[Bug middle-end/64413] New: [AArch64/ARMv7] ICE in copy_to_mode_reg, at explow.c:654

[Bug middle-end/64413] New: [AArch64/ARMv7] ICE in copy_to_mode_reg, at explow.c:654

[david.abdurachmanov at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64413

            Bug ID: 64413
           Summary: [AArch64/ARMv7] ICE in copy_to_mode_reg, at
                    explow.c:654
           Product: gcc
           Version: 4.9.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: david.abdurachmanov at gmail dot com

Created attachment 34338
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34338&action=edit
testcase (AArch64, Fortran)

OpenLoops 1.0.1 package is failing to compile on AArch64 with GCC 4.9.{1,2}
there gfortran ICE'd:

test.f90: In function 'vert_aq_s':
test.f90:38:0: internal compiler error: in copy_to_mode_reg, at explow.c:654
     Jout_S%j(1) = 0
 ^

I managed to extract a minimal testcase (testcase.f90). It should be compiled
like this:

gfortran -o testcase.os -c -ffixed-line-length-0 -ffree-line-length-0 -O0 -O0
-fPIC testcase.f90

A similar ICE was recently posted (Oct 20) with GCC 4.9.1, ARMv7, C++ on
stackoverflow:
http://stackoverflow.com/questions/26474507/compiler-crash-with-arm-neon-datatypes

I.e., the ICE point to the same "copy_to_mode_reg, at explow.c:654". Might be
related.

Have not tested with GCC 5.0 (trunk).

[Bug middle-end/64413] [AArch64/ARMv7] ICE in copy_to_mode_reg, at explow.c:654

[david.abdurachmanov at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64413

David Abdurachmanov <david.abdurachmanov at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
                 CC|                            |jakub at redhat dot com
         Resolution|---                         |DUPLICATE

--- Comment #1 from David Abdurachmanov <david.abdurachmanov at gmail dot com> ---
Worked fine with trunk. Bisecting showed that
5d098163f7aad3bed4124abc929983a620de2a82 / 216765 fixed the issue. Thus it's a
duplicate of PR 63442 ([AArch64] ICE with ubsan/overflow-int128.c test).

Applying the patch on top of GCC 4.9.1 resolved the ICE on OpenLoops 1.0.1
package.

The patch is one liner and resolves an ICE on valid code. Could I suggest RM to
have it backported for 4.9.3 release?

*** This bug has been marked as a duplicate of bug 63442 ***