[Bug target/101058] New: [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[Bug target/101058] New: [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

            Bug ID: 101058
           Summary: [12 Regression] ICE in extract_insn, at recog.c:2770
                    since r12-1215-g8d7dae0eb366a88a
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: marxin at gcc dot gnu.org
                CC: ubizjak at gmail dot com
  Target Milestone: ---
              Host: x86_64-linux-gnu

The following fails:

$ cat vect.i
short add90Hybrid_a_1;
short *add90Hybrid_b, *add90Hybrid_c, *add90Hybrid_d;
void add90Hybrid() {
  for (int i; i < 200; i += 2) {
    add90Hybrid_c[i] = add90Hybrid_b[i];
    add90Hybrid_d[i] = add90Hybrid_a_1 - add90Hybrid_b[i + 1];
  }
}

$ gcc vect.i -c -m32 -O3 -mno-mmx -std=c99
vect.i: In function ‘add90Hybrid’:
vect.i:8:1: error: unrecognizable insn:
    8 | }
      | ^
(insn 474 473 213 13 (set (reg:V4HI 21 xmm1 [orig:321 vect__144.35 ] [321])
        (vec_select:V4HI (reg:V4HI 21 xmm1 [orig:321 vect__144.35 ] [321])
            (parallel [
                    (const_int 2 [0x2])
                    (const_int 3 [0x3])
                    (const_int 0 [0]) repeated x2
                ]))) "vect.i":5:37 -1
     (nil))
during RTL pass: split2
vect.i:8:1: internal compiler error: in extract_insn, at recog.c:2770
0x70345e _fatal_insn(char const*, rtx_def const*, char const*, int, char
const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:108
0x703480 _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:116
0x7015eb extract_insn(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/recog.c:2770
0xdf0094 extract_insn_cached(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/recog.c:2659
0xb1b1dc cleanup_subreg_operands(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/final.c:3099
0xdedeb7 split_insn
        /home/marxin/Programming/gcc/gcc/recog.c:3399
0xdf3387 split_all_insns()
        /home/marxin/Programming/gcc/gcc/recog.c:3467
0xdf342c execute
        /home/marxin/Programming/gcc/gcc/recog.c:4385
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

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

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

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
You forgot to mention -msse2 which is also required to reproduce it.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #2 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #1)
> You forgot to mention -msse2 which is also required to reproduce it.

No, I don't need it for some reason:

$ gcc --version
gcc (GCC) 12.0.0 20210614 (experimental)
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ gcc vect.i -c -m32 -O3 -mno-mmx --verbose
Using built-in specs.
COLLECT_GCC=gcc
Target: x86_64-pc-linux-gnu
Configured with: /home/marxin/Programming/gcc/configure
--enable-languages=c,c++,fortran,lto,jit --prefix=/home/marxin/bin/gcc
--disable-multilib --enable-host-shared --disable-libsanitizer
--enable-valgrind-annotations --disable-bootstrap : (reconfigured)
/home/marxin/Programming/gcc/configure --enable-languages=c,c++,fortran,lto,jit
--prefix=/home/marxin/bin/gcc --disable-multilib --enable-host-shared
--disable-libsanitizer --enable-valgrind-annotations --disable-bootstrap :
(reconfigured)  : (reconfigured) /home/marxin/Programming/gcc/configure
--enable-languages=c,c++,fortran,lto,jit --prefix=/home/marxin/bin/gcc
--disable-multilib --enable-host-shared --disable-libsanitizer
--enable-valgrind-annotations --disable-bootstrap
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 12.0.0 20210614 (experimental) (GCC) 
COLLECT_GCC_OPTIONS='-c' '-m32' '-O3' '-mno-mmx' '-v' '-mtune=generic'
'-march=x86-64'
 /home/marxin/bin/gcc/lib/gcc/x86_64-pc-linux-gnu/12.0.0/cc1 -fpreprocessed
vect.i -quiet -dumpbase vect.i -dumpbase-ext .i -m32 -mno-mmx -mtune=generic
-march=x86-64 -O3 -version -o /tmp/ccC2tr5L.s
GNU C17 (GCC) version 12.0.0 20210614 (experimental) (x86_64-pc-linux-gnu)
        compiled by GNU C version 11.1.1 20210510 [revision
23855a176609fe8dda6abaf2b21846b4517966eb], GMP version 6.2.1, MPFR version
4.1.0-p7, MPC version 1.2.1, isl version isl-0.24-GMP

GGC heuristics: --param ggc-min-expand=30 --param ggc-min-heapsize=4096
GNU C17 (GCC) version 12.0.0 20210614 (experimental) (x86_64-pc-linux-gnu)
        compiled by GNU C version 11.1.1 20210510 [revision
23855a176609fe8dda6abaf2b21846b4517966eb], GMP version 6.2.1, MPFR version
4.1.0-p7, MPC version 1.2.1, isl version isl-0.24-GMP

GGC heuristics: --param ggc-min-expand=30 --param ggc-min-heapsize=4096
Compiler executable checksum: 0f199bfede786f37f70583813ded880d
vect.i: In function ‘add90Hybrid’:
vect.i:8:1: error: unrecognizable insn:
    8 | }
      | ^
(insn 474 473 213 13 (set (reg:V4HI 21 xmm1 [orig:321 vect__144.35 ] [321])
        (vec_select:V4HI (reg:V4HI 21 xmm1 [orig:321 vect__144.35 ] [321])
            (parallel [
                    (const_int 2 [0x2])
                    (const_int 3 [0x3])
                    (const_int 0 [0]) repeated x2
                ]))) "vect.i":5:37 -1
     (nil))
during RTL pass: split2
vect.i:8:1: internal compiler error: in extract_insn, at recog.c:2770
0x70345e _fatal_insn(char const*, rtx_def const*, char const*, int, char
const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:108
0x703480 _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        /home/marxin/Programming/gcc/gcc/rtl-error.c:116
0x7015eb extract_insn(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/recog.c:2770
0xdf0094 extract_insn_cached(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/recog.c:2659
0xb1b1dc cleanup_subreg_operands(rtx_insn*)
        /home/marxin/Programming/gcc/gcc/final.c:3099
0xdedeb7 split_insn
        /home/marxin/Programming/gcc/gcc/recog.c:3399
0xdf3387 split_all_insns()
        /home/marxin/Programming/gcc/gcc/recog.c:3467
0xdf342c execute
        /home/marxin/Programming/gcc/gcc/recog.c:4385
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
You have that -march=x86-64 in there which implies -msse2...

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

Uroš Bizjak <ubizjak at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-06-14
           Assignee|unassigned at gcc dot gnu.org      |ubizjak at gmail dot com
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
Created attachment 51007
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51007&action=edit
Proposed patch

We should not split SSE2 instruction to one that requires TARGET_MMX_WITH_SSE.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I have tried:
--- gcc/config/i386/mmx.md.jj   2021-06-11 11:32:35.813981393 +0200
+++ gcc/config/i386/mmx.md      2021-06-14 13:37:01.483325198 +0200
@@ -3370,7 +3370,9 @@
            (match_operand:V2HI 2 "register_operand" "x,Yw"))
           (parallel [(match_operand 3 "const_0_to_3_operand")
                      (match_operand 4 "const_0_to_3_operand")])))]
-  "TARGET_SSE2"
+  "TARGET_SSE2
+   && ((operands[3] == const0_rtx && INTVAL (operands[4]) == 2)
+       || TARGET_MMX || TARGET_MMX_WITH_SSE)"
   "#"
   "&& reload_completed"
   [(set (match_dup 5)

but it then ICEs on expand_vec_perm_even_odd_1's:
    case E_V2DFmode:
    case E_V4SFmode:
    case E_V2DImode:
    case E_V2SImode:
    case E_V4SImode:
    case E_V2HImode:
      /* These are always directly implementable by expand_vec_perm_1.  */
      gcc_unreachable ();

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #5)

We can split directly to sse2_pshuflw_1, avoiding mmx_pshufw_1. These two
actually generate the same instruction (PSHUFLW) when XMM registers are
involved.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Sure, your patch looks much better.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Though, when this *punpckwd define_insn_and_split handles all possible constant
permutations for V2HImode, shouldn't ix86_vectorize_vec_perm_const say so:
     case E_V2HImode:
       if (!TARGET_SSE2)
         return false;
+      /* All implementable with *punpckwd.  */
+      if (d.testing_p)
+        return true;
       break;
?

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Jakub Jelinek from comment #8)
> Though, when this *punpckwd define_insn_and_split handles all possible
> constant permutations for V2HImode, shouldn't ix86_vectorize_vec_perm_const
> say so:
>      case E_V2HImode:
>        if (!TARGET_SSE2)
>          return false;
> +      /* All implementable with *punpckwd.  */
> +      if (d.testing_p)
> +        return true;
>        break;
> ?

Yes, we can add this optimization. V2HImode is actually mentioned a couple of
lines below, where it is claimed that it is fully implementable for one
operand.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
And maybe for V2SImode too, though haven't verified that.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

--- Comment #11 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:4986946f3b761dd4c3e0d79ca735c90e33f4bb83

commit r12-1436-g4986946f3b761dd4c3e0d79ca735c90e33f4bb83
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Mon Jun 14 20:56:18 2021 +0200

    i386: Split V2HImode *punpckwd to SSE instruction [PR101058]

    V2HImode *punpckwd should not be split to the insn that depends on
    TARGET_MMX_WITH_SSE, since the later is disabled on 32bit targets.

    Also return true early from ix86_vectorize_vec_perm_const when testing
    with V2HI mode.  *punpckwd can be used to implement all permutations.

    2021-06-14  Uroš Bizjak  <ubizjak@gmail.com>

    gcc/
            PR target/101058
            * config/i386/i386-expand.c (ix86_vectorize_vec_perm_const):
            Return true early when testing with V2HImode.
            * config/i386/mmx.md (*punpckwd): Split to sse2_pshuflw_1.

    gcc/testsuite/

            PR target/101058
            * gcc.target/i386/pr101058.c: New test.

[Bug target/101058] [12 Regression] ICE in extract_insn, at recog.c:2770 since r12-1215-g8d7dae0eb366a88a

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101058

Uroš Bizjak <ubizjak at gmail dot com> changed:

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

--- Comment #12 from Uroš Bizjak <ubizjak at gmail dot com> ---
Fixed.