[Bug c/107057] New: [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[Bug c/107057] New: [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[gscfq@t-online.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

            Bug ID: 107057
           Summary: [10/11/12/13 Regression] ICE in
                    extract_constrain_insn, at recog.cc:2692
           Product: gcc
           Version: 13.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: gscfq@t-online.de
  Target Milestone: ---

Started between 20190630 and 20190728, at -O3 or -Ofast :


$ cat z1.c
typedef double v2df __attribute__ ((vector_size (16)));
v2df f (double a, double b)
{
  v2df v;
  double *c = (double *)&v;
  *c = a;
  *(c+1) = b;
  return v;
}
void g ()
{
  v2df x = f (1.0, 1.0);
  v2df y = f (2.0, 2.0);
  for (;*(double *)&x<=8; x+=y)
    g ();
}


$ gcc-13-20220925 -c z1.c -mavx -mcmodel=large -O2
$
$ gcc-13-20220925 -c z1.c -mavx -mcmodel=large -O3
z1.c: In function 'g':
z1.c:16:1: error: insn does not satisfy its constraints:
   16 | }
      | ^
(insn 76 407 101 5 (set (reg/v:V2DF 24 xmm4 [orig:108 x ] [108])
        (vec_concat:V2DF (mem/u/c:DF (reg:DI 0 ax [330]) [0  S8 A64])
            (mem/u/c:DF (reg:DI 4 si [331]) [0  S8 A64]))) "z1.c":7:10 5952
{vec_concatv2df}
     (expr_list:REG_EQUAL (const_vector:V2DF [
                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
            ])
        (nil)))
during RTL pass: reload
z1.c:16:1: internal compiler error: in extract_constrain_insn, at recog.cc:2692
0x69f5c0 _fatal_insn(char const*, rtx_def const*, char const*, int, char
const*)
        ../../gcc/rtl-error.cc:108
0x69f5e9 _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        ../../gcc/rtl-error.cc:118
0xe20add extract_constrain_insn(rtx_insn*)
        ../../gcc/recog.cc:2692
0xce0d77 check_rtl
        ../../gcc/lra.cc:2126
0xce6296 lra(_IO_FILE*)
        ../../gcc/lra.cc:2544
0xc93fc9 do_reload
        ../../gcc/ira.cc:5940
0xc93fc9 execute
        ../../gcc/ira.cc:6126

[Bug target/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |10.5

[Bug target/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2

[Bug target/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

Martin Liška <marxin at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
                 CC|                            |marxin at gcc dot gnu.org
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2022-09-29

--- Comment #1 from Martin Liška <marxin at gcc dot gnu.org> ---
Started with r10-1982-g4eb503965ac70e74 which only exposed it I guess.

[Bug target/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #2 from Uroš Bizjak <ubizjak at gmail dot com> ---
Reload starts with:

(insn 76 67 101 5 (set (reg/v:V2DF 108 [ x ])
        (vec_concat:V2DF (reg:DF 182)
            (reg:DF 182))) "pr107057.c":7:10 5952 {vec_concatv2df}
     (expr_list:REG_EQUAL (const_vector:V2DF [
                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
            ])
        (nil)))

And reloads the insn with:

(insn 392 67 393 5 (set (reg:DI 0 ax [330])
        (symbol_ref/u:DI ("*.LC0") [flags 0x2])) "pr107057.c":7:10 82
{*movdi_internal}
     (nil))
(insn 407 393 76 5 (set (reg:DI 4 si [331])
        (reg:DI 0 ax [330])) "pr107057.c":7:10 82 {*movdi_internal}
     (nil))
(insn 76 407 101 5 (set (reg/v:V2DF 24 xmm4 [orig:108 x ] [108])
        (vec_concat:V2DF (mem/u/c:DF (reg:DI 0 ax [330]) [0  S8 A64])
            (mem/u/c:DF (reg:DI 4 si [331]) [0  S8 A64]))) "pr107057.c":7:10
5952 {vec_concatv2df}
     (expr_list:REG_EQUAL (const_vector:V2DF [
                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
            ])
        (nil)))

Non-matching mem RTXes violate the candidate (x,m,1) constraint (and also insn
constraint):

(define_insn "vec_concatv2df"
  [(set (match_operand:V2DF 0 "register_operand"     "=x,x,v,x,v,x,x, v,x,x")
        (vec_concat:V2DF
          (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,m,m,0,x,vm,0,0")
          (match_operand:DF 2 "nonimm_or_0_operand"  " x,x,v,1,1,m,m,
C,x,m")))]
  "TARGET_SSE
   && (!(MEM_P (operands[1]) && MEM_P (operands[2]))
       || (TARGET_SSE3 && rtx_equal_p (operands[1], operands[2])))"

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |rtl-optimization
           Keywords|                            |ra

--- Comment #3 from Uroš Bizjak <ubizjak at gmail dot com> ---
RA issue.

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #4 from Hongtao.liu <crazylht at gmail dot com> ---
It lookes like after replace equiv value, the pattern turns into

3482(insn 76 67 101 5 (set (reg/v:V2DF 108 [ x ])
3483        (vec_concat:V2DF (const_double:DF 1.0e+0 [0x0.8p+1])
3484            (const_double:DF 1.0e+0 [0x0.8p+1]))) "/:7:10 5956
{vec_concatv2df}
3485     (expr_list:REG_EQUAL (const_vector:V2DF [
3486                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
3487            ])
3488        (nil)))


and in op1 and op2(the same as op1) is forced to mem separately with 


3734(insn 76 67 101 5 (set (reg/v:V2DF 108 [ x ])
3735        (vec_concat:V2DF (mem/u/c:DF (reg:DI 330) [0  S8 A64])
3736            (const_double:DF 1.0e+0 [0x0.8p+1]))) "":7:10 5956
{vec_concatv2df}
3737     (expr_list:REG_EQUAL (const_vector:V2DF [
3738                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
3739            ])
3740        (nil)))
3741(insn 76 67 101 5 (set (reg/v:V2DF 108 [ x ])
3742        (vec_concat:V2DF (mem/u/c:DF (reg:DI 330) [0  S8 A64])
3743            (mem/u/c:DF (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [0  S8
A64]))) "test.c":7:10 5956 {vec_concatv2df}
3744     (expr_list:REG_EQUAL (const_vector:V2DF [
3745                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
3746            ])
3747        (nil)))
3755(insn 76 67 101 5 (set (reg/v:V2DF 108 [ x ])
3756        (vec_concat:V2DF (mem/u/c:DF (reg:DI 330) [0  S8 A64])
3757            (mem/u/c:DF (reg:DI 331) [0  S8 A64]))) "test.c":7:10 5956
{vec_concatv2df}
3758     (expr_list:REG_EQUAL (const_vector:V2DF [
3759                (const_double:DF 1.0e+0 [0x0.8p+1]) repeated x2
3760            ])
3761        (nil)))


but gcc do set goal_alt_win[2] as false, but ignored it since below codes
thought it's already matched before.

4507      /* Operands that match previous ones have already been handled.  */
4508=>    if (goal_alt_matches[i] >= 0)
4509        continue;

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #5 from Hongtao.liu <crazylht at gmail dot com> ---
I'm testing

modified   gcc/lra-constraints.cc
@@ -4382,7 +4382,14 @@ curr_insn_transform (bool check_only_p)
                 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
                || no_input_reloads_p))
          {
-           rtx tem = force_const_mem (mode, op);
+           rtx tem;
+
+           /* PR 107057, for same constant pool, don't bother
+              force_const_mem again, just reuse the matched op.  */
+           if (goal_alt_matches[i] >= 0)
+             tem = *curr_id->operand_loc[goal_alt_matches[i]];
+           else
+             tem = force_const_mem (mode, op);

            change_p = true;
            if (subreg != NULL_RTX)

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #6 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Hongtao.liu from comment #5)
> I'm testing
> 
> modified   gcc/lra-constraints.cc
> @@ -4382,7 +4382,14 @@ curr_insn_transform (bool check_only_p)
>  		 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
>  		|| no_input_reloads_p))
>  	  {
> -	    rtx tem = force_const_mem (mode, op);
> +	    rtx tem;
> +
> +	    /* PR 107057, for same constant pool, don't bother
> +	       force_const_mem again, just reuse the matched op.  */
> +	    if (goal_alt_matches[i] >= 0)
> +	      tem = *curr_id->operand_loc[goal_alt_matches[i]];
> +	    else
> +	      tem = force_const_mem (mode, op);
>  
>  	    change_p = true;
>  	    if (subreg != NULL_RTX)

It failed bootstrap since there's are matches ops with differnce mode

(define_insn "*fop_xf_4_i387"
  [(set (match_operand:XF 0 "register_operand" "=f,f")
        (match_operator:XF 3 "binary_fp_operator"
           [(float_extend:XF
              (match_operand:MODEF 1 "nonimmediate_operand" "fm,0"))
            (match_operand:XF 2 "register_operand" "0,f")]))]

I don't know if there're other usages, so for safety, i'll just invalid
goal_alt_matches when force_const_mem.

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #7 from Hongtao.liu <crazylht at gmail dot com> ---
And it looks like the pattern is wrongly defined since from [1].

----------cut begin------------
Matching constraints are used in these circumstances. More precisely, the two
operands that match must include one input-only operand and one output-only
operand. Moreover, the digit must be a smaller number than the number of the
operand that uses it in the constraint.
----------------cut end----------

(define_insn "vec_concatv2df"
  [(set (match_operand:V2DF 0 "register_operand"     "=x,x,v,x,v,x,x, v,x,x")
        (vec_concat:V2DF
          (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,m,m,0,x,vm,0,0")
          (match_operand:DF 2 "nonimm_or_0_operand"  " x,x,v,1,1,m,m,
C,x,m")))]


For alternatvie 1, the two operands are both input only.

[1]
https://gcc.gnu.org/onlinedocs/gccint/Simple-Constraints.html#Simple-Constraints

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #8 from Hongtao.liu <crazylht at gmail dot com> ---
And it looks like the pattern is wrongly defined since from [1].

----------cut begin------------
Matching constraints are used in these circumstances. More precisely, the two
operands that match must include one input-only operand and one output-only
operand. Moreover, the digit must be a smaller number than the number of the
operand that uses it in the constraint.
----------------cut end----------

(define_insn "vec_concatv2df"
  [(set (match_operand:V2DF 0 "register_operand"     "=x,x,v,x,v,x,x, v,x,x")
        (vec_concat:V2DF
          (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,m,m,0,x,vm,0,0")
          (match_operand:DF 2 "nonimm_or_0_operand"  " x,x,v,1,1,m,m,
C,x,m")))]


For alternatvie 1, the two operands are both input only.

[1]
https://gcc.gnu.org/onlinedocs/gccint/Simple-Constraints.html#Simple-Constraints

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Hongtao.liu from comment #7)
> And it looks like the pattern is wrongly defined since from [1].
> 
> ----------cut begin------------
> Matching constraints are used in these circumstances. More precisely, the
> two operands that match must include one input-only operand and one
> output-only operand. Moreover, the digit must be a smaller number than the
> number of the operand that uses it in the constraint.
> ----------------cut end----------
> 
> (define_insn "vec_concatv2df"
>   [(set (match_operand:V2DF 0 "register_operand"     "=x,x,v,x,v,x,x, v,x,x")
> 	(vec_concat:V2DF
> 	  (match_operand:DF 1 "nonimmediate_operand" " 0,x,v,m,m,0,x,vm,0,0")
> 	  (match_operand:DF 2 "nonimm_or_0_operand"  " x,x,v,1,1,m,m, C,x,m")))]
> 
> 
> For alternatvie 1, the two operands are both input only.
> 
> [1]
> https://gcc.gnu.org/onlinedocs/gccint/Simple-Constraints.html#Simple-
> Constraints

Perhaps an error should be emitted for this situation?

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- 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:4acc4c2be84d66075d60736623c3a7134d129eaa

commit r13-3587-g4acc4c2be84d66075d60736623c3a7134d129eaa
Author: liuhongt <hongtao.liu@intel.com>
Date:   Tue Oct 25 17:10:23 2022 +0800

    Fix incorrect digit constraint

    Matching constraints are used in these circumstances. More precisely,
    the two operands that match must include one input-only operand and
    one output-only operand. Moreover, the digit must be a smaller number
    than the number of the operand that uses it in the constraint.

    In pr107057, the 2 operands in the pattern are both input operands.

    gcc/ChangeLog:

            PR target/107057
            * config/i386/sse.md (*vec_interleave_highv2df): Remove
            constraint 1.
            (*vec_interleave_lowv2df): Ditto.
            (vec_concatv2df): Ditto.
            (*avx512f_unpcklpd512<mask_name>): Ditto and renamed to ..
            (avx512f_unpcklpd512<mask_name>): .. this.
            (avx512f_movddup512<mask_name>): Change to define_insn.
            (avx_movddup256<mask_name>): Ditto.
            (*avx_unpcklpd256<mask_name>): Remove constraint 1 and renamed
            to ..
            (avx_unpcklpd256<mask_name>): .. this.
            * config/i386/i386.cc (ix86_vec_interleave_v2df_operator_ok):
            Disallow MEM_P (op1) && MEM_P (op2).

    gcc/testsuite/ChangeLog:

            * gcc.target/i386/pr107057.c: New test.

[Bug rtl-optimization/107057] [10/11/12/13 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

--- Comment #11 from Hongtao.liu <crazylht at gmail dot com> ---
Fixed in GCC13, and open a separate bug PR107487 for #c9.

[Bug rtl-optimization/107057] [11/12 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|10.5                        |11.5

--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
GCC 10 branch is being closed.

[Bug rtl-optimization/107057] [11/12 Regression] ICE in extract_constrain_insn, at recog.cc:2692

[liuhongt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107057

liuhongt at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
                 CC|                            |liuhongt at gcc dot gnu.org
         Resolution|---                         |FIXED

--- Comment #13 from liuhongt at gcc dot gnu.org ---
.