On Wed, 18 May 2022 19:05:03 GMT, Dmitry Chuyko <dchu...@openjdk.org> wrote:
> On AArch64 it is sometimes convenient to have LSE atomics right from the
> start. Currently they are enabled after feature detection and RR reverse
> debugger works incorrectly.
>
> New build configuration feature 'hardlse' is added. If it is enabled for
> aarch64 type of build, then statically compiled stubs replace the initial
> pessimistic implementation and dynamically generated replacements (when LSE
> support is detected). The feature works for builds of all debug levels.
>
> New file atomic_linux_aarch64_lse.S is derived from atomic_linux_aarch64.S
> and inherits its copyright. This alternative static implementation
> corresponds to the dynamically generated code.
>
> Note, this configuration part is necessary but not sufficient to fully avoid
> strex instructions for practical purposes. Other parts are:
>
> * Run on the OS built without strex family instructions. E.g. Amazon Linux
> 2022.
> * Compile with outline atomics enabled and the configuration flag enabled.
> E.g. configure with
> --with-extra-cflags='-march=armv8.3-a+crc+crypto -moutline-atomics'
> --with-extra-cxxflags='-march=armv8.3-a+crc+crypto -moutline-atomics'
> --with-extra-ldflags='-Wl,--allow-multiple-definition'
> --with-jvm-features=hardlse
>
> Testing: tier1, tier2 on linux-aarch64 release builds with feature off and
> feature on.
I literally just tried this:
.text
#ifdef __ARM_FEATURE_ATOMICS
.globl aarch64_atomic_fetch_add_8_default_impl
.align 5
aarch64_atomic_fetch_add_8_default_impl:
prfm pstl1strm, [x0]
0: ldaddal x1, x2, [x0]
dmb ish
mov x0, x2
with the obvious `#else` and `#endif` around the non-LSE part
and
$ objdump -d
/home/aph/theRealAph-jdk/build/linux-aarch64-server-slowdebug/hotspot/variant-server/libjvm/objs/atomic_linux_aarch64.o
| head -40
-------------
PR: https://git.openjdk.java.net/jdk/pull/8779
