[clang] [clang] Refine preferred alignment for x86 globals (PR #185487)

[Eli Friedman via cfe-commits]

================
@@ -1832,6 +1832,7 @@ CharUnits ASTContext::getDeclAlign(const Decl *D, bool ForAlignof) const {
         uint64_t TypeSize =
             !BaseT->isIncompleteType() ? getTypeSize(T.getTypePtr()) : 0;
         Align = std::max(Align, getMinGlobalAlignOfVar(TypeSize, VD));
+        Align = std::max(Align, getLargeGlobalPreferredAlign(TypeSize, Align));
----------------
efriedma-quic wrote:

I'm really not concerned about the OpenMP cases; we should already generate a warning about misusing atomics like that.

In general, I'd prefer to do alignment optimizations on the IR level for two reasons.  One, we can use better heuristics at the IR level; we have more information about how the global is used.  For example, we could use PGO to figure out if the global is accessed from hot code.  Two, it's a lot easier to check the conditions correctly at the LLVM IR level; your current version doesn't handle inline variables correctly because it's checking the wrong kind of "weak".

In any case, please fix the pull request's description so it's clear this is only intended to increase the alignment of globals; the description makes it sound like you're decreasing the alignment in some cases.

https://github.com/llvm/llvm-project/pull/185487