https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121936
--- Comment #17 from Lénárd Szolnoki <leni536 at gmail dot com> --- (In reply to Andrew Pinski from comment #16) > (In reply to Lénárd Szolnoki from comment #15) > > > We get: > > > depends_on_unpsecified_content(int): > > > stp x29, x30, [sp, #-32]! > > > str x19, [sp, #16] > > > mov x29, sp > > > cbz w0, .LBB1_2 > > > mov w19, w0 > > > > > > If we used gnu::noipa, then GCC does the same as LLVM does on the original > > > testcase. > > > > I don't see the same issue manifesting, this looks like correct inlining and > > optimizations to me. Also, the original issue for GCC manifested in `main`, > > not in `depends_on_unspecified_content` (or `f` in the godbolt). The codegen > > for `depends_on_unspecified_content` is fine in both clang and gcc. > > Try to use depends_on_unpsecified_content or m from a tu compiled with gcc > at -O0 and you would get the same incorrect behavior I still don't follow. Clang's codegen doesn't call either depends_on_unpsecified_content or m anywhere, so how could it call an -O0 version from GCC? https://godbolt.org/z/rKEeqfjhe If you mean that the behavior is different between clang and gcc for each of these functions is itself not the issue. The unspecified behavior is allowed to change from call-to-call. The only issue is with assuming specific manifestations of unspecified behavior within a non-inlined vague linkage function *and* doing IPA optimizations based on that on the caller side. This is manifested in `main` in the issue description, but not in other functions. As the only non-inlined calls in the clang codegen is to `g` and `abort`, I don't see how it could possibly showcase this issue.
