x86inc: automatically insert vzeroupper for YMM functions.

libavcodec/x86/dsputil_yasm.asm

@@ -1158,12 +1158,7 @@ ALIGN 16
     add     src1q, 2*mmsize
     sub     lenq,  2*mmsize
     jge     .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse
@@ -1193,12 +1188,7 @@ ALIGN 16
 
     sub     lenq,   2*mmsize
     jge     .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse
@@ -1243,10 +1233,6 @@ cglobal butterflies_float_interleave, 4,4,3, dst, src0, src1, len
 %endif
     add       lenq, mmsize
     jl .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%endif
 .end:
     REP_RET
 %endmacro

libavcodec/x86/fft_mmx.asm

@@ -749,9 +749,6 @@ section .text
 ; The others pass args in registers and don't spill anything.
 cglobal fft_dispatch%2, 2,5,8, z, nbits
     FFT_DISPATCH fullsuffix, nbits
-%if mmsize == 32
-    vzeroupper
-%endif
     RET
 %endmacro ; DECL_FFT
 
@@ -956,9 +953,6 @@ cglobal imdct_half, 3,12,8; FFTContext *s, FFTSample *output, const FFTSample *i
     %1 r0, r1, r6, rtcos, rtsin
 %if ARCH_X86_64 == 0
     add esp, 12
-%endif
-%if mmsize == 32
-    vzeroupper
 %endif
     RET
 %endmacro

libavresample/x86/audio_convert.asm

@@ -145,12 +145,7 @@ cglobal conv_s32_to_flt, 3,3,3, dst, src, len
     mova  [dstq+lenq+mmsize], m2
     add     lenq, mmsize*2
     jl .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse2
@@ -218,12 +213,7 @@ cglobal conv_flt_to_s32, 3,3,5, dst, src, len
     mova  [dstq+lenq+3*mmsize], m3
     add     lenq, mmsize*4
     jl .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse2

libavresample/x86/audio_mix.asm

@@ -51,12 +51,7 @@ cglobal mix_2_to_1_fltp_flt, 3,4,6, src, matrix, len, src1
     add        srcq, mmsize*2
     sub        lend, mmsize*2/4
     jg .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse
@@ -175,12 +170,7 @@ cglobal mix_1_to_2_fltp_flt, 3,5,4, src0, matrix0, len, src1, matrix1
     add       src0q, mmsize
     sub        lend, mmsize/4
     jg .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse

libavutil/x86/float_dsp.asm

@@ -40,12 +40,7 @@ ALIGN 16
 
     sub       lenq, 2*mmsize
     jge       .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse
@@ -86,12 +81,7 @@ cglobal vector_fmac_scalar, 4,4,3, dst, src, mul, len
     mova  [dstq+lenq+mmsize], m2
     sub    lenq, 2*mmsize
     jge .loop
-%if mmsize == 32
-    vzeroupper
-    RET
-%else
     REP_RET
-%endif
 %endmacro
 
 INIT_XMM sse

libavutil/x86/x86inc.asm

@@ -369,11 +369,14 @@ DECLARE_REG 14, R15, R15D, R15W, R15B, 120
 %macro RET 0
     WIN64_RESTORE_XMM_INTERNAL rsp
     POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+%if mmsize == 32
+    vzeroupper
+%endif
     ret
 %endmacro
 
 %macro REP_RET 0
-    %if regs_used > 7 || xmm_regs_used > 6
+    %if regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
         RET
     %else
         rep ret
@@ -410,11 +413,14 @@ DECLARE_REG 14, R15, R15D, R15W, R15B, 72
 
 %macro RET 0
     POP_IF_USED 14, 13, 12, 11, 10, 9
+%if mmsize == 32
+    vzeroupper
+%endif
     ret
 %endmacro
 
 %macro REP_RET 0
-    %if regs_used > 9
+    %if regs_used > 9 || mmsize == 32
         RET
     %else
         rep ret
@@ -456,11 +462,14 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 
 %macro RET 0
     POP_IF_USED 6, 5, 4, 3
+%if mmsize == 32
+    vzeroupper
+%endif
     ret
 %endmacro
 
 %macro REP_RET 0
-    %if regs_used > 3
+    %if regs_used > 3 || mmsize == 32
         RET
     %else
         rep ret