[Bug middle-end/102464] New: Miss optimization for (_Float16) sqrtf ((float) f16)

[Bug middle-end/102464] New: Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

            Bug ID: 102464
           Summary: Miss optimization for (_Float16) sqrtf ((float) f16)
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: crazylht at gmail dot com
  Target Milestone: ---
            Target: x86_64-*-* i?86-*-*

Related thread
[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574216.html
[2] https://gcc.gnu.org/pipermail/gcc-patches/2021-July/574330.html

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
Similar optimization also applies for
fma
fmax/fmin
fabs
ldexp
ceil
floor
trunc
round
rint
nearbyint
copysign

Since AVX512-FP16 has corresponding instructions.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2021-09-23
           Keywords|                            |internal-improvement

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

fabs and fma I don't think they need to be internal functions as there are
already tree codes for them.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
There's related optimizations in convert () which should ideally move to
match.pd

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[joseph at codesourcery dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #4 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
Note that for fma this would only be valid for 
-funsafe-math-optimizations.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
(gdb) p direct_internal_fn_supported_p (IFN_CEIL, type, OPTIMIZE_FOR_BOTH)
$110 = false

(gdb) p direct_internal_fn_supported_p (IFN_SQRT, type, OPTIMIZE_FOR_BOTH)
$111 = true

hmm, why?

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #6 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #5)
> (gdb) p direct_internal_fn_supported_p (IFN_CEIL, type, OPTIMIZE_FOR_BOTH)
> $110 = false
> 
> (gdb) p direct_internal_fn_supported_p (IFN_SQRT, type, OPTIMIZE_FOR_BOTH)
> $111 = true
> 
> hmm, why?

Hmm, Because in ix86_optab_supported_p, we have

    case rint_optab:
      if ((SSE_FLOAT_MODE_P (mode1)
          && TARGET_SSE_MATH
          && !flag_trapping_math
          && !TARGET_SSE4_1))
        return opt_type == OPTIMIZE_FOR_SPEED;
      return true;

    case floor_optab:
    case ceil_optab:
    case btrunc_optab:
      if ((SSE_FLOAT_MODE_P (mode1)
          && TARGET_SSE_MATH
          && !flag_trapping_math
          && TARGET_SSE4_1)
        return true;
      return opt_type == OPTIMIZE_FOR_SPEED;

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:613196462a62a28de8414b9023ec2be9a29ac3dc

commit r12-4242-g613196462a62a28de8414b9023ec2be9a29ac3dc
Author: liuhongt <hongtao.liu@intel.com>
Date:   Fri Sep 24 19:17:42 2021 +0800

    Simplify (_Float16) ceil ((double) x) to .CEIL (x) when available.

    gcc/ChangeLog:

            PR target/102464
            * config/i386/i386.c (ix86_optab_supported_p):
            Return true for HFmode.
            * match.pd: Simplify (_Float16) ceil ((double) x) to
            __builtin_ceilf16 (a) when a is _Float16 type and
            direct_internal_fn_supported_p.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr102464.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #8 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Richard Biener from comment #3)
> There's related optimizations in convert () which should ideally move to
> match.pd

When i try to mov convert stuffs to match.pd, i find some "mismatch", there's 3
cases
1. math functions are transformed under condition "optimize"
2. math functions are transformed under condition "optimize &&
flag_unsafe_math_optimizations"
3. math functions are transformed under condition "optimize &&
flag_unsafe_math_optimizations flag_errno_maths"

And for logb, it's case 1, which means it can be transformed w/o
!flag_errno_maths, but according to DEF_C99_BUILTIN        (BUILT_IN_LOGB,
"logb", BT_FN_DOUBLE_DOUBLE, ATTR_MATHFN_FPROUNDING_ERRNO), !flag_errno_maths
is needed and the transformation will be prevented by
gimple-match-head.c：maybe_push_res_to_seq

          /* We can't and should not emit calls to non-const functions.  */
          if (!(flags_from_decl_or_type (decl) & ECF_CONST))
            return NULL;


/* fabsl (extend(x)) -> extend(fabsf(x)), etc., if x is a float.  */
(for froms (BUILT_IN_FABS BUILT_IN_FABSL
            BUILT_IN_LOGB BUILT_IN_LOGBL)
     tos (BUILT_IN_FABSF BUILT_IN_FABSF
          BUILT_IN_LOGBF BUILT_IN_LOGBF)
(simplify
  (froms (convert float_value_p@0))
    (if (optimize && canonicalize_math_p ()
         && mathfn_built_in (TREE_TYPE (@0), froms))
      (convert (tos @0)))))

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:1a07bc9cda77b1211e95ae295b30e46c0d9ee222

commit r12-4651-g1a07bc9cda77b1211e95ae295b30e46c0d9ee222
Author: liuhongt <hongtao.liu@intel.com>
Date:   Mon Oct 25 10:51:33 2021 +0800

    Simplify (_Float16) sqrtf((float) a) to .SQRT(a) when a is a _Float16
value.

    Similar for sqrt/sqrtl.

    gcc/ChangeLog:

            PR target/102464
            * match.pd: Simplify (_Float16) sqrtf((float) a) to .SQRT(a)
            when direct_internal_fn_supported_p, similar for sqrt/sqrtl.

    gcc/testsuite/ChangeLog:

            PR target/102464
            * gcc.target/i386/pr102464-sqrtph.c: New test.
            * gcc.target/i386/pr102464-sqrtsh.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:84bcefd5555af6d95e08cd980965098961289215

commit r12-4780-g84bcefd5555af6d95e08cd980965098961289215
Author: liuhongt <hongtao.liu@intel.com>
Date:   Mon Oct 25 15:20:35 2021 +0800

    Enable vectorization for _Float16 floor/ceil/trunc/nearbyint/rint
operations.

    gcc/ChangeLog:

            PR target/102464
            * config/i386/i386-builtin-types.def (V8HF_FTYPE_V8HF): New
            function type.
            (V16HF_FTYPE_V16HF): Ditto.
            (V32HF_FTYPE_V32HF): Ditto.
            (V8HF_FTYPE_V8HF_ROUND): Ditto.
            (V16HF_FTYPE_V16HF_ROUND): Ditto.
            (V32HF_FTYPE_V32HF_ROUND): Ditto.
            * config/i386/i386-builtin.def ( IX86_BUILTIN_FLOORPH,
            IX86_BUILTIN_CEILPH, IX86_BUILTIN_TRUNCPH,
            IX86_BUILTIN_FLOORPH256, IX86_BUILTIN_CEILPH256,
            IX86_BUILTIN_TRUNCPH256, IX86_BUILTIN_FLOORPH512,
            IX86_BUILTIN_CEILPH512, IX86_BUILTIN_TRUNCPH512): New builtin.
            * config/i386/i386-builtins.c
            (ix86_builtin_vectorized_function): Enable vectorization for
            HFmode FLOOR/CEIL/TRUNC operation.
            * config/i386/i386-expand.c (ix86_expand_args_builtin): Handle
            new builtins.
            * config/i386/sse.md (rint<mode>2, nearbyint<mode>2): Extend
            to vector HFmodes.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr102464-vrndscaleph.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:22ce7382fccc15ce2355306b3f5be7afc00f81f4

commit r12-4881-g22ce7382fccc15ce2355306b3f5be7afc00f81f4
Author: liuhongt <hongtao.liu@intel.com>
Date:   Wed Nov 3 16:07:34 2021 +0800

    Simplify (trunc)copysign((extend)a, (extend)b) to .COPYSIGN (a,b).

    a and b are same type as the truncation type and has less precision
    than extend type.

    gcc/ChangeLog:

            PR target/102464
            * match.pd: simplify (trunc)copysign((extend)a, (extend)b) to
            .COPYSIGN (a,b) when a and b are same type as the truncation
            type and has less precision than extend type.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr102464-copysign-1.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:2ad1e8081f4797a99a96b513ffe14c7305e9b3d8

commit r12-4984-g2ad1e8081f4797a99a96b513ffe14c7305e9b3d8
Author: liuhongt <hongtao.liu@intel.com>
Date:   Mon Nov 8 09:19:29 2021 +0800

    [Gimple] Simplify (trunc)fma ((extend)a, (extend)b, (extend)c) to IFN_FMA
(a,b, c).

    a, b, c are same type as truncation type and has less precision than
    extend type, the optimization is guarded under
    flag_unsafe_math_optimizations.

    gcc/ChangeLog:
            PR target/102464
            * match.pd: Simplify
            (trunc)fma ((extend)a, (extend)b, (extend)c) to IFN_FMA (a, b,
            c) under flag_unsafe_math_optimizations.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr102464-fma.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #13 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:a1f7ead09cd41d32e2feffff902eb32e587c36e7

commit r12-4985-ga1f7ead09cd41d32e2feffff902eb32e587c36e7
Author: liuhongt <hongtao.liu@intel.com>
Date:   Mon Nov 8 09:32:17 2021 +0800

    Add !HONOR_SNANS to simplifcation: (trunc)copysign((extend)a, (extend)b) to
copysign (a, b).

    > Note that this is not safe with -fsignaling-nans, so needs to be disabled
    > for that option (if there isn't already logic somewhere with that
effect),
    > because the extend will convert a signaling NaN to quiet (raising
    > "invalid"), but copysign won't, so this transformation could result in a
    > signaling NaN being wrongly returned when the original code would never
    > have returned a signaling NaN.
    >
    > --
    > Joseph S. Myers
    > joseph@codesourcery.com

    gcc/ChangeLog

            PR target/102464
            * match.pd (Simplifcation (trunc)copysign((extend)a, (extend)b)
            to .COPYSIGN (a, b)): Add !HONOR_SNANS.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #14 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>:

https://gcc.gnu.org/g:b879d40a17ec0409f1a2cd9ab6134bb28f53eea8

commit r12-5079-gb879d40a17ec0409f1a2cd9ab6134bb28f53eea8
Author: liuhongt <hongtao.liu@intel.com>
Date:   Thu Nov 4 16:05:45 2021 +0800

    Simplify (trunc)MAX/MIN((extend)a, (extend)b) to MAX/MIN(a,b)

    a and b are same type as trunc type and has less precision than
    extend type.

    gcc/ChangeLog:

            PR target/102464
            * match.pd: Simplify (trunc)fmax/fmin((extend)a, (extend)b) to
            MAX/MIN(a,b)

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr102464-maxmin.c: New test.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

Hongtao.liu <crazylht at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #15 from Hongtao.liu <crazylht at gmail dot com> ---
Fixed in GCC12.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[jbeulich at suse dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

jbeulich at suse dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jbeulich at suse dot com

--- Comment #16 from jbeulich at suse dot com ---
(In reply to Hongtao.liu from comment #15)
> Fixed in GCC12.

Only almost - the new FMA testcase there fails for i?86-*-*. I don't think even
the few uses of VFMA* actually match the expectations. The majority of the
operations are carried in the FPU anyway, despite -mfpmath=sse.

[Bug middle-end/102464] Miss optimization for (_Float16) sqrtf ((float) f16)

[jbeulich at suse dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102464

--- Comment #17 from jbeulich at suse dot com ---
Largely the same is actually true for the RNDSCALEPH test added for the PR
here.