https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62025
bin.cheng <amker.cheng at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amker.cheng at gmail dot com
--- Comment #19 from bin.cheng <amker.cheng at gmail dot com> ---
(In reply to Jeffrey A. Law from comment #16)
> In reference to c#12. I think the ivopts changes are just setting up the
> situation that is mis-handled later. I'd gotten as far as seeing the +128
> increment moving in the scheduler, but hadn't looked to see if it was valid.
>
> Anyway, so yes I think the ivopts stuff is fine.
>
> I should have realized I was chasing something of that nature when the
> bisection settled on the ivopts code as the trigger.
Yes, The scheduling behavior is triggered by specific offset in this case. It
changes below insn sequence:
11405: ......
11406: {%r2:SI=ltu(%cc:CCL1,0)+%r2:SI+[%r8:SI+0xf8];clobber %cc:CC;}
......
2803: {%r8:SI=%r8:SI+0x80;clobber %cc:CC;}
REG_UNUSED %cc:CC
into:
11405: ......
2803: {%r8:SI=%r8:SI+0x80;clobber %cc:CC;}
REG_UNUSED %cc:CC
11406: {%r2:SI=ltu(%cc:CCL1,0)+%r2:SI+[%r8:SI+0x78];clobber %cc:CC;}
by changing the offset in insn 11406.
The problem is why minipass in
http://gcc.gnu.org/ml/gcc-patches/2012-08/msg00155.html would do this
transformation. According to the description, it is to change
rn++
rm=[rn]
into
rm=[rn+4]
rn++
Here it is exactly the opposite tranformation and introducing more dependency.
Thanks.
