This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH][sched-deps] Generalise usage of macro fusion to work on any two insns
- From: Kyrill Tkachov <kyrylo dot tkachov at arm dot com>
- To: Jeff Law <law at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Cc: Ramana Radhakrishnan <Ramana dot Radhakrishnan at arm dot com>, Maxim Kuvyrkov <maxim dot kuvyrkov at linaro dot org>
- Date: Thu, 10 Jul 2014 09:00:04 +0100
- Subject: Re: [PATCH][sched-deps] Generalise usage of macro fusion to work on any two insns
- Authentication-results: sourceware.org; auth=none
- References: <53AD2B65 dot 6030406 at arm dot com> <53B1CB16 dot 10206 at redhat dot com>
On 30/06/14 21:39, Jeff Law wrote:
On 06/27/14 02:29, Kyrill Tkachov wrote:
Hi all,
This patch generalises the TARGET_MACRO_FUSION_PAIR_P hook usage to work
on more than just
compares and conditional branches for which it was initially designed
for (for x86).
There are some instructions in arm and aarch64 that can be fused
together when they're back to back in the instruction stream and I'd
like to use this hook to keep them together.
I'll post an implementation of TARGET_MACRO_FUSION_PAIR_P for arm and
aarch64 shortly...
Bootstrapped and tested on x86, aarch64-none-linux-gnu and
arm-none-linux-gnueabihf.
Ok for trunk?
2014-06-27 Ramana Radhakrishnan <ramana.radhakrishnan@arm.com>
Kyrylo Tkachov <kyrylo.tkachov@arm.com>
* sched-deps.c (try_group_insn): Generalise macro fusion hook usage
to any two insns. Update comment.
Isn't this going to end up calling the x86 specific macro_fusion_pair_p
with a lot more insns than that function was previously prepared to handle?
In particular I'm concerned that the 2nd argument is going to be a
non-jumping insn a lot more often. Of particular concern is this code:
test_if = SET_SRC (pc_set (condjmp));
cond = XEXP (test_if, 0);
ccode = GET_CODE (cond);
if CONDJMP is not a JUMP_INSN, pc_set is going to return NULL and XEXP
(test_if, 0) will then fault.
I realize you bootstrapped on x86, but I suspect that whatever tuning
you need to enable to really exercise this code wasn't on.
I think you can deal with this by putting
if (!any_condjump_p (condjmp)) at the start of the x86 specific
macro_fusion_pair_p is sufficient to address this issue. It also
ensures that we don't do a lot of unnecessary work in that function.
From a general code structure standpoint, can you avoid this kind of
structure:
if (any_condjmp_p (insn))
{
...
goto succ;
}
else
{
...
goto succ
}
return
succ:
Can you structure so that you return for all the cases where you don't
want to set SCHED_GROUP_P from each arm? Or go ahead and duplicate the
SCHED_GROUP_P setting in each arm of the conditional.
Hi Jeff,
Thanks for the pointers, I've reworked the patch and it does look
cleaner. I've made sure to run the x86 bootstrap with Haswell tuning and
instrumented the code to make sure that the x86 macro fusion code was
being exercised and it passed that fine.
How's that?
2014-07-10 Ramana Radhakrishnan <ramana.radhakrishnan@arm.com>
Kyrylo Tkachov <kyrylo.tkachov@arm.com>
* sched-deps.c (try_group_insn): Generalise macro fusion hook usage
to any two insns. Update comment.
* config/i386/i386.c (ix86_macro_fusion_pair_p): Reject 2nd arguments
that are not conditional jumps.
jeff
commit e36b8977738dbe3f63445199710ca627ab37e243
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date: Fri Jun 13 11:41:41 2014 +0100
[sched-deps] Generalise macro fusion hook usage
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 8046c67..7dd2ce5 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -25817,6 +25817,9 @@ ix86_macro_fusion_pair_p (rtx condgen, rtx condjmp)
rtx compare_set = NULL_RTX, test_if, cond;
rtx alu_set = NULL_RTX, addr = NULL_RTX;
+ if (!any_condjump_p (condjmp))
+ return false;
+
if (get_attr_type (condgen) != TYPE_TEST
&& get_attr_type (condgen) != TYPE_ICMP
&& get_attr_type (condgen) != TYPE_INCDEC
diff --git a/gcc/sched-deps.c b/gcc/sched-deps.c
index 7cafc8b..c01a8a6 100644
--- a/gcc/sched-deps.c
+++ b/gcc/sched-deps.c
@@ -2820,35 +2820,48 @@ sched_analyze_2 (struct deps_desc *deps, rtx x, rtx insn)
sched_deps_info->finish_rhs ();
}
-/* Try to group comparison and the following conditional jump INSN if
- they're already adjacent. This is to prevent scheduler from scheduling
- them apart. */
+/* Try to group two fuseable insns together to prevent scheduler
+ from scheduling them apart. */
static void
try_group_insn (rtx insn)
{
- unsigned int condreg1, condreg2;
- rtx cc_reg_1;
rtx prev;
- if (!any_condjump_p (insn))
+ if (!targetm.sched.macro_fusion_p ())
return;
- targetm.fixed_condition_code_regs (&condreg1, &condreg2);
- cc_reg_1 = gen_rtx_REG (CCmode, condreg1);
- prev = prev_nonnote_nondebug_insn (insn);
- if (!reg_referenced_p (cc_reg_1, PATTERN (insn))
- || !prev
- || !modified_in_p (cc_reg_1, prev))
- return;
+ if (any_condjump_p (insn))
+ {
+ unsigned int condreg1, condreg2;
+ rtx cc_reg_1;
+ targetm.fixed_condition_code_regs (&condreg1, &condreg2);
+ cc_reg_1 = gen_rtx_REG (CCmode, condreg1);
+ prev = prev_nonnote_nondebug_insn (insn);
+ if (!reg_referenced_p (cc_reg_1, PATTERN (insn))
+ || !prev
+ || !modified_in_p (cc_reg_1, prev))
+ return;
- /* Different microarchitectures support macro fusions for different
- combinations of insn pairs. */
- if (!targetm.sched.macro_fusion_pair_p
- || !targetm.sched.macro_fusion_pair_p (prev, insn))
- return;
+ if (targetm.sched.macro_fusion_pair_p (prev, insn))
+ SCHED_GROUP_P (insn) = 1;
+ }
+ else
+ {
+ rtx insn_set = single_set (insn);
+
+ prev = prev_nonnote_nondebug_insn (insn);
+ if (prev
+ && insn_set
+ && single_set (prev)
+ && modified_in_p (SET_DEST (insn_set), prev)
+ && targetm.sched.macro_fusion_pair_p (prev, insn))
+ {
+ SCHED_GROUP_P (insn) = 1;
+ }
+
+ }
- SCHED_GROUP_P (insn) = 1;
}
/* Analyze an INSN with pattern X to find all dependencies. */