[Clang] Improve compile times when forming a DeclRef outside of a capturing scope.

[cor3ntin]

Summary

The logic of whether an entity needs to be captured has become
quite complex and the recent changes in https://reviews.llvm.org/D124351
ad a mesurable negative impact on compile times.

However, in the absence of capturing scopes (lambda, block, region)
we usually can avoid running most of that logic
(except that we do need to diagnostic when a nested function
refers to a local variable in the scope of the outer function.).

This patch track whether there is currently an active capturing
scope and exit tryCaptureVariable early when there isn't.

[aaron.ballman]

Do you have some performance measurement numbers for how much benefit we get from the changes?

clang/lib/Sema/SemaExpr.cpp

19195

19199

[cor3ntin]

In D150038#4326549, @aaron.ballman wrote:

Do you have some performance measurement numbers for how much benefit we get from the changes?

Not really.
The PR that changed the scope of trailing return type had a 0.35% regression, I expect this to be in the same ballpark in the other direction
http://llvm-compile-time-tracker.com/compare.php?from=cd173cbd7cca69c29df42cd4b42e60433435c29b&to=d708a186b6a9b050d09558163dd353d9f738c82d&stat=instructions%3Au

[aaron.ballman]

In D150038#4326554, @cor3ntin wrote:

In D150038#4326549, @aaron.ballman wrote:

Do you have some performance measurement numbers for how much benefit we get from the changes?

Not really.
The PR that changed the scope of trailing return type had a 0.35% regression, I expect this to be in the same ballpark in the other direction
http://llvm-compile-time-tracker.com/compare.php?from=cd173cbd7cca69c29df42cd4b42e60433435c29b&to=d708a186b6a9b050d09558163dd353d9f738c82d&stat=instructions%3Au

We usually want performance-related changes to come with some hard measurements because of how painfully easy it is to think something will result in a positive performance change that ends up being a wash. Would you mind putting a branch up on the compile time tracker page to validate your expectations? .35% better performance would be nice to see!

[cor3ntin]

In D150038#4326576, @aaron.ballman wrote:

In D150038#4326554, @cor3ntin wrote:

In D150038#4326549, @aaron.ballman wrote:

Do you have some performance measurement numbers for how much benefit we get from the changes?

Not really.
The PR that changed the scope of trailing return type had a 0.35% regression, I expect this to be in the same ballpark in the other direction
http://llvm-compile-time-tracker.com/compare.php?from=cd173cbd7cca69c29df42cd4b42e60433435c29b&to=d708a186b6a9b050d09558163dd353d9f738c82d&stat=instructions%3Au

We usually want performance-related changes to come with some hard measurements because of how painfully easy it is to think something will result in a positive performance change that ends up being a wash. Would you mind putting a branch up on the compile time tracker page to validate your expectations? .35% better performance would be nice to see!

Ask and you shall receive http://llvm-compile-time-tracker.com/?remote=cor3ntin

[aaron.ballman]

Fantastic, that's some nice green in there. LGTM modulo nits, please add NFC to the patch summary when landing so folks know not to expect test coverage.

clang/include/clang/Sema/Sema.h

public:
   /// Stack containing information about each of the nested
   /// function, block, and method scopes that are currently active.
   SmallVector<sema::FunctionScopeInfo *, 4> FunctionScopes;
 
   /// The index of the first FunctionScope that corresponds to the current
   /// context.
   unsigned FunctionScopesStart = 0;
 
+  /// Track the number of currently active capturing scopes.
+  unsigned CapturingFunctionScopes = 0;
+
   ArrayRef<sema::FunctionScopeInfo*> getFunctionScopes() const {
     return llvm::ArrayRef(FunctionScopes.begin() + FunctionScopesStart,
                           FunctionScopes.end());
   }
 
   /// Stack containing information needed when in C++2a an 'auto' is encountered
   /// in a function declaration parameter type specifier in order to invent a
   /// corresponding template parameter in the enclosing abbreviated function

clang/lib/Sema/Sema.cpp

void Sema::PushFunctionScope() {
   }
   if (LangOpts.OpenMP)
     pushOpenMPFunctionRegion();
 }
 
 void Sema::PushBlockScope(Scope *BlockScope, BlockDecl *Block) {
   FunctionScopes.push_back(new BlockScopeInfo(getDiagnostics(),
                                               BlockScope, Block));
+  CapturingFunctionScopes++;
 }
 
 LambdaScopeInfo *Sema::PushLambdaScope() {
   LambdaScopeInfo *const LSI = new LambdaScopeInfo(getDiagnostics());
   FunctionScopes.push_back(LSI);
+  CapturingFunctionScopes++;
   return LSI;
 }
 
 void Sema::RecordParsingTemplateParameterDepth(unsigned Depth) {
   if (LambdaScopeInfo *const LSI = getCurLambda()) {
     LSI->AutoTemplateParameterDepth = Depth;
     return;
   }
else
     for (const auto &PUD : Scope->PossiblyUnreachableDiags)
       Diag(PUD.Loc, PUD.PD);
 
   return Scope;
 }
 
 void Sema::PoppedFunctionScopeDeleter::
 operator()(sema::FunctionScopeInfo *Scope) const {
+  if(!Scope->isPlainFunction())
+    Self->CapturingFunctionScopes--;
   // Stash the function scope for later reuse if it's for a normal function.
   if (Scope->isPlainFunction() && !Self->CachedFunctionScope)
     Self->CachedFunctionScope.reset(Scope);
   else
     delete Scope;
 }
 
 void Sema::PushCompoundScope(bool IsStmtExpr) {
void Sema::PushCapturedRegionScope(Scope *S, CapturedDecl *CD, RecordDecl *RD,
                                    CapturedRegionKind K,
                                    unsigned OpenMPCaptureLevel) {
   auto *CSI = new CapturedRegionScopeInfo(
       getDiagnostics(), S, CD, RD, CD->getContextParam(), K,
       (getLangOpts().OpenMP && K == CR_OpenMP) ? getOpenMPNestingLevel() : 0,
       OpenMPCaptureLevel);
   CSI->ReturnType = Context.VoidTy;
   FunctionScopes.push_back(CSI);
+  CapturingFunctionScopes++;
 }
 
 CapturedRegionScopeInfo *Sema::getCurCapturedRegion() {
   if (FunctionScopes.empty())
     return nullptr;
 
   return dyn_cast<CapturedRegionScopeInfo>(FunctionScopes.back());
 }

clang/lib/Sema/SemaExpr.cpp

 
 bool Sema::tryCaptureVariable(
     ValueDecl *Var, SourceLocation ExprLoc, TryCaptureKind Kind,
     SourceLocation EllipsisLoc, bool BuildAndDiagnose, QualType &CaptureType,
     QualType &DeclRefType, const unsigned *const FunctionScopeIndexToStopAt) {
   // An init-capture is notionally from the context surrounding its
   // declaration, but its parent DC is the lambda class.
   DeclContext *VarDC = Var->getDeclContext();
+  DeclContext *DC = CurContext;
+
+  // tryCaptureVariable is called ever ytime a DeclRef is formed,
+  // it can therefore have non-negigible impact on performances.
+  // For local variables and when there is no capturing scope,
+  // we can bailout early.
+  if(CapturingFunctionScopes == 0 && (!BuildAndDiagnose || VarDC == DC))
+    return true;
+
   const auto *VD = dyn_cast<VarDecl>(Var);
   if (VD) {
     if (VD->isInitCapture())
       VarDC = VarDC->getParent();
   } else {
     VD = Var->getPotentiallyDecomposedVarDecl();
   }
   assert(VD && "Cannot capture a null variable");
 
-  DeclContext *DC = CurContext;
   const unsigned MaxFunctionScopesIndex = FunctionScopeIndexToStopAt
       ? *FunctionScopeIndexToStopAt : FunctionScopes.size() - 1;
   // We need to sync up the Declaration Context with the
   // FunctionScopeIndexToStopAt
   if (FunctionScopeIndexToStopAt) {
     unsigned FSIndex = FunctionScopes.size() - 1;
     while (FSIndex != MaxFunctionScopesIndex) {
       DC = getLambdaAwareParentOfDeclContext(DC);