https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121999

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Also reproduces w/o -flto.  I do have .gcda and .ii files but there's virtual
function calls in the code making it difficult to do a GIMPLE testcase
(OBJ_TYPE_REF isn't implemented there).  The IL difference does not seem to be
caused by inlining (or speculative inlining).

Reproduces with -fno-ipa-cp -fno-devirtualize -fno-inline -fno-ipa-icf and
the only interesting -fopt-info difference is

textstreambuffer.cpp:271:11: optimized: Indirect call -> direct call
OBJ_TYPE_REF(_2;this_52(D)->4B) => rawoutput (will resolve by ipa-profile)

I'm not sure why that's not done when not profiling.

I'm postponing analyzing/fixing with this large testcase until after Cauldron,
hoping for a non-FDO testcase from fuzzers/rebuilders ;)

Reply via email to