On Wed, Nov 6, 2013 at 11:06 PM, Eric Botcazou <ebotca...@adacore.com> wrote:
>> Attached patch fixes PR 59021, where new vzeroupper instructions are
>> generated for -mavx after Vlad's useless insn removal patch.
>>
>> The problem was, that we depent on (useless) moves to AVX256 function
>> argument registers in front of the function call to switch the mode to
>> DIRTY mode. This is not true anymore, so call_insn has to switch to
>> DIRTY mode by itself.
>
> Thanks for fixing this!
OTOH, looking a bit deeper, it looks that there is a problem in
mode-switching infrastructure. If we have a BB without any mode
requirements, but an insn that switched the mode via MODE_AFTER, we
should not mark the block as transparent. Indeed, we even have a N.B.
comment in mode-switching.c:
/* Check for blocks without ANY mode requirements.
N.B. because of MODE_AFTER, last_mode might still be different
from no_mode. */
But, we do nothing w.r.t to transparency
Attached incremental patch also fixes additional vzeroupper problem.
Calls with AVX argument register in fact do not have any mode
requirements, so we don't need to fake MODE_DIRTY requirement.
I have added Joern to the discussion.
Uros.
Index: mode-switching.c
===================================================================
--- mode-switching.c (revision 204496)
+++ mode-switching.c (working copy)
@@ -577,6 +577,8 @@ optimize_mode_switching (void)
{
ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now);
add_seginfo (info + bb->index, ptr);
+ if (last_mode != no_mode)
+ bitmap_clear_bit (transp[bb->index], j);
}
}
#if defined (MODE_ENTRY) && defined (MODE_EXIT)
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 204496)
+++ config/i386/i386.c (working copy)
@@ -15708,7 +15708,7 @@ ix86_avx_u128_mode_needed (rtx insn)
rtx arg = XEXP (XEXP (link, 0), 0);
if (ix86_check_avx256_register (&arg, NULL))
- return AVX_U128_DIRTY;
+ return AVX_U128_ANY;
}
}
