h264: avoid using uninitialized memory in NEON chroma mc

Adapt commit 982b596e for the arm and
aarch64 NEON asm. 5-10% faster on Cortex-A9.

libavcodec/aarch64/h264cmc_neon.S

@@ -95,9 +95,10 @@ function ff_\type\()_\codec\()_chroma_mc8_neon, export=1
         b.gt            1b
         ret
 
-2:      tst             w6,  w6
-        add             w12, w12, w6
+2:      adds            w12, w12, w6
         dup             v0.8B, w4
+        b.eq            5f
+        tst             w6,  w6
         dup             v1.8B, w12
         b.eq            4f
 
@@ -161,6 +162,33 @@ function ff_\type\()_\codec\()_chroma_mc8_neon, export=1
         st1             {v17.8B}, [x0], x2
         b.gt            4b
         ret
+
+5:      ld1             {v4.8B}, [x1], x2
+        ld1             {v5.8B}, [x1], x2
+        prfm            pldl1strm, [x1]
+        subs            w3,  w3,  #2
+        umull           v16.8H, v4.8B, v0.8B
+        umull           v17.8H, v5.8B, v0.8B
+        prfm            pldl1strm, [x1, x2]
+  .ifc \codec,h264
+        rshrn           v16.8B, v16.8H, #6
+        rshrn           v17.8B, v17.8H, #6
+  .else
+        add             v16.8H, v16.8H, v22.8H
+        add             v17.8H, v17.8H, v22.8H
+        shrn            v16.8B, v16.8H, #6
+        shrn            v17.8B, v17.8H, #6
+  .endif
+  .ifc \type,avg
+        ld1             {v20.8B}, [x8], x2
+        ld1             {v21.8B}, [x8], x2
+        urhadd          v16.8B, v16.8B, v20.8B
+        urhadd          v17.8B, v17.8B, v21.8B
+  .endif
+        st1             {v16.8B}, [x0], x2
+        st1             {v17.8B}, [x0], x2
+        b.gt            5b
+        ret
 endfunc
 .endm
 
@@ -238,9 +266,10 @@ function ff_\type\()_\codec\()_chroma_mc4_neon, export=1
         b.gt            1b
         ret
 
-2:      tst             w6,  w6
-        add             w12, w12, w6
+2:      adds            w12, w12, w6
         dup             v30.8B, w4
+        b.eq            5f
+        tst             w6,  w6
         dup             v31.8B, w12
         trn1            v0.2S,  v30.2S, v31.2S
         trn2            v1.2S,  v30.2S, v31.2S
@@ -303,6 +332,28 @@ function ff_\type\()_\codec\()_chroma_mc4_neon, export=1
         st1             {v16.S}[1], [x0], x2
         b.gt            4b
         ret
+
+5:      ld1             {v4.S}[0], [x1], x2
+        ld1             {v4.S}[1], [x1], x2
+        umull           v18.8H, v4.8B,  v30.8B
+        subs            w3,  w3,  #2
+        prfm            pldl1strm, [x1]
+  .ifc \codec,h264
+        rshrn           v16.8B, v18.8H, #6
+  .else
+        add             v18.8H, v18.8H, v22.8H
+        shrn            v16.8B, v18.8H, #6
+  .endif
+  .ifc \type,avg
+        ld1             {v20.S}[0], [x8], x2
+        ld1             {v20.S}[1], [x8], x2
+        urhadd          v16.8B, v16.8B, v20.8B
+  .endif
+        prfm            pldl1strm, [x1]
+        st1             {v16.S}[0], [x0], x2
+        st1             {v16.S}[1], [x0], x2
+        b.gt            5b
+        ret
 endfunc
 .endm
 

libavcodec/arm/h264cmc_neon.S

@@ -96,9 +96,10 @@ T       cmp             r7,  #0
 
         pop             {r4-r7, pc}
 
-2:      tst             r6,  r6
-        add             r12, r12, r6
+2:      adds            r12, r12, r6
         vdup.8          d0,  r4
+        beq             5f
+        tst             r6,  r6
         vdup.8          d1,  r12
 
         beq             4f
@@ -163,6 +164,33 @@ T       cmp             r7,  #0
         vst1.8          {d17}, [r0,:64], r2
         bgt             4b
 
+        pop             {r4-r7, pc}
+
+5:      vld1.8          {d4}, [r1], r2
+        vld1.8          {d5}, [r1], r2
+        pld             [r1]
+        subs            r3,  r3,  #2
+        vmull.u8        q8,  d4,  d0
+        vmull.u8        q9,  d5,  d0
+        pld             [r1, r2]
+  .ifc \codec,h264
+        vrshrn.u16      d16, q8,  #6
+        vrshrn.u16      d17, q9,  #6
+  .else
+        vadd.u16        q8,  q8,  q11
+        vadd.u16        q9,  q9,  q11
+        vshrn.u16       d16, q8,  #6
+        vshrn.u16       d17, q9,  #6
+  .endif
+  .ifc \type,avg
+        vld1.8          {d20}, [lr,:64], r2
+        vld1.8          {d21}, [lr,:64], r2
+        vrhadd.u8       q8,  q8,  q10
+  .endif
+        vst1.8          {d16}, [r0,:64], r2
+        vst1.8          {d17}, [r0,:64], r2
+        bgt             5b
+
         pop             {r4-r7, pc}
 endfunc
 .endm
@@ -245,9 +273,10 @@ T       cmp             r7,  #0
 
         pop             {r4-r7, pc}
 
-2:      tst             r6,  r6
-        add             r12, r12, r6
+2:      adds            r12, r12, r6
         vdup.8          d0,  r4
+        beq             5f
+        tst             r6,  r6
         vdup.8          d1,  r12
         vtrn.32         d0,  d1
 
@@ -310,6 +339,29 @@ T       cmp             r7,  #0
         vst1.32         {d16[1]}, [r0,:32], r2
         bgt             4b
 
+        pop             {r4-r7, pc}
+
+5:      vld1.32         {d4[0]},  [r1], r2
+        vld1.32         {d4[1]},  [r1], r2
+        vmull.u8        q8,  d4,  d0
+        subs            r3,  r3,  #2
+        pld             [r1]
+  .ifc \codec,h264
+        vrshrn.u16      d16, q8,  #6
+  .else
+        vadd.u16        q8,  q8,  q11
+        vshrn.u16       d16, q8,  #6
+  .endif
+  .ifc \type,avg
+        vld1.32         {d20[0]}, [lr,:32], r2
+        vld1.32         {d20[1]}, [lr,:32], r2
+        vrhadd.u8       d16, d16, d20
+  .endif
+        pld             [r1]
+        vst1.32         {d16[0]}, [r0,:32], r2
+        vst1.32         {d16[1]}, [r0,:32], r2
+        bgt             5b
+
         pop             {r4-r7, pc}
 endfunc
 .endm