Align stack variables manually in yadif x86 code, should fix ticket25

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavfilter/x86/yadif_template.c

@@ -107,10 +107,8 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
                                   uint8_t *prev, uint8_t *cur, uint8_t *next,
                                   int w, int prefs, int mrefs, int parity, int mode)
 {
-    DECLARE_ALIGNED(16, uint8_t, tmp0[16]);
-    DECLARE_ALIGNED(16, uint8_t, tmp1[16]);
-    DECLARE_ALIGNED(16, uint8_t, tmp2[16]);
-    DECLARE_ALIGNED(16, uint8_t, tmp3[16]);
+    uint8_t tmp[5*16];
+    uint8_t *tmpA= (uint8_t*)(((uint64_t)(tmp+15)) & ~15);
     int x;
 
 #define FILTER\
@@ -124,9 +122,9 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             MOVQ"      "MM"3, "MM"4 \n\t"\
             "paddw     "MM"2, "MM"3 \n\t"\
             "psraw     $1,    "MM"3 \n\t" /* d = (prev2[x] + next2[x])>>1 */\
-            MOVQ"      "MM"0, %[tmp0] \n\t" /* c */\
-            MOVQ"      "MM"3, %[tmp1] \n\t" /* d */\
-            MOVQ"      "MM"1, %[tmp2] \n\t" /* e */\
+            MOVQ"      "MM"0, (%[tmpA]) \n\t" /* c */\
+            MOVQ"      "MM"3, 16(%[tmpA]) \n\t" /* d */\
+            MOVQ"      "MM"1, 32(%[tmpA]) \n\t" /* e */\
             "psubw     "MM"4, "MM"2 \n\t"\
             PABS(      MM"4", MM"2") /* temporal_diff0 */\
             LOAD("(%[prev],%[mrefs])", MM"3") /* prev[x-refs] */\
@@ -148,7 +146,7 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             "paddw     "MM"4, "MM"3 \n\t" /* temporal_diff2 */\
             "psrlw     $1,    "MM"3 \n\t"\
             "pmaxsw    "MM"3, "MM"2 \n\t"\
-            MOVQ"      "MM"2, %[tmp3] \n\t" /* diff */\
+            MOVQ"      "MM"2, 48(%[tmpA]) \n\t" /* diff */\
 \
             "paddw     "MM"0, "MM"1 \n\t"\
             "paddw     "MM"0, "MM"0 \n\t"\
@@ -179,7 +177,7 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             CHECK2\
 \
             /* if(p->mode<2) ... */\
-            MOVQ"    %[tmp3], "MM"6 \n\t" /* diff */\
+            MOVQ"    48(%[tmpA]), "MM"6 \n\t" /* diff */\
             "cmpl      $2, %[mode] \n\t"\
             "jge       1f \n\t"\
             LOAD("(%["prev2"],%[mrefs],2)", MM"2") /* prev2[x-2*refs] */\
@@ -190,9 +188,9 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             "paddw     "MM"5, "MM"3 \n\t"\
             "psrlw     $1,    "MM"2 \n\t" /* b */\
             "psrlw     $1,    "MM"3 \n\t" /* f */\
-            MOVQ"    %[tmp0], "MM"4 \n\t" /* c */\
-            MOVQ"    %[tmp1], "MM"5 \n\t" /* d */\
-            MOVQ"    %[tmp2], "MM"7 \n\t" /* e */\
+            MOVQ"    (%[tmpA]), "MM"4 \n\t" /* c */\
+            MOVQ"    16(%[tmpA]), "MM"5 \n\t" /* d */\
+            MOVQ"    32(%[tmpA]), "MM"7 \n\t" /* e */\
             "psubw     "MM"4, "MM"2 \n\t" /* b-c */\
             "psubw     "MM"7, "MM"3 \n\t" /* f-e */\
             MOVQ"      "MM"5, "MM"0 \n\t"\
@@ -211,7 +209,7 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             "pmaxsw    "MM"4, "MM"6 \n\t" /* diff= MAX3(diff, min, -max); */\
             "1: \n\t"\
 \
-            MOVQ"    %[tmp1], "MM"2 \n\t" /* d */\
+            MOVQ"    16(%[tmpA]), "MM"2 \n\t" /* d */\
             MOVQ"      "MM"2, "MM"3 \n\t"\
             "psubw     "MM"6, "MM"2 \n\t" /* d-diff */\
             "paddw     "MM"6, "MM"3 \n\t" /* d+diff */\
@@ -219,11 +217,9 @@ void RENAME(ff_yadif_filter_line)(uint8_t *dst,
             "pminsw    "MM"3, "MM"1 \n\t" /* d = clip(spatial_pred, d-diff, d+diff); */\
             "packuswb  "MM"1, "MM"1 \n\t"\
 \
-            :[tmp0]"=m"(tmp0),\
-             [tmp1]"=m"(tmp1),\
-             [tmp2]"=m"(tmp2),\
-             [tmp3]"=m"(tmp3)\
-            :[prev] "r"(prev),\
+            :\
+            :[tmpA] "r"(tmpA),\
+             [prev] "r"(prev),\
              [cur]  "r"(cur),\
              [next] "r"(next),\
              [prefs]"r"((x86_reg)prefs),\