public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
From: Hongyu Wang <wwwhhhyyy333@gmail.com>
To: Uros Bizjak <ubizjak@gmail.com>
Cc: Hongyu Wang <hongyu.wang@intel.com>,
	hongtao.liu@intel.com, gcc-patches@gcc.gnu.org
Subject: Re: [PATCH] i386: Enable small loop unrolling for O2
Date: Wed, 26 Oct 2022 16:13:31 +0800	[thread overview]
Message-ID: <CA+OydWn4oqLdnjAdUdw7g=W90fM_Jw41uTFS+dVMh6AvUsfqPg@mail.gmail.com> (raw)
In-Reply-To: <CAFULd4a_PJY_t-=FVTtvcacgCu+azWsfX-8rAqFudy93bHXreQ@mail.gmail.com>

> Does this setting benefit all targets?  IIRC, in the past all
> benchmarks also enabled -funroll-loops, so it looks to me that
> unrolling small loops by default is a good compromise.

The idea to unroll small loops can be explained from the x86
micro-architecture. Modern x86 processors has multiple way instruction
decoder (5uops for icelake/zen3). So for small loop with <= 4
instructions (usually has 3 uops with a cmp/jmp pair that can be
macro-fused), the decoder would have 2 uops bubble for each iteration
and the pipeline could not be fully utilized. Therefore we decide to
unroll the 4 insn loop once to at least to full-fill the decoder and
enhance the pipeline utilization.
We are not familiar with micro architecture of other targets, we don't
know whether the unrolling could benefit the instruction decoder, so
the decision could be different.

Uros Bizjak via Gcc-patches <gcc-patches@gcc.gnu.org> 于2022年10月26日周三 14:57写道:

>
> On Wed, Oct 26, 2022 at 7:53 AM Hongyu Wang <hongyu.wang@intel.com> wrote:
> >
> > Hi,
> >
> > Inspired by rs6000 and s390 port changes, this patch
> > enables loop unrolling for small size loop at O2 by default.
> > The default behavior is to unroll loop with unknown trip-count and
> > less than 4 insns by 1 time.
> >
> > This improves 548.exchange2 by 3.5% on icelake and 6% on zen3 with
> > 1.2% codesize increment. For other benchmarks the variants are minor
> > and overall codesize increased by 0.2%.
> >
> > The kernel image size increased by 0.06%, and no impact on eembc.
>
> Does this setting benefit all targets?  IIRC, in the past all
> benchmarks also enabled -funroll-loops, so it looks to me that
> unrolling small loops by default is a good compromise.
>
> The patch is technically OK, but as a tuning default, I would leave
> the final approval to HJ.
>
> Thanks,
> Uros.
>
> >
> > Bootstrapped & regrtested on x86_64-pc-linux-gnu.
> >
> > Ok for trunk?
> >
> > gcc/ChangeLog:
> >
> >         * common/config/i386/i386-common.cc (ix86_optimization_table):
> >         Enable loop unroll and small loop unroll at O2 by default.
> >         * config/i386/i386-options.cc
> >         (ix86_override_options_after_change):
> >         Disable small loop unroll when funroll-loops enabled, reset
> >         cunroll_grow_size when it is not explicitly enabled.
> >         (ix86_option_override_internal): Call
> >         ix86_override_options_after_change instead of calling
> >         ix86_recompute_optlev_based_flags and ix86_default_align
> >         separately.
> >         * config/i386/i386.cc (ix86_loop_unroll_adjust): Adjust unroll
> >         factor if -munroll-only-small-loops enabled.
> >         * config/i386/i386.opt: Add -munroll-only-small-loops,
> >         -param=x86-small-unroll-ninsns= for loop insn limit,
> >         -param=x86-small-unroll-factor= for unroll factor.
> >         * doc/invoke.texi: Document -munroll-only-small-loops,
> >         x86-small-unroll-ninsns and x86-small-unroll-factor.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.target/i386/pr86270.c: Add -mno-unroll-only-small-loops.
> >         * gcc.target/i386/pr93002.c: Likewise.
> > ---
> >  gcc/common/config/i386/i386-common.cc   |  6 ++++
> >  gcc/config/i386/i386-options.cc         | 40 ++++++++++++++++++++++---
> >  gcc/config/i386/i386.cc                 | 13 ++++++++
> >  gcc/config/i386/i386.opt                | 13 ++++++++
> >  gcc/doc/invoke.texi                     | 14 +++++++++
> >  gcc/testsuite/gcc.target/i386/pr86270.c |  2 +-
> >  gcc/testsuite/gcc.target/i386/pr93002.c |  2 +-
> >  7 files changed, 84 insertions(+), 6 deletions(-)
> >
> > diff --git a/gcc/common/config/i386/i386-common.cc b/gcc/common/config/i386/i386-common.cc
> > index d6a68dc9b1d..0e580b39d14 100644
> > --- a/gcc/common/config/i386/i386-common.cc
> > +++ b/gcc/common/config/i386/i386-common.cc
> > @@ -1686,6 +1686,12 @@ static const struct default_options ix86_option_optimization_table[] =
> >      /* The STC algorithm produces the smallest code at -Os, for x86.  */
> >      { OPT_LEVELS_2_PLUS, OPT_freorder_blocks_algorithm_, NULL,
> >        REORDER_BLOCKS_ALGORITHM_STC },
> > +    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_funroll_loops, NULL, 1 },
> > +    { OPT_LEVELS_2_PLUS_SPEED_ONLY, OPT_munroll_only_small_loops, NULL, 1 },
> > +    /* Turns off -frename-registers and -fweb which are enabled by
> > +       funroll-loops.  */
> > +    { OPT_LEVELS_ALL, OPT_frename_registers, NULL, 0 },
> > +    { OPT_LEVELS_ALL, OPT_fweb, NULL, 0 },
> >      /* Turn off -fschedule-insns by default.  It tends to make the
> >         problem with not enough registers even worse.  */
> >      { OPT_LEVELS_ALL, OPT_fschedule_insns, NULL, 0 },
> > diff --git a/gcc/config/i386/i386-options.cc b/gcc/config/i386/i386-options.cc
> > index acb2291e70f..6ea347c32e1 100644
> > --- a/gcc/config/i386/i386-options.cc
> > +++ b/gcc/config/i386/i386-options.cc
> > @@ -1819,8 +1819,43 @@ ix86_recompute_optlev_based_flags (struct gcc_options *opts,
> >  void
> >  ix86_override_options_after_change (void)
> >  {
> > +  /* Default align_* from the processor table.  */
> >    ix86_default_align (&global_options);
> > +
> >    ix86_recompute_optlev_based_flags (&global_options, &global_options_set);
> > +
> > +  /* Disable unrolling small loops when there's explicit
> > +     -f{,no}unroll-loop.  */
> > +  if ((OPTION_SET_P (flag_unroll_loops))
> > +     || (OPTION_SET_P (flag_unroll_all_loops)
> > +        && flag_unroll_all_loops))
> > +    {
> > +      if (!OPTION_SET_P (ix86_unroll_only_small_loops))
> > +       ix86_unroll_only_small_loops = 0;
> > +      /* Re-enable -frename-registers and -fweb if funroll-loops
> > +        enabled.  */
> > +      if (!OPTION_SET_P (flag_web))
> > +       flag_web = flag_unroll_loops;
> > +      if (!OPTION_SET_P (flag_rename_registers))
> > +       flag_rename_registers = flag_unroll_loops;
> > +      if (!OPTION_SET_P (flag_cunroll_grow_size))
> > +       flag_cunroll_grow_size = flag_unroll_loops
> > +                                || flag_peel_loops
> > +                                || optimize >= 3;
> > +    }
> > +  else
> > +    {
> > +      if (!OPTION_SET_P (flag_cunroll_grow_size))
> > +       flag_cunroll_grow_size = flag_peel_loops || optimize >= 3;
> > +      /* Disables loop unrolling if -mno-unroll-only-small-loops is
> > +        explicitly set and -funroll-loops is not enabled.  */
> > +      if (OPTION_SET_P (ix86_unroll_only_small_loops)
> > +         && !ix86_unroll_only_small_loops
> > +         && !(OPTION_SET_P (flag_unroll_loops)
> > +              || OPTION_SET_P (flag_unroll_all_loops)))
> > +       flag_unroll_loops = flag_unroll_all_loops = 0;
> > +    }
> > +
> >  }
> >
> >  /* Clear stack slot assignments remembered from previous functions.
> > @@ -2332,7 +2367,7 @@ ix86_option_override_internal (bool main_args_p,
> >
> >    set_ix86_tune_features (opts, ix86_tune, opts->x_ix86_dump_tunes);
> >
> > -  ix86_recompute_optlev_based_flags (opts, opts_set);
> > +  ix86_override_options_after_change ();
> >
> >    ix86_tune_cost = processor_cost_table[ix86_tune];
> >    /* TODO: ix86_cost should be chosen at instruction or function granuality
> > @@ -2363,9 +2398,6 @@ ix86_option_override_internal (bool main_args_p,
> >        || TARGET_64BIT_P (opts->x_ix86_isa_flags))
> >      opts->x_ix86_regparm = REGPARM_MAX;
> >
> > -  /* Default align_* from the processor table.  */
> > -  ix86_default_align (opts);
> > -
> >    /* Provide default for -mbranch-cost= value.  */
> >    SET_OPTION_IF_UNSET (opts, opts_set, ix86_branch_cost,
> >                        ix86_tune_cost->branch_cost);
> > diff --git a/gcc/config/i386/i386.cc b/gcc/config/i386/i386.cc
> > index 480db35f6cd..75829a5d0f4 100644
> > --- a/gcc/config/i386/i386.cc
> > +++ b/gcc/config/i386/i386.cc
> > @@ -23820,6 +23820,19 @@ ix86_loop_unroll_adjust (unsigned nunroll, class loop *loop)
> >    unsigned i;
> >    unsigned mem_count = 0;
> >
> > +  /* Unroll small size loop when unroll factor is not explicitly
> > +     specified.  */
> > +  if (ix86_unroll_only_small_loops && !loop->unroll)
> > +    {
> > +      int small_unroll = 0;
> > +      if (loop->ninsns <= (unsigned) ix86_small_unroll_ninsns)
> > +       small_unroll = MIN ((unsigned) ix86_small_unroll_factor,
> > +                           nunroll);
> > +      else
> > +       small_unroll = 1;
> > +      return small_unroll;
> > +    }
> > +
> >    if (!TARGET_ADJUST_UNROLL)
> >       return nunroll;
> >
> > diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
> > index 0dbaacb57ed..a724c73c0c4 100644
> > --- a/gcc/config/i386/i386.opt
> > +++ b/gcc/config/i386/i386.opt
> > @@ -1214,3 +1214,16 @@ Do not use GOT to access external symbols.
> >  -param=x86-stlf-window-ninsns=
> >  Target Joined UInteger Var(x86_stlf_window_ninsns) Init(64) Param
> >  Instructions number above which STFL stall penalty can be compensated.
> > +
> > +munroll-only-small-loops
> > +Target Var(ix86_unroll_only_small_loops) Init(0) Save
> > +Enable conservative small loop unrolling.
> > +
> > +-param=x86-small-unroll-ninsns=
> > +Target Joined UInteger Var(ix86_small_unroll_ninsns) Init(4) Param
> > +Insturctions number limit for loop to be unrolled under
> > +-munroll-only-small-loops.
> > +
> > +-param=x86-small-unroll-factor=
> > +Target Joined UInteger Var(ix86_small_unroll_factor) Init(2) Param
> > +Unroll factor for -munroll-only-small-loops.
> > diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
> > index cd4d3c1d72c..b6fa79eccc3 100644
> > --- a/gcc/doc/invoke.texi
> > +++ b/gcc/doc/invoke.texi
> > @@ -15779,6 +15779,14 @@ The following choices of @var{name} are available on i386 and x86_64 targets:
> >  @item x86-stlf-window-ninsns
> >  Instructions number above which STFL stall penalty can be compensated.
> >
> > +@item x86-small-unroll-ninsns
> > +If -munroll-only-small-loops is enabled, only unroll loops with instruction
> > +count less than this parameter. The default value is 4.
> > +
> > +@item x86-small-unroll-factor
> > +If -munroll-only-small-loops is enabled, reset the unroll factor with this
> > +value. The default value is 2 which means the loop will be unrolled once.
> > +
> >  @end table
> >
> >  @end table
> > @@ -25186,6 +25194,12 @@ environments where no dynamic link is performed, like firmwares, OS
> >  kernels, executables linked with @option{-static} or @option{-static-pie}.
> >  @option{-mdirect-extern-access} is not compatible with @option{-fPIC} or
> >  @option{-fpic}.
> > +
> > +@item -munroll-only-small-loops
> > +@itemx -mno-unroll-only-small-loops
> > +@opindex munroll-only-small-loops
> > +Controls conservative small loop unrolling. It is default enbaled by
> > +O2, and unrolls loop with less than 4 insns by 1 time.
> >  @end table
> >
> >  @node M32C Options
> > diff --git a/gcc/testsuite/gcc.target/i386/pr86270.c b/gcc/testsuite/gcc.target/i386/pr86270.c
> > index 81841ef5bd7..cbc9fbb0450 100644
> > --- a/gcc/testsuite/gcc.target/i386/pr86270.c
> > +++ b/gcc/testsuite/gcc.target/i386/pr86270.c
> > @@ -1,5 +1,5 @@
> >  /* { dg-do compile } */
> > -/* { dg-options "-O2" } */
> > +/* { dg-options "-O2 -mno-unroll-only-small-loops" } */
> >
> >  int *a;
> >  long len;
> > diff --git a/gcc/testsuite/gcc.target/i386/pr93002.c b/gcc/testsuite/gcc.target/i386/pr93002.c
> > index 0248fcc00a5..f75a847f75d 100644
> > --- a/gcc/testsuite/gcc.target/i386/pr93002.c
> > +++ b/gcc/testsuite/gcc.target/i386/pr93002.c
> > @@ -1,6 +1,6 @@
> >  /* PR target/93002 */
> >  /* { dg-do compile } */
> > -/* { dg-options "-O2" } */
> > +/* { dg-options "-O2 -mno-unroll-only-small-loops" } */
> >  /* { dg-final { scan-assembler-not "cmp\[^\n\r]*-1" } } */
> >
> >  volatile int sink;
> > --
> > 2.18.1
> >

  reply	other threads:[~2022-10-26  8:18 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2022-10-26  5:52 Hongyu Wang
2022-10-26  6:56 ` Uros Bizjak
2022-10-26  8:13   ` Hongyu Wang [this message]
2022-10-28  7:33 ` Richard Biener
2022-10-28  8:03   ` Hongyu Wang
2022-10-28  8:41     ` Richard Biener

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='CA+OydWn4oqLdnjAdUdw7g=W90fM_Jw41uTFS+dVMh6AvUsfqPg@mail.gmail.com' \
    --to=wwwhhhyyy333@gmail.com \
    --cc=gcc-patches@gcc.gnu.org \
    --cc=hongtao.liu@intel.com \
    --cc=hongyu.wang@intel.com \
    --cc=ubizjak@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).