Re: [PATCH v2] ipa-visibility: Optimize TLS access [PR99619]

[Alexander Monakov <amonakov@ispras.ru>]

On Tue, 30 Aug 2022, Martin Jambor wrote:

> There is still the optimize attribute so in fact no, even in non-LTO
> mode if there is no current function, you cannot trust the "global"
> "optimize" thing.
> 
> Ideally we would assert that no "analysis" phase of an IPA pass reads
> the global optimization flags, so please don't add new places where we
> do.
> 
> You can either add a parameter to decl_default_tls_model to tell it
> what context it is called from and IMHO it would also be acceptable to
> check whether we have a non-NULL cfun and decide based on that (but here
> I only hope it is not something others might object to).

I see, thank you for explaining the issue, and sorry if I was a bit stubborn.

Does the attached patch (incremental change below) look better? It no longer
has the 'shortcut' where iterating over referrers is avoided for the common
case of plain 'gcc -O2' and no 'optimize' attributes, but fortunately TLS
variables are not so numerous to make chasing that worthwhile.

--- a/gcc/varasm.cc
+++ b/gcc/varasm.cc
@@ -6703,8 +6703,8 @@ have_optimized_refs (struct symtab_node *symbol)
 static bool
 optimize_dyn_tls_for_decl_p (const_tree decl)
 {
-  if (optimize)
-    return true;
+  if (cfun)
+    return optimize;
   return symtab->state >= IPA && have_optimized_refs (symtab_node::get (decl));
 }

Alexander