On Fri, 22 Mar 2024 15:30:35 GMT, Hamlin Li <m...@openjdk.org> wrote:
> > > Ah, it'll only be the redirect library that's compiled with > > > -march=armv8-a+sve Forget that. > > > > > > But that raises an interesting question. What happens if you try to load a > > library compiled with `-march=armv8-a+sve` on a non-SVE system? Is the ELF > > flagged to require SVE so it will fail to load? I'm hoping this is the case > > -- if so, everything will work as it should in this PR, but I honestly > > don't know. (After spending like 10 years working with building, I still > > discover things I need to learn...). > > I think we can handle it, when a jdk built with sve support runs on a non-sve > machine, the sve related code will not be executed with the protection of > UseSVE vm flag which is detected at runtime startup. I think we should not have specified any `-march=Xxx` option here in `make/autoconf/flags-cflags.m4` under `AC_DEFUN([FLAGS_SETUP_CFLAGS_CPU_DEP]`. It generates conflicts when a user builds JDK with relevant flags in `--with-extra-cflags=-march=Xxx`. Moreover, above `-march=armv8-a+sve` can accidentally overwrite `-mcpu=Xxx` or `-mtune=Xxx` contents if a user specified in configure cmdline, because `-march` is of higher priority for a toolchain deciding what uarch/cost-model/instructions would be used. The configure sequence can be briefly described as below: > 1. FLAGS_PRE_TOOLCHAIN -> > FLAGS_SETUP_USER_SUPPLIED_FLAGS > AC_ARG_WITH(extra-cflags, [AS_HELP_STRING([--with-extra-cflags], > [extra flags to be used when compiling jdk c-files])]) > (For example, we specify > **--with-extra-cflags="-march=armv8.2a+fp16"**) > > 2. FLAGS_SETUP_FLAGS > -> FLAGS_SETUP_CFLAGS > -> FLAGS_SETUP_CFLAGS_CPU_DEP > CFLAGS="$CFLAGS **-march=armv8-a+sve"** The later one will modify/overwrite the former one that the user is really building for. @Hamlin-Li @magicus @theRealAph ------------- PR Comment: https://git.openjdk.org/jdk/pull/18294#issuecomment-2487793664
