[Bug target/98218] New: [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[Bug target/98218] New: [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

            Bug ID: 98218
           Summary: [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander
                    for 64bit vector
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: crazylht at gmail dot com
  Target Milestone: ---
              Host: x86_64-pc-linux-gnu
            Target: x86_64-*-* i?86-*-*

Refer to https://godbolt.org/z/sYE88f

cat test.c

typedef char v8qi __attribute__ ((vector_size(8)));
v8qi f1(v8qi a, v8qi b) {
  return a == b;
}

gcc -O2 -msse4.1 -S

f1(char __vector(8), char __vector(8)):
        pextrb  edx, xmm0, 0
        pextrb  eax, xmm1, 0
        pextrb  ecx, xmm0, 1
        cmp     dl, al
        pextrb  eax, xmm1, 1
        pextrb  esi, xmm0, 2
        setne   dl
        pextrb  edi, xmm0, 3
        pextrb  r8d, xmm0, 4
        sub     edx, 1
        cmp     cl, al
        pextrb  eax, xmm1, 2
        setne   cl
        pextrb  r9d, xmm0, 5
        movzx   edx, dl
        sub     ecx, 1
        cmp     sil, al
        pextrb  eax, xmm1, 3
        setne   sil
        pextrb  r10d, xmm0, 6
        pextrb  r11d, xmm0, 7
        movzx   ecx, cl
        sub     esi, 1
        cmp     dil, al
        pextrb  eax, xmm1, 4
        setne   dil
        movzx   esi, sil
        sub     edi, 1
        cmp     r8b, al
        pextrb  eax, xmm1, 5
        setne   r8b
        movzx   edi, dil
        sub     r8d, 1
        cmp     r9b, al
        pextrb  eax, xmm1, 6
        setne   r9b
        movzx   r8d, r8b
        sub     r9d, 1
        cmp     r10b, al
        pextrb  eax, xmm1, 7
        setne   r10b
        movzx   r9d, r9b
        sub     r10d, 1
        cmp     r11b, al
        setne   al
        movzx   r10d, r10b
        sub     eax, 1
        movzx   eax, al
        sal     rax, 8
        or      rax, r10
        sal     rax, 8
        or      rax, r9
        sal     rax, 8
        or      rax, r8
        sal     rax, 8
        or      rax, rdi
        sal     rax, 8
        or      rax, rsi
        sal     rax, 8
        or      rax, rcx
        sal     rax, 8
        or      rax, rdx
        movq    xmm0, rax
        ret

It should be better with

f1(char __vector(8), char __vector(8)):                           # @f1(char
__vector(8), char __vector(8))
        pcmpeqb xmm0, xmm1
        ret

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-12-10
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 49796
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49796&action=edit
Proposed patch to implement integer vector compares

Attached patch implements integer vector compares.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #3 from Uroš Bizjak <ubizjak at gmail dot com> ---
Testcase 1:

--cut here--
typedef short vec __attribute__((vector_size(8)));
typedef unsigned short uvec __attribute__((vector_size(8)));

vec lt (vec a, vec b) { return a < b; }
vec le (vec a, vec b) { return a <= b; }
vec eq (vec a, vec b) { return a == b; }
vec ne (vec a, vec b) { return a != b; }
vec ge (vec a, vec b) { return a >= b; }
vec gt (vec a, vec b) { return a > b; }

vec ltu (uvec a, uvec b) { return a < b; }
vec leu (uvec a, uvec b) { return a <= b; }
vec geu (uvec a, uvec b) { return a >= b; }
vec gtu (uvec a, uvec b) { return a > b; }
--cut here--

Testcase 2:

--cut here--
typedef short T;

#define M 4

T a[M];
T b[M];
T s1[M];
T s2[M];
T r[M];

void foo (void)
{
  int j;

  for (j = 0; j < M; j++)
    r[j] = (a[j] < b[j]) ? s1[j] : s2[j];
}
--cut here--

gcc -O3:

foo:
        movq    a(%rip), %xmm1
        movq    b(%rip), %xmm0
        movq    s2(%rip), %xmm2
        pcmpgtw %xmm1, %xmm0
        movq    s1(%rip), %xmm1
        pand    %xmm0, %xmm1
        pandn   %xmm2, %xmm0
        por     %xmm1, %xmm0
        movq    %xmm0, r(%rip)
        ret

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
             Target|x86_64-*-* i?86-*-*         |x86
             Status|NEW                         |ASSIGNED
               Host|x86_64-pc-linux-gnu         |
   Target Milestone|---                         |12.0
           Assignee|unassigned at gcc dot gnu.org      |ubizjak at gmail dot com

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #4 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:f3661f2d63fbc5fd30c24d22137691e16b0a0a17

commit r12-514-gf3661f2d63fbc5fd30c24d22137691e16b0a0a17
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Wed May 5 15:07:25 2021 +0200

    i386: Implement integer vector compares for 64bit vectors [PR98218]

    Implement integer vector compares for 64bit vectors for
TARGET_MMX_WITH_SSE.

    2021-05-05  Uroš Bizjak  <ubizjak@gmail.com>

    gcc/
            PR target/98218
            * config/i386/i386-expand.c (ix86_expand_int_sse_cmp):
            Handle V8QI, V4HI and V2SI modes.
            * config/i386/i386.c (ix86_build_const_vector): Handle V2SImode.
            (ix86_build_signbit_mask): Ditto.
            * config/i386/mmx.md (MMXMODE14): New mode iterator.
            (<smaxmin:code><MMXMODE14:mode>3): New expander.
            (*mmx_<smaxmin:code><MMXMODE14:mode>3): New insn pattern.
            (<umaxmin:code><MMXMODE24:mode>3): New expander.
            (*mmx_<umaxmin:code><MMXMODE24:mode>3): New insn pattern.
            (vec_cmp<MMXMODEI:mode><MMXMODEI:mode>): New expander.
            (vec_cmpu<MMXMODEI:mode><MMXMODEI:mode>): Ditto.
            (vcond<MMXMODEI:mode><MMXMODEI:mode>): Ditto.
            (vcondu<MMXMODEI:mode><MMXMODEI:mode>): Ditto.
            (vcond_mask_<MMXMODEI:mode><MMXMODEI:mode>): Ditto.

    gcc/testsuite/

            PR target/98218
            * gcc.target/i386/pr98218-1.c: New test.
            * gcc.target/i386/pr98218-1a.c: Ditto.
            * gcc.target/i386/pr98218-2.c: Ditto.
            * gcc.target/i386/pr98218-2a.c: Ditto.
            * gcc.target/i386/pr98218-3.c: Ditto.
            * gcc.target/i386/pr98218-3a.c: Ditto.
            * gcc.dg/vect/vect-bool-cmp.c (dg-final):
            Scan vect tree dump for "LOOP VECTORIZED", not VECTORIZED.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

Uroš Bizjak <ubizjak at gmail dot com> changed:

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

--- Comment #5 from Uroš Bizjak <ubizjak at gmail dot com> ---
Implemented for gcc-12.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #6 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:5795ec0edc30e077a9900cf3ca0a04ad8ac5ac97

commit r12-615-g5795ec0edc30e077a9900cf3ca0a04ad8ac5ac97
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Fri May 7 17:14:34 2021 +0200

    i386: Implement mmx_pblendv to optimize SSE conditional moves [PR98218]

    Implement mmx_pblendv to optimize V8HI, V4HI and V2SI mode
    conditional moves for SSE4.1 targets.

    2021-05-07  Uroš Bizjak  <ubizjak@gmail.com>

    gcc/
            PR target/98218
            * config/i386/i386-expand.c (ix86_expand_sse_movcc):
            Handle V8QI, V4HI and V2SI modes.
            * config/i386/mmx.md (mmx_pblendvb): New insn pattern.
            * config/i386/sse.md (unspec): Move UNSPEC_BLENDV ...
            * config/i386/i386.md (unspec): ... here.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Miss vec_cmpmn/vcondmn expander for 64bit vector

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #7 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:b1f7fd8a2a5558da1e101de11bb1cdba081ce010

commit r12-731-gb1f7fd8a2a5558da1e101de11bb1cdba081ce010
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Wed May 12 08:11:18 2021 +0200

    i386: Implement FP vector compares for V2SFmode [PR98218]

    Implement FP vector compares for V2SFmode for TARGET_MMX_WITH_SSE.

    2021-05-12  Uroš Bizjak  <ubizjak@gmail.com>

    gcc/
            PR target/98218
            * config/i386/i386-expand.c (ix86_expand_sse_movcc): Handle V2SF
mode.
            * config/i386/mmx.md (MMXMODE124): New mode iterator.
            (V2FI): Ditto.
            (mmxintvecmode): New mode attribute.
            (mmxintvecmodelower): Ditto.
            (*mmx_maskcmpv2sf3_comm): New insn pattern.
            (*mmx_maskcmpv2sf3): Ditto.
            (vec_cmpv2sfv2si): New expander.
            (vcond<V2FI:mode>v2si): Ditto.
            (mmx_vlendvps): New insn pattern.
            (vcond<MMXMODE124:mode><MMXMODEI:mode>): Also handle V2SFmode.
            (vcondu<MMXMODE124:mode><MMXMODEI:mode>): Ditto.
            (vcond_mask_<mode><mmxintvecmodelower>): Ditto.

    gcc/testsuite/

            PR target/98218
            * g++.target/i386/pr98218-1.C: Ditto.
            * gcc.target/i386/pr98218-4.c: New test.

            * gcc.target/i386/pr98218-1.c: Correct PR number.
            * gcc.target/i386/pr98218-1a.c: Ditto.
            * gcc.target/i386/pr98218-2.c: Ditto.
            * gcc.target/i386/pr98218-2a.c: Ditto.
            * gcc.target/i386/pr98218-3.c: Ditto.
            * gcc.target/i386/pr98218-3a.c: Ditto.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[TARGET_MMX_WITH_SSE] Miss  |[TARGET_MMX_WITH_SSE]
                   |vec_cmpmn/vcondmn expander  |Implement 64bit vector
                   |for 64bit vector            |compares (AVX512 masked
                   |                            |compares missing)
             Status|RESOLVED                    |NEW
         Resolution|FIXED                       |---

--- Comment #8 from Uroš Bizjak <ubizjak at gmail dot com> ---
AVX512 masked compares are currently not implemented and fall back to
vpblendvb/vpblendps.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|ubizjak at gmail dot com           |unassigned at gcc dot gnu.org

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
Un-assigning for AVX512.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #10 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Uroš Bizjak from comment #9)
> Un-assigning for AVX512.

I'll take a look.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

David Binderman <dcb314 at hotmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dcb314 at hotmail dot com

--- Comment #11 from David Binderman <dcb314 at hotmail dot com> ---
I might be seeing something similar:

caxcpy.f: In function 'caxcpy':
caxcpy.f:53:72: error: unrecognizable insn:
   53 |       end subroutine
      |                                                                       
^
(insn 136 135 137 16 (set (reg:V2SF 107 [ vect__96.11 ])
        (if_then_else:V2SF (reg:V2SF 220)
            (reg:V2SF 84 [ _3 ])
            (reg:V2SF 219))) -1
     (nil))
during RTL pass: vregs
caxcpy.f:53:72: internal compiler error: in extract_insn, at recog.c:2770

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #12 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to David Binderman from comment #11)
> I might be seeing something similar:
> 
> caxcpy.f: In function 'caxcpy':
> caxcpy.f:53:72: error: unrecognizable insn:
>    53 |       end subroutine
>       |                                                                     
> ^
> (insn 136 135 137 16 (set (reg:V2SF 107 [ vect__96.11 ])
>         (if_then_else:V2SF (reg:V2SF 220)
>             (reg:V2SF 84 [ _3 ])
>             (reg:V2SF 219))) -1
>      (nil))
> during RTL pass: vregs
> caxcpy.f:53:72: internal compiler error: in extract_insn, at recog.c:2770

Yeah, this is a non-existent SSE "cmove". I tried to find all paths where this
should divert to a sequence of logic instructions or PBLENDB, but due to
plethora of ISAs some cmove expansions fell through the cracks. Fortunately,
easy to fix problem, please just provide a testcase.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #13 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Uroš Bizjak from comment #12)
> Yeah, this is a non-existent SSE "cmove". I tried to find all paths where
> this should divert to a sequence of logic instructions or PBLENDB, but due
> to plethora of ISAs some cmove expansions fell through the cracks.
> Fortunately, easy to fix problem, please just provide a testcase.

A wild guess: you are compiling for TARGET_XOP?

diff --git a/gcc/config/i386/mmx.md b/gcc/config/i386/mmx.md
index d433c524652..7fc2e5d781c 100644
--- a/gcc/config/i386/mmx.md
+++ b/gcc/config/i386/mmx.md
@@ -1816,11 +1816,11 @@

 ;; XOP parallel XMM conditional moves
 (define_insn "*xop_pcmov_<mode>"
-  [(set (match_operand:MMXMODEI 0 "register_operand" "=x")
-        (if_then_else:MMXMODEI
-          (match_operand:MMXMODEI 3 "register_operand" "x")
-          (match_operand:MMXMODEI 1 "register_operand" "x")
-          (match_operand:MMXMODEI 2 "register_operand" "x")))]
+  [(set (match_operand:MMXMODE124 0 "register_operand" "=x")
+        (if_then_else:MMXMODE124
+          (match_operand:MMXMODE124 3 "register_operand" "x")
+          (match_operand:MMXMODE124 1 "register_operand" "x")
+          (match_operand:MMXMODE124 2 "register_operand" "x")))]
   "TARGET_XOP && TARGET_MMX_WITH_SSE"
   "vpcmov\t{%3, %2, %1, %0|%0, %1, %2, %3}"
   [(set_attr "type" "sse4arg")])

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #14 from David Binderman <dcb314 at hotmail dot com> ---
(In reply to Uroš Bizjak from comment #13)
> A wild guess: you are compiling for TARGET_XOP?

Not sure. bdver2.

Reduced C code:

typedef float DBL;
struct {
  DBL c_x, c_y
} Random_Simis_Cur;
Random_Simis_i;
DBL Gauss_Rand() {
  int __trans_tmp_1;
  DBL S, y = exp(S) / 0;
  if (__trans_tmp_1)
    return y;
  return -y;
}
draw_ifs() {
  while (Random_Simis_i--) {
    Random_Simis_Cur.c_x = Gauss_Rand();
    Random_Simis_Cur.c_y = Gauss_Rand();
  }
}

Command line is

$ /home/dcb/gcc/results/bin/gcc -c -w -O3 -march=bdver2 -ffast-math bug717.c

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #15 from David Binderman <dcb314 at hotmail dot com> ---
Bug first appears sometime between git hash 21dfb22920ce32fc,
dated yesterday and git hash 097fde5e7514e909, dated today.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

--- Comment #16 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to David Binderman from comment #15)
> Bug first appears sometime between git hash 21dfb22920ce32fc,
> dated yesterday and git hash 097fde5e7514e909, dated today.

Fixed by PR100581.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|12.0                        |12.2

--- Comment #17 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
GCC 12.1 is being released, retargeting bugs to GCC 12.2.

[Bug target/98218] [TARGET_MMX_WITH_SSE] Implement 64bit vector compares (AVX512 masked compares missing)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98218

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|12.3                        |12.4

--- Comment #19 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 12.3 is being released, retargeting bugs to GCC 12.4.