From: Richard Biener <richard.guenther@gmail.com>
To: liuhongt <hongtao.liu@intel.com>
Cc: gcc-patches@gcc.gnu.org, crazylht@gmail.com, hjl.tools@gmail.com
Subject: Re: [PATCH] Take register pressure into account for vec_construct when the components are not loaded from memory.
Date: Wed, 29 Nov 2023 08:47:01 +0100 [thread overview]
Message-ID: <CAFiYyc3q4cUpsEbGLdAudQf-AdCXN-rmZCUbQzSjiN8Mq3bouQ@mail.gmail.com> (raw)
In-Reply-To: <20231128075212.3526692-1-hongtao.liu@intel.com>
On Tue, Nov 28, 2023 at 8:54 AM liuhongt <hongtao.liu@intel.com> wrote:
>
> For vec_contruct, the components must be live at the same time if
> they're not loaded from memory, when the number of those components
> exceeds available registers, spill happens. Try to account that with a
> rough estimation.
> ??? Ideally, we should have an overall estimation of register pressure
> if we know the live range of all variables.
>
> The patch can avoid regressions due to .i.e. vec_contruct with 32 char.
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
>
> Ok for trunk?
Hmm, I would suggest you put reg_needed into the class and accumulate
over all vec_construct, with your patch you pessimize a single v32qi
over two separate v16qi for example. Also currently the whole block is
gated with INTEGRAL_TYPE_P but register pressure would be also
a concern for floating point vectors. finish_cost would then apply an
adjustment.
'target_avail_regs' is for GENERAL_REGS, does that include APX regs?
I don't see anything similar for FP regs, but I guess the target should know
or maybe there's a #regs in regclass query already.
That said, this kind of adjustment looks somewhat appealing.
Richard.
> gcc/ChangeLog:
>
> * config/i386/i386.cc (ix86_vector_costs::add_stmt_cost): Take
> register pressure into account for vec_construct when the
> components are not loaded from memory.
> ---
> gcc/config/i386/i386.cc | 22 +++++++++++++++++++++-
> 1 file changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> index 683ac643bc8..f8417555930 100644
> --- a/gcc/config/i386/i386.cc
> +++ b/gcc/config/i386/i386.cc
> @@ -24706,6 +24706,7 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
> stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
> unsigned i;
> tree op;
> + unsigned reg_needed = 0;
> FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
> if (TREE_CODE (op) == SSA_NAME)
> TREE_VISITED (op) = 0;
> @@ -24737,11 +24738,30 @@ ix86_vector_costs::add_stmt_cost (int count, vect_cost_for_stmt kind,
> && (gimple_assign_rhs_code (def) != BIT_FIELD_REF
> || !VECTOR_TYPE_P (TREE_TYPE
> (TREE_OPERAND (gimple_assign_rhs1 (def), 0))))))
> - stmt_cost += ix86_cost->sse_to_integer;
> + {
> + stmt_cost += ix86_cost->sse_to_integer;
> + reg_needed++;
> + }
> }
> FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
> if (TREE_CODE (op) == SSA_NAME)
> TREE_VISITED (op) = 0;
> +
> + /* For vec_contruct, the components must be live at the same time if
> + they're not loaded from memory, when the number of those components
> + exceeds available registers, spill happens. Try to account that with a
> + rough estimation. Currently only handle integral modes since scalar fp
> + shares sse_regs with vectors.
> + ??? Ideally, we should have an overall estimation of register pressure
> + if we know the live range of all variables. */
> + if (!fp && kind == vec_construct
> + && reg_needed > target_avail_regs)
> + {
> + unsigned spill_cost = ix86_builtin_vectorization_cost (scalar_store,
> + vectype,
> + misalign);
> + stmt_cost += spill_cost * (reg_needed - target_avail_regs);
> + }
> }
> if (stmt_cost == -1)
> stmt_cost = ix86_builtin_vectorization_cost (kind, vectype, misalign);
> --
> 2.31.1
>
next prev parent reply other threads:[~2023-11-29 7:47 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-28 7:52 liuhongt
2023-11-29 7:47 ` Richard Biener [this message]
2023-11-29 8:00 ` Hongtao Liu
2023-12-01 2:38 ` [PATCH] Take register pressure into account for vec_construct/scalar_to_vec " liuhongt
2023-12-01 14:26 ` Richard Biener
2023-12-04 7:10 ` Hongtao Liu
2023-12-04 7:51 ` Uros Bizjak
2023-12-05 2:47 ` 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=CAFiYyc3q4cUpsEbGLdAudQf-AdCXN-rmZCUbQzSjiN8Mq3bouQ@mail.gmail.com \
--to=richard.guenther@gmail.com \
--cc=crazylht@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=hjl.tools@gmail.com \
--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).