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, ArrayRef<const Attr *> Attrs) {
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 // Ignore all OpenMP directives except for simd if OpenMP with Simd is
79 if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) {
80 if (const auto *D = dyn_cast<OMPExecutableDirective>(S)) {
81 EmitSimpleOMPExecutableDirective(*D);
86 switch (S->getStmtClass()) {
87 case Stmt::NoStmtClass:
88 case Stmt::CXXCatchStmtClass:
89 case Stmt::SEHExceptStmtClass:
90 case Stmt::SEHFinallyStmtClass:
91 case Stmt::MSDependentExistsStmtClass:
92 llvm_unreachable("invalid statement class to emit generically");
93 case Stmt::NullStmtClass:
94 case Stmt::CompoundStmtClass:
95 case Stmt::DeclStmtClass:
96 case Stmt::LabelStmtClass:
97 case Stmt::AttributedStmtClass:
98 case Stmt::GotoStmtClass:
99 case Stmt::BreakStmtClass:
100 case Stmt::ContinueStmtClass:
101 case Stmt::DefaultStmtClass:
102 case Stmt::CaseStmtClass:
103 case Stmt::SEHLeaveStmtClass:
104 llvm_unreachable("should have emitted these statements as simple");
106 #define STMT(Type, Base)
107 #define ABSTRACT_STMT(Op)
108 #define EXPR(Type, Base) \
109 case Stmt::Type##Class:
110 #include "clang/AST/StmtNodes.inc"
112 // Remember the block we came in on.
113 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
114 assert(incoming && "expression emission must have an insertion point");
116 EmitIgnoredExpr(cast<Expr>(S));
118 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
119 assert(outgoing && "expression emission cleared block!");
121 // The expression emitters assume (reasonably!) that the insertion
122 // point is always set. To maintain that, the call-emission code
123 // for noreturn functions has to enter a new block with no
124 // predecessors. We want to kill that block and mark the current
125 // insertion point unreachable in the common case of a call like
126 // "exit();". Since expression emission doesn't otherwise create
127 // blocks with no predecessors, we can just test for that.
128 // However, we must be careful not to do this to our incoming
129 // block, because *statement* emission does sometimes create
130 // reachable blocks which will have no predecessors until later in
131 // the function. This occurs with, e.g., labels that are not
132 // reachable by fallthrough.
133 if (incoming != outgoing && outgoing->use_empty()) {
134 outgoing->eraseFromParent();
135 Builder.ClearInsertionPoint();
140 case Stmt::IndirectGotoStmtClass:
141 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break;
143 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break;
144 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S), Attrs); break;
145 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S), Attrs); break;
146 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S), Attrs); break;
148 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break;
150 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break;
151 case Stmt::GCCAsmStmtClass: // Intentional fall-through.
152 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break;
153 case Stmt::CoroutineBodyStmtClass:
154 EmitCoroutineBody(cast<CoroutineBodyStmt>(*S));
156 case Stmt::CoreturnStmtClass:
157 EmitCoreturnStmt(cast<CoreturnStmt>(*S));
159 case Stmt::CapturedStmtClass: {
160 const CapturedStmt *CS = cast<CapturedStmt>(S);
161 EmitCapturedStmt(*CS, CS->getCapturedRegionKind());
164 case Stmt::ObjCAtTryStmtClass:
165 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S));
167 case Stmt::ObjCAtCatchStmtClass:
169 "@catch statements should be handled by EmitObjCAtTryStmt");
170 case Stmt::ObjCAtFinallyStmtClass:
172 "@finally statements should be handled by EmitObjCAtTryStmt");
173 case Stmt::ObjCAtThrowStmtClass:
174 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S));
176 case Stmt::ObjCAtSynchronizedStmtClass:
177 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S));
179 case Stmt::ObjCForCollectionStmtClass:
180 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S));
182 case Stmt::ObjCAutoreleasePoolStmtClass:
183 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S));
186 case Stmt::CXXTryStmtClass:
187 EmitCXXTryStmt(cast<CXXTryStmt>(*S));
189 case Stmt::CXXForRangeStmtClass:
190 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S), Attrs);
192 case Stmt::SEHTryStmtClass:
193 EmitSEHTryStmt(cast<SEHTryStmt>(*S));
195 case Stmt::OMPParallelDirectiveClass:
196 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S));
198 case Stmt::OMPSimdDirectiveClass:
199 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S));
201 case Stmt::OMPForDirectiveClass:
202 EmitOMPForDirective(cast<OMPForDirective>(*S));
204 case Stmt::OMPForSimdDirectiveClass:
205 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S));
207 case Stmt::OMPSectionsDirectiveClass:
208 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S));
210 case Stmt::OMPSectionDirectiveClass:
211 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S));
213 case Stmt::OMPSingleDirectiveClass:
214 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S));
216 case Stmt::OMPMasterDirectiveClass:
217 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S));
219 case Stmt::OMPCriticalDirectiveClass:
220 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S));
222 case Stmt::OMPParallelForDirectiveClass:
223 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S));
225 case Stmt::OMPParallelForSimdDirectiveClass:
226 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S));
228 case Stmt::OMPParallelSectionsDirectiveClass:
229 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S));
231 case Stmt::OMPTaskDirectiveClass:
232 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S));
234 case Stmt::OMPTaskyieldDirectiveClass:
235 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
237 case Stmt::OMPBarrierDirectiveClass:
238 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
240 case Stmt::OMPTaskwaitDirectiveClass:
241 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S));
243 case Stmt::OMPTaskgroupDirectiveClass:
244 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S));
246 case Stmt::OMPFlushDirectiveClass:
247 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S));
249 case Stmt::OMPOrderedDirectiveClass:
250 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S));
252 case Stmt::OMPAtomicDirectiveClass:
253 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S));
255 case Stmt::OMPTargetDirectiveClass:
256 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S));
258 case Stmt::OMPTeamsDirectiveClass:
259 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S));
261 case Stmt::OMPCancellationPointDirectiveClass:
262 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S));
264 case Stmt::OMPCancelDirectiveClass:
265 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S));
267 case Stmt::OMPTargetDataDirectiveClass:
268 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S));
270 case Stmt::OMPTargetEnterDataDirectiveClass:
271 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S));
273 case Stmt::OMPTargetExitDataDirectiveClass:
274 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S));
276 case Stmt::OMPTargetParallelDirectiveClass:
277 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S));
279 case Stmt::OMPTargetParallelForDirectiveClass:
280 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S));
282 case Stmt::OMPTaskLoopDirectiveClass:
283 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S));
285 case Stmt::OMPTaskLoopSimdDirectiveClass:
286 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S));
288 case Stmt::OMPDistributeDirectiveClass:
289 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S));
291 case Stmt::OMPTargetUpdateDirectiveClass:
292 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S));
294 case Stmt::OMPDistributeParallelForDirectiveClass:
295 EmitOMPDistributeParallelForDirective(
296 cast<OMPDistributeParallelForDirective>(*S));
298 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
299 EmitOMPDistributeParallelForSimdDirective(
300 cast<OMPDistributeParallelForSimdDirective>(*S));
302 case Stmt::OMPDistributeSimdDirectiveClass:
303 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S));
305 case Stmt::OMPTargetParallelForSimdDirectiveClass:
306 EmitOMPTargetParallelForSimdDirective(
307 cast<OMPTargetParallelForSimdDirective>(*S));
309 case Stmt::OMPTargetSimdDirectiveClass:
310 EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S));
312 case Stmt::OMPTeamsDistributeDirectiveClass:
313 EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S));
315 case Stmt::OMPTeamsDistributeSimdDirectiveClass:
316 EmitOMPTeamsDistributeSimdDirective(
317 cast<OMPTeamsDistributeSimdDirective>(*S));
319 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass:
320 EmitOMPTeamsDistributeParallelForSimdDirective(
321 cast<OMPTeamsDistributeParallelForSimdDirective>(*S));
323 case Stmt::OMPTeamsDistributeParallelForDirectiveClass:
324 EmitOMPTeamsDistributeParallelForDirective(
325 cast<OMPTeamsDistributeParallelForDirective>(*S));
327 case Stmt::OMPTargetTeamsDirectiveClass:
328 EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S));
330 case Stmt::OMPTargetTeamsDistributeDirectiveClass:
331 EmitOMPTargetTeamsDistributeDirective(
332 cast<OMPTargetTeamsDistributeDirective>(*S));
334 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass:
335 EmitOMPTargetTeamsDistributeParallelForDirective(
336 cast<OMPTargetTeamsDistributeParallelForDirective>(*S));
338 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass:
339 EmitOMPTargetTeamsDistributeParallelForSimdDirective(
340 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S));
342 case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass:
343 EmitOMPTargetTeamsDistributeSimdDirective(
344 cast<OMPTargetTeamsDistributeSimdDirective>(*S));
349 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) {
350 switch (S->getStmtClass()) {
351 default: return false;
352 case Stmt::NullStmtClass: break;
353 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break;
354 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break;
355 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break;
356 case Stmt::AttributedStmtClass:
357 EmitAttributedStmt(cast<AttributedStmt>(*S)); break;
358 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break;
359 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break;
360 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break;
361 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break;
362 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break;
363 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break;
369 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
370 /// this captures the expression result of the last sub-statement and returns it
371 /// (for use by the statement expression extension).
372 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
373 AggValueSlot AggSlot) {
374 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
375 "LLVM IR generation of compound statement ('{}')");
377 // Keep track of the current cleanup stack depth, including debug scopes.
378 LexicalScope Scope(*this, S.getSourceRange());
380 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
384 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
386 AggValueSlot AggSlot) {
388 for (CompoundStmt::const_body_iterator I = S.body_begin(),
389 E = S.body_end()-GetLast; I != E; ++I)
392 Address RetAlloca = Address::invalid();
394 // We have to special case labels here. They are statements, but when put
395 // at the end of a statement expression, they yield the value of their
396 // subexpression. Handle this by walking through all labels we encounter,
397 // emitting them before we evaluate the subexpr.
398 const Stmt *LastStmt = S.body_back();
399 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) {
400 EmitLabel(LS->getDecl());
401 LastStmt = LS->getSubStmt();
406 QualType ExprTy = cast<Expr>(LastStmt)->getType();
407 if (hasAggregateEvaluationKind(ExprTy)) {
408 EmitAggExpr(cast<Expr>(LastStmt), AggSlot);
410 // We can't return an RValue here because there might be cleanups at
411 // the end of the StmtExpr. Because of that, we have to emit the result
412 // here into a temporary alloca.
413 RetAlloca = CreateMemTemp(ExprTy);
414 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(),
423 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
424 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator());
426 // If there is a cleanup stack, then we it isn't worth trying to
427 // simplify this block (we would need to remove it from the scope map
428 // and cleanup entry).
429 if (!EHStack.empty())
432 // Can only simplify direct branches.
433 if (!BI || !BI->isUnconditional())
436 // Can only simplify empty blocks.
437 if (BI->getIterator() != BB->begin())
440 BB->replaceAllUsesWith(BI->getSuccessor(0));
441 BI->eraseFromParent();
442 BB->eraseFromParent();
445 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
446 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
448 // Fall out of the current block (if necessary).
451 if (IsFinished && BB->use_empty()) {
456 // Place the block after the current block, if possible, or else at
457 // the end of the function.
458 if (CurBB && CurBB->getParent())
459 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB);
461 CurFn->getBasicBlockList().push_back(BB);
462 Builder.SetInsertPoint(BB);
465 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
466 // Emit a branch from the current block to the target one if this
467 // was a real block. If this was just a fall-through block after a
468 // terminator, don't emit it.
469 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
471 if (!CurBB || CurBB->getTerminator()) {
472 // If there is no insert point or the previous block is already
473 // terminated, don't touch it.
475 // Otherwise, create a fall-through branch.
476 Builder.CreateBr(Target);
479 Builder.ClearInsertionPoint();
482 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
483 bool inserted = false;
484 for (llvm::User *u : block->users()) {
485 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) {
486 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(),
494 CurFn->getBasicBlockList().push_back(block);
496 Builder.SetInsertPoint(block);
499 CodeGenFunction::JumpDest
500 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
501 JumpDest &Dest = LabelMap[D];
502 if (Dest.isValid()) return Dest;
504 // Create, but don't insert, the new block.
505 Dest = JumpDest(createBasicBlock(D->getName()),
506 EHScopeStack::stable_iterator::invalid(),
507 NextCleanupDestIndex++);
511 void CodeGenFunction::EmitLabel(const LabelDecl *D) {
512 // Add this label to the current lexical scope if we're within any
513 // normal cleanups. Jumps "in" to this label --- when permitted by
514 // the language --- may need to be routed around such cleanups.
515 if (EHStack.hasNormalCleanups() && CurLexicalScope)
516 CurLexicalScope->addLabel(D);
518 JumpDest &Dest = LabelMap[D];
520 // If we didn't need a forward reference to this label, just go
521 // ahead and create a destination at the current scope.
522 if (!Dest.isValid()) {
523 Dest = getJumpDestInCurrentScope(D->getName());
525 // Otherwise, we need to give this label a target depth and remove
526 // it from the branch-fixups list.
528 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
529 Dest.setScopeDepth(EHStack.stable_begin());
530 ResolveBranchFixups(Dest.getBlock());
533 EmitBlock(Dest.getBlock());
534 incrementProfileCounter(D->getStmt());
537 /// Change the cleanup scope of the labels in this lexical scope to
538 /// match the scope of the enclosing context.
539 void CodeGenFunction::LexicalScope::rescopeLabels() {
540 assert(!Labels.empty());
541 EHScopeStack::stable_iterator innermostScope
542 = CGF.EHStack.getInnermostNormalCleanup();
544 // Change the scope depth of all the labels.
545 for (SmallVectorImpl<const LabelDecl*>::const_iterator
546 i = Labels.begin(), e = Labels.end(); i != e; ++i) {
547 assert(CGF.LabelMap.count(*i));
548 JumpDest &dest = CGF.LabelMap.find(*i)->second;
549 assert(dest.getScopeDepth().isValid());
550 assert(innermostScope.encloses(dest.getScopeDepth()));
551 dest.setScopeDepth(innermostScope);
554 // Reparent the labels if the new scope also has cleanups.
555 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
556 ParentScope->Labels.append(Labels.begin(), Labels.end());
561 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
562 EmitLabel(S.getDecl());
563 EmitStmt(S.getSubStmt());
566 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
567 EmitStmt(S.getSubStmt(), S.getAttrs());
570 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
571 // If this code is reachable then emit a stop point (if generating
572 // debug info). We have to do this ourselves because we are on the
573 // "simple" statement path.
574 if (HaveInsertPoint())
577 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel()));
581 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
582 if (const LabelDecl *Target = S.getConstantTarget()) {
583 EmitBranchThroughCleanup(getJumpDestForLabel(Target));
587 // Ensure that we have an i8* for our PHI node.
588 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()),
590 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
592 // Get the basic block for the indirect goto.
593 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
595 // The first instruction in the block has to be the PHI for the switch dest,
596 // add an entry for this branch.
597 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB);
599 EmitBranch(IndGotoBB);
602 void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
603 // C99 6.8.4.1: The first substatement is executed if the expression compares
604 // unequal to 0. The condition must be a scalar type.
605 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
608 EmitStmt(S.getInit());
610 if (S.getConditionVariable())
611 EmitAutoVarDecl(*S.getConditionVariable());
613 // If the condition constant folds and can be elided, try to avoid emitting
614 // the condition and the dead arm of the if/else.
616 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant,
618 // Figure out which block (then or else) is executed.
619 const Stmt *Executed = S.getThen();
620 const Stmt *Skipped = S.getElse();
621 if (!CondConstant) // Condition false?
622 std::swap(Executed, Skipped);
624 // If the skipped block has no labels in it, just emit the executed block.
625 // This avoids emitting dead code and simplifies the CFG substantially.
626 if (S.isConstexpr() || !ContainsLabel(Skipped)) {
628 incrementProfileCounter(&S);
630 RunCleanupsScope ExecutedScope(*this);
637 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
638 // the conditional branch.
639 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then");
640 llvm::BasicBlock *ContBlock = createBasicBlock("if.end");
641 llvm::BasicBlock *ElseBlock = ContBlock;
643 ElseBlock = createBasicBlock("if.else");
645 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock,
646 getProfileCount(S.getThen()));
648 // Emit the 'then' code.
649 EmitBlock(ThenBlock);
650 incrementProfileCounter(&S);
652 RunCleanupsScope ThenScope(*this);
653 EmitStmt(S.getThen());
655 EmitBranch(ContBlock);
657 // Emit the 'else' code if present.
658 if (const Stmt *Else = S.getElse()) {
660 // There is no need to emit line number for an unconditional branch.
661 auto NL = ApplyDebugLocation::CreateEmpty(*this);
662 EmitBlock(ElseBlock);
665 RunCleanupsScope ElseScope(*this);
669 // There is no need to emit line number for an unconditional branch.
670 auto NL = ApplyDebugLocation::CreateEmpty(*this);
671 EmitBranch(ContBlock);
675 // Emit the continuation block for code after the if.
676 EmitBlock(ContBlock, true);
679 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
680 ArrayRef<const Attr *> WhileAttrs) {
681 // Emit the header for the loop, which will also become
682 // the continue target.
683 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond");
684 EmitBlock(LoopHeader.getBlock());
686 const SourceRange &R = S.getSourceRange();
687 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs,
688 SourceLocToDebugLoc(R.getBegin()),
689 SourceLocToDebugLoc(R.getEnd()));
691 // Create an exit block for when the condition fails, which will
692 // also become the break target.
693 JumpDest LoopExit = getJumpDestInCurrentScope("while.end");
695 // Store the blocks to use for break and continue.
696 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader));
698 // C++ [stmt.while]p2:
699 // When the condition of a while statement is a declaration, the
700 // scope of the variable that is declared extends from its point
701 // of declaration (3.3.2) to the end of the while statement.
703 // The object created in a condition is destroyed and created
704 // with each iteration of the loop.
705 RunCleanupsScope ConditionScope(*this);
707 if (S.getConditionVariable())
708 EmitAutoVarDecl(*S.getConditionVariable());
710 // Evaluate the conditional in the while header. C99 6.8.5.1: The
711 // evaluation of the controlling expression takes place before each
712 // execution of the loop body.
713 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
715 // while(1) is common, avoid extra exit blocks. Be sure
716 // to correctly handle break/continue though.
717 bool EmitBoolCondBranch = true;
718 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
720 EmitBoolCondBranch = false;
722 // As long as the condition is true, go to the loop body.
723 llvm::BasicBlock *LoopBody = createBasicBlock("while.body");
724 if (EmitBoolCondBranch) {
725 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
726 if (ConditionScope.requiresCleanups())
727 ExitBlock = createBasicBlock("while.exit");
728 Builder.CreateCondBr(
729 BoolCondVal, LoopBody, ExitBlock,
730 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
732 if (ExitBlock != LoopExit.getBlock()) {
733 EmitBlock(ExitBlock);
734 EmitBranchThroughCleanup(LoopExit);
738 // Emit the loop body. We have to emit this in a cleanup scope
739 // because it might be a singleton DeclStmt.
741 RunCleanupsScope BodyScope(*this);
743 incrementProfileCounter(&S);
744 EmitStmt(S.getBody());
747 BreakContinueStack.pop_back();
749 // Immediately force cleanup.
750 ConditionScope.ForceCleanup();
753 // Branch to the loop header again.
754 EmitBranch(LoopHeader.getBlock());
758 // Emit the exit block.
759 EmitBlock(LoopExit.getBlock(), true);
761 // The LoopHeader typically is just a branch if we skipped emitting
762 // a branch, try to erase it.
763 if (!EmitBoolCondBranch)
764 SimplifyForwardingBlocks(LoopHeader.getBlock());
767 void CodeGenFunction::EmitDoStmt(const DoStmt &S,
768 ArrayRef<const Attr *> DoAttrs) {
769 JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
770 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
772 uint64_t ParentCount = getCurrentProfileCount();
774 // Store the blocks to use for break and continue.
775 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond));
777 // Emit the body of the loop.
778 llvm::BasicBlock *LoopBody = createBasicBlock("do.body");
780 const SourceRange &R = S.getSourceRange();
781 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs,
782 SourceLocToDebugLoc(R.getBegin()),
783 SourceLocToDebugLoc(R.getEnd()));
785 EmitBlockWithFallThrough(LoopBody, &S);
787 RunCleanupsScope BodyScope(*this);
788 EmitStmt(S.getBody());
791 EmitBlock(LoopCond.getBlock());
793 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
794 // after each execution of the loop body."
796 // Evaluate the conditional in the while header.
797 // C99 6.8.5p2/p4: The first substatement is executed if the expression
798 // compares unequal to 0. The condition must be a scalar type.
799 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
801 BreakContinueStack.pop_back();
803 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
804 // to correctly handle break/continue though.
805 bool EmitBoolCondBranch = true;
806 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
808 EmitBoolCondBranch = false;
810 // As long as the condition is true, iterate the loop.
811 if (EmitBoolCondBranch) {
812 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
813 Builder.CreateCondBr(
814 BoolCondVal, LoopBody, LoopExit.getBlock(),
815 createProfileWeightsForLoop(S.getCond(), BackedgeCount));
820 // Emit the exit block.
821 EmitBlock(LoopExit.getBlock());
823 // The DoCond block typically is just a branch if we skipped
824 // emitting a branch, try to erase it.
825 if (!EmitBoolCondBranch)
826 SimplifyForwardingBlocks(LoopCond.getBlock());
829 void CodeGenFunction::EmitForStmt(const ForStmt &S,
830 ArrayRef<const Attr *> ForAttrs) {
831 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
833 LexicalScope ForScope(*this, S.getSourceRange());
835 // Evaluate the first part before the loop.
837 EmitStmt(S.getInit());
839 // Start the loop with a block that tests the condition.
840 // If there's an increment, the continue scope will be overwritten
842 JumpDest Continue = getJumpDestInCurrentScope("for.cond");
843 llvm::BasicBlock *CondBlock = Continue.getBlock();
844 EmitBlock(CondBlock);
846 const SourceRange &R = S.getSourceRange();
847 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs,
848 SourceLocToDebugLoc(R.getBegin()),
849 SourceLocToDebugLoc(R.getEnd()));
851 // If the for loop doesn't have an increment we can just use the
852 // condition as the continue block. Otherwise we'll need to create
853 // a block for it (in the current scope, i.e. in the scope of the
854 // condition), and that we will become our continue block.
856 Continue = getJumpDestInCurrentScope("for.inc");
858 // Store the blocks to use for break and continue.
859 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
861 // Create a cleanup scope for the condition variable cleanups.
862 LexicalScope ConditionScope(*this, S.getSourceRange());
865 // If the for statement has a condition scope, emit the local variable
867 if (S.getConditionVariable()) {
868 EmitAutoVarDecl(*S.getConditionVariable());
871 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
872 // If there are any cleanups between here and the loop-exit scope,
873 // create a block to stage a loop exit along.
874 if (ForScope.requiresCleanups())
875 ExitBlock = createBasicBlock("for.cond.cleanup");
877 // As long as the condition is true, iterate the loop.
878 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
880 // C99 6.8.5p2/p4: The first substatement is executed if the expression
881 // compares unequal to 0. The condition must be a scalar type.
882 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
883 Builder.CreateCondBr(
884 BoolCondVal, ForBody, ExitBlock,
885 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
887 if (ExitBlock != LoopExit.getBlock()) {
888 EmitBlock(ExitBlock);
889 EmitBranchThroughCleanup(LoopExit);
894 // Treat it as a non-zero constant. Don't even create a new block for the
895 // body, just fall into it.
897 incrementProfileCounter(&S);
900 // Create a separate cleanup scope for the body, in case it is not
901 // a compound statement.
902 RunCleanupsScope BodyScope(*this);
903 EmitStmt(S.getBody());
906 // If there is an increment, emit it next.
908 EmitBlock(Continue.getBlock());
909 EmitStmt(S.getInc());
912 BreakContinueStack.pop_back();
914 ConditionScope.ForceCleanup();
917 EmitBranch(CondBlock);
919 ForScope.ForceCleanup();
923 // Emit the fall-through block.
924 EmitBlock(LoopExit.getBlock(), true);
928 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
929 ArrayRef<const Attr *> ForAttrs) {
930 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
932 LexicalScope ForScope(*this, S.getSourceRange());
934 // Evaluate the first pieces before the loop.
935 EmitStmt(S.getRangeStmt());
936 EmitStmt(S.getBeginStmt());
937 EmitStmt(S.getEndStmt());
939 // Start the loop with a block that tests the condition.
940 // If there's an increment, the continue scope will be overwritten
942 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond");
943 EmitBlock(CondBlock);
945 const SourceRange &R = S.getSourceRange();
946 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs,
947 SourceLocToDebugLoc(R.getBegin()),
948 SourceLocToDebugLoc(R.getEnd()));
950 // If there are any cleanups between here and the loop-exit scope,
951 // create a block to stage a loop exit along.
952 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
953 if (ForScope.requiresCleanups())
954 ExitBlock = createBasicBlock("for.cond.cleanup");
956 // The loop body, consisting of the specified body and the loop variable.
957 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
959 // The body is executed if the expression, contextually converted
961 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
962 Builder.CreateCondBr(
963 BoolCondVal, ForBody, ExitBlock,
964 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
966 if (ExitBlock != LoopExit.getBlock()) {
967 EmitBlock(ExitBlock);
968 EmitBranchThroughCleanup(LoopExit);
972 incrementProfileCounter(&S);
974 // Create a block for the increment. In case of a 'continue', we jump there.
975 JumpDest Continue = getJumpDestInCurrentScope("for.inc");
977 // Store the blocks to use for break and continue.
978 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
981 // Create a separate cleanup scope for the loop variable and body.
982 LexicalScope BodyScope(*this, S.getSourceRange());
983 EmitStmt(S.getLoopVarStmt());
984 EmitStmt(S.getBody());
988 // If there is an increment, emit it next.
989 EmitBlock(Continue.getBlock());
990 EmitStmt(S.getInc());
992 BreakContinueStack.pop_back();
994 EmitBranch(CondBlock);
996 ForScope.ForceCleanup();
1000 // Emit the fall-through block.
1001 EmitBlock(LoopExit.getBlock(), true);
1004 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
1005 if (RV.isScalar()) {
1006 Builder.CreateStore(RV.getScalarVal(), ReturnValue);
1007 } else if (RV.isAggregate()) {
1008 EmitAggregateCopy(ReturnValue, RV.getAggregateAddress(), Ty);
1010 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty),
1013 EmitBranchThroughCleanup(ReturnBlock);
1016 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
1017 /// if the function returns void, or may be missing one if the function returns
1018 /// non-void. Fun stuff :).
1019 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
1020 if (requiresReturnValueCheck()) {
1021 llvm::Constant *SLoc = EmitCheckSourceLocation(S.getLocStart());
1023 new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false,
1024 llvm::GlobalVariable::PrivateLinkage, SLoc);
1025 SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
1026 CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr);
1027 assert(ReturnLocation.isValid() && "No valid return location");
1028 Builder.CreateStore(Builder.CreateBitCast(SLocPtr, Int8PtrTy),
1032 // Returning from an outlined SEH helper is UB, and we already warn on it.
1033 if (IsOutlinedSEHHelper) {
1034 Builder.CreateUnreachable();
1035 Builder.ClearInsertionPoint();
1038 // Emit the result value, even if unused, to evalute the side effects.
1039 const Expr *RV = S.getRetValue();
1041 // Treat block literals in a return expression as if they appeared
1042 // in their own scope. This permits a small, easily-implemented
1043 // exception to our over-conservative rules about not jumping to
1044 // statements following block literals with non-trivial cleanups.
1045 RunCleanupsScope cleanupScope(*this);
1046 if (const ExprWithCleanups *cleanups =
1047 dyn_cast_or_null<ExprWithCleanups>(RV)) {
1048 enterFullExpression(cleanups);
1049 RV = cleanups->getSubExpr();
1052 // FIXME: Clean this up by using an LValue for ReturnTemp,
1053 // EmitStoreThroughLValue, and EmitAnyExpr.
1054 if (getLangOpts().ElideConstructors &&
1055 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) {
1056 // Apply the named return value optimization for this return statement,
1057 // which means doing nothing: the appropriate result has already been
1058 // constructed into the NRVO variable.
1060 // If there is an NRVO flag for this variable, set it to 1 into indicate
1061 // that the cleanup code should not destroy the variable.
1062 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1063 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag);
1064 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1065 // Make sure not to return anything, but evaluate the expression
1066 // for side effects.
1070 // Do nothing (return value is left uninitialized)
1071 } else if (FnRetTy->isReferenceType()) {
1072 // If this function returns a reference, take the address of the expression
1073 // rather than the value.
1074 RValue Result = EmitReferenceBindingToExpr(RV);
1075 Builder.CreateStore(Result.getScalarVal(), ReturnValue);
1077 switch (getEvaluationKind(RV->getType())) {
1079 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue);
1082 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()),
1086 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue,
1088 AggValueSlot::IsDestructed,
1089 AggValueSlot::DoesNotNeedGCBarriers,
1090 AggValueSlot::IsNotAliased));
1096 if (!RV || RV->isEvaluatable(getContext()))
1097 ++NumSimpleReturnExprs;
1099 cleanupScope.ForceCleanup();
1100 EmitBranchThroughCleanup(ReturnBlock);
1103 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1104 // As long as debug info is modeled with instructions, we have to ensure we
1105 // have a place to insert here and write the stop point here.
1106 if (HaveInsertPoint())
1109 for (const auto *I : S.decls())
1113 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1114 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1116 // If this code is reachable then emit a stop point (if generating
1117 // debug info). We have to do this ourselves because we are on the
1118 // "simple" statement path.
1119 if (HaveInsertPoint())
1122 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock);
1125 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1126 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1128 // If this code is reachable then emit a stop point (if generating
1129 // debug info). We have to do this ourselves because we are on the
1130 // "simple" statement path.
1131 if (HaveInsertPoint())
1134 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock);
1137 /// EmitCaseStmtRange - If case statement range is not too big then
1138 /// add multiple cases to switch instruction, one for each value within
1139 /// the range. If range is too big then emit "if" condition check.
1140 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) {
1141 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1143 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext());
1144 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext());
1146 // Emit the code for this case. We do this first to make sure it is
1147 // properly chained from our predecessor before generating the
1148 // switch machinery to enter this block.
1149 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1150 EmitBlockWithFallThrough(CaseDest, &S);
1151 EmitStmt(S.getSubStmt());
1153 // If range is empty, do nothing.
1154 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS))
1157 llvm::APInt Range = RHS - LHS;
1158 // FIXME: parameters such as this should not be hardcoded.
1159 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) {
1160 // Range is small enough to add multiple switch instruction cases.
1161 uint64_t Total = getProfileCount(&S);
1162 unsigned NCases = Range.getZExtValue() + 1;
1163 // We only have one region counter for the entire set of cases here, so we
1164 // need to divide the weights evenly between the generated cases, ensuring
1165 // that the total weight is preserved. E.g., a weight of 5 over three cases
1166 // will be distributed as weights of 2, 2, and 1.
1167 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1168 for (unsigned I = 0; I != NCases; ++I) {
1170 SwitchWeights->push_back(Weight + (Rem ? 1 : 0));
1173 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest);
1179 // The range is too big. Emit "if" condition into a new block,
1180 // making sure to save and restore the current insertion point.
1181 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1183 // Push this test onto the chain of range checks (which terminates
1184 // in the default basic block). The switch's default will be changed
1185 // to the top of this chain after switch emission is complete.
1186 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1187 CaseRangeBlock = createBasicBlock("sw.caserange");
1189 CurFn->getBasicBlockList().push_back(CaseRangeBlock);
1190 Builder.SetInsertPoint(CaseRangeBlock);
1192 // Emit range check.
1194 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS));
1196 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds");
1198 llvm::MDNode *Weights = nullptr;
1199 if (SwitchWeights) {
1200 uint64_t ThisCount = getProfileCount(&S);
1201 uint64_t DefaultCount = (*SwitchWeights)[0];
1202 Weights = createProfileWeights(ThisCount, DefaultCount);
1204 // Since we're chaining the switch default through each large case range, we
1205 // need to update the weight for the default, ie, the first case, to include
1207 (*SwitchWeights)[0] += ThisCount;
1209 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights);
1211 // Restore the appropriate insertion point.
1213 Builder.SetInsertPoint(RestoreBB);
1215 Builder.ClearInsertionPoint();
1218 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) {
1219 // If there is no enclosing switch instance that we're aware of, then this
1220 // case statement and its block can be elided. This situation only happens
1221 // when we've constant-folded the switch, are emitting the constant case,
1222 // and part of the constant case includes another case statement. For
1223 // instance: switch (4) { case 4: do { case 5: } while (1); }
1225 EmitStmt(S.getSubStmt());
1229 // Handle case ranges.
1231 EmitCaseStmtRange(S);
1235 llvm::ConstantInt *CaseVal =
1236 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext()));
1238 // If the body of the case is just a 'break', try to not emit an empty block.
1239 // If we're profiling or we're not optimizing, leave the block in for better
1240 // debug and coverage analysis.
1241 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1242 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1243 isa<BreakStmt>(S.getSubStmt())) {
1244 JumpDest Block = BreakContinueStack.back().BreakBlock;
1246 // Only do this optimization if there are no cleanups that need emitting.
1247 if (isObviouslyBranchWithoutCleanups(Block)) {
1249 SwitchWeights->push_back(getProfileCount(&S));
1250 SwitchInsn->addCase(CaseVal, Block.getBlock());
1252 // If there was a fallthrough into this case, make sure to redirect it to
1253 // the end of the switch as well.
1254 if (Builder.GetInsertBlock()) {
1255 Builder.CreateBr(Block.getBlock());
1256 Builder.ClearInsertionPoint();
1262 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1263 EmitBlockWithFallThrough(CaseDest, &S);
1265 SwitchWeights->push_back(getProfileCount(&S));
1266 SwitchInsn->addCase(CaseVal, CaseDest);
1268 // Recursively emitting the statement is acceptable, but is not wonderful for
1269 // code where we have many case statements nested together, i.e.:
1273 // Handling this recursively will create a new block for each case statement
1274 // that falls through to the next case which is IR intensive. It also causes
1275 // deep recursion which can run into stack depth limitations. Handle
1276 // sequential non-range case statements specially.
1277 const CaseStmt *CurCase = &S;
1278 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt());
1280 // Otherwise, iteratively add consecutive cases to this switch stmt.
1281 while (NextCase && NextCase->getRHS() == nullptr) {
1283 llvm::ConstantInt *CaseVal =
1284 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext()));
1287 SwitchWeights->push_back(getProfileCount(NextCase));
1288 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1289 CaseDest = createBasicBlock("sw.bb");
1290 EmitBlockWithFallThrough(CaseDest, &S);
1293 SwitchInsn->addCase(CaseVal, CaseDest);
1294 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt());
1297 // Normal default recursion for non-cases.
1298 EmitStmt(CurCase->getSubStmt());
1301 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) {
1302 // If there is no enclosing switch instance that we're aware of, then this
1303 // default statement can be elided. This situation only happens when we've
1304 // constant-folded the switch.
1306 EmitStmt(S.getSubStmt());
1310 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1311 assert(DefaultBlock->empty() &&
1312 "EmitDefaultStmt: Default block already defined?");
1314 EmitBlockWithFallThrough(DefaultBlock, &S);
1316 EmitStmt(S.getSubStmt());
1319 /// CollectStatementsForCase - Given the body of a 'switch' statement and a
1320 /// constant value that is being switched on, see if we can dead code eliminate
1321 /// the body of the switch to a simple series of statements to emit. Basically,
1322 /// on a switch (5) we want to find these statements:
1324 /// printf(...); <--
1328 /// and add them to the ResultStmts vector. If it is unsafe to do this
1329 /// transformation (for example, one of the elided statements contains a label
1330 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
1331 /// should include statements after it (e.g. the printf() line is a substmt of
1332 /// the case) then return CSFC_FallThrough. If we handled it and found a break
1333 /// statement, then return CSFC_Success.
1335 /// If Case is non-null, then we are looking for the specified case, checking
1336 /// that nothing we jump over contains labels. If Case is null, then we found
1337 /// the case and are looking for the break.
1339 /// If the recursive walk actually finds our Case, then we set FoundCase to
1342 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
1343 static CSFC_Result CollectStatementsForCase(const Stmt *S,
1344 const SwitchCase *Case,
1346 SmallVectorImpl<const Stmt*> &ResultStmts) {
1347 // If this is a null statement, just succeed.
1349 return Case ? CSFC_Success : CSFC_FallThrough;
1351 // If this is the switchcase (case 4: or default) that we're looking for, then
1352 // we're in business. Just add the substatement.
1353 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) {
1356 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase,
1360 // Otherwise, this is some other case or default statement, just ignore it.
1361 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase,
1365 // If we are in the live part of the code and we found our break statement,
1366 // return a success!
1367 if (!Case && isa<BreakStmt>(S))
1368 return CSFC_Success;
1370 // If this is a switch statement, then it might contain the SwitchCase, the
1371 // break, or neither.
1372 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) {
1373 // Handle this as two cases: we might be looking for the SwitchCase (if so
1374 // the skipped statements must be skippable) or we might already have it.
1375 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
1376 bool StartedInLiveCode = FoundCase;
1377 unsigned StartSize = ResultStmts.size();
1379 // If we've not found the case yet, scan through looking for it.
1381 // Keep track of whether we see a skipped declaration. The code could be
1382 // using the declaration even if it is skipped, so we can't optimize out
1383 // the decl if the kept statements might refer to it.
1384 bool HadSkippedDecl = false;
1386 // If we're looking for the case, just see if we can skip each of the
1388 for (; Case && I != E; ++I) {
1389 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I);
1391 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) {
1392 case CSFC_Failure: return CSFC_Failure;
1394 // A successful result means that either 1) that the statement doesn't
1395 // have the case and is skippable, or 2) does contain the case value
1396 // and also contains the break to exit the switch. In the later case,
1397 // we just verify the rest of the statements are elidable.
1399 // If we found the case and skipped declarations, we can't do the
1402 return CSFC_Failure;
1404 for (++I; I != E; ++I)
1405 if (CodeGenFunction::ContainsLabel(*I, true))
1406 return CSFC_Failure;
1407 return CSFC_Success;
1410 case CSFC_FallThrough:
1411 // If we have a fallthrough condition, then we must have found the
1412 // case started to include statements. Consider the rest of the
1413 // statements in the compound statement as candidates for inclusion.
1414 assert(FoundCase && "Didn't find case but returned fallthrough?");
1415 // We recursively found Case, so we're not looking for it anymore.
1418 // If we found the case and skipped declarations, we can't do the
1421 return CSFC_Failure;
1427 return CSFC_Success;
1429 assert(!HadSkippedDecl && "fallthrough after skipping decl");
1432 // If we have statements in our range, then we know that the statements are
1433 // live and need to be added to the set of statements we're tracking.
1434 bool AnyDecls = false;
1435 for (; I != E; ++I) {
1436 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I);
1438 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) {
1439 case CSFC_Failure: return CSFC_Failure;
1440 case CSFC_FallThrough:
1441 // A fallthrough result means that the statement was simple and just
1442 // included in ResultStmt, keep adding them afterwards.
1445 // A successful result means that we found the break statement and
1446 // stopped statement inclusion. We just ensure that any leftover stmts
1447 // are skippable and return success ourselves.
1448 for (++I; I != E; ++I)
1449 if (CodeGenFunction::ContainsLabel(*I, true))
1450 return CSFC_Failure;
1451 return CSFC_Success;
1455 // If we're about to fall out of a scope without hitting a 'break;', we
1456 // can't perform the optimization if there were any decls in that scope
1457 // (we'd lose their end-of-lifetime).
1459 // If the entire compound statement was live, there's one more thing we
1460 // can try before giving up: emit the whole thing as a single statement.
1461 // We can do that unless the statement contains a 'break;'.
1462 // FIXME: Such a break must be at the end of a construct within this one.
1463 // We could emit this by just ignoring the BreakStmts entirely.
1464 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) {
1465 ResultStmts.resize(StartSize);
1466 ResultStmts.push_back(S);
1468 return CSFC_Failure;
1472 return CSFC_FallThrough;
1475 // Okay, this is some other statement that we don't handle explicitly, like a
1476 // for statement or increment etc. If we are skipping over this statement,
1477 // just verify it doesn't have labels, which would make it invalid to elide.
1479 if (CodeGenFunction::ContainsLabel(S, true))
1480 return CSFC_Failure;
1481 return CSFC_Success;
1484 // Otherwise, we want to include this statement. Everything is cool with that
1485 // so long as it doesn't contain a break out of the switch we're in.
1486 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
1488 // Otherwise, everything is great. Include the statement and tell the caller
1489 // that we fall through and include the next statement as well.
1490 ResultStmts.push_back(S);
1491 return CSFC_FallThrough;
1494 /// FindCaseStatementsForValue - Find the case statement being jumped to and
1495 /// then invoke CollectStatementsForCase to find the list of statements to emit
1496 /// for a switch on constant. See the comment above CollectStatementsForCase
1497 /// for more details.
1498 static bool FindCaseStatementsForValue(const SwitchStmt &S,
1499 const llvm::APSInt &ConstantCondValue,
1500 SmallVectorImpl<const Stmt*> &ResultStmts,
1502 const SwitchCase *&ResultCase) {
1503 // First step, find the switch case that is being branched to. We can do this
1504 // efficiently by scanning the SwitchCase list.
1505 const SwitchCase *Case = S.getSwitchCaseList();
1506 const DefaultStmt *DefaultCase = nullptr;
1508 for (; Case; Case = Case->getNextSwitchCase()) {
1509 // It's either a default or case. Just remember the default statement in
1510 // case we're not jumping to any numbered cases.
1511 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) {
1516 // Check to see if this case is the one we're looking for.
1517 const CaseStmt *CS = cast<CaseStmt>(Case);
1518 // Don't handle case ranges yet.
1519 if (CS->getRHS()) return false;
1521 // If we found our case, remember it as 'case'.
1522 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue)
1526 // If we didn't find a matching case, we use a default if it exists, or we
1527 // elide the whole switch body!
1529 // It is safe to elide the body of the switch if it doesn't contain labels
1530 // etc. If it is safe, return successfully with an empty ResultStmts list.
1532 return !CodeGenFunction::ContainsLabel(&S);
1536 // Ok, we know which case is being jumped to, try to collect all the
1537 // statements that follow it. This can fail for a variety of reasons. Also,
1538 // check to see that the recursive walk actually found our case statement.
1539 // Insane cases like this can fail to find it in the recursive walk since we
1540 // don't handle every stmt kind:
1544 bool FoundCase = false;
1546 return CollectStatementsForCase(S.getBody(), Case, FoundCase,
1547 ResultStmts) != CSFC_Failure &&
1551 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
1552 // Handle nested switch statements.
1553 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
1554 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
1555 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
1557 // See if we can constant fold the condition of the switch and therefore only
1558 // emit the live case statement (if any) of the switch.
1559 llvm::APSInt ConstantCondValue;
1560 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) {
1561 SmallVector<const Stmt*, 4> CaseStmts;
1562 const SwitchCase *Case = nullptr;
1563 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts,
1564 getContext(), Case)) {
1566 incrementProfileCounter(Case);
1567 RunCleanupsScope ExecutedScope(*this);
1570 EmitStmt(S.getInit());
1572 // Emit the condition variable if needed inside the entire cleanup scope
1573 // used by this special case for constant folded switches.
1574 if (S.getConditionVariable())
1575 EmitAutoVarDecl(*S.getConditionVariable());
1577 // At this point, we are no longer "within" a switch instance, so
1578 // we can temporarily enforce this to ensure that any embedded case
1579 // statements are not emitted.
1580 SwitchInsn = nullptr;
1582 // Okay, we can dead code eliminate everything except this case. Emit the
1583 // specified series of statements and we're good.
1584 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i)
1585 EmitStmt(CaseStmts[i]);
1586 incrementProfileCounter(&S);
1588 // Now we want to restore the saved switch instance so that nested
1589 // switches continue to function properly
1590 SwitchInsn = SavedSwitchInsn;
1596 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog");
1598 RunCleanupsScope ConditionScope(*this);
1601 EmitStmt(S.getInit());
1603 if (S.getConditionVariable())
1604 EmitAutoVarDecl(*S.getConditionVariable());
1605 llvm::Value *CondV = EmitScalarExpr(S.getCond());
1607 // Create basic block to hold stuff that comes after switch
1608 // statement. We also need to create a default block now so that
1609 // explicit case ranges tests can have a place to jump to on
1611 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default");
1612 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock);
1613 if (PGO.haveRegionCounts()) {
1614 // Walk the SwitchCase list to find how many there are.
1615 uint64_t DefaultCount = 0;
1616 unsigned NumCases = 0;
1617 for (const SwitchCase *Case = S.getSwitchCaseList();
1619 Case = Case->getNextSwitchCase()) {
1620 if (isa<DefaultStmt>(Case))
1621 DefaultCount = getProfileCount(Case);
1624 SwitchWeights = new SmallVector<uint64_t, 16>();
1625 SwitchWeights->reserve(NumCases);
1626 // The default needs to be first. We store the edge count, so we already
1627 // know the right weight.
1628 SwitchWeights->push_back(DefaultCount);
1630 CaseRangeBlock = DefaultBlock;
1632 // Clear the insertion point to indicate we are in unreachable code.
1633 Builder.ClearInsertionPoint();
1635 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
1636 // then reuse last ContinueBlock.
1637 JumpDest OuterContinue;
1638 if (!BreakContinueStack.empty())
1639 OuterContinue = BreakContinueStack.back().ContinueBlock;
1641 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue));
1643 // Emit switch body.
1644 EmitStmt(S.getBody());
1646 BreakContinueStack.pop_back();
1648 // Update the default block in case explicit case range tests have
1649 // been chained on top.
1650 SwitchInsn->setDefaultDest(CaseRangeBlock);
1652 // If a default was never emitted:
1653 if (!DefaultBlock->getParent()) {
1654 // If we have cleanups, emit the default block so that there's a
1655 // place to jump through the cleanups from.
1656 if (ConditionScope.requiresCleanups()) {
1657 EmitBlock(DefaultBlock);
1659 // Otherwise, just forward the default block to the switch end.
1661 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock());
1662 delete DefaultBlock;
1666 ConditionScope.ForceCleanup();
1668 // Emit continuation.
1669 EmitBlock(SwitchExit.getBlock(), true);
1670 incrementProfileCounter(&S);
1672 // If the switch has a condition wrapped by __builtin_unpredictable,
1673 // create metadata that specifies that the switch is unpredictable.
1674 // Don't bother if not optimizing because that metadata would not be used.
1675 auto *Call = dyn_cast<CallExpr>(S.getCond());
1676 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1677 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1678 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1679 llvm::MDBuilder MDHelper(getLLVMContext());
1680 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable,
1681 MDHelper.createUnpredictable());
1685 if (SwitchWeights) {
1686 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
1687 "switch weights do not match switch cases");
1688 // If there's only one jump destination there's no sense weighting it.
1689 if (SwitchWeights->size() > 1)
1690 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1691 createProfileWeights(*SwitchWeights));
1692 delete SwitchWeights;
1694 SwitchInsn = SavedSwitchInsn;
1695 SwitchWeights = SavedSwitchWeights;
1696 CaseRangeBlock = SavedCRBlock;
1700 SimplifyConstraint(const char *Constraint, const TargetInfo &Target,
1701 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) {
1704 while (*Constraint) {
1705 switch (*Constraint) {
1707 Result += Target.convertConstraint(Constraint);
1713 case '=': // Will see this and the following in mult-alt constraints.
1716 case '#': // Ignore the rest of the constraint alternative.
1717 while (Constraint[1] && Constraint[1] != ',')
1722 Result += *Constraint;
1723 while (Constraint[1] && Constraint[1] == *Constraint)
1734 "Must pass output names to constraints with a symbolic name");
1736 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index);
1737 assert(result && "Could not resolve symbolic name"); (void)result;
1738 Result += llvm::utostr(Index);
1749 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared
1750 /// as using a particular register add that as a constraint that will be used
1751 /// in this asm stmt.
1753 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr,
1754 const TargetInfo &Target, CodeGenModule &CGM,
1755 const AsmStmt &Stmt, const bool EarlyClobber) {
1756 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr);
1759 const ValueDecl &Value = *AsmDeclRef->getDecl();
1760 const VarDecl *Variable = dyn_cast<VarDecl>(&Value);
1763 if (Variable->getStorageClass() != SC_Register)
1765 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>();
1768 StringRef Register = Attr->getLabel();
1769 assert(Target.isValidGCCRegisterName(Register));
1770 // We're using validateOutputConstraint here because we only care if
1771 // this is a register constraint.
1772 TargetInfo::ConstraintInfo Info(Constraint, "");
1773 if (Target.validateOutputConstraint(Info) &&
1774 !Info.allowsRegister()) {
1775 CGM.ErrorUnsupported(&Stmt, "__asm__");
1778 // Canonicalize the register here before returning it.
1779 Register = Target.getNormalizedGCCRegisterName(Register);
1780 return (EarlyClobber ? "&{" : "{") + Register.str() + "}";
1784 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info,
1785 LValue InputValue, QualType InputType,
1786 std::string &ConstraintStr,
1787 SourceLocation Loc) {
1789 if (Info.allowsRegister() || !Info.allowsMemory()) {
1790 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) {
1791 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal();
1793 llvm::Type *Ty = ConvertType(InputType);
1794 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
1795 if (Size <= 64 && llvm::isPowerOf2_64(Size)) {
1796 Ty = llvm::IntegerType::get(getLLVMContext(), Size);
1797 Ty = llvm::PointerType::getUnqual(Ty);
1799 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(),
1802 Arg = InputValue.getPointer();
1803 ConstraintStr += '*';
1807 Arg = InputValue.getPointer();
1808 ConstraintStr += '*';
1814 llvm::Value* CodeGenFunction::EmitAsmInput(
1815 const TargetInfo::ConstraintInfo &Info,
1816 const Expr *InputExpr,
1817 std::string &ConstraintStr) {
1818 // If this can't be a register or memory, i.e., has to be a constant
1819 // (immediate or symbolic), try to emit it as such.
1820 if (!Info.allowsRegister() && !Info.allowsMemory()) {
1821 llvm::APSInt Result;
1822 if (InputExpr->EvaluateAsInt(Result, getContext()))
1823 return llvm::ConstantInt::get(getLLVMContext(), Result);
1824 assert(!Info.requiresImmediateConstant() &&
1825 "Required-immediate inlineasm arg isn't constant?");
1828 if (Info.allowsRegister() || !Info.allowsMemory())
1829 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType()))
1830 return EmitScalarExpr(InputExpr);
1831 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
1832 return EmitScalarExpr(InputExpr);
1833 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext());
1834 LValue Dest = EmitLValue(InputExpr);
1835 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr,
1836 InputExpr->getExprLoc());
1839 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
1840 /// asm call instruction. The !srcloc MDNode contains a list of constant
1841 /// integers which are the source locations of the start of each line in the
1843 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
1844 CodeGenFunction &CGF) {
1845 SmallVector<llvm::Metadata *, 8> Locs;
1846 // Add the location of the first line to the MDNode.
1847 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
1848 CGF.Int32Ty, Str->getLocStart().getRawEncoding())));
1849 StringRef StrVal = Str->getString();
1850 if (!StrVal.empty()) {
1851 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
1852 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
1853 unsigned StartToken = 0;
1854 unsigned ByteOffset = 0;
1856 // Add the location of the start of each subsequent line of the asm to the
1858 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
1859 if (StrVal[i] != '\n') continue;
1860 SourceLocation LineLoc = Str->getLocationOfByte(
1861 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset);
1862 Locs.push_back(llvm::ConstantAsMetadata::get(
1863 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding())));
1867 return llvm::MDNode::get(CGF.getLLVMContext(), Locs);
1870 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
1871 // Assemble the final asm string.
1872 std::string AsmString = S.generateAsmString(getContext());
1874 // Get all the output and input constraints together.
1875 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
1876 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
1878 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1880 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1881 Name = GAS->getOutputName(i);
1882 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
1883 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
1884 assert(IsValid && "Failed to parse output constraint");
1885 OutputConstraintInfos.push_back(Info);
1888 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1890 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1891 Name = GAS->getInputName(i);
1892 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
1894 getTarget().validateInputConstraint(OutputConstraintInfos, Info);
1895 assert(IsValid && "Failed to parse input constraint"); (void)IsValid;
1896 InputConstraintInfos.push_back(Info);
1899 std::string Constraints;
1901 std::vector<LValue> ResultRegDests;
1902 std::vector<QualType> ResultRegQualTys;
1903 std::vector<llvm::Type *> ResultRegTypes;
1904 std::vector<llvm::Type *> ResultTruncRegTypes;
1905 std::vector<llvm::Type *> ArgTypes;
1906 std::vector<llvm::Value*> Args;
1908 // Keep track of inout constraints.
1909 std::string InOutConstraints;
1910 std::vector<llvm::Value*> InOutArgs;
1911 std::vector<llvm::Type*> InOutArgTypes;
1913 // An inline asm can be marked readonly if it meets the following conditions:
1914 // - it doesn't have any sideeffects
1915 // - it doesn't clobber memory
1916 // - it doesn't return a value by-reference
1917 // It can be marked readnone if it doesn't have any input memory constraints
1918 // in addition to meeting the conditions listed above.
1919 bool ReadOnly = true, ReadNone = true;
1921 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1922 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
1924 // Simplify the output constraint.
1925 std::string OutputConstraint(S.getOutputConstraint(i));
1926 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1,
1929 const Expr *OutExpr = S.getOutputExpr(i);
1930 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext());
1932 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr,
1933 getTarget(), CGM, S,
1934 Info.earlyClobber());
1936 LValue Dest = EmitLValue(OutExpr);
1937 if (!Constraints.empty())
1940 // If this is a register output, then make the inline asm return it
1941 // by-value. If this is a memory result, return the value by-reference.
1942 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) {
1943 Constraints += "=" + OutputConstraint;
1944 ResultRegQualTys.push_back(OutExpr->getType());
1945 ResultRegDests.push_back(Dest);
1946 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType()));
1947 ResultTruncRegTypes.push_back(ResultRegTypes.back());
1949 // If this output is tied to an input, and if the input is larger, then
1950 // we need to set the actual result type of the inline asm node to be the
1951 // same as the input type.
1952 if (Info.hasMatchingInput()) {
1954 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
1955 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
1956 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
1959 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
1961 QualType InputTy = S.getInputExpr(InputNo)->getType();
1962 QualType OutputType = OutExpr->getType();
1964 uint64_t InputSize = getContext().getTypeSize(InputTy);
1965 if (getContext().getTypeSize(OutputType) < InputSize) {
1966 // Form the asm to return the value as a larger integer or fp type.
1967 ResultRegTypes.back() = ConvertType(InputTy);
1970 if (llvm::Type* AdjTy =
1971 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1972 ResultRegTypes.back()))
1973 ResultRegTypes.back() = AdjTy;
1975 CGM.getDiags().Report(S.getAsmLoc(),
1976 diag::err_asm_invalid_type_in_input)
1977 << OutExpr->getType() << OutputConstraint;
1980 ArgTypes.push_back(Dest.getAddress().getType());
1981 Args.push_back(Dest.getPointer());
1982 Constraints += "=*";
1983 Constraints += OutputConstraint;
1984 ReadOnly = ReadNone = false;
1987 if (Info.isReadWrite()) {
1988 InOutConstraints += ',';
1990 const Expr *InputExpr = S.getOutputExpr(i);
1991 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
1993 InputExpr->getExprLoc());
1995 if (llvm::Type* AdjTy =
1996 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1998 Arg = Builder.CreateBitCast(Arg, AdjTy);
2000 if (Info.allowsRegister())
2001 InOutConstraints += llvm::utostr(i);
2003 InOutConstraints += OutputConstraint;
2005 InOutArgTypes.push_back(Arg->getType());
2006 InOutArgs.push_back(Arg);
2010 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
2011 // to the return value slot. Only do this when returning in registers.
2012 if (isa<MSAsmStmt>(&S)) {
2013 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
2014 if (RetAI.isDirect() || RetAI.isExtend()) {
2015 // Make a fake lvalue for the return value slot.
2016 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy);
2017 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs(
2018 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes,
2019 ResultRegDests, AsmString, S.getNumOutputs());
2024 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
2025 const Expr *InputExpr = S.getInputExpr(i);
2027 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
2029 if (Info.allowsMemory())
2032 if (!Constraints.empty())
2035 // Simplify the input constraint.
2036 std::string InputConstraint(S.getInputConstraint(i));
2037 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(),
2038 &OutputConstraintInfos);
2040 InputConstraint = AddVariableConstraints(
2041 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()),
2042 getTarget(), CGM, S, false /* No EarlyClobber */);
2044 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints);
2046 // If this input argument is tied to a larger output result, extend the
2047 // input to be the same size as the output. The LLVM backend wants to see
2048 // the input and output of a matching constraint be the same size. Note
2049 // that GCC does not define what the top bits are here. We use zext because
2050 // that is usually cheaper, but LLVM IR should really get an anyext someday.
2051 if (Info.hasTiedOperand()) {
2052 unsigned Output = Info.getTiedOperand();
2053 QualType OutputType = S.getOutputExpr(Output)->getType();
2054 QualType InputTy = InputExpr->getType();
2056 if (getContext().getTypeSize(OutputType) >
2057 getContext().getTypeSize(InputTy)) {
2058 // Use ptrtoint as appropriate so that we can do our extension.
2059 if (isa<llvm::PointerType>(Arg->getType()))
2060 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy);
2061 llvm::Type *OutputTy = ConvertType(OutputType);
2062 if (isa<llvm::IntegerType>(OutputTy))
2063 Arg = Builder.CreateZExt(Arg, OutputTy);
2064 else if (isa<llvm::PointerType>(OutputTy))
2065 Arg = Builder.CreateZExt(Arg, IntPtrTy);
2067 assert(OutputTy->isFloatingPointTy() && "Unexpected output type");
2068 Arg = Builder.CreateFPExt(Arg, OutputTy);
2072 if (llvm::Type* AdjTy =
2073 getTargetHooks().adjustInlineAsmType(*this, InputConstraint,
2075 Arg = Builder.CreateBitCast(Arg, AdjTy);
2077 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input)
2078 << InputExpr->getType() << InputConstraint;
2080 ArgTypes.push_back(Arg->getType());
2081 Args.push_back(Arg);
2082 Constraints += InputConstraint;
2085 // Append the "input" part of inout constraints last.
2086 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) {
2087 ArgTypes.push_back(InOutArgTypes[i]);
2088 Args.push_back(InOutArgs[i]);
2090 Constraints += InOutConstraints;
2093 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) {
2094 StringRef Clobber = S.getClobber(i);
2096 if (Clobber == "memory")
2097 ReadOnly = ReadNone = false;
2098 else if (Clobber != "cc")
2099 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber);
2101 if (!Constraints.empty())
2104 Constraints += "~{";
2105 Constraints += Clobber;
2109 // Add machine specific clobbers
2110 std::string MachineClobbers = getTarget().getClobbers();
2111 if (!MachineClobbers.empty()) {
2112 if (!Constraints.empty())
2114 Constraints += MachineClobbers;
2117 llvm::Type *ResultType;
2118 if (ResultRegTypes.empty())
2119 ResultType = VoidTy;
2120 else if (ResultRegTypes.size() == 1)
2121 ResultType = ResultRegTypes[0];
2123 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes);
2125 llvm::FunctionType *FTy =
2126 llvm::FunctionType::get(ResultType, ArgTypes, false);
2128 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0;
2129 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ?
2130 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT;
2131 llvm::InlineAsm *IA =
2132 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect,
2133 /* IsAlignStack */ false, AsmDialect);
2134 llvm::CallInst *Result = Builder.CreateCall(IA, Args);
2135 Result->addAttribute(llvm::AttributeList::FunctionIndex,
2136 llvm::Attribute::NoUnwind);
2138 // Attach readnone and readonly attributes.
2139 if (!HasSideEffect) {
2141 Result->addAttribute(llvm::AttributeList::FunctionIndex,
2142 llvm::Attribute::ReadNone);
2144 Result->addAttribute(llvm::AttributeList::FunctionIndex,
2145 llvm::Attribute::ReadOnly);
2148 // Slap the source location of the inline asm into a !srcloc metadata on the
2150 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) {
2151 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(),
2154 // At least put the line number on MS inline asm blobs.
2155 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding());
2156 Result->setMetadata("srcloc",
2157 llvm::MDNode::get(getLLVMContext(),
2158 llvm::ConstantAsMetadata::get(Loc)));
2161 if (getLangOpts().assumeFunctionsAreConvergent()) {
2162 // Conservatively, mark all inline asm blocks in CUDA or OpenCL as
2163 // convergent (meaning, they may call an intrinsically convergent op, such
2164 // as bar.sync, and so can't have certain optimizations applied around
2166 Result->addAttribute(llvm::AttributeList::FunctionIndex,
2167 llvm::Attribute::Convergent);
2170 // Extract all of the register value results from the asm.
2171 std::vector<llvm::Value*> RegResults;
2172 if (ResultRegTypes.size() == 1) {
2173 RegResults.push_back(Result);
2175 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2176 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult");
2177 RegResults.push_back(Tmp);
2181 assert(RegResults.size() == ResultRegTypes.size());
2182 assert(RegResults.size() == ResultTruncRegTypes.size());
2183 assert(RegResults.size() == ResultRegDests.size());
2184 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) {
2185 llvm::Value *Tmp = RegResults[i];
2187 // If the result type of the LLVM IR asm doesn't match the result type of
2188 // the expression, do the conversion.
2189 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) {
2190 llvm::Type *TruncTy = ResultTruncRegTypes[i];
2192 // Truncate the integer result to the right size, note that TruncTy can be
2194 if (TruncTy->isFloatingPointTy())
2195 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy);
2196 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) {
2197 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy);
2198 Tmp = Builder.CreateTrunc(Tmp,
2199 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize));
2200 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy);
2201 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) {
2202 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType());
2203 Tmp = Builder.CreatePtrToInt(Tmp,
2204 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize));
2205 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2206 } else if (TruncTy->isIntegerTy()) {
2207 Tmp = Builder.CreateZExtOrTrunc(Tmp, TruncTy);
2208 } else if (TruncTy->isVectorTy()) {
2209 Tmp = Builder.CreateBitCast(Tmp, TruncTy);
2213 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]);
2217 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) {
2218 const RecordDecl *RD = S.getCapturedRecordDecl();
2219 QualType RecordTy = getContext().getRecordType(RD);
2221 // Initialize the captured struct.
2223 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy);
2225 RecordDecl::field_iterator CurField = RD->field_begin();
2226 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
2227 E = S.capture_init_end();
2228 I != E; ++I, ++CurField) {
2229 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField);
2230 if (CurField->hasCapturedVLAType()) {
2231 auto VAT = CurField->getCapturedVLAType();
2232 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV);
2234 EmitInitializerForField(*CurField, LV, *I);
2241 /// Generate an outlined function for the body of a CapturedStmt, store any
2242 /// captured variables into the captured struct, and call the outlined function.
2244 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) {
2245 LValue CapStruct = InitCapturedStruct(S);
2247 // Emit the CapturedDecl
2248 CodeGenFunction CGF(CGM, true);
2249 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K));
2250 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S);
2251 delete CGF.CapturedStmtInfo;
2253 // Emit call to the helper function.
2254 EmitCallOrInvoke(F, CapStruct.getPointer());
2259 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) {
2260 LValue CapStruct = InitCapturedStruct(S);
2261 return CapStruct.getAddress();
2264 /// Creates the outlined function for a CapturedStmt.
2266 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) {
2267 assert(CapturedStmtInfo &&
2268 "CapturedStmtInfo should be set when generating the captured function");
2269 const CapturedDecl *CD = S.getCapturedDecl();
2270 const RecordDecl *RD = S.getCapturedRecordDecl();
2271 SourceLocation Loc = S.getLocStart();
2272 assert(CD->hasBody() && "missing CapturedDecl body");
2274 // Build the argument list.
2275 ASTContext &Ctx = CGM.getContext();
2276 FunctionArgList Args;
2277 Args.append(CD->param_begin(), CD->param_end());
2279 // Create the function declaration.
2280 const CGFunctionInfo &FuncInfo =
2281 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args);
2282 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
2285 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
2286 CapturedStmtInfo->getHelperName(), &CGM.getModule());
2287 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
2288 if (CD->isNothrow())
2289 F->addFnAttr(llvm::Attribute::NoUnwind);
2291 // Generate the function.
2292 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args,
2294 CD->getBody()->getLocStart());
2295 // Set the context parameter in CapturedStmtInfo.
2296 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam());
2297 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr));
2299 // Initialize variable-length arrays.
2300 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(),
2301 Ctx.getTagDeclType(RD));
2302 for (auto *FD : RD->fields()) {
2303 if (FD->hasCapturedVLAType()) {
2304 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD),
2305 S.getLocStart()).getScalarVal();
2306 auto VAT = FD->getCapturedVLAType();
2307 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
2311 // If 'this' is captured, load it into CXXThisValue.
2312 if (CapturedStmtInfo->isCXXThisExprCaptured()) {
2313 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl();
2314 LValue ThisLValue = EmitLValueForField(Base, FD);
2315 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal();
2318 PGO.assignRegionCounters(GlobalDecl(CD), F);
2319 CapturedStmtInfo->EmitBody(*this, CD->getBody());
2320 FinishFunction(CD->getBodyRBrace());