Minor changes that could help the compiler to produce a
better (smaller/faster) code.
Signed-off-by: Frédéric Recoules <frederic.recoules@orange.fr>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Newer versions of GCC (> 5.0) accept that ebx is declared
in the clobber list even in PIC mode. Meanwhile, even
unlikely, the compiler may use ebx as base address of
one of the memory entry, making subsequent access to
them unreliable ('size', 'dst_step', 'src_step', 'sum_step').
Adding ebx in the clobber solves the problem.
By the way, the entry 'old_ebx' is no longer required.
Signed-off-by: Frédéric Recoules <frederic.recoules@orange.fr>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
- add mm0 in the clobber list if the compiler
is aware of the mmx technology;
- otherwise, add the mmx aliased x87 floating point
registers in the clobbers;
- the configure now checks if the compiler is aware of
the MMX technology.
The compiler assumes none of the mmx or x87 registers are used
by the function. If it chooses to store some data in them, they
will be overwritten by the chunk.
Recall that any mmx instruction invalidate the whole set of
x87 floating point registers.
Note: currently does not impact the binary output.
Signed-off-by: Frédéric Recoules <frederic.recoules@orange.fr>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
- move 'size' and 'old_ebx' in the output section
since they are clobbered;
- add the "memory" clobber since input pointers
are accessed;
- (minor) add the "cc" clobber since flags are
clobbered.
The compiler is missing some dataflow information
about the execution of the assembly chunks.
For instance, it assumes that 'size' remains unchanged
and that no input pointer is accessed.
The compiler optimizer may take advantage of these
assumption and produce a wrong code.
Note: currently produces the same binary output.
Signed-off-by: Frédéric Recoules <frederic.recoules@orange.fr>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
It eases the refactoring of assembly chunk since we can now
add/remove/move entries without worrying about maintaining
the token numbering in the template.
Note: does not impact the binary output.
Signed-off-by: Frédéric Recoules <frederic.recoules@orange.fr>
Signed-off-by: Takashi Iwai <tiwai@suse.de>
The i386 and x86-64 dmix may cause segfaults when 3-bytes formats are used
due to btsl asm code, which may overcome the buffer end-boundary.
The patch changes btsl to btsw so that it doesn't happen.
ALSA bug#3341:
https://bugtrack.alsa-project.org/alsa-bug/view.php?id=3341
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Add to the dmix plugin support for the S24_3LE sample format which is
used by 24-bit USB devices.
The optimized assembler version uses only 23 bits for sample data so
that the lowest bit can be used for synchronization because there is no
24-bit cmpxchg instruction.
