From: Hongtao Liu <crazylht@gmail.com>
To: Richard Biener <richard.guenther@gmail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: [PATCH] [i386] Prevent vectorization for load from parm_decl at O2 to avoid STF issue.
Date: Mon, 7 Mar 2022 16:26:01 +0800 [thread overview]
Message-ID: <CAMZc-bzCBr4cbjcC-Ouu8R_-8aggm=Gebr5oqXj03yFcmQ339w@mail.gmail.com> (raw)
In-Reply-To: <20220304072714.55713-1-hongtao.liu@intel.com>
On Fri, Mar 4, 2022 at 3:28 PM liuhongt via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> For parameter passing through stack, vectorized load from parm_decl
> in callee may trigger serious STF issue. This is why GCC12 regresses
> 50% for cray at -O2 compared to GCC11.
>
> The patch add an extremely large number to stmt_cost to prevent
> vectorization for loads from parm_decl under very-cheap cost model,
> this can at least prevent O2 regression due to STF issue, but may lose
> some perf where there's no such issue(1 vector_load vs n scalar_load +
> CTOR).
>
> No impact for SPEC2017 for both plain O2 and native O2 on ICX.
> Bootstrapped and regtested on x86_64-linux-gnu{-m32,}.
> Ok for trunk?
>
> gcc/ChangeLog:
>
> PR target/101908
> * config/i386/i386.cc (ix86_load_maybe_stfs_p): New.
> (ix86_vector_costs::add_stmt_cost): Add extra cost for
> vector_load/unsigned_load which may have stall forward issue.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/pr101908-1.c: New test.
> * gcc.target/i386/pr101908-2.c: New test.
> ---
> gcc/config/i386/i386.cc | 31 ++++++++++++++++++++++
> gcc/testsuite/gcc.target/i386/pr101908-1.c | 12 +++++++++
> gcc/testsuite/gcc.target/i386/pr101908-2.c | 12 +++++++++
> 3 files changed, 55 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/pr101908-1.c
> create mode 100644 gcc/testsuite/gcc.target/i386/pr101908-2.c
>
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index b2bf90576d5..3bbaaf65ea8 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -22976,6 +22976,19 @@ ix86_noce_conversion_profitable_p (rtx_insn *seq, struct noce_if_info *if_info)
> return default_noce_conversion_profitable_p (seq, if_info);
> }
>
> +/* Return true if REF may have STF issue, otherwise false. */
> +static bool
> +ix86_load_maybe_stfs_p (tree ref)
> +{
> + tree addr = get_base_address (ref);
> +
> + if (TREE_CODE (addr) != PARM_DECL
> + || !tree_fits_uhwi_p (TYPE_SIZE (TREE_TYPE (addr)))
> + || tree_to_uhwi (TYPE_SIZE (TREE_TYPE (addr))) <= MAX_BITS_PER_WORD)
> + return false;
> + return true;
> +}
> +
> /* x86-specific vector costs. */
> class ix86_vector_costs : public vector_costs
> {
> @@ -23203,6 +23216,24 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
> if (TREE_CODE (op) == SSA_NAME)
> TREE_VISITED (op) = 0;
> }
> +
> + /* Prevent vectorization for load from parm_decl at O2 to avoid STF issue.
> + Performance may lose when there's no STF issue(1 vector_load vs n
> + scalar_load + CTOR).
> + TODO: both extra cost(2000) and ix86_load_maybe_stfs_p need to be fine
> + tuned. */
> + if ((kind == vector_load || kind == unaligned_load)
> + && flag_vect_cost_model == VECT_COST_MODEL_VERY_CHEAP
> + && stmt_info
> + && stmt_info->slp_type == pure_slp
> + && stmt_info->stmt
> + && gimple_assign_load_p (stmt_info->stmt)
> + && ix86_load_maybe_stfs_p (gimple_assign_rhs1 (stmt_info->stmt)))
> + {
> + stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
> + stmt_cost += 2000;
> + }
> +
> if (stmt_cost == -1)
> stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
>
> diff --git a/gcc/testsuite/gcc.target/i386/pr101908-1.c b/gcc/testsuite/gcc.target/i386/pr101908-1.c
> new file mode 100644
> index 00000000000..f8e0f2e26bb
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr101908-1.c
> @@ -0,0 +1,12 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-slp-details" } */
> +/* { dg-final { scan-tree-dump {(?n)add new stmt:.*MEM \<vector\(2\) double\>} "slp2" } } */
> +
> +struct X { double x[2]; };
> +typedef double v2df __attribute__((vector_size(16)));
> +
> +v2df __attribute__((noipa))
> +foo (struct X* x, struct X* y)
> +{
> + return (v2df) {x->x[1], x->x[0] } + (v2df) { y->x[1], y->x[0] };
> +}
> diff --git a/gcc/testsuite/gcc.target/i386/pr101908-2.c b/gcc/testsuite/gcc.target/i386/pr101908-2.c
> new file mode 100644
> index 00000000000..7f2f00cebab
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/pr101908-2.c
> @@ -0,0 +1,12 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-slp-details" } */
> +/* { dg-final { scan-tree-dump-not {(?n)add new stmt:.*MEM \<vector\(2\) double\>} "slp2" } } */
> +
> +struct X { double x[2]; };
> +typedef double v2df __attribute__((vector_size(16)));
> +
> +v2df __attribute__((noipa))
> +foo (struct X x, struct X y)
> +{
> + return (v2df) {x.x[1], x.x[0] } + (v2df) { y.x[1], y.x[0] };
> +}
> --
> 2.18.1
>
--
BR,
Hongtao
next prev parent reply other threads:[~2022-03-07 8:26 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-03-04 7:27 liuhongt
2022-03-07 8:26 ` Hongtao Liu [this message]
2022-03-07 9:37 ` Richard Biener
2022-03-08 3:39 ` Hongtao Liu
2022-03-15 9:13 ` [PATCH] [i386] Add extra cost for unsigned_load which may have stall forward issue liuhongt
2022-03-16 1:03 ` liuhongt
2022-03-16 2:19 ` liuhongt
2022-03-16 9:54 ` Richard Biener
2022-03-17 7:12 ` Hongtao Liu
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='CAMZc-bzCBr4cbjcC-Ouu8R_-8aggm=Gebr5oqXj03yFcmQ339w@mail.gmail.com' \
--to=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.guenther@gmail.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).