qdm2: avoid integer overflow

[ffmpeg] / libavcodec / arm / dcadsp_neon.S

diff --git a/libavcodec/arm/dcadsp_neon.S b/libavcodec/arm/dcadsp_neon.S
index 735c4c28e5ff7e1137acddcd6f6e380fbf8b164c..70580cdeec8dd6e561d4e70bb2abc689463421aa 100644 (file)
--- a/libavcodec/arm/dcadsp_neon.S
+++ b/libavcodec/arm/dcadsp_neon.S
@@ -20,6 +20,35 @@
 
 #include "libavutil/arm/asm.S"
 
+function ff_decode_hf_neon, export=1
+        push            {r4-r5,lr}
+        add             r2,  r2,  r3
+        ldr             r3,       [sp, #12]
+        ldrd            r4,  r5,  [sp, #16]
+        add             r3,  r3,  r4, lsl #3
+        add             r1,  r1,  r4, lsl #2
+        add             r0,  r0,  r4, lsl #5
+
+1:      ldr_post        lr,  r1,  #4
+        add             r4,  r4,  #1
+        add             lr,  r2,  lr, lsl #5
+        cmp             r4,  r5
+        vld1.32         {d7},     [r3]!
+        vld1.8          {d0},     [lr,:64]
+        vcvt.f32.s32    d7,  d7,  #4
+        vmovl.s8        q1,  d0
+        vmovl.s16       q0,  d2
+        vmovl.s16       q1,  d3
+        vcvt.f32.s32    q0,  q0
+        vcvt.f32.s32    q1,  q1
+        vmul.f32        q0,  q0,  d7[0]
+        vmul.f32        q1,  q1,  d7[0]
+        vst1.32         {q0-q1},  [r0,:128]!
+        bne             1b
+
+        pop             {r4-r5,pc}
+endfunc
+
 function ff_dca_lfe_fir0_neon, export=1
         push            {r4-r6,lr}
         mov             r3,  #32                @ decifactor