[Bug rtl-optimization/51933] [4.7 regression] wrong code due to -free

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51933

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-22 09:42:05 UTC ---
Created attachment 26410
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26410
gcc47-pr51933.patch

Patch I wrote last night.  Passed bootstrap/regtest, the only cases where it
prevented something was in the new testcase and in jcf-parse.c
(rewrite_reflection_indexes) with -m64, where it fixed an miscompilation.  What
happened there is that some earlier bb (but in a loop) was doing a
SImode->DImode zero extension, and later in the loop there was a SImode ior,
followed by HImode->SImode zero extension of that, which fed the earlier
SImode->DImode zero extension.  The SImode->DImode insn was processed first,
removed and changed the HImode->SImode zero extension into HImode->DImode zero
extension.  But as the HImode->SImode extension was changed, its df links were
gone.  When the HImode-> zero extension is later processed as candidate,
get_defs returns NULL and we assume all the dependencies have been adjusted.

Which leads me to thinking that this patch is a wrong approach, and perhaps we
should instead DF_DEFER_INSN_RESCAN in ree.c and adjust, so that we can cope
with the adjusted insns.