`Expr::getExprLoc()` is slow

[tbaederr]

This first came up when the implementation of CallExpr::getBeginLoc() was changed and we saw a surprising drop in compile times.

We call Expr::getExprLoc() (and various other functions returning a SourceLocation) all the time. The assumption is that the call is cheap and that it's fine to do even if the source location isn't used in the end.

Changing the implementation of Expr::getExprLoc() to just return SourceLocation() improves compile times considerably: https://llvm-compile-time-tracker.com/compare.php?from=a71f56d63f569ca0ed543d3d01c684ca6ac7dffd&to=2510892ea46a84a284e0cde9f2a4edae3cdb183d&stat=instructions:u

tl;dr: Up to 15%.

Quick experiment with callgrind when compiling SemaExpr.cpp:

We should improve this.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Timm Baeder (tbaederr)

This first came up when the implementation of `CallExpr::getBeginLoc()` was changed and we saw a surprising drop in compile times.

We call Expr::getExprLoc() (and various other functions returning a SourceLocation) all the time. The assumption is that the call is cheap and that it's fine to do even if the source location isn't used in the end.

Changing the implementation of Expr::getExprLoc() to just return SourceLocation() improves compile times considerably: https://llvm-compile-time-tracker.com/compare.php?from=a71f56d63f569ca0ed543d3d01c684ca6ac7dffd&to=2510892ea46a84a284e0cde9f2a4edae3cdb183d&stat=instructions:u

tl;dr: Up to 15%.

Quick experiment with callgrind when compiling SemaExpr.cpp:

We should improve this.

[zyn0217]

@HighCommander4 @cor3ntin

I wonder if we can have a bit in CallExprBitfields to track whether the callee is an ExplicitObjectMemberFunction, so that we can get rid of the call to getCalleeDecl

[cor3ntin]

@zyn0217 If it helps, sure :) But I don't know how easy it is to track that from construction.
getCalleeDecl is also super inefficient. I push a thing to compile-time-tracker to improve getCalleeDecl, we'll see

I wonder if we can find a better way to handle/remove the special case for explicit object member functions

[cor3ntin]

I'm trying a few different things
https://llvm-compile-time-tracker.com/compare.php?from=f4cebe5d73c24ab53917bd58aedc9db892a164ae&to=16077ed74794645452434c9ec761d642789e09ac&stat=instructions%3Au

Nowhere near 10% though

[zygoloid]

For what it's worth, this is the reason why the code in ExprConstant.cpp generally passes around a const Expr *E instead of a SourceLocation to code that performs some checks and might diagnose -- we used to pass around SourceLocations and it turned out we were spending a significant portion of our time during compile-time evaluation in getExprLoc.

We could generalize the approach there, and create an ExprLoc type that can be used in place of a SourceLocation but is internally represented as a const Expr*. Though it'd obviously be better if we can make getExprLoc itself faster.