Instead of loading the registers one lane at a time, load full
registers and then transpose them.
This is faster, reducing the runtime for the function from about
506 cycles to 434 cycles (tested on a Cortex A8).
This also avoids an issue which seems like a cpu bug, present
on Sony Xperia T (cpu implementer 0x51 architecture 7 variant 0x1
part 0x04d). On such a device, it seemed like the "vswp q9, q10"
could start executing before the previous
vld4.u8 {d20[x],d21[x],d22[x],d23[x]}, [r3], r1
had finished and written back their result. Changing the
"vswp q9, q10" into "vswp q10, q9", or into separate
"vswp d18, d20; vswp d19, d21" (or the other way around) seemed to
avoid the issue. This happened occasionally (a couple times per
100000 invocations or so).
This fixes building with gnu binutils, which don't support this
nonstandard form of the instructions.
Once Apple's tools support the proper standard form of the
instructions, the code should be updated to use that everywhere
instead, and these macros should be removed.
This makes sure the windows version of these functions behave
more like the posix version. The posix *snprintf returns how
much would have been written if the buffer had been large
enough, which we don't know easily in the windows versions.
This basically means that we can assume that the return value is
>= 0 now, which can simplify the calling code.
Windows Phone lacks the old CreateThread/beginthreadex APIs for
creating threads. (Technically, the functions still do exist,
but they aren't officially supported and aren't visible in the
headers when targeting Windows Phone.)
Building code that uses the Windows Runtime language extensions
requires building with the -ZW option.
According to the Win64 ABI, these registers need to be preserved,
and compilers are allowed to rely on their content to stay
available - not only for float usage but for any usage, anywhere,
in the calling C++ code.
This adds a macro which pushes the clobbered registers onto the
stack if targeting win64 (and a matching one which restores them).
The parameter to the macro is the number of xmm registers used
(e.g. if using xmm0 - xmm7, the parameter is 8), or in other
words, the number of the highest xmm register used plus one.
This is similar to how the same issue is handled for the NEON
registers q4-q7 with the vpush instruction, except that they needed
to be preserved on all platforms, not only on one particular platform.
This allows removing the XMMREG_PROTECT_* hacks, which can
easily fail if the compiler chooses to use the callee saved
xmm registers in an unexpected spot.
This is what nasm ended up assembling movsx with 32 bit input to
anyway.
Keep using plain movsx for 16 bit input.
This fixes building with yasm in 64 bit mode.
