Commit adebad0 "arm: intreadwrite: fix inline asm constraints for gcc
4.6 and later" caused some older gcc versions to miscompile code.
This reverts to the old version of the code for these compilers.
Signed-off-by: Mans Rullgard <mans@mansr.com>
Starting with version 4.7, gcc properly supports unaligned
memory accesses on ARM. Not using the inline asm with these
compilers results in better code.
Signed-off-by: Mans Rullgard <mans@mansr.com>
With a dereferenced type-cast pointer as memory operand, gcc 4.6
and later will sometimes copy the data to a temporary location,
the address of which is used as the operand value, if it thinks
the target address might be misaligned. Using a pointer to a
packed struct type instead does the right thing.
The 16-bit case is special since the ldrh instruction addressing
modes are limited compared to ldr. The "Uq" constraint produces a
memory reference suitable for an ldrsb instruction, which supports
the same addressing modes as ldrh. However, the restrictions appear
to apply only when the operand addresses a single byte. The memory
reference must thus be split into two operands each targeting one
byte. Finally, the "Uq" constraint is only available in ARM mode.
The Thumb-2 ldrh instruction supports most addressing modes so the
normal "m" constraint can be used there.
Signed-off-by: Mans Rullgard <mans@mansr.com>
This allows masking CPU features with the -cpuflags avconv option
which is useful for testing different optimisations without rebuilding.
Signed-off-by: Mans Rullgard <mans@mansr.com>
The were broken since August of 2010 without anyone noticing until
three weeks ago. Nobody cares about it anymore and hopefully Marvell
will support NEON like in the PXA978 from now on.
With these changes, gcc 4.5 and later recognise it as a bswap
and use the proper instructions on ARM and x86. On x86, the
16-bit bswap is recognised from gcc 4.1.
Signed-off-by: Mans Rullgard <mans@mansr.com>
The volatile qualifiers are not needed on these statements as
their effects are fully specified by constraints.
Signed-off-by: Mans Rullgard <mans@mansr.com>
On some versions of gcc, these weren't always getting inlined due to hitting
the inline cap limit in some files. This is generally bad, as most of these
functions are smaller inlined than not.
This prevents gcc inserting useless UXTH instructions, at least
in some cases.
Originally committed as revision 25212 to svn://svn.ffmpeg.org/ffmpeg/trunk
Instead of defining functions in per-arch header files included
by the main cpu.c, define them normally and call them from the
generic one.
Originally committed as revision 25084 to svn://svn.ffmpeg.org/ffmpeg/trunk
This reduces the number of false dependencies on header files and
speeds up compilation.
Originally committed as revision 22407 to svn://svn.ffmpeg.org/ffmpeg/trunk
ARMv6 and later support unaligned loads and stores for single
word/halfword but not double/multiple. GCC is ignorant of this and
will always use bytewise accesses for unaligned data. Casting to an
int32_t pointer is dangerous since a load/store double or multiple
instruction might be used (this happens with some code in FFmpeg).
Implementing the AV_[RW]* macros with inline asm using only supported
instructions gives fast and safe unaligned accesses. ARM RVCT does
the right thing with generic code.
This gives an overall speedup of up to 10%.
Originally committed as revision 18601 to svn://svn.ffmpeg.org/ffmpeg/trunk
