gcc 4.8: "invalid expression as operand" in aarch64 inline asm

[Will Newton]

The attached pre-processed source fails with:

# aarch64-linux-gnu-gcc -std=gnu99 -fgnu89-inline -O -Wall -Winline -Wwrite-strings -fmerge-all-constants -frounding-math -g -Wstrict-prototypes malloc.i
malloc.c: In function ‘__libc_mallopt’:
malloc.c:4761:16: error: invalid 'asm': invalid expression as operand

Splitting this inline asm out into a smaller testcase does not seem to reproduce the issue. It's possible this is a bug in the source file, but it's not clear from gcc's output what the problem is.

[Kugan Vivekanandarajah (kugan-vivekanandarajah)]

Seems to be an issue with alignment calculation in
aarch64_classify_address.

for rtl of the form (lo_sum:DI (reg/f:DI 132)
    (symbol_ref:DI ("*.LANCHOR4") [flags 0x182])) we are currently
calculating align of 8bits and due to this the following statement
returns false. This causes output_addr_const (in gcc/final.c) to fail.

 return ((INTVAL (offs) & (ref_size - 1)) == 0
                      && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0);

AFIK, align of 8bits is not correct here.

[Kuganv (kuganv)]

Created attachment 32017
pre processed source

This issue is originally reported in https://bugs.launchpad.net/gcc-linaro/+bug/1270789

The attached pre-processed source fails with:

# aarch64-linux-gnu-gcc -std=gnu99 -fgnu89-inline -O -Wall -Winline -Wwrite-strings -fmerge-all-constants -frounding-math -g -Wstrict-prototypes malloc.i
malloc.c: In function ‘__libc_mallopt’:
malloc.c:4761:16: error: invalid 'asm': invalid expression as operand

Splitting this inline asm out into a smaller testcase does not seem to reproduce the issue. It's possible this is a bug in the source file, but it's not clear from gcc's output what the problem is.

Seems to be an issue with alignment calculation in
aarch64_classify_address.

for rtl of the form (lo_sum:DI (reg/f:DI 132)
    (symbol_ref:DI ("*.LANCHOR4") [flags 0x182])) we are currently
calculating align of 8bits and due to this the following statement
returns false. This causes output_addr_const (in gcc/final.c) to fail.

 return ((INTVAL (offs) & (ref_size - 1)) == 0
                      && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0);

AFIK, align of 8bits is not correct here.

GCC version and config:
arm-none-linux-gnueabi-gcc -v
Using built-in specs.
COLLECT_GCC=/home/kugan/work/builds/gcc-fsf-trunk/tools/bin/arm-none-linux-gnueabi-gcc
COLLECT_LTO_WRAPPER=/home/kugan/work/builds/gcc-fsf-trunk/tools/libexec/gcc/arm-none-linux-gnueabi/4.9.0/lto-wrapper
Target: arm-none-linux-gnueabi
Configured with: /home/kugan/work/sources/gcc-fsf/trunk/configure --target=arm-none-linux-gnueabi --prefix=/home/kugan/work/builds/gcc-fsf-trunk/tools --with-sysroot=/home/kugan/work/builds/gcc-fsf-trunk/sysroot-arm-none-linux-gnueabi --disable-libssp --disable-libgomp --disable-libmudflap --enable-languages=c,c++ --with-arch=armv7-a --with-fpu=vfpv3-d16 --with-float=softfp --with-thumb
Thread model: posix
gcc version 4.9.0 20130922 (experimental) (GCC)

[Kuganv (kuganv)]

Created attachment 32018
Proposed patch

[Kuganv (kuganv)]

I am sorry that I entered wrong gcc -v (cut and paste error). It is an aarch64 issue and the correct gcc -v is below:

aarch64-none-linux-gnu-gcc -v
Using built-in specs.
COLLECT_GCC=/home/kugan/work/builds/gcc-fsf-trunk/tools/bin/aarch64-none-linux-gnu-gcc
COLLECT_LTO_WRAPPER=/home/kugan/work/builds/gcc-fsf-trunk/tools/libexec/gcc/aarch64-none-linux-gnu/4.9.0/lto-wrapper
Target: aarch64-none-linux-gnu
Configured with: /home/kugan/work/sources/gcc-fsf/trunk/configure --target=aarch64-none-linux-gnu --prefix=/home/kugan/work/builds/gcc-fsf-trunk/tools --without-headers --with-newlib --disable-shared --disable-threads --disable-libssp --disable-libgomp --disable-libmudflap --disable-libatomic --without-libquadmath --disable-libquadmath --enable-languages=c
Thread model: single
gcc version 4.9.0 20130922 (experimental) (GCC)

[Pinskia (pinskia)]

Here is a reduced testcase:
static unsigned long global_max_fast;
int __libc_mallopt (int param_number, int value)
{
 __asm__ __volatile__ ("# %[_SDT_A2]" :: [_SDT_A2] "nor" ((global_max_fast)));
 global_max_fast = 1;
}

[Ramana-gcc (ramana-gcc)]

(In reply to Kugan from comment #1)
> Created attachment 32018 [details]
> Proposed patch

Please submit patches on <email address hidden>

[Kuganv (kuganv)]

There is a rgression with this patch in qemu aarch64-none-linux-gnu for pr38151.c.

pr38151.c:(.text+0x10c): relocation truncated to fit:
R_AARCH64_LDST64_ABS_LO12_NC against `.rodata'
collect2: error: ld returned 1 exit status

asm diff is as shown below.

< add x0, x0, :lo12:.LANCHOR0
< ldr x0, [x0]
---
> ldr x0, [x0,#:lo12:.LANCHOR0]

If I however increase the alignment of .rodata where .LANCHOR0 is
defined, this passes. Is alignment of BITS_PER_UNIT valid for
SYMBOL_REF? If I change it as I am doing this attached patch, is there
anything else I need to do. Any tips?

[Kugan (kugan)]

Author: kugan
Date: Sun Mar 30 22:41:59 2014
New Revision: 208949

URL: http://gcc.gnu.org/viewcvs?rev=208949&root=gcc&view=rev
Log:
PR target/60034
* aarch64/aarch64.c (aarch64_classify_address): Fix alignment for
section anchor.

PR target/60034
* gcc.target/aarch64/pr60034.c: New file.

Added:
    trunk/gcc/testsuite/gcc.target/aarch64/pr60034.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/aarch64/aarch64.c
    trunk/gcc/testsuite/ChangeLog

[Launchpad Janitor (janitor)]

This bug was fixed in the package gcc-4.8 - 4.8.2-19ubuntu1

---------------
gcc-4.8 (4.8.2-19ubuntu1) trusty; urgency=medium

  * Merge with Debian; remaining changes:
    - Build from the upstream source.
    - Don't build the libgcc packages, now built from gcc-4.9.

gcc-4.8 (4.8.2-19) unstable; urgency=medium

  * Update to SVN 20140404 (r209122) from the gcc-4_8-branch.
    - Fix PR rtl-optimization/60700 (wrong code on x86),
      PR rtl-optimization/57637 (wrong code on ARM in SPEC2000),
      PR target/60039 (SH), PR ada/60703.

  [ Matthias Klose ]
  * Include include and include-fixed header files into the stage1
    gcc-4.8 package.
  * Explicitly configure with --disable-multilib on sparc64 when no
    multilibs are requested (Helmut Grohne). Closes: #743342.
  * Fix PR target/60034 (AArch64), taken from the trunk. LP: #1270789.
  * Update ada-ppc64.diff from the gnat-4.9 package, fixing the gnat build
    on powerpc.
  * Fix PR target/60609 (ARM), proposed patch (Charles Baylis). LP: #1295653.
  * Build libphobos for armel and armhf.
  * Include the gnu triplet prefixed gcov and gcc-{ar,nm,ranlib} binaries.
  * Stop building ppc64el from the ibm branch, now integrated in the fsf
    branch.

  [ Iain Buclaw ]
  * Update the GDC frontend (20140401).
 -- Matthias Klose <email address hidden> Fri, 04 Apr 2014 19:36:41 +0200

[Ramana-gcc (ramana-gcc)]

Fixed on trunk for 4.9.0.

[Kugan Vivekanandarajah (kugan-vivekanandarajah)]

Fixed in upstream with https://gcc.gnu.org/bugzilla/show_bug.cgi?id=60034