[Bug tree-optimization/109048] New: [13 regression] redundant mask compare generated by vectorizer.

[Bug tree-optimization/109048] New: [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

            Bug ID: 109048
           Summary: [13 regression] redundant mask compare generated by
                    vectorizer.
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: crazylht at gmail dot com
  Target Milestone: ---

#include <math.h>

void tmp2 (float *af, int type, int type2, float *res)
{
    const int Etype = (type == 1 || type2 == 2);
    const float f1 = (type == 3 || type2 == 4) ? 4.f : 2.f;
    const float f2 = (type == 3 || type2 == 4) ? 0.25f : 0.5f;

    for (int i = 0; i < 256; i++)
    {
        float x = af[i];
        int z = (x < 0.f);
        float t1 = (z ? 1.f : f2) + (x < f1 ? 1.f : 0.f);
        float neg_t1 = -fabsf(t1);
        float t2 = Etype ? neg_t1 : t1;
        res[i] += t2 + x;
    }
}

gcc trunk now generates


<bb 58> [local count: 5368707]:
  vect_cst__110 = {iftmp.0_34, iftmp.0_34, iftmp.0_34, iftmp.0_34, iftmp.0_34,
iftmp.0_34, iftmp.0_34, iftmp.0_34};
  vect_cst__119 = {prephitmp_41, prephitmp_41, prephitmp_41, prephitmp_41,
prephitmp_41, prephitmp_41, prephitmp_41, prephitmp_41};
  vect_cst__123 = {iftmp.1_16, iftmp.1_16, iftmp.1_16, iftmp.1_16, iftmp.1_16,
iftmp.1_16, iftmp.1_16, iftmp.1_16};

  <bb 17> [local count: 53687070]:
  # i_18 = PHI <i_47(26), 0(58)>
  # ivtmp_15 = PHI <ivtmp_43(26), 256(58)>
  # vectp_af.11_105 = PHI <vectp_af.11_106(26), af_24(D)(58)>
  # vectp_res.23_125 = PHI <vectp_res.23_126(26), res_28(D)(58)>
  # vectp_res.28_130 = PHI <vectp_res.28_131(26), res_28(D)(58)>
  # ivtmp_133 = PHI <ivtmp_134(26), 0(58)>
  # DEBUG i => NULL
  # DEBUG BEGIN_STMT
  _38 = (long unsigned int) i_18;
  _37 = _38 * 4;
  _36 = af_24(D) + _37;
  vect_x_20.13_107 = MEM <vector(8) float> [(float *)vectp_af.11_105];
  x_20 = *_36;
  # DEBUG x => NULL
  # DEBUG BEGIN_STMT
  # DEBUG D#1 => NULL
  # DEBUG z => NULL
  # DEBUG BEGIN_STMT
  mask__50.14_109 = vect_x_20.13_107 >= { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0 };
  _50 = x_20 >= 0.0;
  mask__52.15_111 = vect_x_20.13_107 < vect_cst__110;
  _52 = x_20 < iftmp.0_34;
  mask__53.16_112 = mask__50.14_109 & mask__52.15_111;
  _53 = _50 & _52;
  mask__55.17_114 = vect_x_20.13_107 >= vect_cst__110;
  _55 = x_20 >= iftmp.0_34;
  mask__56.18_115 = mask__50.14_109 & mask__55.17_114;
  _56 = _50 & _55;
  mask__74.19_117 = vect_x_20.13_107 < { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
};
  _74 = x_20 < 0.0;
  vect__ifc__73.20_120 = VEC_COND_EXPR <mask__74.19_117, { 2.0e+0, 2.0e+0,
2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 }, vect_cst__119>;
  _ifc__73 = _74 ? 2.0e+0 : prephitmp_41;
  _174 = ~mask__53.16_112;
  _175 = mask__74.19_117 & _174;
  vect__ifc__75.21_122 = VEC_COND_EXPR <_175, { 2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0,
2.0e+0, 2.0e+0, 2.0e+0, 2.0e+0 }, vect_cst__119>;
  _ifc__75 = _53 ? prephitmp_41 : _ifc__73;
  vect_prephitmp_17.22_124 = VEC_COND_EXPR <mask__56.18_115, vect_cst__123,
vect__ifc__75.21_122>;
  prephitmp_17 = _56 ? iftmp.1_16 : _ifc__75;
  # DEBUG t1 => D#2
  # DEBUG BEGIN_STMT
  # DEBUG neg_t1 => -D#2
  # DEBUG BEGIN_STMT
  # DEBUG t2 => prephitmp_17
  # DEBUG BEGIN_STMT
  _12 = res_28(D) + _37;
  vect__26.25_127 = MEM <vector(8) float> [(float *)vectp_res.23_125];
  _26 = *_12;
  vect__27.26_128 = vect__26.25_127 + vect_x_20.13_107;
  _27 = _26 + x_20;
  vect__45.27_129 = vect_prephitmp_17.22_124 + vect__27.26_128;
  _45 = prephitmp_17 + _27;
  MEM <vector(8) float> [(float *)vectp_res.28_130] = vect__45.27_129;
  # DEBUG BEGIN_STMT
  i_47 = i_18 + 1;
  # DEBUG i => i_47
  # DEBUG BEGIN_STMT
  ivtmp_43 = ivtmp_15 - 1;
  vectp_af.11_106 = vectp_af.11_105 + 32;
  vectp_res.23_126 = vectp_res.23_125 + 32;
  vectp_res.28_131 = vectp_res.28_130 + 32;
  ivtmp_134 = ivtmp_133 + 1;
  if (ivtmp_134 < 32)
    goto <bb 26>; [90.00%]
  else
    goto <bb 56>; [10.00%]

vs gcc12.2 


<bb 57> [local count: 5368707]:
  vect_cst__128 = {iftmp.1_16, iftmp.1_16, iftmp.1_16, iftmp.1_16, iftmp.1_16,
iftmp.1_16, iftmp.1_16, iftmp.1_16};
  vect_cst__134 = {iftmp.0_33, iftmp.0_33, iftmp.0_33, iftmp.0_33, iftmp.0_33,
iftmp.0_33, iftmp.0_33, iftmp.0_33};

  <bb 5> [local count: 53687070]:
  # i_15 = PHI <i_30(13), 0(57)>
  # ivtmp_20 = PHI <ivtmp_49(13), 256(57)>
  # vectp_af.24_124 = PHI <vectp_af.24_125(13), af_24(D)(57)>
  # vectp_res.31_138 = PHI <vectp_res.31_139(13), res_28(D)(57)>
  # vectp_res.36_143 = PHI <vectp_res.36_144(13), res_28(D)(57)>
  # ivtmp_146 = PHI <ivtmp_147(13), 0(57)>
  # DEBUG i => NULL
  # DEBUG BEGIN_STMT
  _7 = (long unsigned int) i_15;
  _8 = _7 * 4;
  _9 = af_24(D) + _8;
  vect_x_25.26_126 = MEM <vector(8) float> [(float *)vectp_af.24_124];
  x_25 = *_9;
  # DEBUG x => NULL
  # DEBUG BEGIN_STMT
  # DEBUG D#1 => NULL
  # DEBUG z => NULL
  # DEBUG BEGIN_STMT
  _130 = vect_x_25.26_126 >= { 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 };
  vect_iftmp.27_131 = VEC_COND_EXPR <_130, vect_cst__128, { 1.0e+0, 1.0e+0,
1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0 }>;
  iftmp.2_17 = x_25 >= 0.0 ? iftmp.1_16 : 1.0e+0;
  vect__41.28_133 = vect_iftmp.27_131 + { 1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0,
1.0e+0, 1.0e+0, 1.0e+0, 1.0e+0 };
  _41 = iftmp.2_17 + 1.0e+0;
  _135 = vect_x_25.26_126 >= vect_cst__134;
  vect_prephitmp_42.29_136 = VEC_COND_EXPR <_135, vect_iftmp.27_131,
vect__41.28_133>;
  prephitmp_42 = x_25 >= iftmp.0_33 ? iftmp.2_17 : _41;
  # DEBUG t1 => NULL
  # DEBUG BEGIN_STMT
  # DEBUG neg_t1 => -prephitmp_42
  # DEBUG BEGIN_STMT
  vect_neg_t1_27.30_137 = -vect_prephitmp_42.29_136;
  neg_t1_27 = -prephitmp_42;
  # DEBUG t2 => neg_t1_27
  # DEBUG BEGIN_STMT
  _10 = res_28(D) + _8;
  vect__11.33_140 = MEM <vector(8) float> [(float *)vectp_res.31_138];
  _11 = *_10;
  vect__35.34_141 = vect__11.33_140 + vect_x_25.26_126;
  _35 = _11 + x_25;
  vect__13.35_142 = vect_neg_t1_27.30_137 + vect__35.34_141;
  _13 = neg_t1_27 + _35;
  MEM <vector(8) float> [(float *)vectp_res.36_143] = vect__13.35_142;
  # DEBUG BEGIN_STMT
  i_30 = i_15 + 1;
  # DEBUG i => i_30
  # DEBUG BEGIN_STMT
  ivtmp_49 = ivtmp_20 - 1;
  vectp_af.24_125 = vectp_af.24_124 + 32;
  vectp_res.31_139 = vectp_res.31_138 + 32;
  vectp_res.36_144 = vectp_res.36_143 + 32;
  ivtmp_147 = ivtmp_146 + 1;
  if (ivtmp_147 < 32)

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
There's also a backend misoptimization for x86 avx512 mask compare which is
same as PR88570.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
Lots of logic operation for mask in ifcvt

  <bb 17> [local count: 531502205]:
  # i_18 = PHI <i_47(26), 0(42)>
  # ivtmp_15 = PHI <ivtmp_43(26), 256(42)>
  # DEBUG i => NULL
  # DEBUG BEGIN_STMT
  _38 = (long unsigned int) i_18;
  _37 = _38 * 4;
  _36 = af_24(D) + _37;
  x_20 = *_36;
  # DEBUG x => NULL
  # DEBUG BEGIN_STMT
  # DEBUG D#1 => NULL
  # DEBUG z => NULL
  # DEBUG BEGIN_STMT
  _50 = x_20 >= 0.0;
  _52 = x_20 < iftmp.0_34;
  _53 = _50 & _52;
  _55 = x_20 >= iftmp.0_34;
  _56 = _50 & _55;
  _74 = x_20 < 0.0;
  _ifc__73 = _74 ? 2.0e+0 : prephitmp_41;
  _ifc__75 = _53 ? prephitmp_41 : _ifc__73;
  prephitmp_17 = _56 ? iftmp.1_16 : _ifc__75;
  # DEBUG t1 => D#2
  # DEBUG BEGIN_STMT
  # DEBUG neg_t1 => -D#2
  # DEBUG BEGIN_STMT
  # DEBUG t2 => prephitmp_17
  # DEBUG BEGIN_STMT
  _12 = res_28(D) + _37;
  _26 = *_12;
  _27 = _26 + x_20;
  _45 = prephitmp_17 + _27;
  *_12 = _45;
  # DEBUG BEGIN_STMT
  i_47 = i_18 + 1;
  # DEBUG i => i_47
  # DEBUG BEGIN_STMT
  ivtmp_43 = ivtmp_15 - 1;

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
A three-operand phi make ifcvt generate worse code

 <bb 4> [local count: 10737414]:
  # iftmp.1_16 = PHI <2.5e-1(13), 5.0e-1(2)>
  # iftmp.0_34 = PHI <4.0e+0(13), 2.0e+0(2)>
  # prephitmp_40 = PHI <-2.5e-1(13), -5.0e-1(2)>
  # prephitmp_41 = PHI <1.25e+0(13), 1.5e+0(2)>
  # prephitmp_42 = PHI <-1.25e+0(13), -1.5e+0(2)>
  # DEBUG f1 => NULL
  # DEBUG f2 => iftmp.1_16
  # DEBUG BEGIN_STMT
  # DEBUG i => 0
  # DEBUG BEGIN_STMT
  if (_3 != 0)
    goto <bb 28>; [50.00%]
  else
    goto <bb 29>; [50.00%]

  <bb 29> [local count: 5368707]:

  <bb 17> [local count: 531502204]:
  # i_18 = PHI <0(29), i_47(26)>
  # ivtmp_15 = PHI <256(29), ivtmp_43(26)>
  # DEBUG i => i_18
  # DEBUG BEGIN_STMT
  _38 = (long unsigned int) i_18;
  _37 = _38 * 4;
  _36 = af_24(D) + _37;
  x_20 = *_36;
  # DEBUG x => x_20
  # DEBUG BEGIN_STMT
  # DEBUG D#1 => x_20 < 0.0
  # DEBUG z => (int) D#1
  # DEBUG BEGIN_STMT
  if (x_20 >= 0.0)
    goto <bb 19>; [59.00%]
  else
    goto <bb 22>; [41.00%]

  <bb 19> [local count: 313586303]:
  if (x_20 < iftmp.0_34)
    goto <bb 22>; [20.00%]
  else
    goto <bb 21>; [80.00%]

  <bb 21> [local count: 250869041]:

  <bb 22> [local count: 531502205]:
  # prephitmp_17 = PHI <2.0e+0(17), prephitmp_41(19), iftmp.1_16(21)>
  # DEBUG t1 => D#2
  # DEBUG BEGIN_STMT
  # DEBUG neg_t1 => -D#2
  # DEBUG BEGIN_STMT
  # DEBUG t2 => prephitmp_17
  # DEBUG BEGIN_STMT
  _12 = res_28(D) + _37;
  _26 = *_12;
  _27 = _26 + x_20;
  _45 = prephitmp_17 + _27;
  *_12 = _45;
  # DEBUG BEGIN_STMT
  i_47 = i_18 + 1;
  # DEBUG i => i_47
  # DEBUG BEGIN_STMT
  ivtmp_43 = ivtmp_15 - 1;
  if (ivtmp_43 != 0)

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #3)
> A three-operand phi make ifcvt generate worse code
> 
It saved 1 add for potential 1.f + 0.f/1.f since it's constant, but generates
more logic operations for predicates.

Hmm, it looks good for scalar since it's just compare and jump, but hurt
vectorization due to more logical operations for predicates.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization,
                   |                            |needs-bisection
   Target Milestone|---                         |13.0

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
           Keywords|needs-bisection             |
   Last reconfirmed|                            |2023-03-07
                 CC|                            |aldyh at gcc dot gnu.org,
                   |                            |jakub at gcc dot gnu.org

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
In bisection r13-1 through r13-6513 I see only one important change,
r13-2267-g4fbe3e6aa74dae5c75a

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 7 Mar 2023, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048
> 
> Jakub Jelinek <jakub at gcc dot gnu.org> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>              Status|UNCONFIRMED                 |NEW
>      Ever confirmed|0                           |1
>            Keywords|needs-bisection             |
>    Last reconfirmed|                            |2023-03-07
>                  CC|                            |aldyh at gcc dot gnu.org,
>                    |                            |jakub at gcc dot gnu.org
> 
> --- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> In bisection r13-1 through r13-6513 I see only one important change,
> r13-2267-g4fbe3e6aa74dae5c75a

That would probably trigger extra jump threading here.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
The first threadfull pass ends up producing the three-argument PHI:

<bb 7> [local count: 1063004408]:
_7 = (long unsigned int) i_14;
_8 = _7 * 4;
_9 = af_24(D) + _8;
x_25 = *_9;
if (x_25 >= 0.0)
  goto <bb 9>; [59.00%]
else
  goto <bb 8>; [41.00%]

<bb 8> [local count: 435831803]:

<bb 9> [local count: 1063004408]:
# iftmp.2_17 = PHI <iftmp.1_16(7), 1.0e+0(8)>
if (iftmp.0_15 > x_25)
  goto <bb 11>; [50.00%]
else
  goto <bb 10>; [50.00%]

and we thread 9->11 from 8->9 (so when !(x_25 >= 0.0)).  Like a related PR
this is a threading that turns a "conditional move" diamond into something more
complicated.  We could heuristically decide that we don't want to do this
[early?].  But it might be more honest to more aggressively represent
those as COND_EXPRs then.

So I think we have a duplicate for this.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
Created attachment 54762
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=54762&action=edit
heuristic

I am testing this heuristic

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #8)
> Created attachment 54762 [details]
> heuristic
> 
> I am testing this heuristic

FAIL: gcc.dg/tree-ssa/ranger-threader-3.c scan-tree-dump-times ethread
"Register
ing.*jump thread" 1
FAIL: gcc.dg/tree-ssa/ssa-dom-thread-14.c scan-tree-dump-times dom2 "Threaded"
2

is the fallout - both are the same testcase, both can be mitigated by adjusting
the dump to scan for and making sure PROP_loop_opts_done is also provided
for functions without loops.

But it also shows a case where we optimize things significantly later
after the change.

  if ((!unsignedp || (!left && methods == OPTAB_WIDEN)))
    {
      enum optab_methods methods1 = methods;
      if (unsignedp)
        methods1 = OPTAB_MUST_WIDEN;
      expand_binop (left ? lshift_optab : rshift_arith_optab,
                           unsignedp, methods1);

used to be threaded in early threading but now would be delayed to after
loop opts.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=109154

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
PR109154 is another of these cases.  We might want to de-dup them.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
The recent patch improved this to avoid some of the compares.  We still have
the three-argument PHI and thus three VEC_CONDs.

.L10:
        vmovups (%rdi,%rdx), %ymm0
        vcmpltps        %ymm6, %ymm0, %ymm3
        vcmpltps        %ymm2, %ymm0, %ymm1
        vpandn  %ymm1, %ymm3, %ymm1
        vblendvps       %ymm1, %ymm5, %ymm4, %ymm1
        vblendvps       %ymm3, %ymm7, %ymm1, %ymm1
        vaddps  %ymm1, %ymm0, %ymm0
        vaddps  (%rax,%rdx), %ymm0, %ymm0
        vmovups %ymm0, (%rax,%rdx)
        addq    $32, %rdx
        cmpq    $1024, %rdx
        jne     .L10

vs. GCC 12

.L6:
        vmovups (%rdi,%rdx), %ymm1
        vcmpltps        %ymm5, %ymm1, %ymm0
        vcmpltps        %ymm6, %ymm1, %ymm4
        vblendvps       %ymm0, %ymm3, %ymm2, %ymm0
        vandps  %ymm3, %ymm4, %ymm4
        vaddps  %ymm4, %ymm0, %ymm0
        vaddps  %ymm1, %ymm0, %ymm0
        vaddps  (%rax,%rdx), %ymm0, %ymm0
        vmovups %ymm0, (%rax,%rdx)
        addq    $32, %rdx
        cmpq    $1024, %rdx
        jne     .L6

which at least overall looks comparable.

[Bug tree-optimization/109048] [13 regression] redundant mask compare generated by vectorizer.

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109048

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

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

--- Comment #12 from Hongtao.liu <crazylht at gmail dot com> ---
So marked as fixed.