[Bug middle-end/110018] New: Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[Bug middle-end/110018] New: Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

            Bug ID: 110018
           Summary: Missing vectorizable_conversion(unsigned char ->
                    double) for BB vectorizer
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: crazylht at gmail dot com
  Target Milestone: ---

When Looking at PR109812, I noticed there's missing vectorizable_conversion for
BB vectorizer when target doesn't support direct optab for unsigned char to
double. But actually it can be vectorized via unsigned char -> short/int/long
long -> double when vectorizable_conversion is ok for any of the immediate
type.

Currently, when modifier is NONE, vectorizable_conversion doesn't try any
immediate type, it can be extended similar like WIDEN.

 5158    case NONE:
 5159      if (code != FIX_TRUNC_EXPR
 5160          && code != FLOAT_EXPR
 5161          && !CONVERT_EXPR_CODE_P (code))
 5162        return false;
 5163      if (supportable_convert_operation (code, vectype_out, vectype_in,
&code1))
 5164        break;
 5165      /* FALLTHRU */

void
foo (double* __restrict a, unsigned char* b)
{
    a[0] = b[0];
    a[1] = b[1];
    a[2] = b[2];
    a[3] = b[3];
    a[4] = b[4];
    a[5] = b[5];
    a[6] = b[6];
    a[7] = b[7];
}

missed:   conversion not supported by target.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-05-29

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---

> Currently, when modifier is NONE, vectorizable_conversion doesn't try any
> immediate type, it can be extended similar like WIDEN.
> 
After gdb the testcase, the modifier is not NONE, it's widen from V8QI to V4DF,
and failed.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
Even we have direct optab for unsigned char to long long, BB vectorizer failer
to optimize

void
foo1 (long long* __restrict a, unsigned char* b)
{
    a[0] = b[0];
    a[1] = b[1];
    a[2] = b[2];
    a[3] = b[3];
    a[4] = b[4];
    a[5] = b[5];
    a[6] = b[6];
    a[7] = b[7];
}

since we have vectype_out as V4DImode, but vectype_in as V8QImode as *vectype =
SLP_TREE_VECTYPE (child);

If manually change the testcase to

void
foo1 (long long* __restrict a, unsigned char* b)
{
    a[0] = b[0];
    a[1] = b[1];
    a[2] = b[2];
    a[3] = b[3];
}

Then BB vectorizer can do vectorization, but still failed for

void
foo1 (double* __restrict a, unsigned char* b)
{
    a[0] = b[0];
    a[1] = b[1];
    a[2] = b[2];
    a[3] = b[3];
}

So it looks there're 2 separate issue
1. As mentioned in the PR, we can use intermediate type for vectorizable
unsigned char to double.
2. There's mismatch between vectype_out and vectype_in as SLP_TREE_VECTYPE
(child);

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the easiest would be to handle this in pattern recog and piggy-back
on the existing multi-step conversion support for the integral promotion
whilst not relying on such for int<->float conversions.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
Do we know if the stmt belongs to any slp_node in pattern recog? slp_node use
SLP_TREE_LANES for vectype which is different from vinfo->vector_mode.
Different optabs should be checked according to both vectype_in and
vectype_out, for non-slp case, assuming vec_pack/unpack are always used when
lhs has different size from rhs, for slp case, sometimes vec_pack/unpack is
used, somethings direct conversion is used.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- Comment #6 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:6f19cf7526168f840fd22f6af3f0cb67efb90dc8

commit r14-2007-g6f19cf7526168f840fd22f6af3f0cb67efb90dc8
Author: liuhongt <hongtao.liu@intel.com>
Date:   Wed May 31 11:20:46 2023 +0800

    Use intermiediate integer type for float_expr/fix_trunc_expr when direct
optab is not existed.

    We have already use intermidate type in case WIDEN, but not for NONE,
    this patch extended that.

    gcc/ChangeLog:

            PR target/110018
            * tree-vect-stmts.cc (vectorizable_conversion): Use
            intermiediate integer type for float_expr/fix_trunc_expr when
            direct optab is not existed.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr110018-1.c: New test.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

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

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

--- Comment #7 from Hongtao.liu <crazylht at gmail dot com> ---
Fixed for GCC14.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Uros Bizjak <uros@gcc.gnu.org>:

https://gcc.gnu.org/g:b9401c3a323c59705eca177bf72c13c0d2f462b6

commit r14-2020-gb9401c3a323c59705eca177bf72c13c0d2f462b6
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Wed Jun 21 16:34:39 2023 +0200

    vect: Add testcases for unsigned conversions [PR110018]

    Also test convresions with unsigned types.

            PR target/110018

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr110018-1.c: Use explicit signed types.
            * gcc.target/i386/pr110018-2.c: New test.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- 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:77a50c772771f681085922b493922516c3c03e9a

commit r14-2085-g77a50c772771f681085922b493922516c3c03e9a
Author: liuhongt <hongtao.liu@intel.com>
Date:   Sun Jun 25 11:35:09 2023 +0800

    Don't use intermiediate type for FIX_TRUNC_EXPR when ftrapping-math.

    > > Hmm, good question.  GENERIC has a direct truncation to unsigned char
    > > for example, the C standard generally says if the integral part cannot
    > > be represented then the behavior is undefined.  So I think we should be
    > > safe here (0x1.0p32 doesn't fit an int).
    >
    > We should be following Annex F (unspecified value plus "invalid"
exception
    > for out-of-range floating-to-integer conversions rather than undefined
    > behavior).  But we don't achieve that very well at present (see bug 93806
    > comments 27-29 for examples of how such conversions produce wobbly
    > values).

    That would mean guarding this with !flag_trapping_math would be the
appropriate
    thing to do.

    gcc/ChangeLog:

            PR tree-optimization/110371
            PR tree-optimization/110018
            * tree-vect-stmts.cc (vectorizable_conversion): Don't use
            intermiediate type for FIX_TRUNC_EXPR when ftrapping-math.

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr110018-1.c: Add -fno-trapping-math to
dg-options.
            * gcc.target/i386/pr110018-2.c: Ditto.

[Bug middle-end/110018] Missing vectorizable_conversion(unsigned char -> double) for BB vectorizer

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110018

--- 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:1bfe7e5352d1f4ac525317454aca45aa80a517ba

commit r14-2086-g1bfe7e5352d1f4ac525317454aca45aa80a517ba
Author: liuhongt <hongtao.liu@intel.com>
Date:   Sun Jun 25 11:12:29 2023 +0800

    Use cvt_op to save intermediate type operand instead of "subtle" vec_dest.

    When there're multiple operands in vec_oprnds0, vec_dest will be
    overwrited to vectype_out, but in multi_step_cvt case, cvt_type is
    expected. It caused an ICE when verify_gimple_in_cfg.

    gcc/ChangeLog:

            PR tree-optimization/110371
            PR tree-optimization/110018
            * tree-vect-stmts.cc (vectorizable_conversion): Use cvt_op to
            save intermediate type operand instead of "subtle" vec_dest
            for case NONE.

    gcc/testsuite/ChangeLog:

            * gcc.target/aarch64/pr110371.c: New test.