RFR: 8366777: Build fails unknown pseudo-op with old AS on linux-aarch64 [v17]
Erik Joelsson
erikj at openjdk.org
Fri Sep 12 13:13:39 UTC 2025
On Fri, 12 Sep 2025 08:51:50 GMT, SendaoYan <syan at openjdk.org> wrote:
>> Hi all,
>>
>> Build jdk on linux-aarch64 with older assembler version(as 2.30) will report compilation failre, shows as [JDK-8366777](https://bugs.openjdk.org/browse/JDK-8366777).
>>
>> This PR add the expression `svfloat64_t a() {}` to makefile which will trigger the compilation during configure stage. After this PR, at configure stage will check the gcc and assember support `-march=armv8-a+sve` or not more comprehensive.
>>
>> Testing:
>> - [x] On linux-aarch64 gcc12.3.1 as2.37, configure set SVE_CFLAGS as '-march=armv8-a+sve'
>> - [x] On linux-aarch64 gcc10.3.0 as2.30, configure set SVE_CFLAGS as empty
>>
>>
>> The compilation fails demo as below:
>>
>>
>> # cat sve.c
>> #include <arm_sve.h>
>> svfloat64_t a() {}
>> # gcc -march=armv8-a+sve sve.c
>> /tmp/ccVOzMzb.s: Assembler messages:
>> /tmp/ccVOzMzb.s:6: Error: unknown pseudo-op: `.variant_pcs'
>
> SendaoYan has updated the pull request incrementally with one additional commit since the last revision:
>
> Remove unnecessary CFLAGS save and restore
It's very close to ok now.
make/autoconf/flags-other.m4 line 128:
> 126: saved_cc="$CC"
> 127: CFLAGS="$CFLAGS -march=armv8-a+sve $CFLAGS_WARNINGS_ARE_ERRORS ARG_ARGUMENT"
> 128: CC="$ARG_PREFIX[CC]"
You shouldn't need to save `CC` and use a custom one here. `CC` should be the target compiler already.
make/autoconf/flags-other.m4 line 140:
> 138: [AVAILABLE=false]
> 139: )
> 140: AC_MSG_RESULT([$AVAILABLE])
This will now print true/false instead of yes/no, which will look a bit weird. It's a consequence of us preferring `true/false` to express booleans internally. I think I would inline the call to `AC_MSG_RESULT` in each of the branches of `AC_COMPILE_IFELSE` to work around this.
make/autoconf/flags.m4 line 380:
> 378: # for SVE. Set SVE_CFLAGS to -march=armv8-a+sve if it does.
> 379: # ACLE and this flag are required to build the aarch64 SVE related functions
> 380: # in libvectormath.
I think you misunderstood me here, I meant move it to the top of the macro declaration, not the call. I think it makes more sense to document it where it's defined.
-------------
PR Review: https://git.openjdk.org/jdk/pull/27073#pullrequestreview-3216584548
PR Review Comment: https://git.openjdk.org/jdk/pull/27073#discussion_r2344191092
PR Review Comment: https://git.openjdk.org/jdk/pull/27073#discussion_r2344204464
PR Review Comment: https://git.openjdk.org/jdk/pull/27073#discussion_r2344185294
More information about the build-dev
mailing list