[Bug tree-optimization/97706] New: [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[Bug tree-optimization/97706] New: [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[acoplan at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

            Bug ID: 97706
           Summary: [11 Regression] ICE with LTO at -O3: verify_gimple
                    failed (incompatible types in 'PHI' argument 0)
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: acoplan at gcc dot gnu.org
  Target Milestone: ---

Since r11-4428-g4a369d199bf2f34e037030b18d0da923e8a24997 I'm seeing an ICE when
compiling the following two (dubious) source files with LTO at -O3 on AArch64:

$ cat f1.c
void a(_Bool x);
void test(char**);
int var_19;
char **d;
_Bool arr_8;
int main()
{
  test(d);
  a(arr_8);
  return var_19;
}

$ cat f2.c
extern int var_19;
extern _Bool arr_8[][6][11];
void test(char **p)
{
  for (int c; c < 9; c++)
    for (int d = 0; d < 3; d -= 4LL)
      for (int e = 0; e < 10; e++) {
        var_19 = 0;
        arr_8[5][d][e] = p[d][e];
      }
}

$ aarch64-elf-gcc -O3 -flto f1.c f2.c
f1.c: In function 'main':
f1.c:6:5: error: incompatible types in 'PHI' argument 0
    6 | int main()
      |     ^
vector(8) unsigned char

vector(8) <signed-boolean:8>

vect__29.9_48 = PHI <mask__40.8_55(5)>
during GIMPLE pass: slp
f1.c:6:5: internal compiler error: verify_gimple failed
0xcd1728 verify_gimple_in_cfg(function*, bool)
        /home/alecop01/toolchain/src/gcc/gcc/tree-cfg.c:5461
0xb7ea24 execute_function_todo
        /home/alecop01/toolchain/src/gcc/gcc/passes.c:2039
0xb7ec1e do_per_function
        /home/alecop01/toolchain/src/gcc/gcc/passes.c:1687
0xb7edb9 execute_todo
        /home/alecop01/toolchain/src/gcc/gcc/passes.c:2093
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.
lto-wrapper: fatal error: aarch64-elf-gcc returned 1 exit status
compilation terminated.
/home/alecop01/toolchain/build-aarch64-elf/install/lib/gcc/aarch64-elf/11.0.0/../../../../aarch64-elf/bin/ld:
error: lto-wrapper failed
collect2: error: ld returned 1 exit status

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[acoplan at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

Alex Coplan <acoplan at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |11.0

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2020-11-04
     Ever confirmed|0                           |1
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I will have a look.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |WAITING

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Hmm, please specify how you configured.  I do

../src/trunk/configure --target=aarch64-suse-linux

and that doesn't suffice.  Any cpu/fpu selection?

As I'm using a cross I do not have libc and thus no runtime objects to link
against and thus trying

> ./xgcc -B. f1.c f2.c -flto -O3 -r -flinker-output=nolto-rel

which works fine.  Tried with -march=armv8.2-a+sve as well.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu.org,
                   |                            |rsandifo at gcc dot gnu.org

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Eventually a GIMPLE testcase to fix_loops might produce a non-LTO testcase.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[marxin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
                 CC|                            |marxin at gcc dot gnu.org
           Priority|P3                          |P1

--- Comment #4 from Martin Liška <marxin at gcc dot gnu.org> ---
I can reproduce it on an aarch64 machine. GIMPLE FE test-case is not possible,
there are too many errors.
@Richi: Please use any of machines we have: needle.suse.cz, or
oolchainarm.arch.suse.de.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Testcase that fails on x86_64 as well:

_Bool arr[16];
void bar();
void foo(int n, char *p)
{
  _Bool b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15;
  do
    {
      b0 = p[0] != 0;
      b1 = p[1] != 0;
      b2 = p[2] != 0;
      b3 = p[3] != 0;
      b4 = p[4] != 0;
      b5 = p[5] != 0;
      b6 = p[6] != 0;
      b7 = p[7] != 0;
      b8 = p[8] != 0;
      b9 = p[9] != 0;
      b10 = p[10] != 0;
      b11 = p[11] != 0;
      b12 = p[12] != 0;
      b13 = p[13] != 0;
      b14 = p[14] != 0;
      b15 = p[15] != 0;
      arr[0] = b0;
      arr[1] = b1;
      arr[2] = b2;
      arr[3] = b3;
      arr[4] = b4;
      arr[5] = b5;
      arr[6] = b6;
      arr[7] = b7;
      arr[8] = b8;
      arr[9] = b9;
      arr[10] = b10;
      arr[11] = b11;
      arr[12] = b12;
      arr[13] = b13;
      arr[14] = b14;
      arr[15] = b15;
      bar ();
    }
  while (--n);
  arr[0] = b0;
  arr[1] = b1;
  arr[2] = b2;
  arr[3] = b3;
  arr[4] = b4;
  arr[5] = b5;
  arr[6] = b6;
  arr[7] = b7;
  arr[8] = b8;
  arr[9] = b9;
  arr[10] = b10;
  arr[11] = b11;
  arr[12] = b12;
  arr[13] = b13;
  arr[14] = b14;
  arr[15] = b15;
}

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
I guess we're missing what usually is done by bool patterns here, namely
adding a conversion from the mask vector type produced by the comparison
to the value vector type.

A simple workaround would be to reject such PHIs but I wonder how many
other uses prevail.  I don't remember why we chose to assign a mask
vector type to the LHS of the comparisons.

For the SSE case adding compensation via a VIEW_CONVERT_EXPR is possible
but for the AVX512 mask case that's more involved:

> ./cc1 -quiet t.i -O3 -fdump-tree-slp-details -mavx512f -mavx512vl -mavx512bw 
t.i: In function 'foo':
t.i:3:6: error: incompatible types in 'PHI' argument 0
    3 | void foo(int n, char *p)
      |      ^~~
vector(16) unsigned char

vector(16) <signed-boolean:1>

vect_b0_104.7_123 = PHI <mask_b0_37.6_106(3)>

need to think about this.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, so for the stores in the loop we end up with a bool pattern for the store
we don't support (eh):

t.i:24:14: note:   Starting SLP discovery for
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[0]) = patt_88;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[1]) = patt_87;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[2]) = patt_30;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[3]) = patt_28;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[4]) = patt_26;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[5]) = patt_24;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[6]) = patt_22;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[7]) = patt_20;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[8]) = patt_18;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[9]) = patt_16;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[10]) = patt_14;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[11]) = patt_12;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[12]) = patt_10;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[13]) = patt_8;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[14]) = patt_6;
t.i:24:14: note:     VIEW_CONVERT_EXPR<unsigned char>(arr[15]) = patt_4;
t.i:24:14: note:   Build SLP for VIEW_CONVERT_EXPR<unsigned char>(arr[0]) =
patt_88;
t.i:24:14: note:   get vectype for scalar type (group size 16): _Bool
t.i:24:14: note:   vectype: vector(16) unsigned char
t.i:24:14: note:   nunits = 16
t.i:24:14: missed:   Build SLP failed: operation unsupported
VIEW_CONVERT_EXPR<unsigned char>(arr[0]) = patt_88;
t.i:24:14: note:   SLP discovery failed

where the issue is that we have pattern detection follow DR group analysis
for SLP vectorization so the above do not look as DR group to
vect_analyze_slp_instance_1.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, so we'd indeed need to handle PHIs in vect_recog_bool_pattern which might
turn out a bit complicated.  I'm not a fan of pattern detection for SLP
vectorization anyway so I'll see what to do here.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:7307d8e10727aea8069c0e47e64a7a9b8588a22e

commit r11-4776-g7307d8e10727aea8069c0e47e64a7a9b8588a22e
Author: Richard Biener <rguenther@suse.de>
Date:   Fri Nov 6 11:11:42 2020 +0100

    tree-optimization/97706 - part one, refactor vect_determine_mask_precision

    This computes vect_determine_mask_precision in a RPO forward walk
    rather than in a backward walk and using a worklist.  It will make
    fixing PR97706 easier but for bisecting I wanted it to be separate.

    2020-11-06  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97706
            * tree-vect-patterns.c (vect_determine_mask_precision):
            Remove worklist operation.
            (vect_determine_stmt_precisions): Do not call
            vect_determine_mask_precision here.
            (vect_determine_precisions): Compute mask precision
            in a forward walk.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:22175d0dc6a89ddd630f19d0f32a2d1ddb046807

commit r11-4787-g22175d0dc6a89ddd630f19d0f32a2d1ddb046807
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Nov 5 12:34:42 2020 +0100

    tree-optimization/97706 - handle PHIs in pattern recog mask precison

    This adds handling of PHIs to mask precision compute which is
    eventually needed to detect a bool pattern when the def chain
    contains such a PHI node.

    2020-11-06  Richard Biener  <rguenther@suse.de>

            PR tree-optimization/97706
            * tree-vect-patterns.c (possible_vector_mask_operation_p):
            PHIs are possible mask operations.
            (vect_determine_mask_precision): Handle PHIs.
            (vect_determine_precisions): Walk PHIs in BB analysis.

            * gcc.dg/vect/bb-slp-pr97706.c: New testcase.

[Bug tree-optimization/97706] [11 Regression] ICE with LTO at -O3: verify_gimple failed (incompatible types in 'PHI' argument 0)

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97706

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

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

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.