https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127509
--- Comment #2 from Kapelli Rohith <kapellirohith at gmail dot com> --- PR 88010 points the other way: there the attribute is on the alias (noinline) and the target's definition wins. Here the attribute is on the target (noipa) and does not reach a call made through the alias, while a direct call to the same function honours it. Comment 2 there says aliases are transparent for late attribute lookups in IPA, and comment 3 that attributes are part of the definition of a function for IPA passes. Taken together, a noipa target should govern calls through its alias, which is what this report is about. A patch is posted: https://inbox.sourceware.org/gcc-patches/[email protected]/ It makes cgraph_node::get_availability look at the symbol an alias refers to, so it does not change PR 88010, where the alias's own attribute would have to win.
