Get a few more % speedup on ARM (measured on ipod video) - improve imdct full final symmetries using ldm/stm and simple register swapping. Also, add more comments (and improve/update some of the existing ones) regarding the layout of the imdct_half and the imdct_full

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24819 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 5 insertions, 8 deletions

diff --git a/apps/codecs/lib/asm_arm.h b/apps/codecs/lib/asm_arm.h
index 4f31f80c3e..9dcbcef755 100644
--- a/apps/codecs/lib/asm_arm.h
+++ b/apps/codecs/lib/asm_arm.h
@@ -226,14 +226,11 @@ void vect_mult_bw(int32_t *data, int32_t *window, int n)
 #define _V_CLIP_MATH
 
 static inline int32_t CLIP_TO_15(int32_t x) {
-  int tmp;
-  asm volatile("subs    %1, %0, #32768\n\t"
-               "movpl   %0, #0x7f00\n\t"
-               "orrpl   %0, %0, #0xff\n"
-               "adds    %1, %0, #32768\n\t"
-               "movmi   %0, #0x8000"
-               : "+r"(x),"=r"(tmp)
-               :
+  const int32_t mask = 0xffff7fff;
+  asm volatile("teq %0,%0,asr #31\n\t"
+               "eorne %0,%1,%0,asr #31\n\t"
+               : "+r"(x)
+               : "r" (mask)
                : "cc");
   return(x);
 }