* [FFmpeg-devel] [PATCH] Mark C globals with small code model
@ 2025-02-25 21:37 Pranav Kant via ffmpeg-devel
2025-02-25 23:03 ` Andreas Rheinhardt
2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
0 siblings, 2 replies; 5+ messages in thread
From: Pranav Kant via ffmpeg-devel @ 2025-02-25 21:37 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Pranav Kant
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
+
/**
* Disable warnings about deprecated features
* This is useful for sections of code kept for backward compatibility and
--
2.48.1.658.g4767266eb4-goog
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [FFmpeg-devel] [PATCH] Mark C globals with small code model
2025-02-25 21:37 [FFmpeg-devel] [PATCH] Mark C globals with small code model Pranav Kant via ffmpeg-devel
@ 2025-02-25 23:03 ` Andreas Rheinhardt
2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
1 sibling, 0 replies; 5+ messages in thread
From: Andreas Rheinhardt @ 2025-02-25 23:03 UTC (permalink / raw)
To: ffmpeg-devel
Pranav Kant via ffmpeg-devel:
> 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
This should be in attributes_internal.h. Or maybe we should add new
macros in mem_internal.h for the case of variables with external linkage
and static storage duration (one macro for the const case, one for the
non-const case) that would automatically include your proposed attribute?
> +
> /**
> * Disable warnings about deprecated features
> * This is useful for sections of code kept for backward compatibility and
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 5+ messages in thread* [FFmpeg-devel] [PATCH] Mark C globals with small code model
2025-02-25 21:37 [FFmpeg-devel] [PATCH] Mark C globals with small code model Pranav Kant via ffmpeg-devel
2025-02-25 23:03 ` Andreas Rheinhardt
@ 2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
2025-02-26 19:45 ` Pranav Kant via ffmpeg-devel
2025-02-27 3:36 ` James Zern via ffmpeg-devel
1 sibling, 2 replies; 5+ messages in thread
From: Pranav Kant via ffmpeg-devel @ 2025-02-26 19:44 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Pranav Kant
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
Signed-off-by: Pranav Kant <prka@google.com>
---
libavcodec/ac3dsp.c | 2 ++
libavcodec/cabac.c | 2 ++
libavcodec/x86/constants.c | 8 ++++++++
libavutil/attributes_internal.h | 15 +++++++++++++++
4 files changed, 27 insertions(+)
diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
index 730fa70fff..d16b6c24c3 100644
--- a/libavcodec/ac3dsp.c
+++ b/libavcodec/ac3dsp.c
@@ -25,6 +25,7 @@
#include "config.h"
#include "libavutil/attributes.h"
+#include "libavutil/attributes_internal.h"
#include "libavutil/common.h"
#include "libavutil/intmath.h"
#include "libavutil/mem_internal.h"
@@ -104,6 +105,7 @@ static void ac3_update_bap_counts_c(uint16_t mant_cnt[16], uint8_t *bap,
mant_cnt[bap[len]]++;
}
+attribute_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..b8c6db29a2 100644
--- a/libavcodec/cabac.c
+++ b/libavcodec/cabac.c
@@ -24,11 +24,13 @@
* Context Adaptive Binary Arithmetic Coder.
*/
+#include "libavutil/attributes_internal.h"
#include "libavutil/error.h"
#include "libavutil/mem_internal.h"
#include "cabac.h"
+attribute_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,
diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
index bc7f2b17b8..347b7dd1d3 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_internal.h"
#include "libavutil/mem_internal.h"
#include "libavutil/x86/asm.h" // for xmm_reg
#include "constants.h"
+attribute_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 };
+attribute_mcmodel_small
DECLARE_ASM_ALIGNED(32, const ymm_reg, ff_pw_4) = { 0x0004000400040004ULL, 0x0004000400040004ULL,
0x0004000400040004ULL, 0x0004000400040004ULL };
+attribute_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 };
+attribute_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 };
+attribute_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 };
+attribute_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 };
+attribute_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_internal.h b/libavutil/attributes_internal.h
index 3df1ee6af3..0c40e6461f 100644
--- a/libavutil/attributes_internal.h
+++ b/libavutil/attributes_internal.h
@@ -31,4 +31,19 @@
# define FF_VISIBILITY_POP_HIDDEN
#endif
+/**
+ * Some globals defined in C files are used from hardcoded asm that assumes small
+ * code model (that is, accessing these globals without GOT). This is a problem
+ * when FFMpeg is built with medium code model (-mcmodel=medium) which allocates
+ * all globals in a data section that's unreachable with PC relative instructions
+ * (small code model instruction sequence). We mark all such globals with this
+ * attribute_mcmodel_small to ensure assembly accessible globals continue to be
+ * allocated in sections reachable from PC relative instructions.
+ */
+#if ARCH_X86_64 && defined(__ELF__) && __has_attribute(model)
+# define attribute_mcmodel_small __attribute__(model("small"))
+#else
+# define attribute_mcmodel_small
+#endif
+
#endif /* AVUTIL_ATTRIBUTES_INTERNAL_H */
--
2.48.1.658.g4767266eb4-goog
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [FFmpeg-devel] [PATCH] Mark C globals with small code model
2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
@ 2025-02-26 19:45 ` Pranav Kant via ffmpeg-devel
2025-02-27 3:36 ` James Zern via ffmpeg-devel
1 sibling, 0 replies; 5+ messages in thread
From: Pranav Kant via ffmpeg-devel @ 2025-02-26 19:45 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Pranav Kant
I added it to attributes_internal.h. The existing attribute in
attributes_internal.h (attribute_visibility_hidden) is also being used with
DECLARE_ALIGNED macros (see libavcodec/sbrdsp_template.c). My new macro is
similar in nature.
On Wed, Feb 26, 2025 at 11:44 AM Pranav Kant <prka@google.com> 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
>
> Signed-off-by: Pranav Kant <prka@google.com>
> ---
> libavcodec/ac3dsp.c | 2 ++
> libavcodec/cabac.c | 2 ++
> libavcodec/x86/constants.c | 8 ++++++++
> libavutil/attributes_internal.h | 15 +++++++++++++++
> 4 files changed, 27 insertions(+)
>
> diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
> index 730fa70fff..d16b6c24c3 100644
> --- a/libavcodec/ac3dsp.c
> +++ b/libavcodec/ac3dsp.c
> @@ -25,6 +25,7 @@
>
> #include "config.h"
> #include "libavutil/attributes.h"
> +#include "libavutil/attributes_internal.h"
> #include "libavutil/common.h"
> #include "libavutil/intmath.h"
> #include "libavutil/mem_internal.h"
> @@ -104,6 +105,7 @@ static void ac3_update_bap_counts_c(uint16_t
> mant_cnt[16], uint8_t *bap,
> mant_cnt[bap[len]]++;
> }
>
> +attribute_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..b8c6db29a2 100644
> --- a/libavcodec/cabac.c
> +++ b/libavcodec/cabac.c
> @@ -24,11 +24,13 @@
> * Context Adaptive Binary Arithmetic Coder.
> */
>
> +#include "libavutil/attributes_internal.h"
> #include "libavutil/error.h"
> #include "libavutil/mem_internal.h"
>
> #include "cabac.h"
>
> +attribute_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,
> diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
> index bc7f2b17b8..347b7dd1d3 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_internal.h"
> #include "libavutil/mem_internal.h"
> #include "libavutil/x86/asm.h" // for xmm_reg
> #include "constants.h"
>
> +attribute_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 };
> +attribute_mcmodel_small
> DECLARE_ASM_ALIGNED(32, const ymm_reg, ff_pw_4) = {
> 0x0004000400040004ULL, 0x0004000400040004ULL,
>
> 0x0004000400040004ULL, 0x0004000400040004ULL };
> +attribute_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 };
> +attribute_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 };
> +attribute_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 };
> +attribute_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 };
> +attribute_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_internal.h
> b/libavutil/attributes_internal.h
> index 3df1ee6af3..0c40e6461f 100644
> --- a/libavutil/attributes_internal.h
> +++ b/libavutil/attributes_internal.h
> @@ -31,4 +31,19 @@
> # define FF_VISIBILITY_POP_HIDDEN
> #endif
>
> +/**
> + * Some globals defined in C files are used from hardcoded asm that
> assumes small
> + * code model (that is, accessing these globals without GOT). This is a
> problem
> + * when FFMpeg is built with medium code model (-mcmodel=medium) which
> allocates
> + * all globals in a data section that's unreachable with PC relative
> instructions
> + * (small code model instruction sequence). We mark all such globals with
> this
> + * attribute_mcmodel_small to ensure assembly accessible globals continue
> to be
> + * allocated in sections reachable from PC relative instructions.
> + */
> +#if ARCH_X86_64 && defined(__ELF__) && __has_attribute(model)
> +# define attribute_mcmodel_small __attribute__(model("small"))
> +#else
> +# define attribute_mcmodel_small
> +#endif
> +
> #endif /* AVUTIL_ATTRIBUTES_INTERNAL_H */
> --
> 2.48.1.658.g4767266eb4-goog
>
>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 5+ messages in thread* Re: [FFmpeg-devel] [PATCH] Mark C globals with small code model
2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
2025-02-26 19:45 ` Pranav Kant via ffmpeg-devel
@ 2025-02-27 3:36 ` James Zern via ffmpeg-devel
1 sibling, 0 replies; 5+ messages in thread
From: James Zern via ffmpeg-devel @ 2025-02-27 3:36 UTC (permalink / raw)
To: FFmpeg development discussions and patches; +Cc: James Zern
On Wed, Feb 26, 2025 at 11:44 AM Pranav Kant via ffmpeg-devel
<ffmpeg-devel@ffmpeg.org> wrote:
>
> [...]
> --- a/libavutil/attributes_internal.h
> +++ b/libavutil/attributes_internal.h
> @@ -31,4 +31,19 @@
> # define FF_VISIBILITY_POP_HIDDEN
> #endif
>
> +/**
> + * Some globals defined in C files are used from hardcoded asm that assumes small
> + * code model (that is, accessing these globals without GOT). This is a problem
> + * when FFMpeg is built with medium code model (-mcmodel=medium) which allocates
> + * all globals in a data section that's unreachable with PC relative instructions
> + * (small code model instruction sequence). We mark all such globals with this
> + * attribute_mcmodel_small to ensure assembly accessible globals continue to be
> + * allocated in sections reachable from PC relative instructions.
> + */
> +#if ARCH_X86_64 && defined(__ELF__) && __has_attribute(model)
You should check `defined(__has_attribute)` before using it [1], the
preprocessor won't short circuit. See also __has_feature in
libavutil/aarch64/asm.S.
> +# define attribute_mcmodel_small __attribute__(model("small"))
> +#else
> +# define attribute_mcmodel_small
> +#endif
> +
> #endif /* AVUTIL_ATTRIBUTES_INTERNAL_H */
> --
[1]: https://clang.llvm.org/docs/LanguageExtensions.html#has-attribute
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2025-02-27 3:37 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2025-02-25 21:37 [FFmpeg-devel] [PATCH] Mark C globals with small code model Pranav Kant via ffmpeg-devel
2025-02-25 23:03 ` Andreas Rheinhardt
2025-02-26 19:44 ` Pranav Kant via ffmpeg-devel
2025-02-26 19:45 ` Pranav Kant via ffmpeg-devel
2025-02-27 3:36 ` James Zern via ffmpeg-devel
Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
This inbox may be cloned and mirrored by anyone:
git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git
# If you have public-inbox 1.1+ installed, you may
# initialize and index your mirror using the following commands:
public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
ffmpegdev@gitmailbox.com
public-inbox-index ffmpegdev
Example config snippet for mirrors.
AGPL code for this site: git clone https://public-inbox.org/public-inbox.git