Re: [PATCH V2] Enable small loop unrolling for O2

[Hongtao Liu <crazylht@gmail.com>]

On Mon, Nov 7, 2022 at 10:25 PM Richard Biener via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Wed, Nov 2, 2022 at 4:37 AM Hongyu Wang <hongyu.wang@intel.com> wrote:
> >
> > Hi, this is the updated patch of
> > https://gcc.gnu.org/pipermail/gcc-patches/2022-October/604345.html,
> > which uses targetm.loop_unroll_adjust as gate to enable small loop unroll.
> >
> > This patch does not change rs6000/s390 since I don't have machine to
> > test them, but I suppose the default behavior is the same since they
> > enable flag_unroll_loops at O2.
> >
> > Bootstrapped & regrtested on x86_64-pc-linux-gnu.
> >
> > Ok for trunk?
> >
> > ---------- Patch content --------
> >
> > Modern processors has multiple way instruction decoders
> > For x86, icelake/zen3 has 5 uops, 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, this patch enables loop unrolling for small size loop at O2
> > to fullfill the decoder as much as possible. It turns on rtl loop
> > unrolling when targetm.loop_unroll_adjust exists and O2 plus speed only.
> > In x86 backend the default behavior is to unroll small loops with less
> > than 4 insns by 1 time.
> >
> > This improves 548.exchange2 by 9% on icelake and 7.4% on zen3 with
> > 0.9% 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.
> >
> > gcc/ChangeLog:
> >
> >         * common/config/i386/i386-common.cc (ix86_optimization_table):
> >         Enable small loop unroll at O2 by default.
> >         * config/i386/i386.cc (ix86_loop_unroll_adjust): Adjust unroll
> >         factor if -munroll-only-small-loops enabled and -funroll-loops/
> >         -funroll-all-loops are disabled.
> >         * 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.
> >         * loop-init.cc (pass_rtl_unroll_loops::gate): Enable rtl
> >         loop unrolling for -O2-speed and above if target hook
> >         loop_unroll_adjust exists.
> >
> > gcc/testsuite/ChangeLog:
> >
> >         * gcc.dg/guality/loop-1.c: Add additional option
> >           -mno-unroll-only-small-loops.
> >         * gcc.target/i386/pr86270.c: Add -mno-unroll-only-small-loops.
> >         * gcc.target/i386/pr93002.c: Likewise.
> > ---
> >  gcc/common/config/i386/i386-common.cc   |  1 +
> >  gcc/config/i386/i386.cc                 | 18 ++++++++++++++++++
> >  gcc/config/i386/i386.opt                | 13 +++++++++++++
> >  gcc/doc/invoke.texi                     | 16 ++++++++++++++++
> >  gcc/loop-init.cc                        | 10 +++++++---
> >  gcc/testsuite/gcc.dg/guality/loop-1.c   |  2 ++
> >  gcc/testsuite/gcc.target/i386/pr86270.c |  2 +-
> >  gcc/testsuite/gcc.target/i386/pr93002.c |  2 +-
> >  8 files changed, 59 insertions(+), 5 deletions(-)
> >
> > diff --git a/gcc/common/config/i386/i386-common.cc b/gcc/common/config/i386/i386-common.cc
> > index f66bdd5a2af..c6891486078 100644
> > --- a/gcc/common/config/i386/i386-common.cc
> > +++ b/gcc/common/config/i386/i386-common.cc
> > @@ -1724,6 +1724,7 @@ 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_munroll_only_small_loops, NULL, 1 },
> >      /* 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.cc b/gcc/config/i386/i386.cc
> > index c0f37149ed0..0f94a3b609e 100644
> > --- a/gcc/config/i386/i386.cc
> > +++ b/gcc/config/i386/i386.cc
> > @@ -23827,6 +23827,24 @@ 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 (!(flag_unroll_loops
> > +       || flag_unroll_all_loops
> > +       || loop->unroll))
> > +    {
> > +      nunroll = 1;
> > +
> > +      /* Any explicit -f{no-}unroll-{all-}loops turns off
> > +        -munroll-only-small-loops.  */
> > +      if (ix86_unroll_only_small_loops
> > +         && !OPTION_SET_P (flag_unroll_loops))
> > +       if (loop->ninsns <= (unsigned) ix86_small_unroll_ninsns)
>
> either add braces or combine the two if's
>
> Otherwise the middle-end changes look OK.  The target maintainers need to decide
> whether the two --params should be core tunings instead - I would assume that
> given your rationale the decode and issue widths of the core plays an important
> role here.  That might also suggest a single parameter instead and unrolling
> (factor * issue_width) / loop->ninsns times instead of a static unroll_factor?
Although ix86_small_unroll_insns is coming from issue_rate, it's tuned
for codesize.
Make it exact as issue_rate and using factor * issue_width /
loop->ninsns may increase code size too much.
So I prefer to add those 2 parameters to the cost table for core
tunings instead of 1.
>
> Thanks,
> Richard.
>
> > +         nunroll = (unsigned) ix86_small_unroll_factor;
> > +
> > +      return nunroll;
> > +    }
> > +
> >    if (!TARGET_ADJUST_UNROLL)
> >       return nunroll;
> >
> > diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
> > index 53d534f6392..6da9c8d670d 100644
> > --- a/gcc/config/i386/i386.opt
> > +++ b/gcc/config/i386/i386.opt
> > @@ -1224,3 +1224,16 @@ mavxvnniint8
> >  Target Mask(ISA2_AVXVNNIINT8) Var(ix86_isa_flags2) Save
> >  Support MMX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, AVX, AVX2 and
> >  AVXVNNIINT8 built-in functions and code generation.
> > +
> > +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 550aec87809..487218bd0ce 100644
> > --- a/gcc/doc/invoke.texi
> > +++ b/gcc/doc/invoke.texi
> > @@ -15821,6 +15821,14 @@ 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
> > @@ -25232,6 +25240,14 @@ 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. Explicit
> > +-f[no-]unroll-[all-]loops would disable this flag to avoid any
> > +unintended unrolling behavior that user does not want.
> >  @end table
> >
> >  @node M32C Options
> > diff --git a/gcc/loop-init.cc b/gcc/loop-init.cc
> > index b9e07973dd6..9789efa1e11 100644
> > --- a/gcc/loop-init.cc
> > +++ b/gcc/loop-init.cc
> > @@ -565,9 +565,12 @@ public:
> >    {}
> >
> >    /* opt_pass methods: */
> > -  bool gate (function *) final override
> > +  bool gate (function *fun) final override
> >      {
> > -      return (flag_unroll_loops || flag_unroll_all_loops || cfun->has_unroll);
> > +      return (flag_unroll_loops || flag_unroll_all_loops || cfun->has_unroll
> > +             || (targetm.loop_unroll_adjust
> > +                 && optimize >= 2
> > +                 && optimize_function_for_speed_p (fun)));
> >      }
> >
> >    unsigned int execute (function *) final override;
> > @@ -583,7 +586,8 @@ pass_rtl_unroll_loops::execute (function *fun)
> >        if (dump_file)
> >         df_dump (dump_file);
> >
> > -      if (flag_unroll_loops)
> > +      if (flag_unroll_loops
> > +         || targetm.loop_unroll_adjust)
> >         flags |= UAP_UNROLL;
> >        if (flag_unroll_all_loops)
> >         flags |= UAP_UNROLL_ALL;
> > diff --git a/gcc/testsuite/gcc.dg/guality/loop-1.c b/gcc/testsuite/gcc.dg/guality/loop-1.c
> > index 1b1f6d32271..a32ea445a3f 100644
> > --- a/gcc/testsuite/gcc.dg/guality/loop-1.c
> > +++ b/gcc/testsuite/gcc.dg/guality/loop-1.c
> > @@ -1,5 +1,7 @@
> >  /* { dg-do run } */
> >  /* { dg-options "-fno-tree-scev-cprop -fno-tree-vectorize -g" } */
> > +/* { dg-additional-options "-mno-unroll-only-small-loops" { target ia32 } } */
> > +
> >
> >  #include "../nop.h"
> >
> > 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
> >



-- 
BR,
Hongtao