RFR: 8365312: GCC 12 cannot compile SVE on aarch64 with auto-var-init pattern

[Magnus Ihse Bursie]

On Thu, 14 Aug 2025 06:46:46 GMT, Kim Barrett <kbarrett at openjdk.org> wrote:

>> `-ftrivial-auto-var-init=pattern` [1] supported by gcc-12 was added recently to help detect uses of uninitialized memory. [2] Under this extension mode, `__builtin_clear_padding` is called to clear the padding bits in object representation. But the builtin function does not support variable length aggregates [3].
>> 
>> `vfloat2_sve_sleef` is a typedef of `svfloat32x2_t`, which is defined as an opaque sizeless type in ARM C Language Extensions (ACLE) [4]. When `__builtin_clear_padding` is applied to such a type, it triggers the unsupported code path and results in build failures when compiling libsleef on Linux-AArch64, as reported in JDK-8364185 [5].
>> 
>> Switching the initialization mode with gcc-12 from `pattern` to `zero` avoids the use of this unsupported `__builtin_clear_padding` and resolves the issue.
>> 
>> I did not encounter similar problems when building with Clang-16 [6].
>> 
>> This patch changes the initialization mode from `pattern` to `zero` when compiling libsleef SVE component on Linux-AArch64 with gcc-12 to fix the build failure and help avoid potential security issues related to uninitialized padding
>> 
>> Testing (combined with https://github.com/openjdk/jdk/pull/26577):
>> 1. Release and fastdebug build with gcc-11, gcc-12, and clang-16
>> 2. `test/jdk/jdk/incubator/vector` passed on both 256-bit SVE and NEON machines
>> 
>> [1] https://gcc.gnu.org/onlinedocs/gcc-12.1.0/gcc/Optimize-Options.html#index-ftrivial-auto-var-init
>> [2] https://github.com/openjdk/jdk/commit/fae37aaae8b36fd74309b84fa1fdf017c7d932ed
>> [3] https://gcc.gnu.org/pipermail/gcc-patches/2020-November/559214.html
>> [4] https://arm-software.github.io/acle/main/acle.html#arm_sveh
>> [5] https://bugs.openjdk.org/browse/JDK-8364185
>> [6] https://releases.llvm.org/16.0.0/tools/clang/docs/ReleaseNotes.html
>
> make/autoconf/flags-cflags.m4 line 960:
> 
>> 958:             # avoids the use of unsupported `__builtin_clear_padding` for variable
>> 959:             # length aggregates
>> 960:             if test "x$TOOLCHAIN_TYPE" = xgcc ; then
> 
> We only add `-ftrival-auto-var-init=pattern` when "x$DEBUG_LEVEL" != "xrelease", e.g. only in
> debug builds.  I think we should only add the `=zero` flag in debug builds and nothing in release
> builds here too.

I agree. Make sure to use the same test (DEBUG_LEVEL != release) to catch both fast and slow debug builds.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/26755#discussion_r2276126511