InlineFunction.cpp revision 708c1ac077fbc0cb73d489b4f4df3b2718566b05
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 public: 254 InvokeInliningInfo(InvokeInst *II) : 255 OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), 256 InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0) { 257 258 // If there are PHI nodes in the unwind destination block, we 259 // need to keep track of which values came into them from the 260 // invoke before removing the edge from this block. 261 llvm::BasicBlock *invokeBB = II->getParent(); 262 for (BasicBlock::iterator I = OuterUnwindDest->begin(); 263 isa<PHINode>(I); ++I) { 264 // Save the value to use for this edge. 265 PHINode *phi = cast<PHINode>(I); 266 UnwindDestPHIValues.push_back(phi->getIncomingValueForBlock(invokeBB)); 267 } 268 } 269 270 /// The outer unwind destination is the target of unwind edges 271 /// introduced for calls within the inlined function. 272 BasicBlock *getOuterUnwindDest() const { 273 return OuterUnwindDest; 274 } 275 276 EHSelectorInst *getOuterSelector() { 277 if (!OuterSelector) 278 OuterSelector = findSelectorForLandingPad(OuterUnwindDest); 279 return OuterSelector; 280 } 281 282 BasicBlock *getInnerUnwindDest(); 283 284 bool forwardEHResume(CallInst *call, BasicBlock *src); 285 286 /// Add incoming-PHI values to the unwind destination block for 287 /// the given basic block, using the values for the original 288 /// invoke's source block. 289 void addIncomingPHIValuesFor(BasicBlock *BB) const { 290 addIncomingPHIValuesForInto(BB, OuterUnwindDest); 291 } 292 293 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 294 BasicBlock::iterator I = dest->begin(); 295 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 296 PHINode *phi = cast<PHINode>(I); 297 phi->addIncoming(UnwindDestPHIValues[i], src); 298 } 299 } 300 }; 301} 302 303/// Get or create a target for the branch out of rewritten calls to 304/// llvm.eh.resume. 305BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { 306 if (InnerUnwindDest) return InnerUnwindDest; 307 308 // Find and hoist the llvm.eh.exception and llvm.eh.selector calls 309 // in the outer landing pad to immediately following the phis. 310 EHSelectorInst *selector = getOuterSelector(); 311 if (!selector) return 0; 312 313 // The call to llvm.eh.exception *must* be in the landing pad. 314 Instruction *exn = cast<Instruction>(selector->getArgOperand(0)); 315 assert(exn->getParent() == OuterUnwindDest); 316 317 // TODO: recognize when we've already done this, so that we don't 318 // get a linear number of these when inlining calls into lots of 319 // invokes with the same landing pad. 320 321 // Do the hoisting. 322 Instruction *splitPoint = exn->getParent()->getFirstNonPHI(); 323 assert(splitPoint != selector && "selector-on-exception dominance broken!"); 324 if (splitPoint == exn) { 325 selector->removeFromParent(); 326 selector->insertAfter(exn); 327 splitPoint = selector->getNextNode(); 328 } else { 329 exn->moveBefore(splitPoint); 330 selector->moveBefore(splitPoint); 331 } 332 333 // Split the landing pad. 334 InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint, 335 OuterUnwindDest->getName() + ".body"); 336 337 // The number of incoming edges we expect to the inner landing pad. 338 const unsigned phiCapacity = 2; 339 340 // Create corresponding new phis for all the phis in the outer landing pad. 341 BasicBlock::iterator insertPoint = InnerUnwindDest->begin(); 342 BasicBlock::iterator I = OuterUnwindDest->begin(); 343 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 344 PHINode *outerPhi = cast<PHINode>(I); 345 PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity, 346 outerPhi->getName() + ".lpad-body", 347 insertPoint); 348 outerPhi->replaceAllUsesWith(innerPhi); 349 innerPhi->addIncoming(outerPhi, OuterUnwindDest); 350 } 351 352 // Create a phi for the exception value... 353 InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity, 354 "exn.lpad-body", insertPoint); 355 exn->replaceAllUsesWith(InnerExceptionPHI); 356 selector->setArgOperand(0, exn); // restore this use 357 InnerExceptionPHI->addIncoming(exn, OuterUnwindDest); 358 359 // ...and the selector. 360 InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity, 361 "selector.lpad-body", insertPoint); 362 selector->replaceAllUsesWith(InnerSelectorPHI); 363 InnerSelectorPHI->addIncoming(selector, OuterUnwindDest); 364 365 // All done. 366 return InnerUnwindDest; 367} 368 369/// [LIBUNWIND] Try to forward the given call, which logically occurs 370/// at the end of the given block, as a branch to the inner unwind 371/// block. Returns true if the call was forwarded. 372bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) { 373 // First, check whether this is a call to the intrinsic. 374 Function *fn = dyn_cast<Function>(call->getCalledValue()); 375 if (!fn || fn->getName() != "llvm.eh.resume") 376 return false; 377 378 // At this point, we need to return true on all paths, because 379 // otherwise we'll construct an invoke of the intrinsic, which is 380 // not well-formed. 381 382 // Try to find or make an inner unwind dest, which will fail if we 383 // can't find a selector call for the outer unwind dest. 384 BasicBlock *dest = getInnerUnwindDest(); 385 bool hasSelector = (dest != 0); 386 387 // If we failed, just use the outer unwind dest, dropping the 388 // exception and selector on the floor. 389 if (!hasSelector) 390 dest = OuterUnwindDest; 391 392 // Make a branch. 393 BranchInst::Create(dest, src); 394 395 // Update the phis in the destination. They were inserted in an 396 // order which makes this work. 397 addIncomingPHIValuesForInto(src, dest); 398 399 if (hasSelector) { 400 InnerExceptionPHI->addIncoming(call->getArgOperand(0), src); 401 InnerSelectorPHI->addIncoming(call->getArgOperand(1), src); 402 } 403 404 return true; 405} 406 407/// [LIBUNWIND] Check whether this selector is "only cleanups": 408/// call i32 @llvm.eh.selector(blah, blah, i32 0) 409static bool isCleanupOnlySelector(EHSelectorInst *selector) { 410 if (selector->getNumArgOperands() != 3) return false; 411 ConstantInt *val = dyn_cast<ConstantInt>(selector->getArgOperand(2)); 412 return (val && val->isZero()); 413} 414 415/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 416/// an invoke, we have to turn all of the calls that can throw into 417/// invokes. This function analyze BB to see if there are any calls, and if so, 418/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 419/// nodes in that block with the values specified in InvokeDestPHIValues. 420/// 421/// Returns true to indicate that the next block should be skipped. 422static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 423 InvokeInliningInfo &Invoke) { 424 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 425 Instruction *I = BBI++; 426 427 // We only need to check for function calls: inlined invoke 428 // instructions require no special handling. 429 CallInst *CI = dyn_cast<CallInst>(I); 430 if (CI == 0) continue; 431 432 // LIBUNWIND: merge selector instructions. 433 if (EHSelectorInst *Inner = dyn_cast<EHSelectorInst>(CI)) { 434 EHSelectorInst *Outer = Invoke.getOuterSelector(); 435 if (!Outer) continue; 436 437 bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); 438 bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); 439 440 // If both selectors contain only cleanups, we don't need to do 441 // anything. TODO: this is really just a very specific instance 442 // of a much more general optimization. 443 if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; 444 445 // Otherwise, we just append the outer selector to the inner selector. 446 SmallVector<Value*, 16> NewSelector; 447 for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) 448 NewSelector.push_back(Inner->getArgOperand(i)); 449 for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) 450 NewSelector.push_back(Outer->getArgOperand(i)); 451 452 CallInst *NewInner = CallInst::Create(Inner->getCalledValue(), 453 NewSelector.begin(), 454 NewSelector.end(), 455 "", 456 Inner); 457 // No need to copy attributes, calling convention, etc. 458 NewInner->takeName(Inner); 459 Inner->replaceAllUsesWith(NewInner); 460 Inner->eraseFromParent(); 461 continue; 462 } 463 464 // If this call cannot unwind, don't convert it to an invoke. 465 if (CI->doesNotThrow()) 466 continue; 467 468 // Convert this function call into an invoke instruction. 469 // First, split the basic block. 470 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 471 472 // Delete the unconditional branch inserted by splitBasicBlock 473 BB->getInstList().pop_back(); 474 475 // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch 476 // directly to the new landing pad. 477 if (Invoke.forwardEHResume(CI, BB)) { 478 // TODO: 'Split' is now unreachable; clean it up. 479 480 // We want to leave the original call intact so that the call 481 // graph and other structures won't get misled. We also have to 482 // avoid processing the next block, or we'll iterate here forever. 483 return true; 484 } 485 486 // Otherwise, create the new invoke instruction. 487 ImmutableCallSite CS(CI); 488 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 489 InvokeInst *II = 490 InvokeInst::Create(CI->getCalledValue(), Split, 491 Invoke.getOuterUnwindDest(), 492 InvokeArgs.begin(), InvokeArgs.end(), 493 CI->getName(), BB); 494 II->setCallingConv(CI->getCallingConv()); 495 II->setAttributes(CI->getAttributes()); 496 497 // Make sure that anything using the call now uses the invoke! This also 498 // updates the CallGraph if present, because it uses a WeakVH. 499 CI->replaceAllUsesWith(II); 500 501 Split->getInstList().pop_front(); // Delete the original call 502 503 // Update any PHI nodes in the exceptional block to indicate that 504 // there is now a new entry in them. 505 Invoke.addIncomingPHIValuesFor(BB); 506 return false; 507 } 508 509 return false; 510} 511 512 513/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 514/// in the body of the inlined function into invokes and turn unwind 515/// instructions into branches to the invoke unwind dest. 516/// 517/// II is the invoke instruction being inlined. FirstNewBlock is the first 518/// block of the inlined code (the last block is the end of the function), 519/// and InlineCodeInfo is information about the code that got inlined. 520static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 521 ClonedCodeInfo &InlinedCodeInfo) { 522 BasicBlock *InvokeDest = II->getUnwindDest(); 523 524 Function *Caller = FirstNewBlock->getParent(); 525 526 // The inlined code is currently at the end of the function, scan from the 527 // start of the inlined code to its end, checking for stuff we need to 528 // rewrite. If the code doesn't have calls or unwinds, we know there is 529 // nothing to rewrite. 530 if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { 531 // Now that everything is happy, we have one final detail. The PHI nodes in 532 // the exception destination block still have entries due to the original 533 // invoke instruction. Eliminate these entries (which might even delete the 534 // PHI node) now. 535 InvokeDest->removePredecessor(II->getParent()); 536 return; 537 } 538 539 InvokeInliningInfo Invoke(II); 540 541 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 542 if (InlinedCodeInfo.ContainsCalls) 543 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 544 // Honor a request to skip the next block. We don't need to 545 // consider UnwindInsts in this case either. 546 ++BB; 547 continue; 548 } 549 550 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 551 // An UnwindInst requires special handling when it gets inlined into an 552 // invoke site. Once this happens, we know that the unwind would cause 553 // a control transfer to the invoke exception destination, so we can 554 // transform it into a direct branch to the exception destination. 555 BranchInst::Create(InvokeDest, UI); 556 557 // Delete the unwind instruction! 558 UI->eraseFromParent(); 559 560 // Update any PHI nodes in the exceptional block to indicate that 561 // there is now a new entry in them. 562 Invoke.addIncomingPHIValuesFor(BB); 563 } 564 } 565 566 // Now that everything is happy, we have one final detail. The PHI nodes in 567 // the exception destination block still have entries due to the original 568 // invoke instruction. Eliminate these entries (which might even delete the 569 // PHI node) now. 570 InvokeDest->removePredecessor(II->getParent()); 571} 572 573/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 574/// into the caller, update the specified callgraph to reflect the changes we 575/// made. Note that it's possible that not all code was copied over, so only 576/// some edges of the callgraph may remain. 577static void UpdateCallGraphAfterInlining(CallSite CS, 578 Function::iterator FirstNewBlock, 579 ValueToValueMapTy &VMap, 580 InlineFunctionInfo &IFI) { 581 CallGraph &CG = *IFI.CG; 582 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 583 const Function *Callee = CS.getCalledFunction(); 584 CallGraphNode *CalleeNode = CG[Callee]; 585 CallGraphNode *CallerNode = CG[Caller]; 586 587 // Since we inlined some uninlined call sites in the callee into the caller, 588 // add edges from the caller to all of the callees of the callee. 589 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 590 591 // Consider the case where CalleeNode == CallerNode. 592 CallGraphNode::CalledFunctionsVector CallCache; 593 if (CalleeNode == CallerNode) { 594 CallCache.assign(I, E); 595 I = CallCache.begin(); 596 E = CallCache.end(); 597 } 598 599 for (; I != E; ++I) { 600 const Value *OrigCall = I->first; 601 602 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 603 // Only copy the edge if the call was inlined! 604 if (VMI == VMap.end() || VMI->second == 0) 605 continue; 606 607 // If the call was inlined, but then constant folded, there is no edge to 608 // add. Check for this case. 609 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 610 if (NewCall == 0) continue; 611 612 // Remember that this call site got inlined for the client of 613 // InlineFunction. 614 IFI.InlinedCalls.push_back(NewCall); 615 616 // It's possible that inlining the callsite will cause it to go from an 617 // indirect to a direct call by resolving a function pointer. If this 618 // happens, set the callee of the new call site to a more precise 619 // destination. This can also happen if the call graph node of the caller 620 // was just unnecessarily imprecise. 621 if (I->second->getFunction() == 0) 622 if (Function *F = CallSite(NewCall).getCalledFunction()) { 623 // Indirect call site resolved to direct call. 624 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 625 626 continue; 627 } 628 629 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 630 } 631 632 // Update the call graph by deleting the edge from Callee to Caller. We must 633 // do this after the loop above in case Caller and Callee are the same. 634 CallerNode->removeCallEdgeFor(CS); 635} 636 637/// HandleByValArgument - When inlining a call site that has a byval argument, 638/// we have to make the implicit memcpy explicit by adding it. 639static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 640 const Function *CalledFunc, 641 InlineFunctionInfo &IFI, 642 unsigned ByValAlignment) { 643 const Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 644 645 // If the called function is readonly, then it could not mutate the caller's 646 // copy of the byval'd memory. In this case, it is safe to elide the copy and 647 // temporary. 648 if (CalledFunc->onlyReadsMemory()) { 649 // If the byval argument has a specified alignment that is greater than the 650 // passed in pointer, then we either have to round up the input pointer or 651 // give up on this transformation. 652 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 653 return Arg; 654 655 // If the pointer is already known to be sufficiently aligned, or if we can 656 // round it up to a larger alignment, then we don't need a temporary. 657 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 658 IFI.TD) >= ByValAlignment) 659 return Arg; 660 661 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 662 // for code quality, but rarely happens and is required for correctness. 663 } 664 665 LLVMContext &Context = Arg->getContext(); 666 667 const Type *VoidPtrTy = Type::getInt8PtrTy(Context); 668 669 // Create the alloca. If we have TargetData, use nice alignment. 670 unsigned Align = 1; 671 if (IFI.TD) 672 Align = IFI.TD->getPrefTypeAlignment(AggTy); 673 674 // If the byval had an alignment specified, we *must* use at least that 675 // alignment, as it is required by the byval argument (and uses of the 676 // pointer inside the callee). 677 Align = std::max(Align, ByValAlignment); 678 679 Function *Caller = TheCall->getParent()->getParent(); 680 681 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 682 &*Caller->begin()->begin()); 683 // Emit a memcpy. 684 const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 685 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 686 Intrinsic::memcpy, 687 Tys, 3); 688 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 689 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 690 691 Value *Size; 692 if (IFI.TD == 0) 693 Size = ConstantExpr::getSizeOf(AggTy); 694 else 695 Size = ConstantInt::get(Type::getInt64Ty(Context), 696 IFI.TD->getTypeStoreSize(AggTy)); 697 698 // Always generate a memcpy of alignment 1 here because we don't know 699 // the alignment of the src pointer. Other optimizations can infer 700 // better alignment. 701 Value *CallArgs[] = { 702 DestCast, SrcCast, Size, 703 ConstantInt::get(Type::getInt32Ty(Context), 1), 704 ConstantInt::getFalse(Context) // isVolatile 705 }; 706 CallInst::Create(MemCpyFn, CallArgs, CallArgs+5, "", TheCall); 707 708 // Uses of the argument in the function should use our new alloca 709 // instead. 710 return NewAlloca; 711} 712 713// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 714// intrinsic. 715static bool isUsedByLifetimeMarker(Value *V) { 716 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 717 ++UI) { 718 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 719 switch (II->getIntrinsicID()) { 720 default: break; 721 case Intrinsic::lifetime_start: 722 case Intrinsic::lifetime_end: 723 return true; 724 } 725 } 726 } 727 return false; 728} 729 730// hasLifetimeMarkers - Check whether the given alloca already has 731// lifetime.start or lifetime.end intrinsics. 732static bool hasLifetimeMarkers(AllocaInst *AI) { 733 const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 734 if (AI->getType() == Int8PtrTy) 735 return isUsedByLifetimeMarker(AI); 736 737 // Do a scan to find all the casts to i8*. 738 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 739 ++I) { 740 if (I->getType() != Int8PtrTy) continue; 741 if (I->stripPointerCasts() != AI) continue; 742 if (isUsedByLifetimeMarker(*I)) 743 return true; 744 } 745 return false; 746} 747 748// InlineFunction - This function inlines the called function into the basic 749// block of the caller. This returns false if it is not possible to inline this 750// call. The program is still in a well defined state if this occurs though. 751// 752// Note that this only does one level of inlining. For example, if the 753// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 754// exists in the instruction stream. Similarly this will inline a recursive 755// function by one level. 756// 757bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { 758 Instruction *TheCall = CS.getInstruction(); 759 LLVMContext &Context = TheCall->getContext(); 760 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 761 "Instruction not in function!"); 762 763 // If IFI has any state in it, zap it before we fill it in. 764 IFI.reset(); 765 766 const Function *CalledFunc = CS.getCalledFunction(); 767 if (CalledFunc == 0 || // Can't inline external function or indirect 768 CalledFunc->isDeclaration() || // call, or call to a vararg function! 769 CalledFunc->getFunctionType()->isVarArg()) return false; 770 771 // If the call to the callee is not a tail call, we must clear the 'tail' 772 // flags on any calls that we inline. 773 bool MustClearTailCallFlags = 774 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 775 776 // If the call to the callee cannot throw, set the 'nounwind' flag on any 777 // calls that we inline. 778 bool MarkNoUnwind = CS.doesNotThrow(); 779 780 BasicBlock *OrigBB = TheCall->getParent(); 781 Function *Caller = OrigBB->getParent(); 782 783 // GC poses two hazards to inlining, which only occur when the callee has GC: 784 // 1. If the caller has no GC, then the callee's GC must be propagated to the 785 // caller. 786 // 2. If the caller has a differing GC, it is invalid to inline. 787 if (CalledFunc->hasGC()) { 788 if (!Caller->hasGC()) 789 Caller->setGC(CalledFunc->getGC()); 790 else if (CalledFunc->getGC() != Caller->getGC()) 791 return false; 792 } 793 794 // Get an iterator to the last basic block in the function, which will have 795 // the new function inlined after it. 796 // 797 Function::iterator LastBlock = &Caller->back(); 798 799 // Make sure to capture all of the return instructions from the cloned 800 // function. 801 SmallVector<ReturnInst*, 8> Returns; 802 ClonedCodeInfo InlinedFunctionInfo; 803 Function::iterator FirstNewBlock; 804 805 { // Scope to destroy VMap after cloning. 806 ValueToValueMapTy VMap; 807 808 assert(CalledFunc->arg_size() == CS.arg_size() && 809 "No varargs calls can be inlined!"); 810 811 // Calculate the vector of arguments to pass into the function cloner, which 812 // matches up the formal to the actual argument values. 813 CallSite::arg_iterator AI = CS.arg_begin(); 814 unsigned ArgNo = 0; 815 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 816 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 817 Value *ActualArg = *AI; 818 819 // When byval arguments actually inlined, we need to make the copy implied 820 // by them explicit. However, we don't do this if the callee is readonly 821 // or readnone, because the copy would be unneeded: the callee doesn't 822 // modify the struct. 823 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { 824 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 825 CalledFunc->getParamAlignment(ArgNo+1)); 826 827 // Calls that we inline may use the new alloca, so we need to clear 828 // their 'tail' flags if HandleByValArgument introduced a new alloca and 829 // the callee has calls. 830 MustClearTailCallFlags |= ActualArg != *AI; 831 } 832 833 VMap[I] = ActualArg; 834 } 835 836 // We want the inliner to prune the code as it copies. We would LOVE to 837 // have no dead or constant instructions leftover after inlining occurs 838 // (which can happen, e.g., because an argument was constant), but we'll be 839 // happy with whatever the cloner can do. 840 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 841 /*ModuleLevelChanges=*/false, Returns, ".i", 842 &InlinedFunctionInfo, IFI.TD, TheCall); 843 844 // Remember the first block that is newly cloned over. 845 FirstNewBlock = LastBlock; ++FirstNewBlock; 846 847 // Update the callgraph if requested. 848 if (IFI.CG) 849 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 850 } 851 852 // If there are any alloca instructions in the block that used to be the entry 853 // block for the callee, move them to the entry block of the caller. First 854 // calculate which instruction they should be inserted before. We insert the 855 // instructions at the end of the current alloca list. 856 // 857 { 858 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 859 for (BasicBlock::iterator I = FirstNewBlock->begin(), 860 E = FirstNewBlock->end(); I != E; ) { 861 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 862 if (AI == 0) continue; 863 864 // If the alloca is now dead, remove it. This often occurs due to code 865 // specialization. 866 if (AI->use_empty()) { 867 AI->eraseFromParent(); 868 continue; 869 } 870 871 if (!isa<Constant>(AI->getArraySize())) 872 continue; 873 874 // Keep track of the static allocas that we inline into the caller. 875 IFI.StaticAllocas.push_back(AI); 876 877 // Scan for the block of allocas that we can move over, and move them 878 // all at once. 879 while (isa<AllocaInst>(I) && 880 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 881 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 882 ++I; 883 } 884 885 // Transfer all of the allocas over in a block. Using splice means 886 // that the instructions aren't removed from the symbol table, then 887 // reinserted. 888 Caller->getEntryBlock().getInstList().splice(InsertPoint, 889 FirstNewBlock->getInstList(), 890 AI, I); 891 } 892 } 893 894 // Leave lifetime markers for the static alloca's, scoping them to the 895 // function we just inlined. 896 if (!IFI.StaticAllocas.empty()) { 897 IRBuilder<> builder(FirstNewBlock->begin()); 898 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 899 AllocaInst *AI = IFI.StaticAllocas[ai]; 900 901 // If the alloca is already scoped to something smaller than the whole 902 // function then there's no need to add redundant, less accurate markers. 903 if (hasLifetimeMarkers(AI)) 904 continue; 905 906 builder.CreateLifetimeStart(AI); 907 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 908 IRBuilder<> builder(Returns[ri]); 909 builder.CreateLifetimeEnd(AI); 910 } 911 } 912 } 913 914 // If the inlined code contained dynamic alloca instructions, wrap the inlined 915 // code with llvm.stacksave/llvm.stackrestore intrinsics. 916 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 917 Module *M = Caller->getParent(); 918 // Get the two intrinsics we care about. 919 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 920 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 921 922 // Insert the llvm.stacksave. 923 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", 924 FirstNewBlock->begin()); 925 926 // Insert a call to llvm.stackrestore before any return instructions in the 927 // inlined function. 928 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 929 CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); 930 } 931 932 // Count the number of StackRestore calls we insert. 933 unsigned NumStackRestores = Returns.size(); 934 935 // If we are inlining an invoke instruction, insert restores before each 936 // unwind. These unwinds will be rewritten into branches later. 937 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) { 938 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 939 BB != E; ++BB) 940 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { 941 CallInst::Create(StackRestore, SavedPtr, "", UI); 942 ++NumStackRestores; 943 } 944 } 945 } 946 947 // If we are inlining tail call instruction through a call site that isn't 948 // marked 'tail', we must remove the tail marker for any calls in the inlined 949 // code. Also, calls inlined through a 'nounwind' call site should be marked 950 // 'nounwind'. 951 if (InlinedFunctionInfo.ContainsCalls && 952 (MustClearTailCallFlags || MarkNoUnwind)) { 953 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 954 BB != E; ++BB) 955 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 956 if (CallInst *CI = dyn_cast<CallInst>(I)) { 957 if (MustClearTailCallFlags) 958 CI->setTailCall(false); 959 if (MarkNoUnwind) 960 CI->setDoesNotThrow(); 961 } 962 } 963 964 // If we are inlining through a 'nounwind' call site then any inlined 'unwind' 965 // instructions are unreachable. 966 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind) 967 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 968 BB != E; ++BB) { 969 TerminatorInst *Term = BB->getTerminator(); 970 if (isa<UnwindInst>(Term)) { 971 new UnreachableInst(Context, Term); 972 BB->getInstList().erase(Term); 973 } 974 } 975 976 // If we are inlining for an invoke instruction, we must make sure to rewrite 977 // any inlined 'unwind' instructions into branches to the invoke exception 978 // destination, and call instructions into invoke instructions. 979 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 980 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 981 982 // If we cloned in _exactly one_ basic block, and if that block ends in a 983 // return instruction, we splice the body of the inlined callee directly into 984 // the calling basic block. 985 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 986 // Move all of the instructions right before the call. 987 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 988 FirstNewBlock->begin(), FirstNewBlock->end()); 989 // Remove the cloned basic block. 990 Caller->getBasicBlockList().pop_back(); 991 992 // If the call site was an invoke instruction, add a branch to the normal 993 // destination. 994 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 995 BranchInst::Create(II->getNormalDest(), TheCall); 996 997 // If the return instruction returned a value, replace uses of the call with 998 // uses of the returned value. 999 if (!TheCall->use_empty()) { 1000 ReturnInst *R = Returns[0]; 1001 if (TheCall == R->getReturnValue()) 1002 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1003 else 1004 TheCall->replaceAllUsesWith(R->getReturnValue()); 1005 } 1006 // Since we are now done with the Call/Invoke, we can delete it. 1007 TheCall->eraseFromParent(); 1008 1009 // Since we are now done with the return instruction, delete it also. 1010 Returns[0]->eraseFromParent(); 1011 1012 // We are now done with the inlining. 1013 return true; 1014 } 1015 1016 // Otherwise, we have the normal case, of more than one block to inline or 1017 // multiple return sites. 1018 1019 // We want to clone the entire callee function into the hole between the 1020 // "starter" and "ender" blocks. How we accomplish this depends on whether 1021 // this is an invoke instruction or a call instruction. 1022 BasicBlock *AfterCallBB; 1023 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 1024 1025 // Add an unconditional branch to make this look like the CallInst case... 1026 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 1027 1028 // Split the basic block. This guarantees that no PHI nodes will have to be 1029 // updated due to new incoming edges, and make the invoke case more 1030 // symmetric to the call case. 1031 AfterCallBB = OrigBB->splitBasicBlock(NewBr, 1032 CalledFunc->getName()+".exit"); 1033 1034 } else { // It's a call 1035 // If this is a call instruction, we need to split the basic block that 1036 // the call lives in. 1037 // 1038 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 1039 CalledFunc->getName()+".exit"); 1040 } 1041 1042 // Change the branch that used to go to AfterCallBB to branch to the first 1043 // basic block of the inlined function. 1044 // 1045 TerminatorInst *Br = OrigBB->getTerminator(); 1046 assert(Br && Br->getOpcode() == Instruction::Br && 1047 "splitBasicBlock broken!"); 1048 Br->setOperand(0, FirstNewBlock); 1049 1050 1051 // Now that the function is correct, make it a little bit nicer. In 1052 // particular, move the basic blocks inserted from the end of the function 1053 // into the space made by splitting the source basic block. 1054 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 1055 FirstNewBlock, Caller->end()); 1056 1057 // Handle all of the return instructions that we just cloned in, and eliminate 1058 // any users of the original call/invoke instruction. 1059 const Type *RTy = CalledFunc->getReturnType(); 1060 1061 PHINode *PHI = 0; 1062 if (Returns.size() > 1) { 1063 // The PHI node should go at the front of the new basic block to merge all 1064 // possible incoming values. 1065 if (!TheCall->use_empty()) { 1066 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 1067 AfterCallBB->begin()); 1068 // Anything that used the result of the function call should now use the 1069 // PHI node as their operand. 1070 TheCall->replaceAllUsesWith(PHI); 1071 } 1072 1073 // Loop over all of the return instructions adding entries to the PHI node 1074 // as appropriate. 1075 if (PHI) { 1076 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1077 ReturnInst *RI = Returns[i]; 1078 assert(RI->getReturnValue()->getType() == PHI->getType() && 1079 "Ret value not consistent in function!"); 1080 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 1081 } 1082 } 1083 1084 1085 // Add a branch to the merge points and remove return instructions. 1086 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 1087 ReturnInst *RI = Returns[i]; 1088 BranchInst::Create(AfterCallBB, RI); 1089 RI->eraseFromParent(); 1090 } 1091 } else if (!Returns.empty()) { 1092 // Otherwise, if there is exactly one return value, just replace anything 1093 // using the return value of the call with the computed value. 1094 if (!TheCall->use_empty()) { 1095 if (TheCall == Returns[0]->getReturnValue()) 1096 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1097 else 1098 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 1099 } 1100 1101 // Splice the code from the return block into the block that it will return 1102 // to, which contains the code that was after the call. 1103 BasicBlock *ReturnBB = Returns[0]->getParent(); 1104 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 1105 ReturnBB->getInstList()); 1106 1107 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 1108 ReturnBB->replaceAllUsesWith(AfterCallBB); 1109 1110 // Delete the return instruction now and empty ReturnBB now. 1111 Returns[0]->eraseFromParent(); 1112 ReturnBB->eraseFromParent(); 1113 } else if (!TheCall->use_empty()) { 1114 // No returns, but something is using the return value of the call. Just 1115 // nuke the result. 1116 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 1117 } 1118 1119 // Since we are now done with the Call/Invoke, we can delete it. 1120 TheCall->eraseFromParent(); 1121 1122 // We should always be able to fold the entry block of the function into the 1123 // single predecessor of the block... 1124 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 1125 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 1126 1127 // Splice the code entry block into calling block, right before the 1128 // unconditional branch. 1129 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 1130 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 1131 1132 // Remove the unconditional branch. 1133 OrigBB->getInstList().erase(Br); 1134 1135 // Now we can remove the CalleeEntry block, which is now empty. 1136 Caller->getBasicBlockList().erase(CalleeEntry); 1137 1138 // If we inserted a phi node, check to see if it has a single value (e.g. all 1139 // the entries are the same or undef). If so, remove the PHI so it doesn't 1140 // block other optimizations. 1141 if (PHI) 1142 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 1143 PHI->replaceAllUsesWith(V); 1144 PHI->eraseFromParent(); 1145 } 1146 1147 return true; 1148} 1149