Version bump
if's need else's. Fix previous commit
Wrap force_align_arg_pointer in HAVE_I386
Add AltiVec detection support for OpenBSD
Fix double free on failed arm_init
Add SSE detection support for OpenBSD
init: shortcut benchmarking, if there is only one runable implementation.
Add x86_64/amd64 to unaligned access whitelist
Add cris to unaligned access whitelist
remove ltdl.m4
Clean up function prototypes used for casting
bump version
Remove ' DATA' from symbol list
Add srcdir to includes for out-of-source builds
Fix cpu detection on OS/X
Add avg2_32xn_u8
Pkgsrc changes:
o Provide proper detection of altivec on NetBSD powerpc ports.
As an added bonus, this now builds on NetBSD/macppc 3.1 where
it didn't earlier.
This release merges lots of distributions patches and bug fixes
and adds new functionality used by the upcoming (now!) release
of Schroedinger-1.0.0.
pkgsrc note: someone with an arm (pun intended), or who knows how to
detect a "VFP" might want to tune the cpu feature detection code
about ICEs with gcc3.
The difference now is that SSE2+ generally works in i386, due to some
evil hacks in liboil which work around gcc's problems. With gcc4
everything is fine for me. So disable the functions suspected to cause
ICEs only for gcc<4 on i386.
This hopefully fixes PR pkg/36142 by Gilles Gravier.
changes:
- Finally integrated the SSE stack realignment wrappers and reenabled
the SSE intrinsics code.
- Migrated some new classes from Schroedinger for adding/subtracting
s16 and u8 values.
- Improved CPU detection on Solaris
- Remove XScale timestamping function, since it's not available from
user space
- Export oil_debug_print(), so the OIL_DEBUG() macros can be used
outside the library.
- Various new implementations.
changes:
- Lots of new classes and implementations were added for Schrödinger.
- Several i386 implementations also compile on amd64, so these were
copied over to a separate directory and are now enabled on amd64.
- Feature detection on amd64 has been fixed.
- All known startup warnings have been fixed.
- The core now handles arrays that are 'N plus a constant' in length.
- Wrap a bunch of functions in HAVE_UNALIGNED_ACCESS to indicate that
they require an architecture that handles unaligned access.
PR pkg/34886. Just #ifdef out the code on i386; it was disabled at
another layer anyway (due to alignment problems with any gcc on i386).
Thanks to Robert for testing.
-Make sure cpuid is read on amd64, to enable mmx/sse support. (We could
just assume it, but for the future this makes more sense.)
-bump PKGREVISION
