[Bug rtl-optimization/58679] [4.9 regression] ICE in create_pre_exit, at mode-switching.c:421 with -mavx after r202915

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58679

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Eric Botcazou from comment #8)
> > Or should we use a different approach, and use DF to re-create REG_DEAD
> > notes?
> 
> Yes, if you need REG_DEAD/REG_UNUSED notes, just tell DF to compute them:

I think that missing REG_DEAD/REG_UNUSED notes and r202915 just triggered the
problem, fixed with [1] - at least I'm not able to trigger the ICE described in
Comment #0 anymore.

The new problem was introduced by Vladimir's patch that removed final
assignments to return registers in several cases. The postreload mode-switching
pass depends on them, and gcc.target/i386/{avx-1,sse-23,sse-24}.c testcases
start to fail on x86 [2]. Also, bootstrap will be broken when configured with
"--with-arch=core-avx-i --with-cpu=core-avx-i --with-fpmath=avx" due to
post-reload mode switching failures in libjava and libgo.


[1] http://gcc.gnu.org/ml/gcc-patches/2013-10/msg01606.html
[2] http://gcc.gnu.org/ml/gcc-testresults/2013-10/msg02051.html