https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115825
--- Comment #15 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to Segher Boessenkool from comment #14) > Many Power testcases fail since this commit. The commit says it wants to > change > what volatile accesses do, but instead, it changes behaviour for anything > marked > volatil. That is a lot of asm, and a whole lot of builtins; the flag simply > means "there is an unspecified side effect". > > Please repair this? It's intended - stmts with side effects are not going to be eliminated, the heuristic was assuming they'd be, to 1/3, but that's obviously not true. You have to fix the testcases or specifically argue why the heuristic is now bad. This should preferably happen in new missed-optimization or testsuite-fail PRs.
