x/ffmpeg
1
0
Fork 0
mirror of https://git.ffmpeg.org/ffmpeg.git synced 2026-06-11 08:13:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

swscale/x86/ops: use plain ret instruction

Browse Source 
The original intent here was probably to make the ops code agnostic to
which operation is actually last in the list, but the existence of a
divergence between CONTINUE and FINISH already implies that we hard-code
the assumption that the final operation is a write op.

So we can just massively simplify this with a call/ret pair instead of
awkwardly exporting and then jumping back to the return label. This actually
collapses FINISH down into just a plain RET, since the op kernels already
don't set up any extra stack frame.

Signed-off-by: Niklas Haas <git@haasn.dev>
This commit is contained in:
Niklas Haas
2026-04-11 16:30:15 +00:00
committed by Niklas Haas
co-authored by Niklas Haas
parent f7ca6f7481
commit c29465bcb6
3 changed files with 10 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libswscale/x86/ops.c
-3
View File
@@ -1006,9 +1006,6 @@ static int compile(SwsContext *ctx, SwsOpList *ops, SwsCompiledOp *out)
#define ASSIGN_PROCESS_FUNC(NAME) \
do { \
SWS_DECL_FUNC(NAME); \
void NAME##_return(void); \
ret = ff_sws_op_chain_append(chain, NAME##_return, \
NULL, &(SwsOpPriv) {0}); \
out->func = NAME; \
} while (0)
libswscale/x86/ops_common.asm
+1 -19
View File
@@ -26,12 +26,7 @@
; function is responsible for the block loop, as well as initializing the
; plane pointers. It will jump directly into the first operation kernel,
; and each operation kernel will jump directly into the next one, with the
; final kernel jumping back into the sws_process return point. (See label
; `sws_process.return` in ops_int.asm)
;
; To handle the jump back to the return point, we append an extra address
; corresponding to the correct sws_process.return label into the SwsOpChain,
; and have the WRITE kernel jump into it as usual. (See the FINISH macro)
; final kernel returning back into the entry point.
;
; Inside an operation chain, we use a custom calling convention to preserve
; registers between kernels. The exact register allocation is found further
@@ -291,19 +286,6 @@ endstruc
CONTINUE tmp0q
%endmacro
; Final macro to end the operation chain, used by WRITE kernels to jump back
; to the process function return point. Very similar to CONTINUE, but skips
; incrementing the implq pointer, and also clears AVX registers to avoid
; phantom dependencies between loop iterations.
%macro FINISH 1 ; reg
%if vzeroupper_required
; we may jump back into an SSE read, so always zero upper regs here
vzeroupper
%endif
jmp %1
annotate_function_size
%endmacro
; Helper for inline conditionals; used to conditionally include single lines
%macro IF 2+ ; cond, body
%if %1
libswscale/x86/ops_int.asm
+9 -29
View File
@@ -93,27 +93,12 @@ IF %1 > 3, mov in3q, [execq + SwsOpExec.in3]
IF %1 > 1, mov out1q, [execq + SwsOpExec.out1]
IF %1 > 2, mov out2q, [execq + SwsOpExec.out2]
IF %1 > 3, mov out3q, [execq + SwsOpExec.out3]
jmp [rsp] ; call into op chain
; Declare a separate global label for the return point, so that we can append
; it to the list of op function pointers from the C code, effectively ensuring
; that we end up here again after the op chain finishes processing a line.
; (See also: cglobal_label in x86inc.asm)
%if FORMAT_ELF
global current_function %+ _return:function hidden
%elif FORMAT_MACHO && HAVE_PRIVATE_EXTERN
global current_function %+ _return:private_extern
%else
global current_function %+ _return
%endif
align function_align
current_function %+ _return:
; op chain always returns back here
.loop:
call [rsp] ; call into op chain
mov implq, [rsp + 8]
inc bxd
cmp bxd, [rsp + 20]
jne .continue
jne .loop
; end of line
inc yd
cmp yd, [rsp + 24]
@@ -131,7 +116,7 @@ IF %1 > 3, add out3q, [execq + SwsOpExec.out_bump3]
; conditionally apply y bump (if non-NULL)
mov tmp0q, [execq + SwsOpExec.in_bump_y]
test tmp0q, tmp0q
jz .continue
jz .loop
movsxd tmp0q, [tmp0q + yq * 4 - 4] ; load (signed) y bump
%if %1 > 3
mov tmp1q, tmp0q
@@ -150,8 +135,7 @@ IF %1 > 3, add out3q, [execq + SwsOpExec.out_bump3]
%endif
imul tmp0q, [execq + SwsOpExec.in_stride0]
add in0q, tmp0q
.continue:
jmp [rsp]
jmp .loop
.end:
add rsp, 32
RET
@@ -271,7 +255,6 @@ IF %1 > 3, add in3q, mmsize * (1 + V2)
%macro write_planar 1 ; elems
op write_planar%1
LOAD_CONT tmp0q
movu [out0q], mx
IF %1 > 1, movu [out1q], my
IF %1 > 2, movu [out2q], mz
@@ -286,7 +269,7 @@ IF %1 > 3, movu [out3q + mmsize], mw2
IF %1 > 1, add out1q, mmsize * (1 + V2)
IF %1 > 2, add out2q, mmsize * (1 + V2)
IF %1 > 3, add out3q, mmsize * (1 + V2)
FINISH tmp0q
RET
%endmacro
%macro read_packed2 1 ; depth
@@ -325,7 +308,6 @@ IF %1 < 32, VBROADCASTI128 m12, [read%1_unpack2]
%macro write_packed2 1 ; depth
op write%1_packed2
IF %1 < 32, VBROADCASTI128 m12, [write%1_pack2]
LOAD_CONT tmp0q
%if cpuflag(avx2)
vpermq mx, mx, q3120 ; { X0 X2 | X1 X3 }
vpermq my, my, q3120 ; { Y0 Y2 | Y1 Y3 }
@@ -352,7 +334,7 @@ IF %1 < 32, VBROADCASTI128 m12, [write%1_pack2]
IF V2, movu [out0q + 2*mmsize], m10
IF V2, movu [out0q + 3*mmsize], m11
add out0q, mmsize * (2 + V2 * 2)
FINISH tmp0q
RET
%endmacro
; helper macro reused for both 3 and 4 component packed reads
@@ -433,11 +415,10 @@ IF1 V2, read_packed_inner mx2, my2, mz2, mw2, in0q + %1 * mmsize, %1, %2
%macro write_packed 2 ; num, depth
op write%2_packed%1
IF %2 < 32, VBROADCASTI128 m12, [write%2_pack%1]
LOAD_CONT tmp0q
write_packed_inner mx, my, mz, mw, out0q, %1, %2
IF1 V2, write_packed_inner mx2, my2, mz2, mw2, out0q + %1 * mmsize, %1, %2
add out0q, %1 * mmsize * (1 + V2)
FINISH tmp0q
RET
%endmacro
%macro rw_packed 1 ; depth
@@ -512,9 +493,8 @@ IF V2, pshufb mx2, m8
IF V2, pmovmskb tmp1d, mx2
mov [out0q], tmp0d
IF V2, mov [out0q + (mmsize >> 3)], tmp1d
LOAD_CONT tmp0q
add out0q, (mmsize >> 3) * (1 + V2)
FINISH tmp0q
RET
%endmacro
;--------------------------