[Bug c++/60675] New: [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[Bug c++/60675] New: [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[david.abdurachmanov at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

            Bug ID: 60675
           Summary: [4.9 regression][aarch64] internal compiler error:
                    Max. number of generated reload insns per insn is
                    achieved (90)
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: david.abdurachmanov at gmail dot com

$ rpm --eval "%{_build}"
aarch64-redhat-linux-gnu

Compiles fine with system GCC (gcc version 4.8.1 20130920 (Red Hat 4.8.1-10)
(GCC)).

Fails with pre-4.9.0 (tested 208693 and trunk [208843]).

Compiles fine with trunk on x86_64.

Compiles fine on aarch64 is `-fPIC` is removed from compilation line.

Attaching minimized pre-processed file.

## COMPILE ##

g++ -std=c++11 -O2 -fPIC -o G4ErrorFreeTrajState.cc.o -c
G4ErrorFreeTrajState2.ii

## ICE ##

G4ErrorFreeTrajState2.ii: In member function 'virtual G4int
G4ErrorFreeTrajState::PropagateError(const G4Track*)':
G4ErrorFreeTrajState2.ii:35892:1: internal compiler error: Max. number of
generated reload insns per insn is achieved (90)

 }
 ^

[Bug c++/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[david.abdurachmanov at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #1 from David Abdurachmanov <david.abdurachmanov at gmail dot com> ---
Seems the testcase is too large. Trimming it more usually causes it not to ICE,
but will try to trim more.

[Bug c++/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[ktkachov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

ktkachov at gcc dot gnu.org changed:

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

--- Comment #2 from ktkachov at gcc dot gnu.org ---
Can you try compiling the testcase with -mno-lra?
Seems like something that could be related to LRA.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ice-on-valid-code
             Target|                            |aarch64-none-linux-gnu
          Component|c++                         |target
   Target Milestone|---                         |4.9.0

[Bug c++/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to David Abdurachmanov from comment #1)
> Seems the testcase is too large. Trimming it more usually causes it not to
> ICE, but will try to trim more.

Does compressing it help with the size to attach it?

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[david.abdurachmanov at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #5 from David Abdurachmanov <david.abdurachmanov at gmail dot com> ---
Created attachment 32460
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32460&action=edit
Minimized pre-processed file

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #6 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I can reproduce it, I am trying to reduce it.  Note using -mcpu=cortex-a57 also
causes the internal error to go away.  So I think the scheduler before the
register allocator is causing some differences.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Created attachment 32461
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32461&action=edit
More reduced (below 14k)

>From this reduced testcase here is the most important info from the reload
dump:
  344: r173:DF=r594:DF-r693:DF
      REG_DEAD r594:DF
      REG_EQUAL 1.0e+0-r324:DF
    Inserting insn reload before:
  749: r693:DF=r324:DF

            Staticly defined alt reject+=6
            0 Non pseudo reload: reject++
          alt=0,overall=7,losers=0,rld_nregs=0
     Choosing alt 0 in insn 749:  (0) =w  (1) ?rY {*movdf_aarch64}
            Staticly defined alt reject+=6
            0 Non input pseudo reload: reject++
            1 Non-pseudo reload: reject+=2
            1 Non input pseudo reload: reject++
          alt=0,overall=22,losers=2,rld_nregs=2
            Staticly defined alt reject+=6
            1 Non-pseudo reload: reject+=2
            1 Non input pseudo reload: reject++
          alt=1,overall=15,losers=1,rld_nregs=1
            0 Non input pseudo reload: reject++
            1 Non-pseudo reload: reject+=2
            1 Non input pseudo reload: reject++
            alt=2,overall=16,losers=2 -- refuse
            0 Non input pseudo reload: reject++
            alt=3: Bad operand -- refuse
            0 Non input pseudo reload: reject++
          alt=4,overall=7,losers=1,rld_nregs=1
            0 Spill pseudo into memory: reject+=3
            0 Non input pseudo reload: reject++
            alt=5,overall=10,losers=1 -- refuse
          alt=6,overall=0,losers=0,rld_nregs=0
     Choosing alt 6 in insn 266:  (0) r  (1) m {*movdf_aarch64}
            Staticly defined alt reject+=6
          alt=0,overall=12,losers=1,rld_nregs=1
            Staticly defined alt reject+=6
            0 Non-pseudo reload: reject+=2
            0 Non input pseudo reload: reject++
            alt=1,overall=15,losers=1 -- refuse
          alt=2,overall=0,losers=0,rld_nregs=0
     Choosing alt 2 in insn 262:  (0) w  (1) w {*movdf_aarch64}
            Staticly defined alt reject+=6
          alt=0,overall=12,losers=1,rld_nregs=1
            Staticly defined alt reject+=6
            0 Non input pseudo reload: reject++
            alt=1,overall=13,losers=1 -- refuse
          alt=2,overall=606,losers=1,rld_nregs=1
            alt=3: Bad operand -- refuse
            1 Non pseudo reload: reject++
          alt=4,overall=1,losers=0,rld_nregs=0
     Choosing alt 4 in insn 744:  (0) w  (1) m {*movdf_aarch64}
            Staticly defined alt reject+=6
            0 Non input pseudo reload: reject++
          alt=0,overall=619,losers=2,rld_nregs=2
            Staticly defined alt reject+=6
            0 Non input pseudo reload: reject++
          alt=1,overall=13,losers=1,rld_nregs=1
            0 Non input pseudo reload: reject++
          alt=2,overall=607,losers=1,rld_nregs=1
            0 Non input pseudo reload: reject++
            alt=3: Bad operand -- refuse
            0 Non input pseudo reload: reject++
            1 Spill pseudo into memory: reject+=3
            alt=4,overall=16,losers=2 -- refuse
            0 Non pseudo reload: reject++
          alt=5,overall=1,losers=0,rld_nregs=0
     Choosing alt 5 in insn 743:  (0) m  (1) w {*movdf_aarch64}
            0 Non input pseudo reload: reject++
          alt=0,overall=613,losers=2,rld_nregs=4
            0 Costly loser: reject++
            0 Non input pseudo reload: reject++
          alt=1,overall=14,losers=2,rld_nregs=4
            0 Non input pseudo reload: reject++
            1 Costly loser: reject++
          alt=2,overall=614,losers=2,rld_nregs=4
            0 Costly loser: reject++
            0 Non input pseudo reload: reject++
            1 Costly loser: reject++
            alt=3,overall=15,losers=2 -- refuse
            0 Non input pseudo reload: reject++
            1 Spill pseudo into memory: reject+=3
            alt=4,overall=16,losers=2 -- refuse
            0 Non pseudo reload: reject++
          alt=5,overall=7,losers=1,rld_nregs=2
            0 Non pseudo reload: reject++
            alt=6: Bad operand -- refuse
            0 Costly loser: reject++
            0 Non input pseudo reload: reject++
            alt=7,overall=8,losers=1 -- refuse
            0 Non pseudo reload: reject++
            1 Costly loser: reject++
            alt=8,overall=8,losers=1 -- refuse
     Choosing alt 5 in insn 737:  (0) Ump  (1) r {*movti_aarch64}
      Creating newreg=694, assigning class CORE_REGS to r694

It decides to use movti_aarch64 for some reason even though we don't have a
TImode here, only DFmode or DImode.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-03-27
     Ever confirmed|0                           |1

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Confirmed by my reduced testcase.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Note I couldn't reproduce this with a cross from x86_64-linux to aarch64-linux,
wonder what is the important difference.  While I don't have cross binutils
installed, I made sure I have HAVE_GAS_HIDDEN and HAVE_AS_TLS enabled in
auto-host.h.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[ktkachov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #10 from ktkachov at gcc dot gnu.org ---
(In reply to Jakub Jelinek from comment #9)
> Note I couldn't reproduce this with a cross from x86_64-linux to
> aarch64-linux, wonder what is the important difference.  While I don't have
> cross binutils installed, I made sure I have HAVE_GAS_HIDDEN and HAVE_AS_TLS
> enabled in auto-host.h.

I can reproduce it with an aarch64-none-elf cross compiler at revision 208889
configured with:
--with-pkgversion=unknown --disable-shared --disable-nls --disable-threads
--disable-tls --enable-checking=yes --enable-languages=c,c++,fortran
--with-newlib --with-cpu=cortex-a53 --with-arch=armv8-a

I haven't tried a linux compiler, I'll try that next...

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[vmakarov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #11 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
I've been working on it.  I hope the patch will be ready today.

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[vmakarov at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

--- Comment #12 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
Author: vmakarov
Date: Fri Mar 28 15:27:58 2014
New Revision: 208900

URL: http://gcc.gnu.org/viewcvs?rev=208900&root=gcc&view=rev
Log:
2014-03-28  Vladimir Makarov  <vmakarov@redhat.com>

    PR target/60675
    * lra-assigns.c (find_hard_regno_for): Remove unavailable hard
    regs from checking multi-reg pseudos.

2014-03-28  Vladimir Makarov  <vmakarov@redhat.com>

    PR target/60675
    * gcc.target/aarch64/pr60675.C: New.


Added:
    trunk/gcc/testsuite/gcc.target/aarch64/pr60675.C
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/lra-assigns.c
    trunk/gcc/testsuite/ChangeLog

[Bug target/60675] [4.9 regression][aarch64] internal compiler error: Max. number of generated reload insns per insn is achieved (90)

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60675

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

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

--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Fixed then.