Hi,
The patch enhances ifcvt to allow_cc_mode if HAVE_cbranchcc4.
Bootstrap and no make check regression on X86-64.
Will add new test cases after ccmp is enabled.
Ok for trunk?
Thanks!
-Zhenqiang
ChangeLog:
2014-10-29 Zhenqiang Chen <zhenqiang.c...@arm.com>
* ifcvt.c (noce_emit_cmove, noce_get_alt_condition,
noce_get_condition):
Allow CC mode if HAVE_cbranchcc4.
diff --git a/gcc/ifcvt.c b/gcc/ifcvt.c
index a28f5c1..5cd0ac0 100644
--- a/gcc/ifcvt.c
+++ b/gcc/ifcvt.c
@@ -1441,10 +1441,17 @@ noce_emit_cmove (struct noce_if_info *if_info, rtx
x, enum rtx_code code,
end_sequence ();
}
- /* Don't even try if the comparison operands are weird. */
+ /* Don't even try if the comparison operands are weird
+ except that the target supports cbranchcc4. */
if (! general_operand (cmp_a, GET_MODE (cmp_a))
|| ! general_operand (cmp_b, GET_MODE (cmp_b)))
- return NULL_RTX;
+ {
+#if HAVE_cbranchcc4
+ if (GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC
+ || cmp_b != const0_rtx)
+#endif
+ return NULL_RTX;
+ }
#if HAVE_conditional_move
unsignedp = (code == LTU || code == GEU
@@ -1770,6 +1777,11 @@ noce_get_alt_condition (struct noce_if_info *if_info,
rtx target,
rtx cond, set;
rtx_insn *insn;
int reverse;
+ int allow_cc_mode = false;
+#if HAVE_cbranchcc4
+ allow_cc_mode = true;
+#endif
+
/* If target is already mentioned in the known condition, return it. */
if (reg_mentioned_p (target, if_info->cond))
@@ -1891,7 +1903,7 @@ noce_get_alt_condition (struct noce_if_info *if_info,
rtx target,
}
cond = canonicalize_condition (if_info->jump, cond, reverse,
- earliest, target, false, true);
+ earliest, target, allow_cc_mode, true);
if (! cond || ! reg_mentioned_p (target, cond))
return NULL;
@@ -2347,6 +2359,10 @@ noce_get_condition (rtx_insn *jump, rtx_insn
**earliest, bool then_else_reversed
{
rtx cond, set, tmp;
bool reverse;
+ int allow_cc_mode = false;
+#if HAVE_cbranchcc4
+ allow_cc_mode = true;
+#endif
if (! any_condjump_p (jump))
return NULL_RTX;
@@ -2383,7 +2399,7 @@ noce_get_condition (rtx_insn *jump, rtx_insn
**earliest, bool then_else_reversed
/* Otherwise, fall back on canonicalize_condition to do the dirty
work of manipulating MODE_CC values and COMPARE rtx codes. */
tmp = canonicalize_condition (jump, cond, reverse, earliest,
- NULL_RTX, false, true);
+ NULL_RTX, allow_cc_mode, true);
/* We don't handle side-effects in the condition, like handling
REG_INC notes and making sure no duplicate conditions are emitted. */
> -----Original Message-----
> From: Richard Henderson [mailto:r...@redhat.com]
> Sent: Tuesday, October 28, 2014 12:03 AM
> To: Zhenqiang Chen
> Cc: gcc-patches@gcc.gnu.org
> Subject: Re: [Ping] [PATCH, 10/10] aarch64: Handle ccmp in ifcvt to make
it
> work with cmov
>
> On 10/27/2014 12:50 AM, Zhenqiang Chen wrote:
> > Good point. It is not ccmp special. It is cbranchcc4 related. If I
> > understand correct, without cbranchcc4, we need put the result to a
> > tmp register and generate additional compares, which is not good for
> > performance.
>
> It won't be an additional compare. It is regenerating the compare at a
new
> location. The old comparison would be deleted via dead code elimination.
>
> That said,
>
> > +#if HAVE_cbranchcc4
> > + allow_cc_mode = true;
> > +#endif
>
> does seem to be the right solution.
>
> If a target has this, we ought to be able to reasonably expect that it's
got all
> the other patterns that this implies. If a target is missing them,
hopefully we
> can get that sorted fairly quickly after this patch is included.
>
> > +#if HAVE_cbranchcc4
> > + if (!(GET_MODE_CLASS (GET_MODE (cmp_a)) == MODE_CC
> > + || GET_MODE_CLASS (GET_MODE (cmp_b)) == MODE_CC)) #endif
>
> This test looks weird, considering what we're looking for.
> I think a better test is
>
> if (GET_MODE_CLASS (GET_MODE (cmp_a)) != MODE_CC
> || cmp_b != const0_rtx)
>
> Accepting something like (compare (reg:CC a) (reg:CC b)) is definitely
non-
> canonical. Even (compare (const_int 0) (reg:CC flags)) is odd.
>
> The ifcvt.c change should go in by itself.
> The expr.c change should also be standalone.
> The ccmp.c change should probably be merged with the initial commit of
> ccmp.c.
> The aaarch64.md change should probably be merged with the patch that
> adds cbranchcc.
>
>
> r~
