[armel] ICE immed_double_const at emit-rtl.c (-mfpu=neon -g)

[Juha Kallioinen]

I've tried this with linaro 2010.09 and 2011.03 and the one that is currently available from Ubuntu Natty repo:

kaltsi@kaltsnatty:~/tmp/ice$ arm-linux-gnueabi-gcc-4.5 --version
arm-linux-gnueabi-gcc-4.5 (Ubuntu/Linaro 4.5.2-3ubuntu3) 4.5.2
Copyright (C) 2010 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

The problem comes when using -mfpu=neon and -g (and at least -O1).

kaltsi@kaltsnatty:~/tmp/ice$ arm-linux-gnueabi-gcc-4.5 -mfpu=neon -O1 -g -c emit-rtl-ice.i
optimized.c: In function ‘move_16bit_to_32bit’:
optimized.c:4:6: internal compiler error: in immed_double_const, at emit-rtl.c:552
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-4.5/README.Bugs> for instructions.

I do not get the ICE with linaro gcc-4.4 available from Natty.

[Michael Hope (michaelh1)]

Confirmed in gcc-linaro-4.5+bzr99489:

ichaelh@ursa1:/scratch/michaelh/bugs$ ../toolchains/gcc-linaro-4.5+bzr99489-armv7l-maverick-cbuild79-carina5-cortexa8r1/bin/gcc -mfpu=neon -O1 -g -c emit-rtl-ice.i
optimized.c: In function 'move_16bit_to_32bit':
optimized.c:4:6: internal compiler error: in immed_double_const, at emit-rtl.c:552

Also appears on a cross compiler. The backtrace is:

#1 0x0000000000580d19 in immed_double_const (i0=0, i1=0, mode=OImode)
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/emit-rtl.c:552
#2 0x00000000005a34c3 in expand_expr_real_1 (exp=0x7ffff4dd1ca8, target=0x0,
    tmode=<value optimized out>, modifier=<value optimized out>, alt_rtl=0x0)
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/expr.c:8465
#3 0x0000000000521425 in expand_expr (exp=0x7ffff4dd1ca8)
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/expr.h:558
#4 expand_debug_expr (exp=0x7ffff4dd1ca8)
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/cfgexpand.c:2329
#5 0x00000000005237d8 in expand_debug_locations ()
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/cfgexpand.c:3122
#6 gimple_expand_cfg () at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/cfgexpand.c:3885
#7 0x00000000006715dc in execute_one_pass (pass=0xf8a720)
    at ../../../src/gcc-linaro-4.5-2011.03-0/gcc/passes.c:1572

Note that mode is OImode (a 32 byte integer) which makes sense with the int32x4x2 types in the example code.

[Chung-Lin Tang (cltang)]

Also confirmed on upstream trunk.

[Chung-Lin Tang (cltang)]

Testing a patch, would you mind also submitting to upstream bugzilla too? Thanks.

[Juha Kallioinen (juha-kallioinen)]

Created attachment 23707
pre-processed source

When compiling the attached pre-processed source for arm (-march=armv7-a -mtune=cortex-a8) and using options -mfpu=neon -g -O1 I get an ICE.

kaltsi@kaltsnatty:~/tmp/ice$ arm-linux-gnueabi-gcc-4.5 -mfpu=neon -O1 -g -c emit-rtl-ice.i
optimized.c: In function ‘move_16bit_to_32bit’:
optimized.c:4:6: internal compiler error: in immed_double_const, at emit-rtl.c:552
Please submit a full bug report,
with preprocessed source if appropriate.
See <file:///usr/share/doc/gcc-4.5/README.Bugs> for instructions.

This does not happen if I leave the -g option out.

Also reported to linaro: https://bugs.launchpad.net/gcc-linaro/+bug/736007

[Juha Kallioinen (juha-kallioinen)]

Upstream bug submitted http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48183

[Chung-Lin Tang (cltang)]

Upstream patch posted:
http://gcc.gnu.org/ml/gcc-patches/2011-03/msg01220.html

[Michael Hope (michaelh1)]

Richard, Chung-Lin: could you follow up on this one please?

[Ira Rosen (irar)]

I am out of the office until 17/04/2011.

Note: This is an automated response to your message "[Bug 736007] Re:
[armel] ICE immed_double_const at emit-rtl.c (-mfpu=neon -g)" sent on
15/4/2011 4:50:25.

This is the only notification you will receive while this person is away.

[Richard Sandiford (rsandifo)]

Michael Hope <email address hidden> writes:
> Richard, Chung-Lin: could you follow up on this one please?

I think the status here is that we (Chung-Lin, Julian and myself)
have proposed three different patches on the list. They all work
around the underlying problem rather than fix it.

I was hoping the maintainers would pick one, but I suppose the
hackish nature of the patches means that no-one really wants to.

FWIW, I'm happy to defer to Chung-Lin's change for Linaro. I think
his patch has received the most testing and I agree it should be safe.

Richard

[Chung-Lin Tang (cltang)]

IMHO, I think that among the proposals, Richard Guenther's actually seems the most elegant (avoid the seemingly unneeded zero-set completely), though I think his patch is not yet really a intended tested fix currently...

[Michael Hope (michaelh1)]

OK. Can you move this bug forward then please Chung-Lin? That might be by pinging Richard Guenther and asking him to finish the patch or taking it over yourself.

[Michael Hope (michaelh1)]

Richard, could you follow this one up please?

[Michael Hope (michaelh1)]

Still present in gcc-linaro-4.6+bzr106768:

michaelh@ursa1:~/linaro/bugs$ /scratch/michaelh/toolchains/gcc-linaro-4.6+bzr106768-armv7l-natty-cbuild151-ursa4-cortexa9r1/bin/gcc -mfpu=neon -O1 -g -c emit-rtl-ice.i
optimized.c: In function 'move_16bit_to_32bit':
optimized.c:4:6: internal compiler error: in immed_double_const, at emit-rtl.c:550

[Michael K. Edwards (m-k-edwards)]

Adding a test case which Chung-Lin's patch does not appear to fix.

[Michael K. Edwards (m-k-edwards)]

Correction: Chung-Lin's patch does fix this test case. (I was using a stale compiler build when I re-checked.)

[Michael Hope (michaelh1)]

Withdrawn from stable due to the size of the change and it not affecting a user of the stable branch.

[Havard-graff-l (havard-graff-l)]

I have some additional info related to this bug:

If I compile the following:

#include "arm_neon.h"
#include "stdlib.h"

int main ()
{
   float r [] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
   float32x4x2_t d;
   d = vld2q_f32 (r);
   vst2q_f32 (r, d);
   return 0;
}

using: arm-linux-androideabi-gcc test.c -mfloat-abi=softfp -mfpu=neon -g -O1

I get:
test.c: In function 'main':
test.c:4:5: internal compiler error: in immed_double_const, at emit-rtl.c:550
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

However, if I comment out vst2q_f32 (r, d);, the problem goes away. Also by not using -g or -O1.

$ arm-linux-androideabi-gcc -v
Using built-in specs.
COLLECT_GCC=arm-linux-androideabi-gcc
COLLECT_LTO_WRAPPER=/Users/havard.graff/Code/android-ndk-r7/toolchains/arm-linux-androideabi-4.6.2/prebuilt/darwin-x86/libexec/gcc/arm-linux-androideabi/4.6.2/lto-wrapper
Target: arm-linux-androideabi
Configured with: /Users/havard.graff/Code/selfbuilt_android_toolchain/build/../gcc/gcc-4.6.2/configure --prefix=/Users/havard.graff/Code/android-ndk-r7/toolchains/arm-linux-androideabi-4.6.2/prebuilt/darwin-x86 --target=arm-linux-androideabi --host=i686-apple-darwin --build=i686-apple-darwin --with-gnu-as --with-gnu-ld --enable-languages=c,c++,objc --with-gmp=/tmp/ndk-havard.graff/build/toolchain/temp-install --with-mpfr=/tmp/ndk-havard.graff/build/toolchain/temp-install --with-mpc=/tmp/ndk-havard.graff/build/toolchain/temp-install --disable-libssp --enable-threads --disable-nls --disable-libmudflap --disable-libgomp --disable-sjlj-exceptions --disable-shared --disable-tls --with-float=soft --with-fpu=vfp --with-arch=armv5te --enable-target-optspace --disable-hosted-libstdcxx --enable-cxx-flags='-fexceptions -frtti' --enable-initfini-array --disable-nls --prefix=/Users/havard.graff/Code/android-ndk-r7/toolchains/arm-linux-androideabi-4.6.2/prebuilt/darwin-x86 --with-sysroot=/Users/havard.graff/Code/android-ndk-r7/toolchains/arm-linux-androideabi-4.6.2/prebuilt/darwin-x86/sysroot --with-binutils-version=2.22.51 --with-mpfr-version=2.4.1 --with-gmp-version=4.2.4 --with-gcc-version=4.6.2 --with-gdb-version=6.6 --with-arch=armv5te --program-transform-name='s,^,arm-linux-androideabi-,'
Thread model: posix
gcc version 4.6.2 (GCC)

[Andrew Stubbs (ams-codesourcery)]

Related: lp:gcc-linaro/4.6,revno=106781

[Andrew Stubbs (ams-codesourcery)]

Shouldn't the upstream bug be resolved now?