internal compiler error: in expand_debug_locations, at cfgexpand.c:5118

[Colin Hercus]

Hi,
I'm getting subject error when compiling attached code with options

g++ -finline-functions -m64 -O3 -c -g -MMD -MP -MF "main.o.d" -o main.o main.cpp

Using: gcc (Ubuntu 5.4.0-6ubuntu1~16.04.4) 5.4.0 20160609
System: Ubuntu 16.04

The error asked me to "Please submit a full bug report,"

This is a cut down example from a larger program. I'm compiling larger program with -m64 and have SIMD routines for SSE, AVX & AVX2

The problem goes away if I
1. Remove -O option or,
2. Remove the -g option (but I do want both of these for profiling) or,
3. Add explicit inline to
    inline __attribute__ ((target ("avx2"))) __m256i _mm256_shift_left(__m256i a, int n) and
    inline __attribute__ ((target ("avx2"))) __m256i _mm256_shift_right(__m256i a, int n)

I love gcc and was so happy when I found __attribute__ ((target ("avx2"))) option. Keep up the good work.

Thanks, Colin
PS. The routines __m256i _mm256_shift_[left|right](__m256i a, int n) are used in initialisation preceding a large loop and performance doesn't really matter. I'm sure they can be done better.

[Matthias Klose (doko)]

not reproducible:

$ g++ -finline-functions -m64 -O3 -c -g -MMD -MP -MF "main.o.d" -o main.o main.cpp
main.cpp: In function 'void p(AV)':
main.cpp:80:6: note: The ABI for passing parameters with 32-byte alignment has changed in GCC 4.6
 void p(union AV a) {
      ^

[Doko-v (doko-v)]

Created attachment 40223
test case

seen on trunk, 6 and 5 branches on x86_64-linux-gnu, works with -O1 and up:

$ g++ -c -O0 main.cpp
main.cpp: In function 'void p(AV)':
main.cpp:80:6: note: The ABI for passing parameters with 32-byte alignment has changed in GCC 4.6
 void p(union AV a) {
      ^
main.cpp: In function 'void test(AV, int)':
main.cpp:101:29: internal compiler error: in convert_move, at expr.c:230
     r.av = _mm256_shift_left(a.av,n);
            ~~~~~~~~~~~~~~~~~^~~~~~~~
0x898da0 convert_move(rtx_def*, rtx_def*, int)
        ../../src/gcc/expr.c:230
0x89f5cb store_expr_with_bounds(tree_node*, rtx_def*, int, bool, bool, tree_node*)
        ../../src/gcc/expr.c:5628
0x89fa3e expand_assignment(tree_node*, tree_node*, bool)
        ../../src/gcc/expr.c:5320
0x7b3dcd expand_gimple_stmt_1
        ../../src/gcc/cfgexpand.c:3641
0x7b3dcd expand_gimple_stmt
        ../../src/gcc/cfgexpand.c:3737
0x7b538f expand_gimple_basic_block
        ../../src/gcc/cfgexpand.c:5744
0x7ba596 execute
        ../../src/gcc/cfgexpand.c:6358
Please submit a full bug report,
with preprocessed source if appropriate.

[Matthias Klose (doko)]

but reproduces with -O0

[Marxin-m (marxin-m)]

Confirmed, started with r220609.https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78643#

[Jakub-gcc (jakub-gcc)]

I think the problem is that while TYPE_MODE has the hacks for vector types (vector_type_mode call), DECL_MODE doesn't have something similar.
We have some hacks for this here and there.

[Jakub-gcc (jakub-gcc)]

E.g. the following untested patch fixes it (or we could do it in the callers):
--- gcc/expr.c.jj 2017-08-29 19:03:09.000000000 +0200
+++ gcc/expr.c 2017-08-30 02:09:33.150618229 +0200
@@ -7010,7 +7010,11 @@ get_inner_reference (tree exp, HOST_WIDE
       size. */
  mode = TYPE_MODE (DECL_BIT_FIELD_TYPE (field));
       else if (!DECL_BIT_FIELD (field))
- mode = DECL_MODE (field);
+ {
+ mode = DECL_MODE (field);
+ if (mode == BLKmode && VECTOR_TYPE_P (TREE_TYPE (exp)))
+ mode = TYPE_MODE (TREE_TYPE (exp));
+ }
       else if (DECL_MODE (field) == BLKmode)
  blkmode_bitfield = true;

[Jakub-gcc (jakub-gcc)]

GCC 5 branch is being closed

[Jakub-gcc (jakub-gcc)]

Created attachment 42764
gcc8-pr80583.patch

Untested fix.

[Jakub-gcc (jakub-gcc)]

Author: jakub
Date: Sat Dec 2 07:54:47 2017
New Revision: 255353

URL: https://gcc.gnu.org/viewcvs?rev=255353&root=gcc&view=rev
Log:
 PR target/78643
 PR target/80583
 * expr.c (get_inner_reference): If DECL_MODE of a non-bitfield
 is BLKmode for vector field with vector raw mode, use TYPE_MODE
 instead of DECL_MODE.

 * gcc.target/i386/pr80583.c: New test.

Added:
    trunk/gcc/testsuite/gcc.target/i386/pr80583.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/expr.c
    trunk/gcc/testsuite/ChangeLog

[Jakub-gcc (jakub-gcc)]

Fixed on the trunk so far.

[Jakub-gcc (jakub-gcc)]

Author: jakub
Date: Fri Dec 15 22:09:07 2017
New Revision: 255721

URL: https://gcc.gnu.org/viewcvs?rev=255721&root=gcc&view=rev
Log:
 Backported from mainline
 2017-12-02 Jakub Jelinek <email address hidden>

 PR target/78643
 PR target/80583
 * expr.c (get_inner_reference): If DECL_MODE of a non-bitfield
 is BLKmode for vector field with vector raw mode, use TYPE_MODE
 instead of DECL_MODE.

 * gcc.target/i386/pr80583.c: New test.

Added:
    branches/gcc-7-branch/gcc/testsuite/gcc.target/i386/pr80583.c
Modified:
    branches/gcc-7-branch/gcc/ChangeLog
    branches/gcc-7-branch/gcc/expr.c
    branches/gcc-7-branch/gcc/testsuite/ChangeLog

[Jakub-gcc (jakub-gcc)]

Fixed for 7.3+ too.

[Jakub-gcc (jakub-gcc)]

Author: jakub
Date: Mon Jun 25 16:57:13 2018
New Revision: 262041

URL: https://gcc.gnu.org/viewcvs?rev=262041&root=gcc&view=rev
Log:
 Backported from mainline
 2017-12-02 Jakub Jelinek <email address hidden>

 PR target/78643
 PR target/80583
 * expr.c (get_inner_reference): If DECL_MODE of a non-bitfield
 is BLKmode for vector field with vector raw mode, use TYPE_MODE
 instead of DECL_MODE.

 * gcc.target/i386/pr80583.c: New test.

Added:
    branches/gcc-6-branch/gcc/testsuite/gcc.target/i386/pr80583.c
Modified:
    branches/gcc-6-branch/gcc/ChangeLog
    branches/gcc-6-branch/gcc/expr.c
    branches/gcc-6-branch/gcc/testsuite/ChangeLog

[Jakub-gcc (jakub-gcc)]

Fixed for 6.5.

[Cvs-commit (cvs-commit)]

The master branch has been updated by H.J. Lu <email address hidden>:

https://gcc.gnu.org/g:1c64aba8cdf6509533f554ad86640f274cdbe37f

commit r13-3490-g1c64aba8cdf6509533f554ad86640f274cdbe37f
Author: H.J. Lu <email address hidden>
Date: Wed Oct 19 12:53:35 2022 -0700

    Always use TYPE_MODE instead of DECL_MODE for vector field

    e034c5c8957 re PR target/78643 (ICE in convert_move, at expr.c:230)

    fixed the case where DECL_MODE of a vector field is BLKmode and its
    TYPE_MODE is a vector mode because of target attribute. Remove the
    BLKmode check for the case where DECL_MODE of a vector field is a vector
    mode and its TYPE_MODE isn't a vector mode because of target attribute.

    gcc/

            PR target/107304
            * expr.cc (get_inner_reference): Always use TYPE_MODE for vector
            field with vector raw mode.

    gcc/testsuite/

            PR target/107304
            * gcc.target/i386/pr107304.c: New test.

[Cvs-commit (cvs-commit)]

The releases/gcc-12 branch has been updated by H.J. Lu <email address hidden>:

https://gcc.gnu.org/g:0138ebcd482c71f55d74eb9fa533fbb54a4391d6

commit r12-8899-g0138ebcd482c71f55d74eb9fa533fbb54a4391d6
Author: H.J. Lu <email address hidden>
Date: Wed Oct 19 12:53:35 2022 -0700

    Always use TYPE_MODE instead of DECL_MODE for vector field

    e034c5c8957 re PR target/78643 (ICE in convert_move, at expr.c:230)

    fixed the case where DECL_MODE of a vector field is BLKmode and its
    TYPE_MODE is a vector mode because of target attribute. Remove the
    BLKmode check for the case where DECL_MODE of a vector field is a vector
    mode and its TYPE_MODE isn't a vector mode because of target attribute.

    gcc/

            PR target/107304
            * expr.cc (get_inner_reference): Always use TYPE_MODE for vector
            field with vector raw mode.

    gcc/testsuite/

            PR target/107304
            * gcc.target/i386/pr107304.c: New test.

    (cherry picked from commit 1c64aba8cdf6509533f554ad86640f274cdbe37f)

[Cvs-commit (cvs-commit)]

The releases/gcc-11 branch has been updated by H.J. Lu <email address hidden>:

https://gcc.gnu.org/g:c97c6569c826731910459812e3608255962dab78

commit r11-10370-gc97c6569c826731910459812e3608255962dab78
Author: H.J. Lu <email address hidden>
Date: Wed Oct 19 12:53:35 2022 -0700

    Always use TYPE_MODE instead of DECL_MODE for vector field

    e034c5c8957 re PR target/78643 (ICE in convert_move, at expr.c:230)

    fixed the case where DECL_MODE of a vector field is BLKmode and its
    TYPE_MODE is a vector mode because of target attribute. Remove the
    BLKmode check for the case where DECL_MODE of a vector field is a vector
    mode and its TYPE_MODE isn't a vector mode because of target attribute.

    gcc/

            PR target/107304
            * expr.c (get_inner_reference): Always use TYPE_MODE for vector
            field with vector raw mode.

    gcc/testsuite/

            PR target/107304
            * gcc.target/i386/pr107304.c: New test.

    (cherry picked from commit 1c64aba8cdf6509533f554ad86640f274cdbe37f)

[Cvs-commit (cvs-commit)]

The releases/gcc-10 branch has been updated by H.J. Lu <email address hidden>:

https://gcc.gnu.org/g:bb08940b30bd50c6e860bb8ac72d6f2ce7c1b25d

commit r10-11076-gbb08940b30bd50c6e860bb8ac72d6f2ce7c1b25d
Author: H.J. Lu <email address hidden>
Date: Wed Oct 19 12:53:35 2022 -0700

    Always use TYPE_MODE instead of DECL_MODE for vector field

    e034c5c8957 re PR target/78643 (ICE in convert_move, at expr.c:230)

    fixed the case where DECL_MODE of a vector field is BLKmode and its
    TYPE_MODE is a vector mode because of target attribute. Remove the
    BLKmode check for the case where DECL_MODE of a vector field is a vector
    mode and its TYPE_MODE isn't a vector mode because of target attribute.

    gcc/

            PR target/107304
            * expr.c (get_inner_reference): Always use TYPE_MODE for vector
            field with vector raw mode.

    gcc/testsuite/

            PR target/107304
            * gcc.target/i386/pr107304.c: New test.

    (cherry picked from commit 1c64aba8cdf6509533f554ad86640f274cdbe37f)