[cfe-dev] Stmt.getEndLoc() vs semicolon

Alexander Lanin via cfe-dev cfe-dev at lists.llvm.org
Wed Feb 26 13:53:49 PST 2020


thanks, that’s interesting, but I’m not sure how this relates to trailing comments + semicolon?
Why does it matter whether Expr is wrapped or not? It can contain comments in general. But doesn’t contain those at the end.
At least for such examples I find it strange: http://ce.steveire.com/z/KdXBDg

Is Expr the only Stmt that doesn’t include „it’s end“?
Would it make sense to have some unifyStmtRange() in SourceCode.h with special handling for Expr which would extend it with (n x comment) + (optional comma or semi) [+ all the special cases I’m not thinking about right now]?

Currently the check mentioned below always subtracts one char and then checks whether it’s a semicolon. Might be nicer to distinguish by isa<Expr>?


Von: Yitzhak Mandelbaum
Gesendet: Mittwoch, 26. Februar 2020 16:40
An: Alexander Lanin
Cc: cfe-dev at lists.llvm.org
Betreff: Re: [cfe-dev] Stmt.getEndLoc() vs semicolon


I agree that this behavior is rather confusing, especially to first-time users. The cause of the issue is that Expr derives from Stmt so that Exprs can appear directly in, for example, compound statements, rather than being wrapped in an explicit node to represent statements that consist only of an expression.  As far as I understand, this choice to avoid an explicit node had some (positive) performance implications for the AST when it was originally designed.  While I'd love to see this choice revisited, I think it would be a significant effort and don't expect it to happen.

Instead, we tend to work around the issue.  For getting proper ranges of statements, you might find clang::tooling::getExtendedText/getExtendedRange meet your needs:

For declarations, the (just added!) getAssociatedRange might be even better: https://github.com/llvm/llvm-project/blob/master/clang/include/clang/Tooling/Transformer/SourceCode.h#L39


On Tue, Feb 25, 2020 at 5:15 PM Alexander Lanin via cfe-dev <cfe-dev at lists.llvm.org> wrote:
I’m having trouble with the locations returned by Stmt getEndLoc()/getSourceRange() and am wondering whether this is a bug/flaw:
the returned location sometimes includes and sometimes excludes the semicolon in the end.
This gets even more interesting when there are (multiple) comments before the semicolon, simply because the difference between the returned values gets even bigger.
Here is an example showing how DeclStmt includes the semicolon, while CallExpr and BinaryOperator exclude the semicolon http://ce.steveire.com/z/TW3IAG.
So Stmt behaves “completely differently” depending on it’s type, which is no way suggested by it’s interface.
Wouldn’t it be better for Stmt-Users if it would always be the same?
(Not sure whether it should always be included or excluded. It’s not even always a semicolon as in that 3rd foo() in the example)
Here is one of the effects of this complexity for users of Stmt as they struggle to find that semicolon:
https://bugs.llvm.org/show_bug.cgi?id=25970 / https://reviews.llvm.org/D16267
Of course it’s fixable there, but that would imply working around the issue in multiple places.
Alexander Lanin
cfe-dev mailing list
cfe-dev at lists.llvm.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20200226/c63db385/attachment.html>

More information about the cfe-dev mailing list