From: Vladimir Makarov <vmakarov@redhat.com>
To: liuhongt <hongtao.liu@intel.com>, gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] Add a bit dislike for separate mem alternative when op is REG_P.
Date: Thu, 26 May 2022 17:12:02 -0400 [thread overview]
Message-ID: <8b505a07-64bb-f483-63cc-cef6e8e4642c@redhat.com> (raw)
In-Reply-To: <20220525033920.77449-1-hongtao.liu@intel.com>
On 2022-05-24 23:39, liuhongt wrote:
> Rigt now, mem_cost for separate mem alternative is 1 * frequency which
> is pretty small and caused the unnecessary SSE spill in the PR, I've tried
> to rework backend cost model, but RA still not happy with that(regress
> somewhere else). I think the root cause of this is cost for separate 'm'
> alternative cost is too small, especially considering that the mov cost
> of gpr are 2(default for REGISTER_MOVE_COST). So this patch increase mem_cost
> to 2*frequency, also increase 1 for reg_class cost when m alternative.
>
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> Ok for trunk?
Thank you for addressing this problem. And sorry I can not approve this
patch at least w/o your additional work on benchmarking this change.
This code is very old. It is coming from older RA (former file
regclass.c) and existed practically since GCC day 1. People tried many
times to improve this code. The code also affects many targets.
I can approve this patch if you show that there is no regression at
least on x86-64 on some credible benchmark, e.g. SPEC2006 or SPEC2017.
I know it is a big work but when I myself do such changes I check
SPEC2017. I rejected my changes like this one several times when I
benchmarked them on SPEC2017 although at the first glance they looked
reasonable.
> gcc/ChangeLog:
>
> PR target/105513
> * ira-costs.cc (record_reg_classes): Increase both mem_cost
> and reg class cost by 1 for separate mem alternative when
> REG_P (op).
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr105513-1.c: New test.
> ---
> gcc/ira-costs.cc | 26 +++++++++++++---------
> gcc/testsuite/gcc.target/i386/pr105513-1.c | 16 +++++++++++++
> 2 files changed, 31 insertions(+), 11 deletions(-)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr105513-1.c
>
> diff --git a/gcc/ira-costs.cc b/gcc/ira-costs.cc
> index 964c94a06ef..f7b8325e195 100644
> --- a/gcc/ira-costs.cc
> +++ b/gcc/ira-costs.cc
> @@ -625,7 +625,8 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = mem_cost[rclass][0] * frequency;
> + pp_costs[k] = (mem_cost[rclass][0]
> + + 1) * frequency;
> }
> }
> else
> @@ -648,7 +649,8 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = mem_cost[rclass][1] * frequency;
> + pp_costs[k] = (mem_cost[rclass][1]
> + + 1) * frequency;
> }
> }
> else
> @@ -670,9 +672,9 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = ((mem_cost[rclass][0]
> - + mem_cost[rclass][1])
> - * frequency);
> + pp_costs[k] = (mem_cost[rclass][0]
> + + mem_cost[rclass][1]
> + + 2) * frequency;
> }
> }
> else
> @@ -861,7 +863,8 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = mem_cost[rclass][0] * frequency;
> + pp_costs[k] = (mem_cost[rclass][0]
> + + 1) * frequency;
> }
> }
> else
> @@ -884,7 +887,8 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = mem_cost[rclass][1] * frequency;
> + pp_costs[k] = (mem_cost[rclass][1]
> + + 1) * frequency;
> }
> }
> else
> @@ -906,9 +910,9 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> for (k = cost_classes_ptr->num - 1; k >= 0; k--)
> {
> rclass = cost_classes[k];
> - pp_costs[k] = ((mem_cost[rclass][0]
> - + mem_cost[rclass][1])
> - * frequency);
> + pp_costs[k] = (mem_cost[rclass][0]
> + + mem_cost[rclass][1]
> + + 2) * frequency;
> }
> }
> else
> @@ -929,7 +933,7 @@ record_reg_classes (int n_alts, int n_ops, rtx *ops,
> /* Although we don't need insn to reload from
> memory, still accessing memory is usually more
> expensive than a register. */
> - pp->mem_cost = frequency;
> + pp->mem_cost = 2 * frequency;
> else
> /* If the alternative actually allows memory, make
> things a bit cheaper since we won't need an
> diff --git a/gcc/testsuite/gcc.target/i386/pr105513-1.c b/gcc/testsuite/gcc.target/i386/pr105513-1.c
> new file mode 100644
> index 00000000000..530f5292252
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr105513-1.c
> @@ -0,0 +1,16 @@
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-O2 -msse2 -mtune=skylake -mfpmath=sse" } */
> +/* { dg-final { scan-assembler-not "\\(%rsp\\)" } } */
> +
> +static int as_int(float x)
> +{
> + return (union{float x; int i;}){x}.i;
> +}
> +
> +float f(double y, float x)
> +{
> + int i = as_int(x);
> + if (__builtin_expect(i > 99, 0)) return 0;
> + if (i*2u < 77) if (i==2) return 0;
> + return y*x;
> +}
next prev parent reply other threads:[~2022-05-26 21:12 UTC|newest]
Thread overview: 14+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-05-25 3:39 liuhongt
2022-05-25 5:17 ` Hongtao Liu
2022-05-26 21:12 ` Vladimir Makarov [this message]
2022-05-30 3:05 ` Hongtao Liu
2022-05-31 16:28 ` Vladimir Makarov
2022-05-31 16:40 ` Richard Sandiford
2022-05-31 23:51 ` Hongtao Liu
2022-05-27 9:39 ` Alexander Monakov
2022-05-30 2:52 ` Liu, Hongtao
2022-05-30 6:22 ` Alexander Monakov
2022-05-30 7:14 ` Hongtao Liu
2022-05-30 7:44 ` Alexander Monakov
2022-05-30 8:34 ` Hongtao Liu
2022-05-30 9:41 ` Alexander Monakov
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=8b505a07-64bb-f483-63cc-cef6e8e4642c@redhat.com \
--to=vmakarov@redhat.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hongtao.liu@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).