[FFmpeg-devel] [PATCH] Mark C globals with small code model
Pranav Kant
prka at google.com
Fri Mar 7 04:14:05 EET 2025
I think you were looking at an older version of the patch. Newer version
didn't have this. Anyhow, there's a new version I uploaded (v3).
On Thu, Feb 27, 2025 at 6:31 PM Lynne <dev at lynne.ee> wrote:
> On 25/02/2025 22:37, Pranav Kant via ffmpeg-devel wrote:
> > By default, all globals in C/C++ compiled by clang are allocated
> > in non-large data sections. See [1] for background on code models.
> > For PIC (Position independent code), this is fine as long as binary is
> > small but as binary size increases, users maybe want to use medium/large
> > code models (-mcmodel=medium) which moves data in to large sections.
> > As data in these large sections cannot be accessed using PIC code
> > anymore (as it may be too far away), compiler ends up using a different
> > instruction sequence when building C/C++ code -- using GOT to access
> > these globals (which can be relaxed by linker at link time if binary
> > ends up being smaller). However, assembly files continue to access these
> > globals defined in C/C++ files using older (and invalid instruction
> > sequence). So, we mark all such globals with an attribute that forces
> > them to be allocated in small sections allowing them to validly be
> > accessed from the assembly code.
> >
> > This patch should not have any affect on builds that use small code
> > model, which is the default mode.
> >
> > [1]
> https://eli.thegreenplace.net/2012/01/03/understanding-the-x64-code-models
> > ---
> > libavcodec/ac3dsp.c | 2 +-
> > libavcodec/cabac.c | 3 ++-
> > libavcodec/x86/constants.c | 8 ++++++++
> > libavutil/attributes.h | 6 ++++++
> > 4 files changed, 17 insertions(+), 2 deletions(-)
> >
> > diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
> > index 730fa70fff..43b4fcbda9 100644
> > --- a/libavcodec/ac3dsp.c
> > +++ b/libavcodec/ac3dsp.c
> > @@ -104,7 +104,7 @@ static void ac3_update_bap_counts_c(uint16_t
> mant_cnt[16], uint8_t *bap,
> > mant_cnt[bap[len]]++;
> > }
> >
> > -DECLARE_ALIGNED(16, const uint16_t, ff_ac3_bap_bits)[16] = {
> > +av_mcmodel_small DECLARE_ALIGNED(16, const uint16_t,
> ff_ac3_bap_bits)[16] = {
> > 0, 0, 0, 3, 0, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16
> > };
> >
> > diff --git a/libavcodec/cabac.c b/libavcodec/cabac.c
> > index 7d41cd2ae6..dfc3ba135a 100644
> > --- a/libavcodec/cabac.c
> > +++ b/libavcodec/cabac.c
> > @@ -24,12 +24,13 @@
> > * Context Adaptive Binary Arithmetic Coder.
> > */
> >
> > +#include "libavutil/attributes.h"
> > #include "libavutil/error.h"
> > #include "libavutil/mem_internal.h"
> >
> > #include "cabac.h"
> >
> > -DECLARE_ASM_ALIGNED(1, const uint8_t, ff_h264_cabac_tables)[512 +
> 4*2*64 + 4*64 + 63] = {
> > +av_mcmodel_small DECLARE_ASM_ALIGNED(1, const uint8_t,
> ff_h264_cabac_tables)[512 + 4*2*64 + 4*64 + 63] = {
> > 9,8,7,7,6,6,6,6,5,5,5,5,5,5,5,5,
> > 4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,
> > 3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,
> > diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
> > index bc7f2b17b8..9a5af2871c 100644
> > --- a/libavcodec/x86/constants.c
> > +++ b/libavcodec/x86/constants.c
> > @@ -18,17 +18,21 @@
> > * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
> 02110-1301 USA
> > */
> >
> > +#include "libavutil/attributes.h"
> > #include "libavutil/mem_internal.h"
> > #include "libavutil/x86/asm.h" // for xmm_reg
> > #include "constants.h"
> >
> > +av_mcmodel_small
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pw_1) = {
> 0x0001000100010001ULL, 0x0001000100010001ULL,
> >
> 0x0001000100010001ULL, 0x0001000100010001ULL };
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pw_2) = {
> 0x0002000200020002ULL, 0x0002000200020002ULL,
> >
> 0x0002000200020002ULL, 0x0002000200020002ULL };
> > DECLARE_ASM_ALIGNED(16, const xmm_reg, ff_pw_3) = {
> 0x0003000300030003ULL, 0x0003000300030003ULL };
> > +av_mcmodel_small
> > DECLARE_ASM_ALIGNED(32, const ymm_reg, ff_pw_4) = {
> 0x0004000400040004ULL, 0x0004000400040004ULL,
> >
> 0x0004000400040004ULL, 0x0004000400040004ULL };
> > +av_mcmodel_small
> > DECLARE_ASM_ALIGNED(16, const xmm_reg, ff_pw_5) = {
> 0x0005000500050005ULL, 0x0005000500050005ULL };
> > DECLARE_ALIGNED(16, const xmm_reg, ff_pw_8) = {
> 0x0008000800080008ULL, 0x0008000800080008ULL };
> > DECLARE_ASM_ALIGNED(16, const xmm_reg, ff_pw_9) = {
> 0x0009000900090009ULL, 0x0009000900090009ULL };
> > @@ -49,6 +53,7 @@ DECLARE_ALIGNED(32, const ymm_reg, ff_pw_256) = {
> 0x0100010001000100ULL, 0x010
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pw_512) = {
> 0x0200020002000200ULL, 0x0200020002000200ULL,
> >
> 0x0200020002000200ULL, 0x0200020002000200ULL };
> > DECLARE_ALIGNED(16, const xmm_reg, ff_pw_1019) = {
> 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
> > +av_mcmodel_small
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pw_1023) = {
> 0x03ff03ff03ff03ffULL, 0x03ff03ff03ff03ffULL,
> >
> 0x03ff03ff03ff03ffULL, 0x03ff03ff03ff03ffULL};
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pw_1024) = {
> 0x0400040004000400ULL, 0x0400040004000400ULL,
> > @@ -66,13 +71,16 @@ DECLARE_ALIGNED(32, const ymm_reg, ff_pw_m1) = {
> 0xFFFFFFFFFFFFFFFFULL, 0xFFF
> >
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_0) = {
> 0x0000000000000000ULL, 0x0000000000000000ULL,
> >
> 0x0000000000000000ULL, 0x0000000000000000ULL };
> > +av_mcmodel_small
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_1) = {
> 0x0101010101010101ULL, 0x0101010101010101ULL,
> >
> 0x0101010101010101ULL, 0x0101010101010101ULL };
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_2) = {
> 0x0202020202020202ULL, 0x0202020202020202ULL,
> >
> 0x0202020202020202ULL, 0x0202020202020202ULL };
> > +av_mcmodel_small
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_3) = {
> 0x0303030303030303ULL, 0x0303030303030303ULL,
> >
> 0x0303030303030303ULL, 0x0303030303030303ULL };
> > DECLARE_ALIGNED(32, const xmm_reg, ff_pb_15) = {
> 0x0F0F0F0F0F0F0F0FULL, 0x0F0F0F0F0F0F0F0FULL };
> > +av_mcmodel_small
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_80) = {
> 0x8080808080808080ULL, 0x8080808080808080ULL,
> >
> 0x8080808080808080ULL, 0x8080808080808080ULL };
> > DECLARE_ALIGNED(32, const ymm_reg, ff_pb_FE) = {
> 0xFEFEFEFEFEFEFEFEULL, 0xFEFEFEFEFEFEFEFEULL,
> > diff --git a/libavutil/attributes.h b/libavutil/attributes.h
> > index 04c615c952..704a7070db 100644
> > --- a/libavutil/attributes.h
> > +++ b/libavutil/attributes.h
> > @@ -104,6 +104,12 @@
> > # define attribute_deprecated
> > #endif
> >
> > +#if defined(__clang__) && __has_attribute(model)
> > +# define av_mcmodel_small __attribute__((model("small")))
> > +#else
> > +# define av_mcmodel_small
> > +#endif
>
> That's a bad macro name, moreover being a compiler-specific name.
> Could you consider calling it simply av_global?
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
>
More information about the ffmpeg-devel
mailing list