[Bug target/100445] New: ice during RTL pass: vregs

[Bug target/100445] New: ice during RTL pass: vregs

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

            Bug ID: 100445
           Summary: ice during RTL pass: vregs
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dcb314 at hotmail dot com
  Target Milestone: ---

For this C code:

int kgd_get_dense_grid_point_double_mesh_mesh_i;
int kgd_get_dense_grid_point_double_mesh_mesh_address[3];
kgd_get_dense_grid_point_double_mesh_mesh() {
  for (; kgd_get_dense_grid_point_double_mesh_mesh_i;
       kgd_get_dense_grid_point_double_mesh_mesh_i++) {
    kgd_get_dense_grid_point_double_mesh_mesh_address
        [kgd_get_dense_grid_point_double_mesh_mesh_i] =
            (kgd_get_dense_grid_point_double_mesh_mesh_i - 1) / 2;
  }
 
get_grid_point_single_mesh(kgd_get_dense_grid_point_double_mesh_mesh_address);
}

compiled like this:

$ /home/dcb/gcc/results/bin/gcc -c -w -O3 -march=bdver2 bug716.c 2>&1 | more
bug716.c: In function ‘kgd_get_dense_grid_point_double_mesh_mesh’:
bug716.c:11:1: error: unrecognizable insn:
   11 | }
      | ^
(insn 22 21 23 5 (set (reg:V2SI 95 [ vect_patt_5.16 ])
        (if_then_else:V2SI (reg:V2SI 105)
            (reg:V2SI 93 [ _40 ])
            (reg:V2SI 94 [ vect__1.14 ]))) -1
     (nil))
during RTL pass: vregs
bug716.c:11:1: internal compiler error: in extract_insn, at recog.c:2770

The bug seems to occur sometime between git hash 61d48b1e2b5bae1e
and e1fcf14f33e4f371. 

My best guess, from a bit bisect, is that Uros's commit with hash
f3661f2d63fbc5fd30c24d22137691e16b0a0a17 is the culprit.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |12.0
                 CC|                            |jakub at gcc dot gnu.org
            Summary|ice during RTL pass: vregs  |[12 Regression] ice during
                   |                            |RTL pass: vregs
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-05-06
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I can confirm it started with r12-512-gd0d6ca019717305df0ef41e3fe1da48f7f561fac

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Sorry, r12-514-gf3661f2d63fbc5fd30c24d22137691e16b0a0a17

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Slightly more readable testcase without warnings, -O3 -mxop:
int a, b[3];
void bar (int *);

void
foo (void)
{
  for (; a < 3; a++)
    b[a] = (a - 1) / 2;
  bar (b);
}

IMHO we can either avoid the TARGET_XOP stuff for 8-byte vector modes, like:
--- gcc/config/i386/i386-expand.c.jj    2021-05-06 10:14:55.486063301 +0200
+++ gcc/config/i386/i386-expand.c       2021-05-06 14:37:58.104025120 +0200
@@ -3653,7 +3653,7 @@ ix86_expand_sse_movcc (rtx dest, rtx cmp
       emit_insn (gen_rtx_SET (dest, x));
       return;
     }
-  else if (TARGET_XOP)
+  else if (TARGET_XOP && (!VECTOR_MODE_P (mode) || GET_MODE_SIZE (mode) != 8))
     {
       op_true = force_reg (mode, op_true);

or add the MMXMODEI xop_pcmov_* patterns.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Created attachment 50765
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50765&action=edit
gcc12-pr100445.patch

Untested patch that implements the latter.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #5 from Uroš Bizjak <ubizjak at gmail dot com> ---
ix86_expand_sse_movcc has special TARGET_XOP path, so the following patch is
needed:

diff --git a/gcc/config/i386/mmx.md b/gcc/config/i386/mmx.md
index 347295afbb5..667dd057e0d 100644
--- a/gcc/config/i386/mmx.md
+++ b/gcc/config/i386/mmx.md
@@ -1687,6 +1687,17 @@ (define_expand "vcondu<mode><mode>"
   DONE;
 })

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

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #6 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Uroš Bizjak from comment #5)
> ix86_expand_sse_movcc has special TARGET_XOP path, so the following patch is
> needed:

Ah, you beat me by the second ;)

Anyway, I have no XOP target, so probably you can test the patch on a real XOP
target.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I have no XOP CPU either, normally I bootstrap on Intel i9-7960X and on my
desktop I have AMD Ryzen 5 3600, but that doesn't have XOP either.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[hjl.tools at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hjl.tools at gmail dot com

--- Comment #8 from H.J. Lu <hjl.tools at gmail dot com> ---
r514 also caused

FAIL: gcc.dg/pr97238.c (internal compiler error)
FAIL: gcc.dg/pr97238.c (test for excess errors)
FAIL: gcc.dg/vect/bb-slp-49.c -flto -ffat-lto-objects (internal compiler error)
FAIL: gcc.dg/vect/bb-slp-49.c -flto -ffat-lto-objects (test for excess errors)
FAIL: gcc.dg/vect/bb-slp-49.c (internal compiler error)
FAIL: gcc.dg/vect/bb-slp-49.c (test for excess errors)
FAIL: gcc.target/i386/pr96827.c (internal compiler error)
FAIL: gcc.target/i386/pr96827.c (test for excess errors)

with

$ cd {build_dir}/gcc && make check RUNTESTFLAGS="dg.exp=gcc.dg/pr97238.c
--target_board='unix{-m64\ -march=cascadelake}'"
$ cd {build_dir}/gcc && make check
RUNTESTFLAGS="vect.exp=gcc.dg/vect/bb-slp-49.c --target_board='unix{-m64\
-march=cascadelake}'"
$ cd {build_dir}/gcc && make check
RUNTESTFLAGS="i386.exp=gcc.target/i386/pr96827.c --target_board='unix{-m64\
-march=cascadelake}'"

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #9 from Uroš Bizjak <ubizjak at gmail dot com> ---
ix86_use_mask_cmp_p should be refined, it has an early return for 64bit modes:

  if (GET_MODE_SIZE (mode) == 64)
    return true;

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[ubizjak at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #10 from Uroš Bizjak <ubizjak at gmail dot com> ---
Following patch fixes the failures:

--cut here--
diff --git a/gcc/config/i386/i386-expand.c b/gcc/config/i386/i386-expand.c
index 4dfe7d6c282..61b2f921f41 100644
--- a/gcc/config/i386/i386-expand.c
+++ b/gcc/config/i386/i386-expand.c
@@ -3490,7 +3490,11 @@ static bool
 ix86_use_mask_cmp_p (machine_mode mode, machine_mode cmp_mode,
                     rtx op_true, rtx op_false)
 {
-  if (GET_MODE_SIZE (mode) == 64)
+  int vector_size = GET_MODE_SIZE (mode);
+
+  if (vector_size < 16)
+    return false;
+  else if (vector_size == 64)
     return true;

   /* When op_true is NULL, op_false must be NULL, or vice versa.  */
--cut here--

But perhaps 64bit V2SI, V4HI and V8QI vpcmp insns can be emulated via their
128bit versions?

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:323b18d54b960d3ef64f60ad20838ef958334dc0

commit r12-582-g323b18d54b960d3ef64f60ad20838ef958334dc0
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Fri May 7 10:37:52 2021 +0200

    i386: Fix up 8-byte vcond* with -mxop [PR100445]

    ix86_expand_sse_movcc has special TARGET_XOP handling and the recent
    addition of support of v*cond* patterns for MMXMODEI modes results in
    ICEs because the expected pattern doesn't exist.  We can handle it
    using 128-bit vpcmov (if we ignore the upper 64 bits like we ignore in
    other TARGET_MMX_WITH_SSE support).

    2021-05-07  Jakub Jelinek  <jakub@redhat.com>

            PR target/100445
            * config/i386/mmx.md (*xop_pcmov_<mode>): New define_insn.

            * gcc.target/i386/pr100445.c: New test.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #12 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:92f372f00936a549de2cb9764eee722bb07959ba

commit r12-583-g92f372f00936a549de2cb9764eee722bb07959ba
Author: Uros Bizjak <ubizjak@gmail.com>
Date:   Fri May 7 11:15:07 2021 +0200

    i386: Do not emit mask compares for mode sizes < 16 [PR100445]

    Recent addition of v*cond* patterns for MMXMODEI modes allows 64bit MMX
    modes to enter ix86_expand_sse_cmp. ix86_use_mask_cmp_p was not prepared
    to reject mode sizes < 16, resulting in ICE due to unavailability of 64bit
    masked PCOM instructions.

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

    gcc/
            PR target/100445
            * config/i386/i386-expand.c (ix86_use_mask_cmp_p):
            Return false for mode sizes < 16.

    gcc/testsuite/

            PR target/100445
            * gcc.target/i386/pr100445-1.c: New test.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

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

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

--- Comment #13 from Martin Liška <marxin at gcc dot gnu.org> ---
Is it fixed now?

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[dcb314 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

--- Comment #14 from David Binderman <dcb314 at hotmail dot com> ---
AFAIK, yes. I didn't try out the gcc/testsuite, but I did try out
all the places (about 20) where fedora rawhide failed to build.

[Bug target/100445] [12 Regression] ice during RTL pass: vregs

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100445

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

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

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
.