1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This contains code to emit Stmt nodes as LLVM code.
12 //===----------------------------------------------------------------------===//
14 #include "CodeGenFunction.h"
15 #include "CGDebugInfo.h"
16 #include "CodeGenModule.h"
17 #include "TargetInfo.h"
18 #include "clang/AST/StmtVisitor.h"
19 #include "clang/Basic/Builtins.h"
20 #include "clang/Basic/PrettyStackTrace.h"
21 #include "clang/Basic/TargetInfo.h"
22 #include "clang/Sema/LoopHint.h"
23 #include "clang/Sema/SemaDiagnostic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/InlineAsm.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/MDBuilder.h"
31 using namespace clang;
32 using namespace CodeGen;
34 //===----------------------------------------------------------------------===//
36 //===----------------------------------------------------------------------===//
38 void CodeGenFunction::EmitStopPoint(const Stmt *S) {
39 if (CGDebugInfo *DI = getDebugInfo()) {
41 Loc = S->getLocStart();
42 DI->EmitLocation(Builder, Loc);
48 void CodeGenFunction::EmitStmt(const Stmt *S) {
49 assert(S && "Null statement?");
50 PGO.setCurrentStmt(S);
52 // These statements have their own debug info handling.
53 if (EmitSimpleStmt(S))
56 // Check if we are generating unreachable code.
57 if (!HaveInsertPoint()) {
58 // If so, and the statement doesn't contain a label, then we do not need to
59 // generate actual code. This is safe because (1) the current point is
60 // unreachable, so we don't need to execute the code, and (2) we've already
61 // handled the statements which update internal data structures (like the
62 // local variable map) which could be used by subsequent statements.
63 if (!ContainsLabel(S)) {
64 // Verify that any decl statements were handled as simple, they may be in
65 // scope of subsequent reachable statements.
66 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!");
70 // Otherwise, make a new block to hold the code.
74 // Generate a stoppoint if we are emitting debug info.
77 switch (S->getStmtClass()) {
78 case Stmt::NoStmtClass:
79 case Stmt::CXXCatchStmtClass:
80 case Stmt::SEHExceptStmtClass:
81 case Stmt::SEHFinallyStmtClass:
82 case Stmt::MSDependentExistsStmtClass:
83 llvm_unreachable("invalid statement class to emit generically");
84 case Stmt::NullStmtClass:
85 case Stmt::CompoundStmtClass:
86 case Stmt::DeclStmtClass:
87 case Stmt::LabelStmtClass:
88 case Stmt::AttributedStmtClass:
89 case Stmt::GotoStmtClass:
90 case Stmt::BreakStmtClass:
91 case Stmt::ContinueStmtClass:
92 case Stmt::DefaultStmtClass:
93 case Stmt::CaseStmtClass:
94 case Stmt::SEHLeaveStmtClass:
95 llvm_unreachable("should have emitted these statements as simple");
97 #define STMT(Type, Base)
98 #define ABSTRACT_STMT(Op)
99 #define EXPR(Type, Base) \
100 case Stmt::Type##Class:
101 #include "clang/AST/StmtNodes.inc"
103 // Remember the block we came in on.
104 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
105 assert(incoming && "expression emission must have an insertion point");
107 EmitIgnoredExpr(cast<Expr>(S));
109 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
110 assert(outgoing && "expression emission cleared block!");
112 // The expression emitters assume (reasonably!) that the insertion
113 // point is always set. To maintain that, the call-emission code
114 // for noreturn functions has to enter a new block with no
115 // predecessors. We want to kill that block and mark the current
116 // insertion point unreachable in the common case of a call like
117 // "exit();". Since expression emission doesn't otherwise create
118 // blocks with no predecessors, we can just test for that.
119 // However, we must be careful not to do this to our incoming
120 // block, because *statement* emission does sometimes create
121 // reachable blocks which will have no predecessors until later in
122 // the function. This occurs with, e.g., labels that are not
123 // reachable by fallthrough.
124 if (incoming != outgoing && outgoing->use_empty()) {
125 outgoing->eraseFromParent();
126 Builder.ClearInsertionPoint();
131 case Stmt::IndirectGotoStmtClass:
132 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break;
134 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break;
135 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S)); break;
136 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S)); break;
137 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S)); break;
139 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break;
141 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break;
142 case Stmt::GCCAsmStmtClass: // Intentional fall-through.
143 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break;
144 case Stmt::CoroutineBodyStmtClass:
145 case Stmt::CoreturnStmtClass:
146 CGM.ErrorUnsupported(S, "coroutine");
148 case Stmt::CapturedStmtClass: {
149 const CapturedStmt *CS = cast<CapturedStmt>(S);
150 EmitCapturedStmt(*CS, CS->getCapturedRegionKind());
153 case Stmt::ObjCAtTryStmtClass:
154 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S));
156 case Stmt::ObjCAtCatchStmtClass:
158 "@catch statements should be handled by EmitObjCAtTryStmt");
159 case Stmt::ObjCAtFinallyStmtClass:
161 "@finally statements should be handled by EmitObjCAtTryStmt");
162 case Stmt::ObjCAtThrowStmtClass:
163 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S));
165 case Stmt::ObjCAtSynchronizedStmtClass:
166 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S));
168 case Stmt::ObjCForCollectionStmtClass:
169 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S));
171 case Stmt::ObjCAutoreleasePoolStmtClass:
172 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S));
175 case Stmt::CXXTryStmtClass:
176 EmitCXXTryStmt(cast<CXXTryStmt>(*S));
178 case Stmt::CXXForRangeStmtClass:
179 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S));
181 case Stmt::SEHTryStmtClass:
182 EmitSEHTryStmt(cast<SEHTryStmt>(*S));
184 case Stmt::OMPParallelDirectiveClass:
185 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S));
187 case Stmt::OMPSimdDirectiveClass:
188 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S));
190 case Stmt::OMPForDirectiveClass:
191 EmitOMPForDirective(cast<OMPForDirective>(*S));
193 case Stmt::OMPForSimdDirectiveClass:
194 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S));
196 case Stmt::OMPSectionsDirectiveClass:
197 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S));
199 case Stmt::OMPSectionDirectiveClass:
200 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S));
202 case Stmt::OMPSingleDirectiveClass:
203 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S));
205 case Stmt::OMPMasterDirectiveClass:
206 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S));
208 case Stmt::OMPCriticalDirectiveClass:
209 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S));
211 case Stmt::OMPParallelForDirectiveClass:
212 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S));
214 case Stmt::OMPParallelForSimdDirectiveClass:
215 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S));
217 case Stmt::OMPParallelSectionsDirectiveClass:
218 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S));
220 case Stmt::OMPTaskDirectiveClass:
221 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S));
223 case Stmt::OMPTaskyieldDirectiveClass:
224 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
226 case Stmt::OMPBarrierDirectiveClass:
227 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
229 case Stmt::OMPTaskwaitDirectiveClass:
230 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S));
232 case Stmt::OMPTaskgroupDirectiveClass:
233 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S));
235 case Stmt::OMPFlushDirectiveClass:
236 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S));
238 case Stmt::OMPOrderedDirectiveClass:
239 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S));
241 case Stmt::OMPAtomicDirectiveClass:
242 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S));
244 case Stmt::OMPTargetDirectiveClass:
245 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S));
247 case Stmt::OMPTeamsDirectiveClass:
248 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S));
250 case Stmt::OMPCancellationPointDirectiveClass:
251 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S));
253 case Stmt::OMPCancelDirectiveClass:
254 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S));
256 case Stmt::OMPTargetDataDirectiveClass:
257 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S));
259 case Stmt::OMPTargetEnterDataDirectiveClass:
260 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S));
262 case Stmt::OMPTargetExitDataDirectiveClass:
263 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S));
265 case Stmt::OMPTargetParallelDirectiveClass:
266 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S));
268 case Stmt::OMPTargetParallelForDirectiveClass:
269 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S));
271 case Stmt::OMPTaskLoopDirectiveClass:
272 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S));
274 case Stmt::OMPTaskLoopSimdDirectiveClass:
275 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S));
277 case Stmt::OMPDistributeDirectiveClass:
278 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S));
280 case Stmt::OMPTargetUpdateDirectiveClass:
281 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S));
283 case Stmt::OMPDistributeParallelForDirectiveClass:
284 EmitOMPDistributeParallelForDirective(
285 cast<OMPDistributeParallelForDirective>(*S));
287 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
288 EmitOMPDistributeParallelForSimdDirective(
289 cast<OMPDistributeParallelForSimdDirective>(*S));
291 case Stmt::OMPDistributeSimdDirectiveClass:
292 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S));
294 case Stmt::OMPTargetParallelForSimdDirectiveClass:
295 EmitOMPTargetParallelForSimdDirective(
296 cast<OMPTargetParallelForSimdDirective>(*S));
301 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) {
302 switch (S->getStmtClass()) {
303 default: return false;
304 case Stmt::NullStmtClass: break;
305 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break;
306 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break;
307 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break;
308 case Stmt::AttributedStmtClass:
309 EmitAttributedStmt(cast<AttributedStmt>(*S)); break;
310 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break;
311 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break;
312 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break;
313 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break;
314 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break;
315 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break;
321 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
322 /// this captures the expression result of the last sub-statement and returns it
323 /// (for use by the statement expression extension).
324 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
325 AggValueSlot AggSlot) {
326 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
327 "LLVM IR generation of compound statement ('{}')");
329 // Keep track of the current cleanup stack depth, including debug scopes.
330 LexicalScope Scope(*this, S.getSourceRange());
332 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
336 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
338 AggValueSlot AggSlot) {
340 for (CompoundStmt::const_body_iterator I = S.body_begin(),
341 E = S.body_end()-GetLast; I != E; ++I)
344 Address RetAlloca = Address::invalid();
346 // We have to special case labels here. They are statements, but when put
347 // at the end of a statement expression, they yield the value of their
348 // subexpression. Handle this by walking through all labels we encounter,
349 // emitting them before we evaluate the subexpr.
350 const Stmt *LastStmt = S.body_back();
351 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) {
352 EmitLabel(LS->getDecl());
353 LastStmt = LS->getSubStmt();
358 QualType ExprTy = cast<Expr>(LastStmt)->getType();
359 if (hasAggregateEvaluationKind(ExprTy)) {
360 EmitAggExpr(cast<Expr>(LastStmt), AggSlot);
362 // We can't return an RValue here because there might be cleanups at
363 // the end of the StmtExpr. Because of that, we have to emit the result
364 // here into a temporary alloca.
365 RetAlloca = CreateMemTemp(ExprTy);
366 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(),
375 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
376 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator());
378 // If there is a cleanup stack, then we it isn't worth trying to
379 // simplify this block (we would need to remove it from the scope map
380 // and cleanup entry).
381 if (!EHStack.empty())
384 // Can only simplify direct branches.
385 if (!BI || !BI->isUnconditional())
388 // Can only simplify empty blocks.
389 if (BI->getIterator() != BB->begin())
392 BB->replaceAllUsesWith(BI->getSuccessor(0));
393 BI->eraseFromParent();
394 BB->eraseFromParent();
397 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
398 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
400 // Fall out of the current block (if necessary).
403 if (IsFinished && BB->use_empty()) {
408 // Place the block after the current block, if possible, or else at
409 // the end of the function.
410 if (CurBB && CurBB->getParent())
411 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB);
413 CurFn->getBasicBlockList().push_back(BB);
414 Builder.SetInsertPoint(BB);
417 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
418 // Emit a branch from the current block to the target one if this
419 // was a real block. If this was just a fall-through block after a
420 // terminator, don't emit it.
421 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
423 if (!CurBB || CurBB->getTerminator()) {
424 // If there is no insert point or the previous block is already
425 // terminated, don't touch it.
427 // Otherwise, create a fall-through branch.
428 Builder.CreateBr(Target);
431 Builder.ClearInsertionPoint();
434 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
435 bool inserted = false;
436 for (llvm::User *u : block->users()) {
437 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) {
438 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(),
446 CurFn->getBasicBlockList().push_back(block);
448 Builder.SetInsertPoint(block);
451 CodeGenFunction::JumpDest
452 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
453 JumpDest &Dest = LabelMap[D];
454 if (Dest.isValid()) return Dest;
456 // Create, but don't insert, the new block.
457 Dest = JumpDest(createBasicBlock(D->getName()),
458 EHScopeStack::stable_iterator::invalid(),
459 NextCleanupDestIndex++);
463 void CodeGenFunction::EmitLabel(const LabelDecl *D) {
464 // Add this label to the current lexical scope if we're within any
465 // normal cleanups. Jumps "in" to this label --- when permitted by
466 // the language --- may need to be routed around such cleanups.
467 if (EHStack.hasNormalCleanups() && CurLexicalScope)
468 CurLexicalScope->addLabel(D);
470 JumpDest &Dest = LabelMap[D];
472 // If we didn't need a forward reference to this label, just go
473 // ahead and create a destination at the current scope.
474 if (!Dest.isValid()) {
475 Dest = getJumpDestInCurrentScope(D->getName());
477 // Otherwise, we need to give this label a target depth and remove
478 // it from the branch-fixups list.
480 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
481 Dest.setScopeDepth(EHStack.stable_begin());
482 ResolveBranchFixups(Dest.getBlock());
485 EmitBlock(Dest.getBlock());
486 incrementProfileCounter(D->getStmt());
489 /// Change the cleanup scope of the labels in this lexical scope to
490 /// match the scope of the enclosing context.
491 void CodeGenFunction::LexicalScope::rescopeLabels() {
492 assert(!Labels.empty());
493 EHScopeStack::stable_iterator innermostScope
494 = CGF.EHStack.getInnermostNormalCleanup();
496 // Change the scope depth of all the labels.
497 for (SmallVectorImpl<const LabelDecl*>::const_iterator
498 i = Labels.begin(), e = Labels.end(); i != e; ++i) {
499 assert(CGF.LabelMap.count(*i));
500 JumpDest &dest = CGF.LabelMap.find(*i)->second;
501 assert(dest.getScopeDepth().isValid());
502 assert(innermostScope.encloses(dest.getScopeDepth()));
503 dest.setScopeDepth(innermostScope);
506 // Reparent the labels if the new scope also has cleanups.
507 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
508 ParentScope->Labels.append(Labels.begin(), Labels.end());
513 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
514 EmitLabel(S.getDecl());
515 EmitStmt(S.getSubStmt());
518 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
519 const Stmt *SubStmt = S.getSubStmt();
520 switch (SubStmt->getStmtClass()) {
521 case Stmt::DoStmtClass:
522 EmitDoStmt(cast<DoStmt>(*SubStmt), S.getAttrs());
524 case Stmt::ForStmtClass:
525 EmitForStmt(cast<ForStmt>(*SubStmt), S.getAttrs());
527 case Stmt::WhileStmtClass:
528 EmitWhileStmt(cast<WhileStmt>(*SubStmt), S.getAttrs());
530 case Stmt::CXXForRangeStmtClass:
531 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*SubStmt), S.getAttrs());
538 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
539 // If this code is reachable then emit a stop point (if generating
540 // debug info). We have to do this ourselves because we are on the
541 // "simple" statement path.
542 if (HaveInsertPoint())
545 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel()));
549 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
550 if (const LabelDecl *Target = S.getConstantTarget()) {
551 EmitBranchThroughCleanup(getJumpDestForLabel(Target));
555 // Ensure that we have an i8* for our PHI node.
556 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()),
558 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
560 // Get the basic block for the indirect goto.
561 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
563 // The first instruction in the block has to be the PHI for the switch dest,
564 // add an entry for this branch.
565 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB);
567 EmitBranch(IndGotoBB);
570 void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
571 // C99 6.8.4.1: The first substatement is executed if the expression compares
572 // unequal to 0. The condition must be a scalar type.
573 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
576 EmitStmt(S.getInit());
578 if (S.getConditionVariable())
579 EmitAutoVarDecl(*S.getConditionVariable());
581 // If the condition constant folds and can be elided, try to avoid emitting
582 // the condition and the dead arm of the if/else.
584 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant,
586 // Figure out which block (then or else) is executed.
587 const Stmt *Executed = S.getThen();
588 const Stmt *Skipped = S.getElse();
589 if (!CondConstant) // Condition false?
590 std::swap(Executed, Skipped);
592 // If the skipped block has no labels in it, just emit the executed block.
593 // This avoids emitting dead code and simplifies the CFG substantially.
594 if (S.isConstexpr() || !ContainsLabel(Skipped)) {
596 incrementProfileCounter(&S);
598 RunCleanupsScope ExecutedScope(*this);
605 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
606 // the conditional branch.
607 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then");
608 llvm::BasicBlock *ContBlock = createBasicBlock("if.end");
609 llvm::BasicBlock *ElseBlock = ContBlock;
611 ElseBlock = createBasicBlock("if.else");
613 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock,
614 getProfileCount(S.getThen()));
616 // Emit the 'then' code.
617 EmitBlock(ThenBlock);
618 incrementProfileCounter(&S);
620 RunCleanupsScope ThenScope(*this);
621 EmitStmt(S.getThen());
623 EmitBranch(ContBlock);
625 // Emit the 'else' code if present.
626 if (const Stmt *Else = S.getElse()) {
628 // There is no need to emit line number for an unconditional branch.
629 auto NL = ApplyDebugLocation::CreateEmpty(*this);
630 EmitBlock(ElseBlock);
633 RunCleanupsScope ElseScope(*this);
637 // There is no need to emit line number for an unconditional branch.
638 auto NL = ApplyDebugLocation::CreateEmpty(*this);
639 EmitBranch(ContBlock);
643 // Emit the continuation block for code after the if.
644 EmitBlock(ContBlock, true);
647 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
648 ArrayRef<const Attr *> WhileAttrs) {
649 // Emit the header for the loop, which will also become
650 // the continue target.
651 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond");
652 EmitBlock(LoopHeader.getBlock());
654 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs,
655 Builder.getCurrentDebugLocation());
657 // Create an exit block for when the condition fails, which will
658 // also become the break target.
659 JumpDest LoopExit = getJumpDestInCurrentScope("while.end");
661 // Store the blocks to use for break and continue.
662 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader));
664 // C++ [stmt.while]p2:
665 // When the condition of a while statement is a declaration, the
666 // scope of the variable that is declared extends from its point
667 // of declaration (3.3.2) to the end of the while statement.
669 // The object created in a condition is destroyed and created
670 // with each iteration of the loop.
671 RunCleanupsScope ConditionScope(*this);
673 if (S.getConditionVariable())
674 EmitAutoVarDecl(*S.getConditionVariable());
676 // Evaluate the conditional in the while header. C99 6.8.5.1: The
677 // evaluation of the controlling expression takes place before each
678 // execution of the loop body.
679 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
681 // while(1) is common, avoid extra exit blocks. Be sure
682 // to correctly handle break/continue though.
683 bool EmitBoolCondBranch = true;
684 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
686 EmitBoolCondBranch = false;
688 // As long as the condition is true, go to the loop body.
689 llvm::BasicBlock *LoopBody = createBasicBlock("while.body");
690 if (EmitBoolCondBranch) {
691 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
692 if (ConditionScope.requiresCleanups())
693 ExitBlock = createBasicBlock("while.exit");
694 Builder.CreateCondBr(
695 BoolCondVal, LoopBody, ExitBlock,
696 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
698 if (ExitBlock != LoopExit.getBlock()) {
699 EmitBlock(ExitBlock);
700 EmitBranchThroughCleanup(LoopExit);
704 // Emit the loop body. We have to emit this in a cleanup scope
705 // because it might be a singleton DeclStmt.
707 RunCleanupsScope BodyScope(*this);
709 incrementProfileCounter(&S);
710 EmitStmt(S.getBody());
713 BreakContinueStack.pop_back();
715 // Immediately force cleanup.
716 ConditionScope.ForceCleanup();
719 // Branch to the loop header again.
720 EmitBranch(LoopHeader.getBlock());
724 // Emit the exit block.
725 EmitBlock(LoopExit.getBlock(), true);
727 // The LoopHeader typically is just a branch if we skipped emitting
728 // a branch, try to erase it.
729 if (!EmitBoolCondBranch)
730 SimplifyForwardingBlocks(LoopHeader.getBlock());
733 void CodeGenFunction::EmitDoStmt(const DoStmt &S,
734 ArrayRef<const Attr *> DoAttrs) {
735 JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
736 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
738 uint64_t ParentCount = getCurrentProfileCount();
740 // Store the blocks to use for break and continue.
741 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond));
743 // Emit the body of the loop.
744 llvm::BasicBlock *LoopBody = createBasicBlock("do.body");
746 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs,
747 Builder.getCurrentDebugLocation());
749 EmitBlockWithFallThrough(LoopBody, &S);
751 RunCleanupsScope BodyScope(*this);
752 EmitStmt(S.getBody());
755 EmitBlock(LoopCond.getBlock());
757 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
758 // after each execution of the loop body."
760 // Evaluate the conditional in the while header.
761 // C99 6.8.5p2/p4: The first substatement is executed if the expression
762 // compares unequal to 0. The condition must be a scalar type.
763 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
765 BreakContinueStack.pop_back();
767 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
768 // to correctly handle break/continue though.
769 bool EmitBoolCondBranch = true;
770 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
772 EmitBoolCondBranch = false;
774 // As long as the condition is true, iterate the loop.
775 if (EmitBoolCondBranch) {
776 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
777 Builder.CreateCondBr(
778 BoolCondVal, LoopBody, LoopExit.getBlock(),
779 createProfileWeightsForLoop(S.getCond(), BackedgeCount));
784 // Emit the exit block.
785 EmitBlock(LoopExit.getBlock());
787 // The DoCond block typically is just a branch if we skipped
788 // emitting a branch, try to erase it.
789 if (!EmitBoolCondBranch)
790 SimplifyForwardingBlocks(LoopCond.getBlock());
793 void CodeGenFunction::EmitForStmt(const ForStmt &S,
794 ArrayRef<const Attr *> ForAttrs) {
795 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
797 LexicalScope ForScope(*this, S.getSourceRange());
799 llvm::DebugLoc DL = Builder.getCurrentDebugLocation();
801 // Evaluate the first part before the loop.
803 EmitStmt(S.getInit());
805 // Start the loop with a block that tests the condition.
806 // If there's an increment, the continue scope will be overwritten
808 JumpDest Continue = getJumpDestInCurrentScope("for.cond");
809 llvm::BasicBlock *CondBlock = Continue.getBlock();
810 EmitBlock(CondBlock);
812 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, DL);
814 // If the for loop doesn't have an increment we can just use the
815 // condition as the continue block. Otherwise we'll need to create
816 // a block for it (in the current scope, i.e. in the scope of the
817 // condition), and that we will become our continue block.
819 Continue = getJumpDestInCurrentScope("for.inc");
821 // Store the blocks to use for break and continue.
822 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
824 // Create a cleanup scope for the condition variable cleanups.
825 LexicalScope ConditionScope(*this, S.getSourceRange());
828 // If the for statement has a condition scope, emit the local variable
830 if (S.getConditionVariable()) {
831 EmitAutoVarDecl(*S.getConditionVariable());
834 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
835 // If there are any cleanups between here and the loop-exit scope,
836 // create a block to stage a loop exit along.
837 if (ForScope.requiresCleanups())
838 ExitBlock = createBasicBlock("for.cond.cleanup");
840 // As long as the condition is true, iterate the loop.
841 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
843 // C99 6.8.5p2/p4: The first substatement is executed if the expression
844 // compares unequal to 0. The condition must be a scalar type.
845 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
846 Builder.CreateCondBr(
847 BoolCondVal, ForBody, ExitBlock,
848 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
850 if (ExitBlock != LoopExit.getBlock()) {
851 EmitBlock(ExitBlock);
852 EmitBranchThroughCleanup(LoopExit);
857 // Treat it as a non-zero constant. Don't even create a new block for the
858 // body, just fall into it.
860 incrementProfileCounter(&S);
863 // Create a separate cleanup scope for the body, in case it is not
864 // a compound statement.
865 RunCleanupsScope BodyScope(*this);
866 EmitStmt(S.getBody());
869 // If there is an increment, emit it next.
871 EmitBlock(Continue.getBlock());
872 EmitStmt(S.getInc());
875 BreakContinueStack.pop_back();
877 ConditionScope.ForceCleanup();
880 EmitBranch(CondBlock);
882 ForScope.ForceCleanup();
886 // Emit the fall-through block.
887 EmitBlock(LoopExit.getBlock(), true);
891 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
892 ArrayRef<const Attr *> ForAttrs) {
893 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
895 LexicalScope ForScope(*this, S.getSourceRange());
897 llvm::DebugLoc DL = Builder.getCurrentDebugLocation();
899 // Evaluate the first pieces before the loop.
900 EmitStmt(S.getRangeStmt());
901 EmitStmt(S.getBeginStmt());
902 EmitStmt(S.getEndStmt());
904 // Start the loop with a block that tests the condition.
905 // If there's an increment, the continue scope will be overwritten
907 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond");
908 EmitBlock(CondBlock);
910 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, DL);
912 // If there are any cleanups between here and the loop-exit scope,
913 // create a block to stage a loop exit along.
914 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
915 if (ForScope.requiresCleanups())
916 ExitBlock = createBasicBlock("for.cond.cleanup");
918 // The loop body, consisting of the specified body and the loop variable.
919 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
921 // The body is executed if the expression, contextually converted
923 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
924 Builder.CreateCondBr(
925 BoolCondVal, ForBody, ExitBlock,
926 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
928 if (ExitBlock != LoopExit.getBlock()) {
929 EmitBlock(ExitBlock);
930 EmitBranchThroughCleanup(LoopExit);
934 incrementProfileCounter(&S);
936 // Create a block for the increment. In case of a 'continue', we jump there.
937 JumpDest Continue = getJumpDestInCurrentScope("for.inc");
939 // Store the blocks to use for break and continue.
940 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
943 // Create a separate cleanup scope for the loop variable and body.
944 LexicalScope BodyScope(*this, S.getSourceRange());
945 EmitStmt(S.getLoopVarStmt());
946 EmitStmt(S.getBody());
950 // If there is an increment, emit it next.
951 EmitBlock(Continue.getBlock());
952 EmitStmt(S.getInc());
954 BreakContinueStack.pop_back();
956 EmitBranch(CondBlock);
958 ForScope.ForceCleanup();
962 // Emit the fall-through block.
963 EmitBlock(LoopExit.getBlock(), true);
966 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
968 Builder.CreateStore(RV.getScalarVal(), ReturnValue);
969 } else if (RV.isAggregate()) {
970 EmitAggregateCopy(ReturnValue, RV.getAggregateAddress(), Ty);
972 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty),
975 EmitBranchThroughCleanup(ReturnBlock);
978 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
979 /// if the function returns void, or may be missing one if the function returns
980 /// non-void. Fun stuff :).
981 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
982 // Returning from an outlined SEH helper is UB, and we already warn on it.
983 if (IsOutlinedSEHHelper) {
984 Builder.CreateUnreachable();
985 Builder.ClearInsertionPoint();
988 // Emit the result value, even if unused, to evalute the side effects.
989 const Expr *RV = S.getRetValue();
991 // Treat block literals in a return expression as if they appeared
992 // in their own scope. This permits a small, easily-implemented
993 // exception to our over-conservative rules about not jumping to
994 // statements following block literals with non-trivial cleanups.
995 RunCleanupsScope cleanupScope(*this);
996 if (const ExprWithCleanups *cleanups =
997 dyn_cast_or_null<ExprWithCleanups>(RV)) {
998 enterFullExpression(cleanups);
999 RV = cleanups->getSubExpr();
1002 // FIXME: Clean this up by using an LValue for ReturnTemp,
1003 // EmitStoreThroughLValue, and EmitAnyExpr.
1004 if (getLangOpts().ElideConstructors &&
1005 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) {
1006 // Apply the named return value optimization for this return statement,
1007 // which means doing nothing: the appropriate result has already been
1008 // constructed into the NRVO variable.
1010 // If there is an NRVO flag for this variable, set it to 1 into indicate
1011 // that the cleanup code should not destroy the variable.
1012 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1013 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag);
1014 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1015 // Make sure not to return anything, but evaluate the expression
1016 // for side effects.
1020 // Do nothing (return value is left uninitialized)
1021 } else if (FnRetTy->isReferenceType()) {
1022 // If this function returns a reference, take the address of the expression
1023 // rather than the value.
1024 RValue Result = EmitReferenceBindingToExpr(RV);
1025 Builder.CreateStore(Result.getScalarVal(), ReturnValue);
1027 switch (getEvaluationKind(RV->getType())) {
1029 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue);
1032 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()),
1036 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue,
1038 AggValueSlot::IsDestructed,
1039 AggValueSlot::DoesNotNeedGCBarriers,
1040 AggValueSlot::IsNotAliased));
1046 if (!RV || RV->isEvaluatable(getContext()))
1047 ++NumSimpleReturnExprs;
1049 cleanupScope.ForceCleanup();
1050 EmitBranchThroughCleanup(ReturnBlock);
1053 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1054 // As long as debug info is modeled with instructions, we have to ensure we
1055 // have a place to insert here and write the stop point here.
1056 if (HaveInsertPoint())
1059 for (const auto *I : S.decls())
1063 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1064 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1066 // If this code is reachable then emit a stop point (if generating
1067 // debug info). We have to do this ourselves because we are on the
1068 // "simple" statement path.
1069 if (HaveInsertPoint())
1072 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock);
1075 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1076 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1078 // If this code is reachable then emit a stop point (if generating
1079 // debug info). We have to do this ourselves because we are on the
1080 // "simple" statement path.
1081 if (HaveInsertPoint())
1084 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock);
1087 /// EmitCaseStmtRange - If case statement range is not too big then
1088 /// add multiple cases to switch instruction, one for each value within
1089 /// the range. If range is too big then emit "if" condition check.
1090 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) {
1091 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1093 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext());
1094 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext());
1096 // Emit the code for this case. We do this first to make sure it is
1097 // properly chained from our predecessor before generating the
1098 // switch machinery to enter this block.
1099 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1100 EmitBlockWithFallThrough(CaseDest, &S);
1101 EmitStmt(S.getSubStmt());
1103 // If range is empty, do nothing.
1104 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS))
1107 llvm::APInt Range = RHS - LHS;
1108 // FIXME: parameters such as this should not be hardcoded.
1109 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) {
1110 // Range is small enough to add multiple switch instruction cases.
1111 uint64_t Total = getProfileCount(&S);
1112 unsigned NCases = Range.getZExtValue() + 1;
1113 // We only have one region counter for the entire set of cases here, so we
1114 // need to divide the weights evenly between the generated cases, ensuring
1115 // that the total weight is preserved. E.g., a weight of 5 over three cases
1116 // will be distributed as weights of 2, 2, and 1.
1117 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1118 for (unsigned I = 0; I != NCases; ++I) {
1120 SwitchWeights->push_back(Weight + (Rem ? 1 : 0));
1123 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest);
1129 // The range is too big. Emit "if" condition into a new block,
1130 // making sure to save and restore the current insertion point.
1131 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1133 // Push this test onto the chain of range checks (which terminates
1134 // in the default basic block). The switch's default will be changed
1135 // to the top of this chain after switch emission is complete.
1136 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1137 CaseRangeBlock = createBasicBlock("sw.caserange");
1139 CurFn->getBasicBlockList().push_back(CaseRangeBlock);
1140 Builder.SetInsertPoint(CaseRangeBlock);
1142 // Emit range check.
1144 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS));
1146 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds");
1148 llvm::MDNode *Weights = nullptr;
1149 if (SwitchWeights) {
1150 uint64_t ThisCount = getProfileCount(&S);
1151 uint64_t DefaultCount = (*SwitchWeights)[0];
1152 Weights = createProfileWeights(ThisCount, DefaultCount);
1154 // Since we're chaining the switch default through each large case range, we
1155 // need to update the weight for the default, ie, the first case, to include
1157 (*SwitchWeights)[0] += ThisCount;
1159 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights);
1161 // Restore the appropriate insertion point.
1163 Builder.SetInsertPoint(RestoreBB);
1165 Builder.ClearInsertionPoint();
1168 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) {
1169 // If there is no enclosing switch instance that we're aware of, then this
1170 // case statement and its block can be elided. This situation only happens
1171 // when we've constant-folded the switch, are emitting the constant case,
1172 // and part of the constant case includes another case statement. For
1173 // instance: switch (4) { case 4: do { case 5: } while (1); }
1175 EmitStmt(S.getSubStmt());
1179 // Handle case ranges.
1181 EmitCaseStmtRange(S);
1185 llvm::ConstantInt *CaseVal =
1186 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext()));
1188 // If the body of the case is just a 'break', try to not emit an empty block.
1189 // If we're profiling or we're not optimizing, leave the block in for better
1190 // debug and coverage analysis.
1191 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1192 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1193 isa<BreakStmt>(S.getSubStmt())) {
1194 JumpDest Block = BreakContinueStack.back().BreakBlock;
1196 // Only do this optimization if there are no cleanups that need emitting.
1197 if (isObviouslyBranchWithoutCleanups(Block)) {
1199 SwitchWeights->push_back(getProfileCount(&S));
1200 SwitchInsn->addCase(CaseVal, Block.getBlock());
1202 // If there was a fallthrough into this case, make sure to redirect it to
1203 // the end of the switch as well.
1204 if (Builder.GetInsertBlock()) {
1205 Builder.CreateBr(Block.getBlock());
1206 Builder.ClearInsertionPoint();
1212 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1213 EmitBlockWithFallThrough(CaseDest, &S);
1215 SwitchWeights->push_back(getProfileCount(&S));
1216 SwitchInsn->addCase(CaseVal, CaseDest);
1218 // Recursively emitting the statement is acceptable, but is not wonderful for
1219 // code where we have many case statements nested together, i.e.:
1223 // Handling this recursively will create a new block for each case statement
1224 // that falls through to the next case which is IR intensive. It also causes
1225 // deep recursion which can run into stack depth limitations. Handle
1226 // sequential non-range case statements specially.
1227 const CaseStmt *CurCase = &S;
1228 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt());
1230 // Otherwise, iteratively add consecutive cases to this switch stmt.
1231 while (NextCase && NextCase->getRHS() == nullptr) {
1233 llvm::ConstantInt *CaseVal =
1234 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext()));
1237 SwitchWeights->push_back(getProfileCount(NextCase));
1238 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1239 CaseDest = createBasicBlock("sw.bb");
1240 EmitBlockWithFallThrough(CaseDest, &S);
1243 SwitchInsn->addCase(CaseVal, CaseDest);
1244 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt());
1247 // Normal default recursion for non-cases.
1248 EmitStmt(CurCase->getSubStmt());
1251 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) {
1252 // If there is no enclosing switch instance that we're aware of, then this
1253 // default statement can be elided. This situation only happens when we've
1254 // constant-folded the switch.
1256 EmitStmt(S.getSubStmt());
1260 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1261 assert(DefaultBlock->empty() &&
1262 "EmitDefaultStmt: Default block already defined?");
1264 EmitBlockWithFallThrough(DefaultBlock, &S);
1266 EmitStmt(S.getSubStmt());
1269 /// CollectStatementsForCase - Given the body of a 'switch' statement and a
1270 /// constant value that is being switched on, see if we can dead code eliminate
1271 /// the body of the switch to a simple series of statements to emit. Basically,
1272 /// on a switch (5) we want to find these statements:
1274 /// printf(...); <--
1278 /// and add them to the ResultStmts vector. If it is unsafe to do this
1279 /// transformation (for example, one of the elided statements contains a label
1280 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
1281 /// should include statements after it (e.g. the printf() line is a substmt of
1282 /// the case) then return CSFC_FallThrough. If we handled it and found a break
1283 /// statement, then return CSFC_Success.
1285 /// If Case is non-null, then we are looking for the specified case, checking
1286 /// that nothing we jump over contains labels. If Case is null, then we found
1287 /// the case and are looking for the break.
1289 /// If the recursive walk actually finds our Case, then we set FoundCase to
1292 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
1293 static CSFC_Result CollectStatementsForCase(const Stmt *S,
1294 const SwitchCase *Case,
1296 SmallVectorImpl<const Stmt*> &ResultStmts) {
1297 // If this is a null statement, just succeed.
1299 return Case ? CSFC_Success : CSFC_FallThrough;
1301 // If this is the switchcase (case 4: or default) that we're looking for, then
1302 // we're in business. Just add the substatement.
1303 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) {
1306 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase,
1310 // Otherwise, this is some other case or default statement, just ignore it.
1311 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase,
1315 // If we are in the live part of the code and we found our break statement,
1316 // return a success!
1317 if (!Case && isa<BreakStmt>(S))
1318 return CSFC_Success;
1320 // If this is a switch statement, then it might contain the SwitchCase, the
1321 // break, or neither.
1322 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) {
1323 // Handle this as two cases: we might be looking for the SwitchCase (if so
1324 // the skipped statements must be skippable) or we might already have it.
1325 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
1326 bool StartedInLiveCode = FoundCase;
1327 unsigned StartSize = ResultStmts.size();
1329 // If we've not found the case yet, scan through looking for it.
1331 // Keep track of whether we see a skipped declaration. The code could be
1332 // using the declaration even if it is skipped, so we can't optimize out
1333 // the decl if the kept statements might refer to it.
1334 bool HadSkippedDecl = false;
1336 // If we're looking for the case, just see if we can skip each of the
1338 for (; Case && I != E; ++I) {
1339 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I);
1341 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) {
1342 case CSFC_Failure: return CSFC_Failure;
1344 // A successful result means that either 1) that the statement doesn't
1345 // have the case and is skippable, or 2) does contain the case value
1346 // and also contains the break to exit the switch. In the later case,
1347 // we just verify the rest of the statements are elidable.
1349 // If we found the case and skipped declarations, we can't do the
1352 return CSFC_Failure;
1354 for (++I; I != E; ++I)
1355 if (CodeGenFunction::ContainsLabel(*I, true))
1356 return CSFC_Failure;
1357 return CSFC_Success;
1360 case CSFC_FallThrough:
1361 // If we have a fallthrough condition, then we must have found the
1362 // case started to include statements. Consider the rest of the
1363 // statements in the compound statement as candidates for inclusion.
1364 assert(FoundCase && "Didn't find case but returned fallthrough?");
1365 // We recursively found Case, so we're not looking for it anymore.
1368 // If we found the case and skipped declarations, we can't do the
1371 return CSFC_Failure;
1377 return CSFC_Success;
1379 assert(!HadSkippedDecl && "fallthrough after skipping decl");
1382 // If we have statements in our range, then we know that the statements are
1383 // live and need to be added to the set of statements we're tracking.
1384 bool AnyDecls = false;
1385 for (; I != E; ++I) {
1386 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I);
1388 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) {
1389 case CSFC_Failure: return CSFC_Failure;
1390 case CSFC_FallThrough:
1391 // A fallthrough result means that the statement was simple and just
1392 // included in ResultStmt, keep adding them afterwards.
1395 // A successful result means that we found the break statement and
1396 // stopped statement inclusion. We just ensure that any leftover stmts
1397 // are skippable and return success ourselves.
1398 for (++I; I != E; ++I)
1399 if (CodeGenFunction::ContainsLabel(*I, true))
1400 return CSFC_Failure;
1401 return CSFC_Success;
1405 // If we're about to fall out of a scope without hitting a 'break;', we
1406 // can't perform the optimization if there were any decls in that scope
1407 // (we'd lose their end-of-lifetime).
1409 // If the entire compound statement was live, there's one more thing we
1410 // can try before giving up: emit the whole thing as a single statement.
1411 // We can do that unless the statement contains a 'break;'.
1412 // FIXME: Such a break must be at the end of a construct within this one.
1413 // We could emit this by just ignoring the BreakStmts entirely.
1414 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) {
1415 ResultStmts.resize(StartSize);
1416 ResultStmts.push_back(S);
1418 return CSFC_Failure;
1422 return CSFC_FallThrough;
1425 // Okay, this is some other statement that we don't handle explicitly, like a
1426 // for statement or increment etc. If we are skipping over this statement,
1427 // just verify it doesn't have labels, which would make it invalid to elide.
1429 if (CodeGenFunction::ContainsLabel(S, true))
1430 return CSFC_Failure;
1431 return CSFC_Success;
1434 // Otherwise, we want to include this statement. Everything is cool with that
1435 // so long as it doesn't contain a break out of the switch we're in.
1436 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
1438 // Otherwise, everything is great. Include the statement and tell the caller
1439 // that we fall through and include the next statement as well.
1440 ResultStmts.push_back(S);
1441 return CSFC_FallThrough;
1444 /// FindCaseStatementsForValue - Find the case statement being jumped to and
1445 /// then invoke CollectStatementsForCase to find the list of statements to emit
1446 /// for a switch on constant. See the comment above CollectStatementsForCase
1447 /// for more details.
1448 static bool FindCaseStatementsForValue(const SwitchStmt &S,
1449 const llvm::APSInt &ConstantCondValue,
1450 SmallVectorImpl<const Stmt*> &ResultStmts,
1452 const SwitchCase *&ResultCase) {
1453 // First step, find the switch case that is being branched to. We can do this
1454 // efficiently by scanning the SwitchCase list.
1455 const SwitchCase *Case = S.getSwitchCaseList();
1456 const DefaultStmt *DefaultCase = nullptr;
1458 for (; Case; Case = Case->getNextSwitchCase()) {
1459 // It's either a default or case. Just remember the default statement in
1460 // case we're not jumping to any numbered cases.
1461 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) {
1466 // Check to see if this case is the one we're looking for.
1467 const CaseStmt *CS = cast<CaseStmt>(Case);
1468 // Don't handle case ranges yet.
1469 if (CS->getRHS()) return false;
1471 // If we found our case, remember it as 'case'.
1472 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue)
1476 // If we didn't find a matching case, we use a default if it exists, or we
1477 // elide the whole switch body!
1479 // It is safe to elide the body of the switch if it doesn't contain labels
1480 // etc. If it is safe, return successfully with an empty ResultStmts list.
1482 return !CodeGenFunction::ContainsLabel(&S);
1486 // Ok, we know which case is being jumped to, try to collect all the
1487 // statements that follow it. This can fail for a variety of reasons. Also,
1488 // check to see that the recursive walk actually found our case statement.
1489 // Insane cases like this can fail to find it in the recursive walk since we
1490 // don't handle every stmt kind:
1494 bool FoundCase = false;
1496 return CollectStatementsForCase(S.getBody(), Case, FoundCase,
1497 ResultStmts) != CSFC_Failure &&
1501 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
1502 // Handle nested switch statements.
1503 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
1504 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
1505 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
1507 // See if we can constant fold the condition of the switch and therefore only
1508 // emit the live case statement (if any) of the switch.
1509 llvm::APSInt ConstantCondValue;
1510 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) {
1511 SmallVector<const Stmt*, 4> CaseStmts;
1512 const SwitchCase *Case = nullptr;
1513 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts,
1514 getContext(), Case)) {
1516 incrementProfileCounter(Case);
1517 RunCleanupsScope ExecutedScope(*this);
1520 EmitStmt(S.getInit());
1522 // Emit the condition variable if needed inside the entire cleanup scope
1523 // used by this special case for constant folded switches.
1524 if (S.getConditionVariable())
1525 EmitAutoVarDecl(*S.getConditionVariable());
1527 // At this point, we are no longer "within" a switch instance, so
1528 // we can temporarily enforce this to ensure that any embedded case
1529 // statements are not emitted.
1530 SwitchInsn = nullptr;
1532 // Okay, we can dead code eliminate everything except this case. Emit the
1533 // specified series of statements and we're good.
1534 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i)
1535 EmitStmt(CaseStmts[i]);
1536 incrementProfileCounter(&S);
1538 // Now we want to restore the saved switch instance so that nested
1539 // switches continue to function properly
1540 SwitchInsn = SavedSwitchInsn;
1546 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog");
1548 RunCleanupsScope ConditionScope(*this);
1551 EmitStmt(S.getInit());
1553 if (S.getConditionVariable())
1554 EmitAutoVarDecl(*S.getConditionVariable());
1555 llvm::Value *CondV = EmitScalarExpr(S.getCond());
1557 // Create basic block to hold stuff that comes after switch
1558 // statement. We also need to create a default block now so that
1559 // explicit case ranges tests can have a place to jump to on
1561 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default");
1562 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock);
1563 if (PGO.haveRegionCounts()) {
1564 // Walk the SwitchCase list to find how many there are.
1565 uint64_t DefaultCount = 0;
1566 unsigned NumCases = 0;
1567 for (const SwitchCase *Case = S.getSwitchCaseList();
1569 Case = Case->getNextSwitchCase()) {
1570 if (isa<DefaultStmt>(Case))
1571 DefaultCount = getProfileCount(Case);
1574 SwitchWeights = new SmallVector<uint64_t, 16>();
1575 SwitchWeights->reserve(NumCases);
1576 // The default needs to be first. We store the edge count, so we already
1577 // know the right weight.
1578 SwitchWeights->push_back(DefaultCount);
1580 CaseRangeBlock = DefaultBlock;
1582 // Clear the insertion point to indicate we are in unreachable code.
1583 Builder.ClearInsertionPoint();
1585 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
1586 // then reuse last ContinueBlock.
1587 JumpDest OuterContinue;
1588 if (!BreakContinueStack.empty())
1589 OuterContinue = BreakContinueStack.back().ContinueBlock;
1591 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue));
1593 // Emit switch body.
1594 EmitStmt(S.getBody());
1596 BreakContinueStack.pop_back();
1598 // Update the default block in case explicit case range tests have
1599 // been chained on top.
1600 SwitchInsn->setDefaultDest(CaseRangeBlock);
1602 // If a default was never emitted:
1603 if (!DefaultBlock->getParent()) {
1604 // If we have cleanups, emit the default block so that there's a
1605 // place to jump through the cleanups from.
1606 if (ConditionScope.requiresCleanups()) {
1607 EmitBlock(DefaultBlock);
1609 // Otherwise, just forward the default block to the switch end.
1611 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock());
1612 delete DefaultBlock;
1616 ConditionScope.ForceCleanup();
1618 // Emit continuation.
1619 EmitBlock(SwitchExit.getBlock(), true);
1620 incrementProfileCounter(&S);
1622 // If the switch has a condition wrapped by __builtin_unpredictable,
1623 // create metadata that specifies that the switch is unpredictable.
1624 // Don't bother if not optimizing because that metadata would not be used.
1625 auto *Call = dyn_cast<CallExpr>(S.getCond());
1626 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1627 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1628 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1629 llvm::MDBuilder MDHelper(getLLVMContext());
1630 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable,
1631 MDHelper.createUnpredictable());
1635 if (SwitchWeights) {
1636 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
1637 "switch weights do not match switch cases");
1638 // If there's only one jump destination there's no sense weighting it.
1639 if (SwitchWeights->size() > 1)
1640 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1641 createProfileWeights(*SwitchWeights));
1642 delete SwitchWeights;
1644 SwitchInsn = SavedSwitchInsn;
1645 SwitchWeights = SavedSwitchWeights;
1646 CaseRangeBlock = SavedCRBlock;
1650 SimplifyConstraint(const char *Constraint, const TargetInfo &Target,
1651 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) {
1654 while (*Constraint) {
1655 switch (*Constraint) {
1657 Result += Target.convertConstraint(Constraint);
1663 case '=': // Will see this and the following in mult-alt constraints.
1666 case '#': // Ignore the rest of the constraint alternative.
1667 while (Constraint[1] && Constraint[1] != ',')
1672 Result += *Constraint;
1673 while (Constraint[1] && Constraint[1] == *Constraint)
1684 "Must pass output names to constraints with a symbolic name");
1686 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index);
1687 assert(result && "Could not resolve symbolic name"); (void)result;
1688 Result += llvm::utostr(Index);
1699 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared
1700 /// as using a particular register add that as a constraint that will be used
1701 /// in this asm stmt.
1703 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr,
1704 const TargetInfo &Target, CodeGenModule &CGM,
1705 const AsmStmt &Stmt, const bool EarlyClobber) {
1706 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr);
1709 const ValueDecl &Value = *AsmDeclRef->getDecl();
1710 const VarDecl *Variable = dyn_cast<VarDecl>(&Value);
1713 if (Variable->getStorageClass() != SC_Register)
1715 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>();
1718 StringRef Register = Attr->getLabel();
1719 assert(Target.isValidGCCRegisterName(Register));
1720 // We're using validateOutputConstraint here because we only care if
1721 // this is a register constraint.
1722 TargetInfo::ConstraintInfo Info(Constraint, "");
1723 if (Target.validateOutputConstraint(Info) &&
1724 !Info.allowsRegister()) {
1725 CGM.ErrorUnsupported(&Stmt, "__asm__");
1728 // Canonicalize the register here before returning it.
1729 Register = Target.getNormalizedGCCRegisterName(Register);
1730 return (EarlyClobber ? "&{" : "{") + Register.str() + "}";
1734 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info,
1735 LValue InputValue, QualType InputType,
1736 std::string &ConstraintStr,
1737 SourceLocation Loc) {
1739 if (Info.allowsRegister() || !Info.allowsMemory()) {
1740 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) {
1741 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal();
1743 llvm::Type *Ty = ConvertType(InputType);
1744 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
1745 if (Size <= 64 && llvm::isPowerOf2_64(Size)) {
1746 Ty = llvm::IntegerType::get(getLLVMContext(), Size);
1747 Ty = llvm::PointerType::getUnqual(Ty);
1749 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(),
1752 Arg = InputValue.getPointer();
1753 ConstraintStr += '*';
1757 Arg = InputValue.getPointer();
1758 ConstraintStr += '*';
1764 llvm::Value* CodeGenFunction::EmitAsmInput(
1765 const TargetInfo::ConstraintInfo &Info,
1766 const Expr *InputExpr,
1767 std::string &ConstraintStr) {
1768 // If this can't be a register or memory, i.e., has to be a constant
1769 // (immediate or symbolic), try to emit it as such.
1770 if (!Info.allowsRegister() && !Info.allowsMemory()) {
1771 llvm::APSInt Result;
1772 if (InputExpr->EvaluateAsInt(Result, getContext()))
1773 return llvm::ConstantInt::get(getLLVMContext(), Result);
1774 assert(!Info.requiresImmediateConstant() &&
1775 "Required-immediate inlineasm arg isn't constant?");
1778 if (Info.allowsRegister() || !Info.allowsMemory())
1779 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType()))
1780 return EmitScalarExpr(InputExpr);
1781 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
1782 return EmitScalarExpr(InputExpr);
1783 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext());
1784 LValue Dest = EmitLValue(InputExpr);
1785 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr,
1786 InputExpr->getExprLoc());
1789 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
1790 /// asm call instruction. The !srcloc MDNode contains a list of constant
1791 /// integers which are the source locations of the start of each line in the
1793 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
1794 CodeGenFunction &CGF) {
1795 SmallVector<llvm::Metadata *, 8> Locs;
1796 // Add the location of the first line to the MDNode.
1797 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
1798 CGF.Int32Ty, Str->getLocStart().getRawEncoding())));
1799 StringRef StrVal = Str->getString();
1800 if (!StrVal.empty()) {
1801 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
1802 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
1803 unsigned StartToken = 0;
1804 unsigned ByteOffset = 0;
1806 // Add the location of the start of each subsequent line of the asm to the
1808 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
1809 if (StrVal[i] != '\n') continue;
1810 SourceLocation LineLoc = Str->getLocationOfByte(
1811 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset);
1812 Locs.push_back(llvm::ConstantAsMetadata::get(
1813 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding())));
1817 return llvm::MDNode::get(CGF.getLLVMContext(), Locs);
1820 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
1821 // Assemble the final asm string.
1822 std::string AsmString = S.generateAsmString(getContext());
1824 // Get all the output and input constraints together.
1825 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
1826 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
1828 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1830 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1831 Name = GAS->getOutputName(i);
1832 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
1833 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
1834 assert(IsValid && "Failed to parse output constraint");
1835 OutputConstraintInfos.push_back(Info);
1838 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1840 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1841 Name = GAS->getInputName(i);
1842 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
1844 getTarget().validateInputConstraint(OutputConstraintInfos, Info);
1845 assert(IsValid && "Failed to parse input constraint"); (void)IsValid;
1846 InputConstraintInfos.push_back(Info);
1849 std::string Constraints;
1851 std::vector<LValue> ResultRegDests;
1852 std::vector<QualType> ResultRegQualTys;
1853 std::vector<llvm::Type *> ResultRegTypes;
1854 std::vector<llvm::Type *> ResultTruncRegTypes;
1855 std::vector<llvm::Type *> ArgTypes;
1856 std::vector<llvm::Value*> Args;
1858 // Keep track of inout constraints.
1859 std::string InOutConstraints;
1860 std::vector<llvm::Value*> InOutArgs;
1861 std::vector<llvm::Type*> InOutArgTypes;
1863 // An inline asm can be marked readonly if it meets the following conditions:
1864 // - it doesn't have any sideeffects
1865 // - it doesn't clobber memory
1866 // - it doesn't return a value by-reference
1867 // It can be marked readnone if it doesn't have any input memory constraints
1868 // in addition to meeting the conditions listed above.
1869 bool ReadOnly = true, ReadNone = true;
1871 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1872 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
1874 // Simplify the output constraint.
1875 std::string OutputConstraint(S.getOutputConstraint(i));
1876 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1,
1879 const Expr *OutExpr = S.getOutputExpr(i);
1880 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext());
1882 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr,
1883 getTarget(), CGM, S,
1884 Info.earlyClobber());
1886 LValue Dest = EmitLValue(OutExpr);
1887 if (!Constraints.empty())
1890 // If this is a register output, then make the inline asm return it
1891 // by-value. If this is a memory result, return the value by-reference.
1892 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) {
1893 Constraints += "=" + OutputConstraint;
1894 ResultRegQualTys.push_back(OutExpr->getType());
1895 ResultRegDests.push_back(Dest);
1896 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType()));
1897 ResultTruncRegTypes.push_back(ResultRegTypes.back());
1899 // If this output is tied to an input, and if the input is larger, then
1900 // we need to set the actual result type of the inline asm node to be the
1901 // same as the input type.
1902 if (Info.hasMatchingInput()) {
1904 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
1905 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
1906 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
1909 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
1911 QualType InputTy = S.getInputExpr(InputNo)->getType();
1912 QualType OutputType = OutExpr->getType();
1914 uint64_t InputSize = getContext().getTypeSize(InputTy);
1915 if (getContext().getTypeSize(OutputType) < InputSize) {
1916 // Form the asm to return the value as a larger integer or fp type.
1917 ResultRegTypes.back() = ConvertType(InputTy);
1920 if (llvm::Type* AdjTy =
1921 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1922 ResultRegTypes.back()))
1923 ResultRegTypes.back() = AdjTy;
1925 CGM.getDiags().Report(S.getAsmLoc(),
1926 diag::err_asm_invalid_type_in_input)
1927 << OutExpr->getType() << OutputConstraint;
1930 ArgTypes.push_back(Dest.getAddress().getType());
1931 Args.push_back(Dest.getPointer());
1932 Constraints += "=*";
1933 Constraints += OutputConstraint;
1934 ReadOnly = ReadNone = false;
1937 if (Info.isReadWrite()) {
1938 InOutConstraints += ',';
1940 const Expr *InputExpr = S.getOutputExpr(i);
1941 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
1943 InputExpr->getExprLoc());
1945 if (llvm::Type* AdjTy =
1946 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1948 Arg = Builder.CreateBitCast(Arg, AdjTy);
1950 if (Info.allowsRegister())
1951 InOutConstraints += llvm::utostr(i);
1953 InOutConstraints += OutputConstraint;
1955 InOutArgTypes.push_back(Arg->getType());
1956 InOutArgs.push_back(Arg);
1960 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
1961 // to the return value slot. Only do this when returning in registers.
1962 if (isa<MSAsmStmt>(&S)) {
1963 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
1964 if (RetAI.isDirect() || RetAI.isExtend()) {
1965 // Make a fake lvalue for the return value slot.
1966 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy);
1967 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs(
1968 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes,
1969 ResultRegDests, AsmString, S.getNumOutputs());
1974 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1975 const Expr *InputExpr = S.getInputExpr(i);
1977 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
1979 if (Info.allowsMemory())
1982 if (!Constraints.empty())
1985 // Simplify the input constraint.
1986 std::string InputConstraint(S.getInputConstraint(i));
1987 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(),
1988 &OutputConstraintInfos);
1990 InputConstraint = AddVariableConstraints(
1991 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()),
1992 getTarget(), CGM, S, false /* No EarlyClobber */);
1994 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints);
1996 // If this input argument is tied to a larger output result, extend the
1997 // input to be the same size as the output. The LLVM backend wants to see
1998 // the input and output of a matching constraint be the same size. Note
1999 // that GCC does not define what the top bits are here. We use zext because
2000 // that is usually cheaper, but LLVM IR should really get an anyext someday.
2001 if (Info.hasTiedOperand()) {
2002 unsigned Output = Info.getTiedOperand();
2003 QualType OutputType = S.getOutputExpr(Output)->getType();
2004 QualType InputTy = InputExpr->getType();
2006 if (getContext().getTypeSize(OutputType) >
2007 getContext().getTypeSize(InputTy)) {
2008 // Use ptrtoint as appropriate so that we can do our extension.
2009 if (isa<llvm::PointerType>(Arg->getType()))
2010 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy);
2011 llvm::Type *OutputTy = ConvertType(OutputType);
2012 if (isa<llvm::IntegerType>(OutputTy))
2013 Arg = Builder.CreateZExt(Arg, OutputTy);
2014 else if (isa<llvm::PointerType>(OutputTy))
2015 Arg = Builder.CreateZExt(Arg, IntPtrTy);
2017 assert(OutputTy->isFloatingPointTy() && "Unexpected output type");
2018 Arg = Builder.CreateFPExt(Arg, OutputTy);
2022 if (llvm::Type* AdjTy =
2023 getTargetHooks().adjustInlineAsmType(*this, InputConstraint,
2025 Arg = Builder.CreateBitCast(Arg, AdjTy);
2027 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input)
2028 << InputExpr->getType() << InputConstraint;
2030 ArgTypes.push_back(Arg->getType());
2031 Args.push_back(Arg);
2032 Constraints += InputConstraint;
2035 // Append the "input" part of inout constraints last.
2036 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) {
2037 ArgTypes.push_back(InOutArgTypes[i]);
2038 Args.push_back(InOutArgs[i]);
2040 Constraints += InOutConstraints;
2043 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) {
2044 StringRef Clobber = S.getClobber(i);
2046 if (Clobber == "memory")
2047 ReadOnly = ReadNone = false;
2048 else if (Clobber != "cc")
2049 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber);
2051 if (!Constraints.empty())
2054 Constraints += "~{";
2055 Constraints += Clobber;
2059 // Add machine specific clobbers
2060 std::string MachineClobbers = getTarget().getClobbers();
2061 if (!MachineClobbers.empty()) {
2062 if (!Constraints.empty())
2064 Constraints += MachineClobbers;
2067 llvm::Type *ResultType;
2068 if (ResultRegTypes.empty())
2069 ResultType = VoidTy;
2070 else if (ResultRegTypes.size() == 1)
2071 ResultType = ResultRegTypes[0];
2073 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes);
2075 llvm::FunctionType *FTy =
2076 llvm::FunctionType::get(ResultType, ArgTypes, false);
2078 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0;
2079 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ?
2080 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT;
2081 llvm::InlineAsm *IA =
2082 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect,
2083 /* IsAlignStack */ false, AsmDialect);
2084 llvm::CallInst *Result = Builder.CreateCall(IA, Args);
2085 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2086 llvm::Attribute::NoUnwind);
2088 if (isa<MSAsmStmt>(&S)) {
2089 // If the assembly contains any labels, mark the call noduplicate to prevent
2090 // defining the same ASM label twice (PR23715). This is pretty hacky, but it
2092 if (AsmString.find("__MSASMLABEL_") != std::string::npos)
2093 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2094 llvm::Attribute::NoDuplicate);
2097 // Attach readnone and readonly attributes.
2098 if (!HasSideEffect) {
2100 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2101 llvm::Attribute::ReadNone);
2103 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2104 llvm::Attribute::ReadOnly);
2107 // Slap the source location of the inline asm into a !srcloc metadata on the
2109 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) {
2110 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(),
2113 // At least put the line number on MS inline asm blobs.
2114 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding());
2115 Result->setMetadata("srcloc",
2116 llvm::MDNode::get(getLLVMContext(),
2117 llvm::ConstantAsMetadata::get(Loc)));
2120 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
2121 // Conservatively, mark all inline asm blocks in CUDA as convergent
2122 // (meaning, they may call an intrinsically convergent op, such as bar.sync,
2123 // and so can't have certain optimizations applied around them).
2124 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2125 llvm::Attribute::Convergent);
2128 // Extract all of the register value results from the asm.
2129 std::vector<llvm::Value*> RegResults;
2130 if (ResultRegTypes.size() == 1) {
2131 RegResults.push_back(Result);
2133 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2134 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult");
2135 RegResults.push_back(Tmp);
2139 assert(RegResults.size() == ResultRegTypes.size());
2140 assert(RegResults.size() == ResultTruncRegTypes.size());
2141 assert(RegResults.size() == ResultRegDests.size());
2142 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) {
2143 llvm::Value *Tmp = RegResults[i];
2145 // If the result type of the LLVM IR asm doesn't match the result type of
2146 // the expression, do the conversion.
2147 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) {
2148 llvm::Type *TruncTy = ResultTruncRegTypes[i];
2150 // Truncate the integer result to the right size, note that TruncTy can be
2152 if (TruncTy->isFloatingPointTy())
2153 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy);
2154 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) {
2155 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy);
2156 Tmp = Builder.CreateTrunc(Tmp,
2157 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize));
2158 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy);
2159 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) {
2160 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType());
2161 Tmp = Builder.CreatePtrToInt(Tmp,
2162 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize));
2163 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2164 } else if (TruncTy->isIntegerTy()) {
2165 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2166 } else if (TruncTy->isVectorTy()) {
2167 Tmp = Builder.CreateBitCast(Tmp, TruncTy);
2171 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]);
2175 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) {
2176 const RecordDecl *RD = S.getCapturedRecordDecl();
2177 QualType RecordTy = getContext().getRecordType(RD);
2179 // Initialize the captured struct.
2181 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy);
2183 RecordDecl::field_iterator CurField = RD->field_begin();
2184 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
2185 E = S.capture_init_end();
2186 I != E; ++I, ++CurField) {
2187 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField);
2188 if (CurField->hasCapturedVLAType()) {
2189 auto VAT = CurField->getCapturedVLAType();
2190 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV);
2192 EmitInitializerForField(*CurField, LV, *I, None);
2199 /// Generate an outlined function for the body of a CapturedStmt, store any
2200 /// captured variables into the captured struct, and call the outlined function.
2202 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) {
2203 LValue CapStruct = InitCapturedStruct(S);
2205 // Emit the CapturedDecl
2206 CodeGenFunction CGF(CGM, true);
2207 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K));
2208 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S);
2209 delete CGF.CapturedStmtInfo;
2211 // Emit call to the helper function.
2212 EmitCallOrInvoke(F, CapStruct.getPointer());
2217 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) {
2218 LValue CapStruct = InitCapturedStruct(S);
2219 return CapStruct.getAddress();
2222 /// Creates the outlined function for a CapturedStmt.
2224 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) {
2225 assert(CapturedStmtInfo &&
2226 "CapturedStmtInfo should be set when generating the captured function");
2227 const CapturedDecl *CD = S.getCapturedDecl();
2228 const RecordDecl *RD = S.getCapturedRecordDecl();
2229 SourceLocation Loc = S.getLocStart();
2230 assert(CD->hasBody() && "missing CapturedDecl body");
2232 // Build the argument list.
2233 ASTContext &Ctx = CGM.getContext();
2234 FunctionArgList Args;
2235 Args.append(CD->param_begin(), CD->param_end());
2237 // Create the function declaration.
2238 FunctionType::ExtInfo ExtInfo;
2239 const CGFunctionInfo &FuncInfo =
2240 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args);
2241 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
2244 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
2245 CapturedStmtInfo->getHelperName(), &CGM.getModule());
2246 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
2247 if (CD->isNothrow())
2248 F->addFnAttr(llvm::Attribute::NoUnwind);
2250 // Generate the function.
2251 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args,
2253 CD->getBody()->getLocStart());
2254 // Set the context parameter in CapturedStmtInfo.
2255 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam());
2256 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr));
2258 // Initialize variable-length arrays.
2259 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(),
2260 Ctx.getTagDeclType(RD));
2261 for (auto *FD : RD->fields()) {
2262 if (FD->hasCapturedVLAType()) {
2263 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD),
2264 S.getLocStart()).getScalarVal();
2265 auto VAT = FD->getCapturedVLAType();
2266 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
2270 // If 'this' is captured, load it into CXXThisValue.
2271 if (CapturedStmtInfo->isCXXThisExprCaptured()) {
2272 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl();
2273 LValue ThisLValue = EmitLValueForField(Base, FD);
2274 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal();
2277 PGO.assignRegionCounters(GlobalDecl(CD), F);
2278 CapturedStmtInfo->EmitBody(*this, CD->getBody());
2279 FinishFunction(CD->getBodyRBrace());