On 4/22/2014 6:41 PM, Robert O'Callahan wrote:
On B2G and Android we're compiling with -fvisibility=hidden. configure
output says:
checking For gcc visibility bug with class-level attributes (GCC bug 26905)...
yes
checking For x86_64 gcc visibility bug with builtins (GCC bug 20297)... yes
Is that a problem?
Yes! We can probably get codesize, startup perf, and runtime perf gains
of 3-8% by fixing this. There was a recent thread in dev.builds about
this, actually, which I didn't realize was related but probably is:
I found a bug in configure.in around line 2640:
ac_cv_have_visibility_builtin_bug=no
if ! ${CC-cc} ${CFLAGS} ${DSO_PIC_CFLAGS}
${DSO_LDOPTS} -O2 -S -o conftest.S conftest.c > /dev/null 2>&1 ; then
ac_cv_have_visibility_builtin_bug=yes
else
if test `grep -c "@PLT" conftest.S` = 0; then
ac_cv_visibility_builtin_bug=yes
fi
fi
The second 'if' uses a wrong variable name, so the result of 'grep' is
effectively ignored. The problem is that if I fix this bug, clang
(3.4.1) fails this test (it optimizes all code out with -O2), the
configuration variables (related to -fvisibility) are changed
accordingly, and then a simple firefox build fails. The same bug is
also in js/src/configure.in.
We should either fix or remove this configure test. I think that
removing it and always using pragmas is probably correct, but we might
need to opt mac out of that because our list of system headers on mac
isn't up to date.
--BDS
_______________________________________________
dev-platform mailing list
dev-platform@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-platform
