It turns out that even if you stub out the rounding and exception
support and use the ieee754 version, it's still much better than
the generic version that just uses normal multiply and add.
The resulting functions have only 1 ULP of error according to the tests.
In general we just try to include <not-cancel.h> once, thus avoiding
any issues with multiple inclusion.
However, the getaddrinfo sources are somewhat tangled, and .c files
include each other to get the final result. Each .c file currently has
its own include of <not-cancel.h>. While you could properly re-factor
this for the case of gai.c, the posix/tst-rfc3484.c tests directly include
one of the sub-files (sysdeps/posix/getaddrinfo.c), and it starts to
feel fragile.
The latest main tree has split out stackguard-macros.h so we no
longer have to override it with a stub.
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Installed headers should always use the forms of attributes with __
prefixes and suffixes, so __attribute__ ((__aligned__ (8))) instead of
the plain "aligned" form, to avoid taking identifiers from the user's
namespace.
URL: http://sourceware.org/bugzilla/show_bug.cgi?id=14109
Reported-by: Joseph Myers <joseph@codesourcery.com>
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Previously we weren't re-aligning the stack pointer during the
call to _dl_init(), so for tilegx32 and an odd value in _dl_skip_args
and kernel unaligned access fixups disabled, we would die with SIGBUS.
We now handle this case properly by aligning before calling _dl_init().
We were multiplying a byte by 0x0101010101010101ULL to create a
constant for SIMD ops, but the compiler isn't good at optimizing
this case (the fact that one operand is a byte is lost by the time
it would be possible to do the optimization). So instead we add
a helper routine that explicitly uses SIMD ops to create the constant.
Although this is not required by the definition of memcpy(),
in practice this sort of thing does happen, and it's easy to make
the code robust by doing nothing in this case. (Since structure
copy causes the compiler to emit a memcpy, in the case where the
target structure is the same as the destination, we were seeing
corruption.)
This patches fixes up the tile startup files, moving elf/start.S up a
directory level and implementing the required crti.S and crtn.S files
based on the old initfini.c compiler output (hand-optimized to bum a
couple of cycles).
