save one cycle

Originally committed as revision 7799 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/armv4l/simple_idct_armv6.S

@@ -51,13 +51,13 @@ w57:    .long W57
   Compute partial IDCT of single row.
   shift = left-shift amount
   a1 = source address
-  a3 = row[2,0]
+  a3 = row[2,0] <= 2 cycles
   a4 = row[3,1]
+  ip = w42      <= 2 cycles
 
   Output in registers v1--v8
 */
         .macro idct_row shift
-        ldr    ip, [pc, #(w42-.-8)]  /* ip = W4 | (W2 << 16) */
         ldr    lr, [pc, #(w46-.-8)]  /* lr = W4 | (W6 << 16) */
         mov    a2, #(1<<(\shift-1))
         smlad  v1, a3, ip, a2
@@ -204,6 +204,7 @@ idct_row_armv6:
         cmpeq  lr, a3, lsr #16
         beq    1f
         str    a2, [sp, #-4]!
+        ldr    ip, [pc, #(w42-.-8)]  /* ip = W4 | (W2 << 16) */
         cmp    lr, #0
         beq    2f
 
@@ -249,6 +250,7 @@ idct_col_armv6:
         stmfd  sp!, {a2, lr}
 
         ldr    a3, [a1]              /* a3 = row[2,0] */
+        ldr    ip, [pc, #(w42-.-8)]  /* ip = W4 | (W2 << 16) */
         ldr    a4, [a1, #8]          /* a4 = row[3,1] */
         idct_row COL_SHIFT
         ldr    a2, [sp], #4
@@ -278,6 +280,7 @@ idct_col_put_armv6:
         stmfd  sp!, {a2, a3, lr}
 
         ldr    a3, [a1]              /* a3 = row[2,0] */
+        ldr    ip, [pc, #(w42-.-8)]  /* ip = W4 | (W2 << 16) */
         ldr    a4, [a1, #8]          /* a4 = row[3,1] */
         idct_row COL_SHIFT
         ldmfd  sp!, {a2, a3}
@@ -309,6 +312,7 @@ idct_col_add_armv6:
         stmfd  sp!, {a2, a3, lr}
 
         ldr    a3, [a1]              /* a3 = row[2,0] */
+        ldr    ip, [pc, #(w42-.-8)]  /* ip = W4 | (W2 << 16) */
         ldr    a4, [a1, #8]          /* a4 = row[3,1] */
         idct_row COL_SHIFT
         ldmfd  sp!, {a2, a3}