[Bug tree-optimization/98537] New: [11 Regression] ICE in emit_move_insn since r11-5839

[Bug tree-optimization/98537] New: [11 Regression] ICE in emit_move_insn since r11-5839

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

            Bug ID: 98537
           Summary: [11 Regression] ICE in emit_move_insn since r11-5839
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

Since r11-5839-g3a6e3ad38a17a03ee0139b49a0946e7b9ded1eb1 we ICE on the
following testcase on x86_64-linux with -O2:
typedef int V __attribute__((__vector_size__(32)));
typedef long W __attribute__((__vector_size__(32)));
W c, d;
struct B {};
B e;
struct C { W i; };
void foo (C);

C
operator== (B, B)
{
  W r = (V)c == (V)d;
  return {r};
}

__attribute__((target ("avx512vl"))) void
bar ()
{
  B a;
  foo (a == e);
}
entropy_coder.cc.ii: In function ‘void bar()’:
entropy_coder.cc.ii:13:12: internal compiler error: in emit_move_insn, at
expr.c:3820
   13 |   return {r};
      |            ^
0x11b183f emit_move_insn(rtx_def*, rtx_def*)
        ../../gcc/expr.c:3820
0x1e51e63 ix86_expand_int_vec_cmp(rtx_def**)
        ../../gcc/config/i386/i386-expand.c:4304
0x23f8be7 gen_vec_cmpv8siv8si(rtx_def*, rtx_def*, rtx_def*, rtx_def*)
        ../../gcc/config/i386/sse.md:3226
0x15acff0 rtx_insn* insn_gen_fn::operator()<rtx_def*, rtx_def*, rtx_def*,
rtx_def*>(rtx_def*, rtx_def*, rtx_def*, rtx_def*) const
        ../../gcc/recog.h:407
0x15aa82a maybe_gen_insn(insn_code, unsigned int, expand_operand*)
        ../../gcc/optabs.c:7787
0x15aab20 maybe_expand_insn(insn_code, unsigned int, expand_operand*)
        ../../gcc/optabs.c:7818
0x15aabb4 expand_insn(insn_code, unsigned int, expand_operand*)
        ../../gcc/optabs.c:7849
0x15a6952 expand_vec_cmp_expr(tree_node*, tree_node*, rtx_def*)
        ../../gcc/optabs.c:6285
0x11d746d do_store_flag
        ../../gcc/expr.c:12373

[Bug tree-optimization/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Target Milestone|---                         |11.0
   Last reconfirmed|                            |2021-01-05
           Priority|P3                          |P1
                 CC|                            |prathamesh3492 at gcc dot gnu.org
     Ever confirmed|0                           |1

[Bug tree-optimization/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #1 from Hongtao.liu <crazylht at gmail dot com> ---
Mine.

[Bug tree-optimization/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #1)
> Mine.

For vec_cmpmn, it goes into ix86_expand_fp_vec_cmp/ix86_expand_int_vec_cmp,
when cmp is integer mask, we need to use ix86_expand_sse_movcc instead of
emit_move_insn.

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
BTW(In reply to Jakub Jelinek from comment #0)
> Since r11-5839-g3a6e3ad38a17a03ee0139b49a0946e7b9ded1eb1 we ICE on the
BTW, for testcase

typedef int v4si __attribute__ ((vector_size (16)));
v4si
foo (v4si b, v4si a)
{
  v4si c = a == b;
  return c;
}

r11-5839 failed to transform vcond to vec_cmp in gimple.

test.c.244t.optimized

v4si foo (v4si b, v4si a)
{
  v4si c;

  <bb 2> [local count: 1073741824]:
  c_4 = .VCOND (a_2(D), b_3(D), { -1, -1, -1, -1 }, { 0, 0, 0, 0 }, 113);
  return c_4;
}

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[prathamesh3492 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #4 from prathamesh3492 at gcc dot gnu.org ---
Hi,
It seems to work on my machine for x86_64.
Compiling with -O3 (or -O2),
.optimized dump shows:

v4si foo (v4si b, v4si a)
{
  v4si c;
  vector(4) <signed-boolean:32> _1;

  <bb 2> [local count: 1073741824]:
  _1 = a_2(D) == b_3(D);
  c_4 = VIEW_CONVERT_EXPR<v4si>(_1);
  return c_4;

}

I tried on top of af362af18f405c34840d820143aa3a94f72fce4d.

Btw, on ARM it seems to "scalarize" the code,
.optimized dump shows:

  _6 = BIT_FIELD_REF <a_2(D), 32, 0>;
  _7 = BIT_FIELD_REF <b_3(D), 32, 0>;
  _8 = _6 == _7 ? -1 : 0;
  _9 = BIT_FIELD_REF <a_2(D), 32, 32>;
  _10 = BIT_FIELD_REF <b_3(D), 32, 32>;
  _11 = _9 == _10 ? -1 : 0;
  _12 = BIT_FIELD_REF <a_2(D), 32, 64>;
  _13 = BIT_FIELD_REF <b_3(D), 32, 64>;
  _14 = _12 == _13 ? -1 : 0;
  _15 = BIT_FIELD_REF <a_2(D), 32, 96>;
  _16 = BIT_FIELD_REF <b_3(D), 32, 96>;
  _17 = _15 == _16 ? -1 : 0;
  c_4 = {_8, _11, _14, _17};
  return c_4;

Thanks,
Prathamesh

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to prathamesh3492 from comment #4)
> Hi,
> It seems to work on my machine for x86_64.
> Compiling with -O3 (or -O2),
> .optimized dump shows:
> 
> v4si foo (v4si b, v4si a)
> {
>   v4si c;
>   vector(4) <signed-boolean:32> _1;
> 
>   <bb 2> [local count: 1073741824]:
>   _1 = a_2(D) == b_3(D);
>   c_4 = VIEW_CONVERT_EXPR<v4si>(_1);
>   return c_4;
> 
> }
> 
> I tried on top of af362af18f405c34840d820143aa3a94f72fce4d.
> 
> Btw, on ARM it seems to "scalarize" the code,
> .optimized dump shows:
> 
>   _6 = BIT_FIELD_REF <a_2(D), 32, 0>;
>   _7 = BIT_FIELD_REF <b_3(D), 32, 0>;
>   _8 = _6 == _7 ? -1 : 0;
>   _9 = BIT_FIELD_REF <a_2(D), 32, 32>;
>   _10 = BIT_FIELD_REF <b_3(D), 32, 32>;
>   _11 = _9 == _10 ? -1 : 0;
>   _12 = BIT_FIELD_REF <a_2(D), 32, 64>;
>   _13 = BIT_FIELD_REF <b_3(D), 32, 64>;
>   _14 = _12 == _13 ? -1 : 0;
>   _15 = BIT_FIELD_REF <a_2(D), 32, 96>;
>   _16 = BIT_FIELD_REF <b_3(D), 32, 96>;
>   _17 = _15 == _16 ? -1 : 0;
>   c_4 = {_8, _11, _14, _17};
>   return c_4;
> 
> Thanks,
> Prathamesh

try -march=skylake-avx512(In reply to prathamesh3492 from comment #4)
> Hi,
> It seems to work on my machine for x86_64.
> Compiling with -O3 (or -O2),
> .optimized dump shows:
> 
> v4si foo (v4si b, v4si a)
> {
>   v4si c;
>   vector(4) <signed-boolean:32> _1;
> 
>   <bb 2> [local count: 1073741824]:
>   _1 = a_2(D) == b_3(D);
>   c_4 = VIEW_CONVERT_EXPR<v4si>(_1);
>   return c_4;
> 
> }
> 
> I tried on top of af362af18f405c34840d820143aa3a94f72fce4d.
> 
> Btw, on ARM it seems to "scalarize" the code,
> .optimized dump shows:
> 
>   _6 = BIT_FIELD_REF <a_2(D), 32, 0>;
>   _7 = BIT_FIELD_REF <b_3(D), 32, 0>;
>   _8 = _6 == _7 ? -1 : 0;
>   _9 = BIT_FIELD_REF <a_2(D), 32, 32>;
>   _10 = BIT_FIELD_REF <b_3(D), 32, 32>;
>   _11 = _9 == _10 ? -1 : 0;
>   _12 = BIT_FIELD_REF <a_2(D), 32, 64>;
>   _13 = BIT_FIELD_REF <b_3(D), 32, 64>;
>   _14 = _12 == _13 ? -1 : 0;
>   _15 = BIT_FIELD_REF <a_2(D), 32, 96>;
>   _16 = BIT_FIELD_REF <b_3(D), 32, 96>;
>   _17 = _15 == _16 ? -1 : 0;
>   c_4 = {_8, _11, _14, _17};
>   return c_4;
> 
> Thanks,
> Prathamesh

It need avx512, try -march=skylake-avx512

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[prathamesh3492 at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #6 from prathamesh3492 at gcc dot gnu.org ---
Thanks for the suggestions, I could reproduce it now.

Input to isel is:
  _1 = a_2(D) == b_3(D);
  c_4 = VEC_COND_EXPR <_1, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }>;
  return c_4;

For the following check added in r11-5839:
          if (integer_minus_onep (op1)
              && integer_zerop (op2)
              && TYPE_MODE (TREE_TYPE (lhs)) == TYPE_MODE (TREE_TYPE (op0))
              && expand_vec_cmp_expr_p (op0a_type, op0_type, tcode))

With -march=skylake-avx512, it seems the TYPE_MODE (TREE_TYPE (lhs))
and TYPE_MODE (TREE_TYPE (op0)) do not agree, and we bail out.

That happens because, lhs (c_4) mode is V4SI while op0 (_1) mode is QI.
Without -march=skylake-avx512, the type mode is V4SI for both lhs and op0.

With -march=skylake-avx512, c_4's type is vector(4) <signed-boolean:1>
and without it, the type is vector(4) <signed-boolean:32>.

Thanks,
Prathamesh

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #7 from Hongtao.liu <crazylht at gmail dot com> ---
A patch is posted at
https://gcc.gnu.org/pipermail/gcc-patches/2021-January/562904.html

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #8 from Hongtao.liu <crazylht at gmail dot com> ---
Created attachment 50045
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50045&action=edit
Fix ICE: Don't generate integer mask comparision for 128/256-bits vector when
op_true/op_false are NULL or constm1_rtx/const0_rtx [PR98537]

I'm retesting this patch which is supposed to fix all performance and
correctness regressions caused by r10-5250.


in ix86_expand_sse_cmp/ix86_expand_int_sse_cmp

-  if (ix86_valid_mask_cmp_mode (cmp_ops_mode))
+  if (GET_MODE_SIZE (mode) == 64
+      || (ix86_valid_mask_cmp_mode (cmp_ops_mode)
+         /* When op_true and op_false is NULL, vector dest is required.  */
+         && op_true && op_false
+         /* Gimple sometimes transforms vec_cmpmn to vcondmn with
+            op_true/op_false as constm1_rtx/const0_rtx.
+            Don't generate integer mask comparison then.  */
+         && !((vector_all_ones_operand (op_true, GET_MODE (op_true))
+               && CONST0_RTX (GET_MODE (op_false)) == op_false)
+              || (vector_all_ones_operand (op_false, GET_MODE (op_false))
+                  && CONST0_RTX (GET_MODE (op_true)) == op_true))))

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- 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:8d0737d8f4b10bffe0411507ad2dc21ba7679883

commit r11-7116-g8d0737d8f4b10bffe0411507ad2dc21ba7679883
Author: liuhongt <hongtao.liu@intel.com>
Date:   Thu Jan 7 10:15:33 2021 +0800

    Fix ICE: Don't generate integer mask comparision for 128/256-bits vector
when op_true/op_false are NULL or constm1_rtx/const0_rtx [PR98537]

    avx512vl-pr92686-vpcmp-{1,2,intelasm-1}.c are used to guard code
    generation of integer mask comparison, but for vector comparison to
    vector dest, integer mask comparison is disliked, so delete these
    useless tests.

    gcc/ChangeLog:

            PR target/98537
            * config/i386/i386-expand.c (ix86_expand_sse_cmp): Don't
            generate integer mask comparison for 128/256-bits vector when
            op_true/op_false is NULL_RTX or CONSTM1_RTX/CONST0_RTX. Also
            delete redundant !maskcmp condition.
            (ix86_expand_int_vec_cmp): Ditto but no redundant deletion
            here.
            (ix86_expand_sse_movcc): Delete definition of maskcmp, add the
            condition directly to if (maskcmp), add extra check for
            cmpmode, it should be MODE_INT.
            (ix86_expand_fp_vec_cmp): Pass NULL to ix86_expand_sse_cmp's
            parameters op_true/op_false.
            (ix86_use_mask_cmp_p): New.

    gcc/testsuite/ChangeLog:

            PR target/98537
            * g++.target/i386/avx512bw-pr98537-1.C: New test.
            * g++.target/i386/avx512vl-pr98537-1.C: New test.
            * g++.target/i386/avx512vl-pr98537-2.C: New test.
            * gcc.target/i386/avx512vl-pr88547-1.c: Adjust testcase,
            integer mask comparison should not be generated.
            * gcc.target/i386/avx512vl-pr92686-vpcmp-1.c: Remove.
            * gcc.target/i386/avx512vl-pr92686-vpcmp-2.c: Ditto.
            * gcc.target/i386/avx512vl-pr92686-vpcmp-intelasm-1.c: Ditto.

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

--- Comment #10 from Hongtao.liu <crazylht at gmail dot com> ---
Fixed in GCC11, not sure backport

[Bug target/98537] [11 Regression] ICE in emit_move_insn since r11-5839

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98537

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

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

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
.