ICEs and infinite loops in neon_disambiguate_copy()

[Michael K. Edwards]

neon_disambiguate_copy() fails on some inputs, either with an ICE (mismatched number of register moves) or an infinite loop. Unreduced testcase attached, along with a patch that fixes the ICE, exposing the infinite loop.

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

This patch fixes the ICE, leaving the infinite loop when src and dest are permutations of one another.

Compile test case with -O3 -mfpu=neon. (If you have the fix for 736007, you can add -g as well.)

Before this patch:

make[3]: Entering directory `/home/build/CTBU/MonoLake-foreign/opensource-core/build-trees/libjpeg-turbo/simd'
/bin/bash ../libtool --tag=CC --mode=compile arm-cortex_a9-linux-gcc -DHAVE_CONFIG_H -I. -I.. -I.. -Wall -O3 -mfpu=neon --save-temps -MT jsimd_arm.lo -MD -MP -MF .deps/jsimd_arm.Tpo -c -o jsimd_arm.lo jsimd_arm.c
libtool: compile: arm-cortex_a9-linux-gcc -DHAVE_CONFIG_H -I. -I.. -I.. -Wall -O3 -mfpu=neon --save-temps -MT jsimd_arm.lo -MD -MP -MF .deps/jsimd_arm.Tpo -c jsimd_arm.c -fPIC -DPIC -o .libs/jsimd_arm.o
jsimd_arm.c:526:1: warning: 'parse_proc_cpuinfo' defined but not used [-Wunused-function]
jsimd_arm.c: In function 'jsimd_quantize':
jsimd_arm.c:1334:1: internal compiler error: in neon_disambiguate_copy, at config/arm/arm.c:19841

[Michael Hope (michaelh1)]

Thank you for the bug report. I've confirmed this with gcc-linaro-4.6-2011.06-0 on ARM:

michaelh@ursa1:~/linaro/bugs$ /tools/toolchains/arch/armv7l/gcc-linaro-4.6-2011.06-0-armv7l-maverick-cbuild130-ursa3-cortexa9r1/bin/gcc -O2 -mfpu=neon -S jsimd_arm.i
jsimd_arm.c: In function 'jsimd_quantize':
jsimd_arm.c:1334:1: internal compiler error: in smallest_mode_for_size, at stor-layout.c:451

The fault occurs at all optimisation levels. There is no known work-around partly as this is NEON specific code. The fault does not occur in gcc-linaro-4.5-2011.06 or gcc trunk r176026.

It fails in a dfferent way in gcc 4.5.3 and gcc 4.6.1:

/tools/toolchains/arch/armv7l/gcc-4.5.3-armv7l-maverick-cbuild121-ursa4-cortexa9r1
jsimd_arm.c: In function 'jsimd_ycc_rgb_convert':
jsimd_arm.c:831:1: error: insn does not satisfy its constraints:
(insn 577 490 578 30 /opt/x-tools/arm-cortex_a9-linux-gnueabi/lib/gcc/arm-cortex_a9-linux-gnueabi/4.6.1/include/arm_neon.h:9670 (set (reg:EI 95 d16 [orig:566 __b ] [566])
        (mem/s/c:EI (pre_dec:SI (reg/f:SI 3 r3 [565])) [11 __b+0 S24 A64])) 730 {*neon_movei} (expr_list:REG_INC (reg/f:SI 3 r3 [565])
        (nil)))
jsimd_arm.c:831:1: internal compiler error: in reload_cse_simplify_operands, at postreload.c:396

This has been fixed in later gcc-linaro releases.

I've set this to medium priority as it is a ftbfs and occurs in NEON specific code.

[Michael Hope (michaelh1)]

The fault seems to be in the splits for EI, CI, OI, and XI size moves. The splitter assumes you can decompose these into a set of TI moves, but that's not always true. Be interesting to work around it by changing the condition from "TARGET_NEON && reload_completed" to ""TARGET_NEON && reload_completed && 0"" and see what happens.

[Michael Hope (michaelh1)]

Hang on, the fault in the last release is different. This requires the patch for LP: #803232 before it occurs.

[Rmansfield-t (rmansfield-t)]

Created attachment 25347
preprocessed source

$ ./xgcc -v
Using built-in specs.
COLLECT_GCC=./xgcc
Target: arm-unknown-linux-gnueabi
Configured with: ../configure --target=arm-unknown-linux-gnueabi --prefix=/home/ryan/x-tools/arm-unknown-linux-gnueabi --with-sysroot=/home/ryan/x-tools/arm-unknown-linux-gnueabi/arm-unknown-linux-gnueabi//sys-root --disable-multilib --with-local-prefix=/home/ryan/x-tools/arm-unknown-linux-gnueabi/arm-unknown-linux-gnueabi/sys-root --disable-nls --enable-threads=posix --enable-symvers=gnu --enable-c99 --enable-long-long --enable-target-optspace target_alias=arm-unknown-linux-gnueabi --enable-languages=c++ --disable-shared --disable-libmudflap --disable-libssp
Thread model: posix
gcc version 4.7.0 20110923 (experimental) [trunk revision 179117] (GCC)

$ ./xgcc -B. -O3 -mfpu=neon -mfloat-abi=softfp ~/ice.i -fpic
/home/ryan/ice.i: In function 'SetupUSEVertexShader':
/home/ryan/ice.i:342:38: warning: incompatible implicit declaration of built-in function 'calloc' [enabled by default]
/home/ryan/ice.i:353:1: internal compiler error: in neon_disambiguate_copy, at config/arm/arm.c:20388
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Mikpe (mikpe)]

This test case started ICEing on trunk with Ira's r178588:

http://gcc.gnu.org/ml/gcc-cvs/2011-09/msg00199.html

[Michael Hope (michaelh1)]

Thank you for the bug report. I've confirmed this with gcc-linaro-4.6-2011.09-1 on ARM:

michaelh@ursa1:~/linaro/bugs$ /tools/toolchains/arch/armv7l/gcc-linaro-4.6-2011.09-armv7l-natty-cbuild181-ursa4-cortexa9r1/bin/gcc -c -O3 jsimd_arm.i
jsimd_arm.c: In function 'jsimd_quantize':
jsimd_arm.c:1334:1: internal compiler error: in neon_disambiguate_copy, at config/arm/arm.c:20315

The work-around is to compile at -O1. Which is a bit boring :)

I've set it to medium priority as it is a ftbfs, occurs at high optimisation levels, and the fault exists upstream.

[Ira Rosen (irar)]

Looks like the problem is with neon_movv4si.

My patch enabled the use of quad-word vectors by default, the test doesn't fail with double-words.

[Michael Hope (michaelh1)]

I'm seeing similar at:
 https://bugs.launchpad.net/gcc/+bug/809446

and:
 https://bugs.launchpad.net/gcc-linaro/+bug/885393

but it's a bit random.

Richard Earnshaw was looking at this recently. I'll ping him.

[Rearnsha (rearnsha)]

Has gone latent again on trunk, but confirmed that the code here is both broken and unnecessarily complex. Testing a fix.

[Rearnsha (rearnsha)]

Author: rearnsha
Date: Sat Nov 19 13:47:35 2011
New Revision: 181508

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=181508
Log:
PR target/50493
* arm.c (neon_disambiguate_copy): Correctly handle partial overlap
of src and dest operands.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/arm/arm.c

[Rearnsha (rearnsha)]

Author: rearnsha
Date: Sat Nov 19 16:47:02 2011
New Revision: 181510

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=181510
Log:
PR target/50493
* arm.c (neon_disambiguate_copy): Correctly handle partial overlap
of src and dest operands.

Modified:
    branches/gcc-4_6-branch/gcc/ChangeLog
    branches/gcc-4_6-branch/gcc/config/arm/arm.c

[Rearnsha (rearnsha)]

Fixed on trunk and 4.6 branch

[Michael Hope (michaelh1)]

Was fixed upstream and in the 4.6 branch. Assigning to Andrew as it will be picked up in the next merge.