true; 1312} 1313