InlineFunction.cpp revision e8ce880f1a66dffcad7955868be5193e782c09ca
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 PHINode *InnerEHValuesPHI; 254 LandingPadInst *CallerLPad; 255 BasicBlock *SplitLPad; 256 257 public: 258 InvokeInliningInfo(InvokeInst *II) 259 : OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), 260 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0), 261 InnerEHValuesPHI(0), CallerLPad(0), SplitLPad(0) { 262 // If there are PHI nodes in the unwind destination block, we 263 // need to keep track of which values came into them from the 264 // invoke before removing the edge from this block. 265 llvm::BasicBlock *invokeBB = II->getParent(); 266 BasicBlock::iterator I = OuterUnwindDest->begin(); 267 for (; isa<PHINode>(I); ++I) { 268 // Save the value to use for this edge. 269 PHINode *phi = cast<PHINode>(I); 270 UnwindDestPHIValues.push_back(phi->getIncomingValueForBlock(invokeBB)); 271 } 272 273 // FIXME: With the new EH, this if/dyn_cast should be a 'cast'. 274 if (LandingPadInst *LPI = dyn_cast<LandingPadInst>(I)) 275 CallerLPad = LPI; 276 } 277 278 /// The outer unwind destination is the target of unwind edges 279 /// introduced for calls within the inlined function. 280 BasicBlock *getOuterUnwindDest() const { 281 return OuterUnwindDest; 282 } 283 284 EHSelectorInst *getOuterSelector() { 285 if (!OuterSelector) 286 OuterSelector = findSelectorForLandingPad(OuterUnwindDest); 287 return OuterSelector; 288 } 289 290 BasicBlock *getInnerUnwindDest(); 291 292 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 293 BasicBlock *getSplitLandingPad() { 294 if (SplitLPad) return SplitLPad; 295 assert(CallerLPad && "Trying to split a block that isn't a landing pad!"); 296 BasicBlock::iterator I = CallerLPad; ++I; 297 SplitLPad = CallerLPad->getParent()->splitBasicBlock(I, "split.lpad"); 298 return SplitLPad; 299 } 300 301 bool forwardEHResume(CallInst *call, BasicBlock *src); 302 303 /// forwardResume - Forward the 'resume' instruction to the caller's landing 304 /// pad block. When the landing pad block has only one predecessor, this is 305 /// a simple branch. When there is more than one predecessor, we need to 306 /// split the landing pad block after the landingpad instruction and jump 307 /// to there. 308 void forwardResume(ResumeInst *RI); 309 310 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind 311 /// destination block for the given basic block, using the values for the 312 /// original invoke's source block. 313 void addIncomingPHIValuesFor(BasicBlock *BB) const { 314 addIncomingPHIValuesForInto(BB, OuterUnwindDest); 315 } 316 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 317 BasicBlock::iterator I = dest->begin(); 318 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 319 PHINode *phi = cast<PHINode>(I); 320 phi->addIncoming(UnwindDestPHIValues[i], src); 321 } 322 } 323 }; 324} 325 326/// Get or create a target for the branch out of rewritten calls to 327/// llvm.eh.resume. 328BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { 329 if (InnerUnwindDest) return InnerUnwindDest; 330 331 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls 332 // in the outer landing pad to immediately following the phis. 333 EHSelectorInst *selector = getOuterSelector(); 334 if (!selector) return 0; 335 336 // The call to llvm.eh.exception *must* be in the landing pad. 337 Instruction *exn = cast<Instruction>(selector->getArgOperand(0)); 338 assert(exn->getParent() == OuterUnwindDest); 339 340 // TODO: recognize when we've already done this, so that we don't 341 // get a linear number of these when inlining calls into lots of 342 // invokes with the same landing pad. 343 344 // Do the hoisting. 345 Instruction *splitPoint = exn->getParent()->getFirstNonPHI(); 346 assert(splitPoint != selector && "selector-on-exception dominance broken!"); 347 if (splitPoint == exn) { 348 selector->removeFromParent(); 349 selector->insertAfter(exn); 350 splitPoint = selector->getNextNode(); 351 } else { 352 exn->moveBefore(splitPoint); 353 selector->moveBefore(splitPoint); 354 } 355 356 // Split the landing pad. 357 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint, 358 OuterUnwindDest->getName() + ".body"); 359 360 // The number of incoming edges we expect to the inner landing pad. 361 const unsigned phiCapacity = 2; 362 363 // Create corresponding new phis for all the phis in the outer landing pad. 364 BasicBlock::iterator insertPoint = InnerUnwindDest->begin(); 365 BasicBlock::iterator I = OuterUnwindDest->begin(); 366 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 367 PHINode *outerPhi = cast<PHINode>(I); 368 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity, 369 outerPhi->getName() + ".lpad-body", 370 insertPoint); 371 outerPhi->replaceAllUsesWith(innerPhi); 372 innerPhi->addIncoming(outerPhi, OuterUnwindDest); 373 } 374 375 // Create a phi for the exception value... 376 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity, 377 "exn.lpad-body", insertPoint); 378 exn->replaceAllUsesWith(InnerExceptionPHI); 379 selector->setArgOperand(0, exn); // restore this use 380 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest); 381 382 // ...and the selector. 383 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity, 384 "selector.lpad-body", insertPoint); 385 selector->replaceAllUsesWith(InnerSelectorPHI); 386 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest); 387 388 // All done. 389 return InnerUnwindDest; 390} 391 392/// [LIBUNWIND] Try to forward the given call, which logically occurs 393/// at the end of the given block, as a branch to the inner unwind 394/// block. Returns true if the call was forwarded. 395bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) { 396 // First, check whether this is a call to the intrinsic. 397 Function *fn = dyn_cast<Function>(call->getCalledValue()); 398 if (!fn || fn->getName() != "llvm.eh.resume") 399 return false; 400 401 // At this point, we need to return true on all paths, because 402 // otherwise we'll construct an invoke of the intrinsic, which is 403 // not well-formed. 404 405 // Try to find or make an inner unwind dest, which will fail if we 406 // can't find a selector call for the outer unwind dest. 407 BasicBlock *dest = getInnerUnwindDest(); 408 bool hasSelector = (dest != 0); 409 410 // If we failed, just use the outer unwind dest, dropping the 411 // exception and selector on the floor. 412 if (!hasSelector) 413 dest = OuterUnwindDest; 414 415 // Make a branch. 416 BranchInst::Create(dest, src); 417 418 // Update the phis in the destination. They were inserted in an 419 // order which makes this work. 420 addIncomingPHIValuesForInto(src, dest); 421 422 if (hasSelector) { 423 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src); 424 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src); 425 } 426 427 return true; 428} 429 430/// forwardResume - Forward the 'resume' instruction to the caller's landing pad 431/// block. When the landing pad block has only one predecessor, this is a simple 432/// branch. When there is more than one predecessor, we need to split the 433/// landing pad block after the landingpad instruction and jump to there. 434void InvokeInliningInfo::forwardResume(ResumeInst *RI) { 435 BasicBlock *LPadBB = CallerLPad->getParent(); 436 Value *ResumeOp = RI->getOperand(0); 437 438 if (!LPadBB->getSinglePredecessor()) { 439 // There are multiple predecessors to this landing pad block. Split this 440 // landing pad block and jump to the new BB. 441 BasicBlock *SplitLPad = getSplitLandingPad(); 442 BranchInst::Create(SplitLPad, RI->getParent()); 443 444 if (CallerLPad->hasOneUse() && isa<PHINode>(CallerLPad->use_back())) { 445 PHINode *PN = cast<PHINode>(CallerLPad->use_back()); 446 PN->addIncoming(ResumeOp, RI->getParent()); 447 } else { 448 PHINode *PN = PHINode::Create(ResumeOp->getType(), 0, "lpad.phi", 449 &SplitLPad->front()); 450 CallerLPad->replaceAllUsesWith(PN); 451 PN->addIncoming(ResumeOp, RI->getParent()); 452 PN->addIncoming(CallerLPad, LPadBB); 453 } 454 455 RI->eraseFromParent(); 456 return; 457 } 458 459 BranchInst::Create(LPadBB, RI->getParent()); 460 CallerLPad->replaceAllUsesWith(ResumeOp); 461 CallerLPad->eraseFromParent(); 462 RI->eraseFromParent(); 463} 464 465/// [LIBUNWIND] Check whether this selector is "only cleanups": 466/// call i32 @llvm.eh.selector(blah, blah, i32 0) 467static bool isCleanupOnlySelector(EHSelectorInst *selector) { 468 if (selector->getNumArgOperands() != 3) return false; 469 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2)); 470 return (val && val->isZero()); 471} 472 473/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 474/// an invoke, we have to turn all of the calls that can throw into 475/// invokes. This function analyze BB to see if there are any calls, and if so, 476/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 477/// nodes in that block with the values specified in InvokeDestPHIValues. 478/// 479/// Returns true to indicate that the next block should be skipped. 480static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 481 InvokeInliningInfo &Invoke) { 482 LandingPadInst *LPI = Invoke.getLandingPadInst(); 483 484 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 485 Instruction *I = BBI++; 486 487 if (LPI) // FIXME: This won't be NULL in the new EH. 488 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) { 489 unsigned NumClauses = LPI->getNumClauses(); 490 L->reserveClauses(NumClauses); 491 for (unsigned i = 0; i != NumClauses; ++i) 492 L->addClause(LPI->getClauseType(i), LPI->getClauseValue(i)); 493 } 494 495 // We only need to check for function calls: inlined invoke 496 // instructions require no special handling. 497 CallInst *CI = dyn_cast<CallInst>(I); 498 if (CI == 0) continue; 499 500 // LIBUNWIND: merge selector instructions. 501 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) { 502 EHSelectorInst *Outer = Invoke.getOuterSelector(); 503 if (!Outer) continue; 504 505 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); 506 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); 507 508 // If both selectors contain only cleanups, we don't need to do 509 // anything. TODO: this is really just a very specific instance 510 // of a much more general optimization. 511 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; 512 513 // Otherwise, we just append the outer selector to the inner selector. 514 SmallVector<Value*, 16> NewSelector; 515 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) 516 NewSelector.push_back(Inner->getArgOperand(i)); 517 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) 518 NewSelector.push_back(Outer->getArgOperand(i)); 519 520 CallInst *NewInner = 521 IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), NewSelector); 522 // No need to copy attributes, calling convention, etc. 523 NewInner->takeName(Inner); 524 Inner->replaceAllUsesWith(NewInner); 525 Inner->eraseFromParent(); 526 continue; 527 } 528 529 // If this call cannot unwind, don't convert it to an invoke. 530 if (CI->doesNotThrow()) 531 continue; 532 533 // Convert this function call into an invoke instruction. 534 // First, split the basic block. 535 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 536 537 // Delete the unconditional branch inserted by splitBasicBlock 538 BB->getInstList().pop_back(); 539 540 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch 541 // directly to the new landing pad. 542 if (Invoke.forwardEHResume(CI, BB)) { 543 // TODO: 'Split' is now unreachable; clean it up. 544 545 // We want to leave the original call intact so that the call 546 // graph and other structures won't get misled. We also have to 547 // avoid processing the next block, or we'll iterate here forever. 548 return true; 549 } 550 551 // Otherwise, create the new invoke instruction. 552 ImmutableCallSite CS(CI); 553 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 554 InvokeInst *II = 555 InvokeInst::Create(CI->getCalledValue(), Split, 556 Invoke.getOuterUnwindDest(), 557 InvokeArgs, CI->getName(), BB); 558 II->setCallingConv(CI->getCallingConv()); 559 II->setAttributes(CI->getAttributes()); 560 561 // Make sure that anything using the call now uses the invoke! This also 562 // updates the CallGraph if present, because it uses a WeakVH. 563 CI->replaceAllUsesWith(II); 564 565 Split->getInstList().pop_front(); // Delete the original call 566 567 // Update any PHI nodes in the exceptional block to indicate that 568 // there is now a new entry in them. 569 Invoke.addIncomingPHIValuesFor(BB); 570 return false; 571 } 572 573 return false; 574} 575 576 577/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 578/// in the body of the inlined function into invokes and turn unwind 579/// instructions into branches to the invoke unwind dest. 580/// 581/// II is the invoke instruction being inlined. FirstNewBlock is the first 582/// block of the inlined code (the last block is the end of the function), 583/// and InlineCodeInfo is information about the code that got inlined. 584static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 585 ClonedCodeInfo &InlinedCodeInfo) { 586 BasicBlock *InvokeDest = II->getUnwindDest(); 587 588 Function *Caller = FirstNewBlock->getParent(); 589 590 // The inlined code is currently at the end of the function, scan from the 591 // start of the inlined code to its end, checking for stuff we need to 592 // rewrite. If the code doesn't have calls or unwinds, we know there is 593 // nothing to rewrite. 594 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 595 // Now that everything is happy, we have one final detail. The PHI nodes in 596 // the exception destination block still have entries due to the original 597 // invoke instruction. Eliminate these entries (which might even delete the 598 // PHI node) now. 599 InvokeDest->removePredecessor(II->getParent()); 600 return; 601 } 602 603 InvokeInliningInfo Invoke(II); 604 605 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 606 if (InlinedCodeInfo.ContainsCalls) 607 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 608 // Honor a request to skip the next block. We don't need to 609 // consider UnwindInsts in this case either. 610 ++BB; 611 continue; 612 } 613 614 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 615 // An UnwindInst requires special handling when it gets inlined into an 616 // invoke site. Once this happens, we know that the unwind would cause 617 // a control transfer to the invoke exception destination, so we can 618 // transform it into a direct branch to the exception destination. 619 BranchInst::Create(InvokeDest, UI); 620 621 // Delete the unwind instruction! 622 UI->eraseFromParent(); 623 624 // Update any PHI nodes in the exceptional block to indicate that 625 // there is now a new entry in them. 626 Invoke.addIncomingPHIValuesFor(BB); 627 } 628 629 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) { 630 Invoke.forwardResume(RI); 631 } 632 } 633 634 // Now that everything is happy, we have one final detail. The PHI nodes in 635 // the exception destination block still have entries due to the original 636 // invoke instruction. Eliminate these entries (which might even delete the 637 // PHI node) now. 638 InvokeDest->removePredecessor(II->getParent()); 639} 640 641/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 642/// into the caller, update the specified callgraph to reflect the changes we 643/// made. Note that it's possible that not all code was copied over, so only 644/// some edges of the callgraph may remain. 645static void UpdateCallGraphAfterInlining(CallSite CS, 646 Function::iterator FirstNewBlock, 647 ValueToValueMapTy &VMap, 648 InlineFunctionInfo &IFI) { 649 CallGraph &CG = *IFI.CG; 650 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 651 const Function *Callee = CS.getCalledFunction(); 652 CallGraphNode *CalleeNode = CG[Callee]; 653 CallGraphNode *CallerNode = CG[Caller]; 654 655 // Since we inlined some uninlined call sites in the callee into the caller, 656 // add edges from the caller to all of the callees of the callee. 657 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 658 659 // Consider the case where CalleeNode == CallerNode. 660 CallGraphNode::CalledFunctionsVector CallCache; 661 if (CalleeNode == CallerNode) { 662 CallCache.assign(I, E); 663 I = CallCache.begin(); 664 E = CallCache.end(); 665 } 666 667 for (; I != E; ++I) { 668 const Value *OrigCall = I->first; 669 670 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 671 // Only copy the edge if the call was inlined! 672 if (VMI == VMap.end() || VMI->second == 0) 673 continue; 674 675 // If the call was inlined, but then constant folded, there is no edge to 676 // add. Check for this case. 677 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 678 if (NewCall == 0) continue; 679 680 // Remember that this call site got inlined for the client of 681 // InlineFunction. 682 IFI.InlinedCalls.push_back(NewCall); 683 684 // It's possible that inlining the callsite will cause it to go from an 685 // indirect to a direct call by resolving a function pointer. If this 686 // happens, set the callee of the new call site to a more precise 687 // destination. This can also happen if the call graph node of the caller 688 // was just unnecessarily imprecise. 689 if (I->second->getFunction() == 0) 690 if (Function *F = CallSite(NewCall).getCalledFunction()) { 691 // Indirect call site resolved to direct call. 692 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 693 694 continue; 695 } 696 697 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 698 } 699 700 // Update the call graph by deleting the edge from Callee to Caller. We must 701 // do this after the loop above in case Caller and Callee are the same. 702 CallerNode->removeCallEdgeFor(CS); 703} 704 705/// HandleByValArgument - When inlining a call site that has a byval argument, 706/// we have to make the implicit memcpy explicit by adding it. 707static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 708 const Function *CalledFunc, 709 InlineFunctionInfo &IFI, 710 unsigned ByValAlignment) { 711 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 712 713 // If the called function is readonly, then it could not mutate the caller's 714 // copy of the byval'd memory. In this case, it is safe to elide the copy and 715 // temporary. 716 if (CalledFunc->onlyReadsMemory()) { 717 // If the byval argument has a specified alignment that is greater than the 718 // passed in pointer, then we either have to round up the input pointer or 719 // give up on this transformation. 720 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 721 return Arg; 722 723 // If the pointer is already known to be sufficiently aligned, or if we can 724 // round it up to a larger alignment, then we don't need a temporary. 725 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 726 IFI.TD) >= ByValAlignment) 727 return Arg; 728 729 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 730 // for code quality, but rarely happens and is required for correctness. 731 } 732 733 LLVMContext &Context = Arg->getContext(); 734 735 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 736 737 // Create the alloca. If we have TargetData, use nice alignment. 738 unsigned Align = 1; 739 if (IFI.TD) 740 Align = IFI.TD->getPrefTypeAlignment(AggTy); 741 742 // If the byval had an alignment specified, we *must* use at least that 743 // alignment, as it is required by the byval argument (and uses of the 744 // pointer inside the callee). 745 Align = std::max(Align, ByValAlignment); 746 747 Function *Caller = TheCall->getParent()->getParent(); 748 749 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 750 &*Caller->begin()->begin()); 751 // Emit a memcpy. 752 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 753 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 754 Intrinsic::memcpy, 755 Tys); 756 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 757 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 758 759 Value *Size; 760 if (IFI.TD == 0) 761 Size = ConstantExpr::getSizeOf(AggTy); 762 else 763 Size = ConstantInt::get(Type::getInt64Ty(Context), 764 IFI.TD->getTypeStoreSize(AggTy)); 765 766 // Always generate a memcpy of alignment 1 here because we don't know 767 // the alignment of the src pointer. Other optimizations can infer 768 // better alignment. 769 Value *CallArgs[] = { 770 DestCast, SrcCast, Size, 771 ConstantInt::get(Type::getInt32Ty(Context), 1), 772 ConstantInt::getFalse(Context) // isVolatile 773 }; 774 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs); 775 776 // Uses of the argument in the function should use our new alloca 777 // instead. 778 return NewAlloca; 779} 780 781// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 782// intrinsic. 783static bool isUsedByLifetimeMarker(Value *V) { 784 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 785 ++UI) { 786 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 787 switch (II->getIntrinsicID()) { 788 default: break; 789 case Intrinsic::lifetime_start: 790 case Intrinsic::lifetime_end: 791 return true; 792 } 793 } 794 } 795 return false; 796} 797 798// hasLifetimeMarkers - Check whether the given alloca already has 799// lifetime.start or lifetime.end intrinsics. 800static bool hasLifetimeMarkers(AllocaInst *AI) { 801 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 802 if (AI->getType() == Int8PtrTy) 803 return isUsedByLifetimeMarker(AI); 804 805 // Do a scan to find all the casts to i8*. 806 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 807 ++I) { 808 if (I->getType() != Int8PtrTy) continue; 809 if (I->stripPointerCasts() != AI) continue; 810 if (isUsedByLifetimeMarker(*I)) 811 return true; 812 } 813 return false; 814} 815 816/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively 817/// update InlinedAtEntry of a DebugLoc. 818static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 819 const DebugLoc &InlinedAtDL, 820 LLVMContext &Ctx) { 821 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 822 DebugLoc NewInlinedAtDL 823 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 824 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 825 NewInlinedAtDL.getAsMDNode(Ctx)); 826 } 827 828 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 829 InlinedAtDL.getAsMDNode(Ctx)); 830} 831 832 833/// fixupLineNumbers - Update inlined instructions' line numbers to 834/// to encode location where these instructions are inlined. 835static void fixupLineNumbers(Function *Fn, Function::iterator FI, 836 Instruction *TheCall) { 837 DebugLoc TheCallDL = TheCall->getDebugLoc(); 838 if (TheCallDL.isUnknown()) 839 return; 840 841 for (; FI != Fn->end(); ++FI) { 842 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 843 BI != BE; ++BI) { 844 DebugLoc DL = BI->getDebugLoc(); 845 if (!DL.isUnknown()) { 846 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 847 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 848 LLVMContext &Ctx = BI->getContext(); 849 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 850 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 851 InlinedAt, Ctx)); 852 } 853 } 854 } 855 } 856} 857 858// InlineFunction - This function inlines the called function into the basic 859// block of the caller. This returns false if it is not possible to inline this 860// call. The program is still in a well defined state if this occurs though. 861// 862// Note that this only does one level of inlining. For example, if the 863// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 864// exists in the instruction stream. Similarly this will inline a recursive 865// function by one level. 866// 867bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 868 Instruction *TheCall = CS.getInstruction(); 869 LLVMContext &Context = TheCall->getContext(); 870 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 871 "Instruction not in function!"); 872 873 // If IFI has any state in it, zap it before we fill it in. 874 IFI.reset(); 875 876 const Function *CalledFunc = CS.getCalledFunction(); 877 if (CalledFunc == 0 || // Can't inline external function or indirect 878 CalledFunc->isDeclaration() || // call, or call to a vararg function! 879 CalledFunc->getFunctionType()->isVarArg()) return false; 880 881 // If the call to the callee is not a tail call, we must clear the 'tail' 882 // flags on any calls that we inline. 883 bool MustClearTailCallFlags = 884 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 885 886 // If the call to the callee cannot throw, set the 'nounwind' flag on any 887 // calls that we inline. 888 bool MarkNoUnwind = CS.doesNotThrow(); 889 890 BasicBlock *OrigBB = TheCall->getParent(); 891 Function *Caller = OrigBB->getParent(); 892 893 // GC poses two hazards to inlining, which only occur when the callee has GC: 894 // 1. If the caller has no GC, then the callee's GC must be propagated to the 895 // caller. 896 // 2. If the caller has a differing GC, it is invalid to inline. 897 if (CalledFunc->hasGC()) { 898 if (!Caller->hasGC()) 899 Caller->setGC(CalledFunc->getGC()); 900 else if (CalledFunc->getGC() != Caller->getGC()) 901 return false; 902 } 903 904 // Find the personality function used by the landing pads of the caller. If it 905 // exists, then check to see that it matches the personality function used in 906 // the callee. 907 for (Function::const_iterator 908 I = Caller->begin(), E = Caller->end(); I != E; ++I) 909 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 910 const BasicBlock *BB = II->getUnwindDest(); 911 // FIXME: This 'isa' here should become go away once the new EH system is 912 // in place. 913 if (!isa<LandingPadInst>(BB->getFirstNonPHI())) 914 continue; 915 const LandingPadInst *LP = cast<LandingPadInst>(BB->getFirstNonPHI()); 916 const Value *CallerPersFn = LP->getPersonalityFn(); 917 918 // If the personality functions match, then we can perform the 919 // inlining. Otherwise, we can't inline. 920 // TODO: This isn't 100% true. Some personality functions are proper 921 // supersets of others and can be used in place of the other. 922 for (Function::const_iterator 923 I = CalledFunc->begin(), E = CalledFunc->end(); I != E; ++I) 924 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 925 const BasicBlock *BB = II->getUnwindDest(); 926 // FIXME: This 'if/dyn_cast' here should become a normal 'cast' once 927 // the new EH system is in place. 928 if (const LandingPadInst *LP = 929 dyn_cast<LandingPadInst>(BB->getFirstNonPHI())) 930 if (CallerPersFn != LP->getPersonalityFn()) 931 return false; 932 break; 933 } 934 935 break; 936 } 937 938 // Get an iterator to the last basic block in the function, which will have 939 // the new function inlined after it. 940 // 941 Function::iterator LastBlock = &Caller->back(); 942 943 // Make sure to capture all of the return instructions from the cloned 944 // function. 945 SmallVector<ReturnInst*, 8> Returns; 946 ClonedCodeInfo InlinedFunctionInfo; 947 Function::iterator FirstNewBlock; 948 949 { // Scope to destroy VMap after cloning. 950 ValueToValueMapTy VMap; 951 952 assert(CalledFunc->arg_size() == CS.arg_size() && 953 "No varargs calls can be inlined!"); 954 955 // Calculate the vector of arguments to pass into the function cloner, which 956 // matches up the formal to the actual argument values. 957 CallSite::arg_iterator AI = CS.arg_begin(); 958 unsigned ArgNo = 0; 959 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 960 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 961 Value *ActualArg = *AI; 962 963 // When byval arguments actually inlined, we need to make the copy implied 964 // by them explicit. However, we don't do this if the callee is readonly 965 // or readnone, because the copy would be unneeded: the callee doesn't 966 // modify the struct. 967 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 968 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 969 CalledFunc->getParamAlignment(ArgNo+1)); 970 971 // Calls that we inline may use the new alloca, so we need to clear 972 // their 'tail' flags if HandleByValArgument introduced a new alloca and 973 // the callee has calls. 974 MustClearTailCallFlags |= ActualArg != *AI; 975 } 976 977 VMap[I] = ActualArg; 978 } 979 980 // We want the inliner to prune the code as it copies. We would LOVE to 981 // have no dead or constant instructions leftover after inlining occurs 982 // (which can happen, e.g., because an argument was constant), but we'll be 983 // happy with whatever the cloner can do. 984 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 985 /*ModuleLevelChanges=*/false, Returns, ".i", 986 &InlinedFunctionInfo, IFI.TD, TheCall); 987 988 // Remember the first block that is newly cloned over. 989 FirstNewBlock = LastBlock; ++FirstNewBlock; 990 991 // Update the callgraph if requested. 992 if (IFI.CG) 993 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 994 995 // Update inlined instructions' line number information. 996 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 997 } 998 999 // If there are any alloca instructions in the block that used to be the entry 1000 // block for the callee, move them to the entry block of the caller. First 1001 // calculate which instruction they should be inserted before. We insert the 1002 // instructions at the end of the current alloca list. 1003 // 1004 { 1005 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 1006 for (BasicBlock::iterator I = FirstNewBlock->begin(), 1007 E = FirstNewBlock->end(); I != E; ) { 1008 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 1009 if (AI == 0) continue; 1010 1011 // If the alloca is now dead, remove it. This often occurs due to code 1012 // specialization. 1013 if (AI->use_empty()) { 1014 AI->eraseFromParent(); 1015 continue; 1016 } 1017 1018 if (!isa<Constant>(AI->getArraySize())) 1019 continue; 1020 1021 // Keep track of the static allocas that we inline into the caller. 1022 IFI.StaticAllocas.push_back(AI); 1023 1024 // Scan for the block of allocas that we can move over, and move them 1025 // all at once. 1026 while (isa<AllocaInst>(I) && 1027 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 1028 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 1029 ++I; 1030 } 1031 1032 // Transfer all of the allocas over in a block. Using splice means 1033 // that the instructions aren't removed from the symbol table, then 1034 // reinserted. 1035 Caller->getEntryBlock().getInstList().splice(InsertPoint, 1036 FirstNewBlock->getInstList(), 1037 AI, I); 1038 } 1039 } 1040 1041 // Leave lifetime markers for the static alloca's, scoping them to the 1042 // function we just inlined. 1043 if (!IFI.StaticAllocas.empty()) { 1044 IRBuilder<> builder(FirstNewBlock->begin()); 1045 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 1046 AllocaInst *AI = IFI.StaticAllocas[ai]; 1047 1048 // If the alloca is already scoped to something smaller than the whole 1049 // function then there's no need to add redundant, less accurate markers. 1050 if (hasLifetimeMarkers(AI)) 1051 continue; 1052 1053 builder.CreateLifetimeStart(AI); 1054 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 1055 IRBuilder<> builder(Returns[ri]); 1056 builder.CreateLifetimeEnd(AI); 1057 } 1058 } 1059 } 1060 1061 // If the inlined code contained dynamic alloca instructions, wrap the inlined 1062 // code with llvm.stacksave/llvm.stackrestore intrinsics. 1063 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 1064 Module *M = Caller->getParent(); 1065 // Get the two intrinsics we care about. 1066 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 1067 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 1068 1069 // Insert the llvm.stacksave. 1070 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 1071 .CreateCall(StackSave, "savedstack"); 1072 1073 // Insert a call to llvm.stackrestore before any return instructions in the 1074 // inlined function. 1075 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1076 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); 1077 } 1078 1079 // Count the number of StackRestore calls we insert. 1080 unsigned NumStackRestores = Returns.size(); 1081 1082 // If we are inlining an invoke instruction, insert restores before each 1083 // unwind. These unwinds will be rewritten into branches later. 1084 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 1085 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1086 BB != E; ++BB) 1087 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 1088 IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr); 1089 ++NumStackRestores; 1090 } 1091 } 1092 } 1093 1094 // If we are inlining tail call instruction through a call site that isn't 1095 // marked 'tail', we must remove the tail marker for any calls in the inlined 1096 // code. Also, calls inlined through a 'nounwind' call site should be marked 1097 // 'nounwind'. 1098 if (InlinedFunctionInfo.ContainsCalls && 1099 (MustClearTailCallFlags || MarkNoUnwind)) { 1100 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1101 BB != E; ++BB) 1102 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 1103 if (CallInst *CI = dyn_cast<CallInst>(I)) { 1104 if (MustClearTailCallFlags) 1105 CI->setTailCall(false); 1106 if (MarkNoUnwind) 1107 CI->setDoesNotThrow(); 1108 } 1109 } 1110 1111 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 1112 // instructions are unreachable. 1113 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 1114 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 1115 BB != E; ++BB) { 1116 TerminatorInst *Term = BB->getTerminator(); 1117 if (isa<UnwindInst>(Term)) { 1118 new UnreachableInst(Context, Term); 1119 BB->getInstList().erase(Term); 1120 } 1121 } 1122 1123 // If we are inlining for an invoke instruction, we must make sure to rewrite 1124 // any inlined 'unwind' instructions into branches to the invoke exception 1125 // destination, and call instructions into invoke instructions. 1126 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 1127 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 1128 1129 // If we cloned in _exactly one_ basic block, and if that block ends in a 1130 // return instruction, we splice the body of the inlined callee directly into 1131 // the calling basic block. 1132 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 1133 // Move all of the instructions right before the call. 1134 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 1135 FirstNewBlock->begin(), FirstNewBlock->end()); 1136 // Remove the cloned basic block. 1137 Caller->getBasicBlockList().pop_back(); 1138 1139 // If the call site was an invoke instruction, add a branch to the normal 1140 // destination. 1141 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 1142 BranchInst::Create(II->getNormalDest(), TheCall); 1143 1144 // If the return instruction returned a value, replace uses of the call with 1145 // uses of the returned value. 1146 if (!TheCall->use_empty()) { 1147 ReturnInst *R = Returns[0]; 1148 if (TheCall == R->getReturnValue()) 1149 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1150 else 1151 TheCall->replaceAllUsesWith(R->getReturnValue()); 1152 } 1153 // Since we are now done with the Call/Invoke, we can delete it. 1154 TheCall->eraseFromParent(); 1155 1156 // Since we are now done with the return instruction, delete it also. 1157 Returns[0]->eraseFromParent(); 1158 1159 // We are now done with the inlining. 1160 return true; 1161 } 1162 1163 // Otherwise, we have the normal case, of more than one block to inline or 1164 // multiple return sites. 1165 1166 // We want to clone the entire callee function into the hole between the 1167 // "starter" and "ender" blocks. How we accomplish this depends on whether 1168 // this is an invoke instruction or a call instruction. 1169 BasicBlock *AfterCallBB; 1170 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 1171 1172 // Add an unconditional branch to make this look like the CallInst case... 1173 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 1174 1175 // Split the basic block. This guarantees that no PHI nodes will have to be 1176 // updated due to new incoming edges, and make the invoke case more 1177 // symmetric to the call case. 1178 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 1179 CalledFunc->getName()+".exit"); 1180 1181 } else { // It's a call 1182 // If this is a call instruction, we need to split the basic block that 1183 // the call lives in. 1184 // 1185 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 1186 CalledFunc->getName()+".exit"); 1187 } 1188 1189 // Change the branch that used to go to AfterCallBB to branch to the first 1190 // basic block of the inlined function. 1191 // 1192 TerminatorInst *Br = OrigBB->getTerminator(); 1193 assert(Br && Br->getOpcode() == Instruction::Br && 1194 "splitBasicBlock broken!"); 1195 Br->setOperand(0, FirstNewBlock); 1196 1197 1198 // Now that the function is correct, make it a little bit nicer. In 1199 // particular, move the basic blocks inserted from the end of the function 1200 // into the space made by splitting the source basic block. 1201 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 1202 FirstNewBlock, Caller->end()); 1203 1204 // Handle all of the return instructions that we just cloned in, and eliminate 1205 // any users of the original call/invoke instruction. 1206 Type *RTy = CalledFunc->getReturnType(); 1207 1208 PHINode *PHI = 0; 1209 if (Returns.size() > 1) { 1210 // The PHI node should go at the front of the new basic block to merge all 1211 // possible incoming values. 1212 if (!TheCall->use_empty()) { 1213 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 1214 AfterCallBB->begin()); 1215 // Anything that used the result of the function call should now use the 1216 // PHI node as their operand. 1217 TheCall->replaceAllUsesWith(PHI); 1218 } 1219 1220 // Loop over all of the return instructions adding entries to the PHI node 1221 // as appropriate. 1222 if (PHI) { 1223 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1224 ReturnInst *RI = Returns[i]; 1225 assert(RI->getReturnValue()->getType() == PHI->getType() && 1226 "Ret value not consistent in function!"); 1227 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 1228 } 1229 } 1230 1231 1232 // Add a branch to the merge points and remove return instructions. 1233 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1234 ReturnInst *RI = Returns[i]; 1235 BranchInst::Create(AfterCallBB, RI); 1236 RI->eraseFromParent(); 1237 } 1238 } else if (!Returns.empty()) { 1239 // Otherwise, if there is exactly one return value, just replace anything 1240 // using the return value of the call with the computed value. 1241 if (!TheCall->use_empty()) { 1242 if (TheCall == Returns[0]->getReturnValue()) 1243 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1244 else 1245 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 1246 } 1247 1248 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 1249 BasicBlock *ReturnBB = Returns[0]->getParent(); 1250 ReturnBB->replaceAllUsesWith(AfterCallBB); 1251 1252 // Splice the code from the return block into the block that it will return 1253 // to, which contains the code that was after the call. 1254 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 1255 ReturnBB->getInstList()); 1256 1257 // Delete the return instruction now and empty ReturnBB now. 1258 Returns[0]->eraseFromParent(); 1259 ReturnBB->eraseFromParent(); 1260 } else if (!TheCall->use_empty()) { 1261 // No returns, but something is using the return value of the call. Just 1262 // nuke the result. 1263 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1264 } 1265 1266 // Since we are now done with the Call/Invoke, we can delete it. 1267 TheCall->eraseFromParent(); 1268 1269 // We should always be able to fold the entry block of the function into the 1270 // single predecessor of the block... 1271 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 1272 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 1273 1274 // Splice the code entry block into calling block, right before the 1275 // unconditional branch. 1276 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 1277 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 1278 1279 // Remove the unconditional branch. 1280 OrigBB->getInstList().erase(Br); 1281 1282 // Now we can remove the CalleeEntry block, which is now empty. 1283 Caller->getBasicBlockList().erase(CalleeEntry); 1284 1285 // If we inserted a phi node, check to see if it has a single value (e.g. all 1286 // the entries are the same or undef). If so, remove the PHI so it doesn't 1287 // block other optimizations. 1288 if (PHI) 1289 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 1290 PHI->replaceAllUsesWith(V); 1291 PHI->eraseFromParent(); 1292 } 1293 1294 return true; 1295} 1296