================
@@ -147,6 +147,20 @@ bool BugSuppression::isSuppressed(const
PathDiagnosticLocation &Location,
// done as well as perform a lot of work we'll never need.
// Gladly, none of our on-by-default checkers currently need it.
DeclWithIssue = ACtx.getTranslationUnitDecl();
+ } else {
+ // This is the fast path. However, we should still consider the topmost
+ // declaration that isn't TranslationUnitDecl, because we should respect
+ // attributes on the entire declaration chain.
+ while (true) {
----------------
haoNoQ wrote:
This is, uh, a somewhat premature performance optimization. I think I want to
get rid of it, as discussed in #79398 (it's incorrect as well), but it probably
requires a separate discussion.
The static analyzer normally always takes the fast path as every single
"supported" checker fills in the data correctly.
This isn't just about scanning the TU *here* for discovering suppressed ranges,
it's more about the fact that the static analyzer is so careful about not
analyzing non-user code that normally it doesn't even deserialize PCHs if they
aren't directly required for analysis. This is demonstrated by the lovely test
case in
[`clang/test/Analysis/check-deserialization.cpp`](https://github.com/llvm/llvm-project/blob/llvmorg-19-init/clang/test/Analysis/check-deserialization.cpp).
A full TU scan here will cause these PCHs to be deserialized even though the
rest of the analysis never needed to look at them.
I have some data that this can have significant performance impact. However,
other tools such as clang-tidy never had such optimization (they always scan
the entire TU indiscriminately) and this never stopped them from satisfying the
usual performance expectations of static analysis tools. We've also abandoned
this optimization in a downstream fork for a while and it seems perfectly
livable.
So I want to either abandon this non-deserialization guarantee (simply always
scan the entire TU), or find a different way to satisfy it (eg., collect those
ranges eagerly during AST construction so that to avoid the after-the-fact
scan; then if PCHs/modules are used, suppression ranges will make it into PCHs
and get deserialized together with them as-needed). I'll probably try to do
some of that in the coming weeks but I wanted this patch to be independent from
that so that to get the attribute's behavior to reasonable shape without
breaking anything else.
https://github.com/llvm/llvm-project/pull/80371
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
