Some debuggers/profilers use this metadata to determine which function a
given instruction is in; without it they get can confused by local labels
(if you haven't stripped those). On the other hand, some tools are still
confused even with this metadata. e.g. this fixes `gdb`, but not `perf`.
Currently only implemented for ELF.
The REP_RET workaround is only needed on old AMD cpus, and the labels clutter
up the symbol table and confuse debugging/profiling tools, so use EQU to
create SHN_ABS symbols instead of creating local labels. Furthermore, skip
the workaround completely in functions that definitely won't run on such cpus.
Note that EQU is just creating a local label when using nasm instead of yasm.
This is probably a bug, but at least it doesn't break anything.
When allocating stack space with a larger alignment than the known stack
alignment a temporary register is used for storing the stack pointer.
Ensure that this isn't one of the registers used for passing arguments.
* Correctly handle FMA instructions with memory operands.
* Print a warning if FMA instructions are used without the correct cpuflag.
* Simplify the instantiation code.
* Clarify documentation.
Only the last operand in FMA3 instructions can be a memory operand. When
converting FMA4 instructions to FMA3 instructions we can utilize the fact
that multiply is a commutative operation and reorder operands if necessary
to ensure that a memory operand is used only as the last operand.
It seems to miscompile them
Should fix fate-ra-288 and fate-twinvq
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: James Almer <jamrial@gmail.com>
The function documentation explicitly mentions it needs to be a multiple of 4.
Reviewed-by: Christophe Gisquet <christophe.gisquet@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
This adds msvc optimisations as well as fixing an error in icl whereby it will generate invalid code otherwise.
Signed-off-by: Matt Oliver <protogonoi@gmail.com>
ICC versions older than atleast 12.1.6 dont have the tzcnt intrinsics.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Matt Oliver <protogonoi@gmail.com>
The .text section is already 16-byte aligned by default on all supported
platforms so `SECTION_TEXT` isn't any different from `SECTION .text`.
Signed-off-by: Anton Khirnov <anton@khirnov.net>
Change ALLOC_STACK to always align the stack before allocating stack space for
consistency. Previously alignment would occur either before or after allocating
stack space depending on whether manual alignment was required or not.
Signed-off-by: Anton Khirnov <anton@khirnov.net>
Emulation requires a temporary register if arguments 1 and 4 are the same; this
doesn't obey the semantics of the original instruction, so we can't emulate
that in x86inc.
Also add pmacsdql emulation.
Signed-off-by: Henrik Gramner <henrik@gramner.com>
Signed-off-by: Anton Khirnov <anton@khirnov.net>
Change ALLOC_STACK to always align the stack before allocating stack space for
consistency. Previously alignment would occur either before or after allocating
stack space depending on whether manual alignment was required or not.
Only two functions that use xop multiply-accumulate instructions where the
first operand is the same as the fourth actually took advantage of the macros.
This further reduces differences with x264's x86inc.
Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
The bug was fixed in 1.3.0, so only perform the workaround in earlier versions.
Reviewed-by: "Ronald S. Bultje" <rsbultje@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Recent ICC versions that define GCC as >= 4.5 (like ICC 13) apparently can't
optimize the generic C versions of av_bswap*() on their own.
Reviewed-by: Michael Niedermayer <michaelni@gmx.at>
Signed-off-by: James Almer <jamrial@gmail.com>
Improves the accuracy of measurements, especially in short sections.
To quote the Intel 64 and IA-32 Architectures Software Developer's Manual:
"The RDTSC instruction is not a serializing instruction. It does not necessarily
wait until all previous instructions have been executed before reading the counter.
Similarly, subsequent instructions may begin execution before the read operation
is performed. If software requires RDTSC to be executed only after all previous
instructions have completed locally, it can either use RDTSCP (if the processor
supports that instruction) or execute the sequence LFENCE;RDTSC."
SSE2 is a requirement for lfence so only use it on SSE2-capable systems.
Prefer lfence;rdtsc over rdtscp since rdtscp is supported on fewer systems.
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>
Silences warning(s) like:
libavcodec/x86/fft.asm:93: warning: section flags ignored on
section redeclaration
The cause of this warning is that because `struc` and `endstruc`
attempts to revert to the previous section state [1].
The section state is stored in the macro __SECT__, defined by
x86inc.asm to be `.note.GNU-stack ...`, through the `SECTION`
directive [2].
Thus, the `.note.GNU-stack` section is defined twice
(once in x86inc.asm, once during `endstruc`), causing the warning.
That is the first part of the commit: using the primitive `[section]` format
for .note.GNU-stack etc., which does not update `__SECT__` [2].
That fixes only half of the problem. Even without any `SECTION` directives,
`__SECT__` is predefined as `.text`, which conflicting with the later
`SECTION_TEXT` (which expands to `.text align=16`).
[1]: http://www.nasm.us/doc/nasmdoc6.html#section-6.4
[2]: http://www.nasm.us/doc/nasmdoc6.html#section-6.3
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>
This commit silences warning(s) like:
libavcodec/x86/fft.asm:93: warning: section flags ignored on section
redeclaration
The cause of this warning is that because `struc` and `endstruc` attempts to
revert to the previous section state [1]. The section state is stored in the
macro __SECT__, defined by x86inc.asm to be `.note.GNU-stack ...`, through the
`SECTION` directive [2]. Thus, the `.note.GNU-stack` section is defined twice
(once in x86inc.asm, once during `endstruc`), causing the warning.
That is the first part of the commit: using the primitive `[section]` format
for .note.GNU-stack etc., which does not update `__SECT__` [2].
That fixes only half of the problem. Even without any `SECTION` directives,
`__SECT__` is predefined as `.text`, which conflicting with the later
`SECTION_TEXT` (which expands to `.text align=16`).
[1]: http://www.nasm.us/doc/nasmdoc6.html#section-6.4
[2]: http://www.nasm.us/doc/nasmdoc6.html#section-6.3
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
SSE2 instructions that are XMM-implementations of pre-existing MMX/MMX2
instructions did not issue warnings when used in SSE functions. Handle
it by also checking the register type when such instructions are used.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This mimicks what is done for the other instruction sets.
Tested-by: James Almer <jamrial@gmail.com>
Tested-by: Mickaël Raulet <mraulet@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
~560 → ~500 decicycles
This is following the comments from Michael in
https://ffmpeg.org/pipermail/ffmpeg-devel/2014-August/160599.html
Using 2 registers for accumulator didn't help. On the other hand,
some re-ordering between the movs and psadbw allowed going ~538 to ~500.
Up to four instructions less depending on function and instruction set.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Only 8-bit and 10-bit idct_dc() functions are included (adding others should be trivial).
Benchmarks on an Intel Core i5-4200U:
idct8x8_dc
SSE2 MMXEXT C
cycles 22 26 57
idct16x16_dc
AVX2 SSE2 C
cycles 27 32 249
idct32x32_dc
AVX2 SSE2 C
cycles 62 126 1375
Signed-off-by: James Almer <jamrial@gmail.com>
Reviewed-by: Mickaël Raulet <mraulet@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Those macros take a byte number as shift argument, as this argument
differs between MMX and SSE2 instructions.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
It was lost during the port.
Should fix fate on 3dnowext machines.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Use the xm# and ym# aliases as they remain in sync with m# after a SWAP.
No actual changes to the assembly.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Also port relevant AVX2/XOP optimizations from x264 with permission
to relicense to LGPL from the corresponding authors
Signed-off-by: James Almer <jamrial@gmail.com>
Reviewed-by: "Ronald S. Bultje" <rsbultje@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
AV_CPU_FLAG_AVX is enabled at this point only if there's OS support.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
We need the emulation to support the cases where the first
argument is the same as the fourth. To achieve this a fifth
argument working as a temporary may be needed.
Emulation that doesn't obey the original instruction semantics
can't be in x86inc.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
vector_fmul and vector_fmac_scalar are guaranteed that they can process in
batch of 16 elements, but their SSE versions only does 8 at a time.
Therefore, unroll them a bit.
299 to 261c for 256 elements in vector_fmac_scalar on Arrandale/Win64.
Signed-off-by: Janne Grunau <janne-libav@jannau.net>
vector_fmul and vector_fmac_scalar are guaranteed that they can process in
batch of 16 elements, but their SSE versions only does 8 at a time.
Therefore, unroll them a bit.
299 to 261c for 256 elements in vector_fmac_scalar on Arrandale/Win64.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Support the cases where the first and last operand of
the XOP instruction are the same.
Also add vpmacsdql emulation.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
* commit 'c108ba0175d4fc3a3253a8b0f782fbfb96ba5098':
x86inc: Use VEX-encoded instructions in AVX functions
Merged-by: Michael Niedermayer <michaelni@gmx.at>
This is so we can sync to x264's version of FMA4 support.
This partialy reverts commit 79687079a9.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Automatically use VEX-encoding in AVX/AVX2/XOP/FMA3/FMA4
functions for all instructions that exists in a VEX-encoded
version.
This change makes it easier to extend existing code to use AVX2.
Also add support for AVX emulation of a few instructions that
were missing before.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
* commit '47f9d7ce5493e119e09d1227d017414feaaf8d97':
x86inc: Check for __OUTPUT_FORMAT__ having a value of "x64"
Merged-by: Michael Niedermayer <michaelni@gmx.at>
* commit 'bbe4a6db44f0b55b424a5cc9d3e89cd88e250450':
x86inc: Utilize the shadow space on 64-bit Windows
Merged-by: Michael Niedermayer <michaelni@gmx.at>
* commit '63f0d623100bdb0c6081456127f4b6713e83d3db':
x86inc: Use SSE instead of SSE2 for copying data
Merged-by: Michael Niedermayer <michaelni@gmx.at>
* commit 'ad76e6e7e193b98e7335156422d35467816f9ef1':
x86inc: Set ELF hidden visibility for global constants
Merged-by: Michael Niedermayer <michaelni@gmx.at>
Prevents a crash if the misaligned exception mask bit is
cleared for some reason.
Misaligned SSE functions are only used on AMD Phenom CPUs
and the benefit is miniscule. They also require modifying
the MXCSR control register and by removing those functions
we can get rid of that complexity altogether.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Store XMM6 and XMM7 in the shadow space in functions that
clobbers them. This way we don't have to adjust the stack
pointer as often, reducing the number of instructions as
well as code size.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
SWAP with >=3 named (rather than numbered) args
PERMUTE followed by SWAP with 2 named args
used to produce the wrong permutation
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
Now RET checks whether it immediately follows a branch, so the
programmer dosen't have to keep track of that condition. REP_RET
is still needed manually when it's a branch target, but that's
much rarer.
The implementation involves lots of spurious labels, but that's OK
because we strip them.
Signed-off-by: Derek Buitenhuis <derek.buitenhuis@gmail.com>
