Shuffle some instructions around for that extra percent of performance. Fix a bunch of comments.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@15216 a1c6a512-1295-4272-9138-f99709370657
author: Jens Arnold <amiconn@rockbox.org> 2007-10-19 22:57:19 +0000
committer: Jens Arnold <amiconn@rockbox.org> 2007-10-19 22:57:19 +0000
commit: 87f5359d604a3d51526965a97c74896f392ed444 (patch)
tree: b24323088f7b2b7e470d5b61feca19794f7b22a8
parent: 152f405cce90a285b9c4f1bcaba27dbccc51fdf2 (diff)
download: rockbox-87f5359d604a3d51526965a97c74896f392ed444.tar.gz
rockbox-87f5359d604a3d51526965a97c74896f392ed444.zip
1 files changed, 123 insertions, 120 deletions
diff --git a/apps/codecs/demac/libdemac/predictor-cf.S b/apps/codecs/demac/libdemac/predictor-cf.S
index 19873420c3..3b9489e791 100644
--- a/apps/codecs/demac/libdemac/predictor-cf.S
+++ b/apps/codecs/demac/libdemac/predictor-cf.S
@@ -97,6 +97,8 @@ predictor_decode_stereo:
    sub.l   %d3, %d2
    neg.l   %d2                         | %d2 = %d3 - %d2
+    move.l  %d2, (YDELAYA-4,%a5)        | p->buf[YDELAYA-1]  = %d2
    movem.l (YcoeffsA,%a6), %a0-%a3     | %a0  = p->YcoeffsA[0]
                                        | %a1  = p->YcoeffsA[1]
                                        | %a2  = p->YcoeffsA[2]
@@ -107,10 +109,7 @@ predictor_decode_stereo:
    mac.l   %d1, %a2, %acc0             | %acc0 += p->buf[YDELAYA-2] * p->YcoeffsA[2]
    mac.l   %d0, %a3, %acc0             | %acc0 += p->buf[YDELAYA-3] * p->YcoeffsA[3]
-    move.l  %d2, (YDELAYA-4,%a5)        | p->buf[YDELAYA-1]  = %d2
    move.l  %d3, (YDELAYA,%a5)          | p->buf[YDELAYA]  = %d3
-    
-    movclr.l %acc0, %d0
    tst.l   %d2
    beq.s   1f
@@ -118,16 +117,18 @@ predictor_decode_stereo:
    extb.l  %d2                         | pos: 0xffffffff, neg: 0x00000000
    or.l    #1, %d2                     | pos: 0xffffffff, neg: 0x00000001
 1:                                      | %d2 = SIGN(%d2)
+    move.l  %d2, (YADAPTCOEFFSA-4,%a5)  | p->buf[YADAPTCOEFFSA-1]  = %d2
    tst.l   %d3
    beq.s   1f
    spl.b   %d3
    extb.l  %d3
    or.l    #1, %d3
 1:                                      | %d3 = SIGN(%d3)
-    move.l  %d2, (YADAPTCOEFFSA-4,%a5)  | p->buf[YADAPTCOEFFSA-1]  = %d2
    move.l  %d3, (YADAPTCOEFFSA,%a5)    | p->buf[YADAPTCOEFFSA]  = %d3
-    
+    movclr.l %acc0, %d0
    | NOTE: %d0 now contains predictionA - don't overwrite.
    
    | Predictor Y, Filter B
@@ -149,11 +150,13 @@ predictor_decode_stereo:
    sub.l   %d3, %d7
    neg.l   %d7                         | %d7  = %d3 - %d7
-    movem.l (YcoeffsB,%a6), %d2/%a0-%a3     | %d2  = p->YcoeffsB[0]
+    move.l  %d7, (YDELAYB-4,%a5)        | p->buf[YDELAYB-1]  = %d7
-                                            | %a0  = p->YcoeffsB[1]
-                                            | %a1  = p->YcoeffsB[2]
+    movem.l (YcoeffsB,%a6), %d2/%a0-%a3 | %d2  = p->YcoeffsB[0]
-                                            | %a2  = p->YcoeffsB[3]
+                                        | %a0  = p->YcoeffsB[1]
-                                            | %a3  = p->YcoeffsB[4]
+                                        | %a1  = p->YcoeffsB[2]
+                                        | %a2  = p->YcoeffsB[3]
+                                        | %a3  = p->YcoeffsB[4]
    mac.l   %d3, %d2, %acc0             | %acc0  = p->buf[YDELAYB] * p->YcoeffsB[0]
    mac.l   %d7, %a0, %acc0             | %acc0 += p->buf[YDELAYB-1] * p->YcoeffsB[1]
@@ -161,26 +164,24 @@ predictor_decode_stereo:
    mac.l   %d5, %a2, %acc0             | %acc0 += p->buf[YDELAYB-3] * p->YcoeffsB[3]
    mac.l   %d4, %a3, %acc0             | %acc0 += p->buf[YDELAYB-4] * p->YcoeffsB[4]
    
-    move.l  %d7, (YDELAYB-4,%a5)        | p->buf[YDELAYB-1]  = %d7
    move.l  %d3, (YDELAYB, %a5)         | p->buf[YDELAYB]  = %d3
    
-    movclr.l %acc0, %d1
-    
    tst.l   %d7
    beq.s   1f
    spl.b   %d7
    extb.l  %d7
    or.l    #1, %d7
 1:                                      | %d7 = SIGN(%d7)
+    move.l  %d7, (YADAPTCOEFFSB-4,%a5)  | p->buf[YADAPTCOEFFSB-1]  = %d7
    tst.l   %d3
    beq.s   1f
    spl.b   %d3
    extb.l  %d3
    or.l    #1, %d3
 1:                                      | %d3 = SIGN(%d3)
-    move.l  %d7, (YADAPTCOEFFSB-4,%a5)  | p->buf[YADAPTCOEFFSB-1]  = %d7
    move.l  %d3, (YADAPTCOEFFSB, %a5)   | p->buf[YADAPTCOEFFSB]  = %d3
+    movclr.l %acc0, %d1
    
    | %d0 still contains predictionA
    | %d1 contains predictionB
@@ -215,70 +216,70 @@ predictor_decode_stereo:
    move.l  %d1, (%a4)+                 | *(decoded0++)  = %d1 (p->YfilterA)
    move.l  %a4, (%sp)                  | save decoded0    
    tst.l   %d5
-    beq.s   2f
+    beq.s   3f
-    movem.l (YADAPTCOEFFSB-16,%a5), %d4-%d6  | d4  = p->buf[YADAPTCOEFFSB-4]
+    movem.l (YADAPTCOEFFSB-16,%a5), %d4-%d6 | %d4  = p->buf[YADAPTCOEFFSB-4]
-                                            | d5  = p->buf[YADAPTCOEFFSB-3]
+                                            | %d5  = p->buf[YADAPTCOEFFSB-3]
-                                            | d6  = p->buf[YADAPTCOEFFSB-2]
+                                            | %d6  = p->buf[YADAPTCOEFFSB-2]
    bmi.s   1f                          | flags still valid here
    | *decoded0 > 0
-    sub.l   %d3, %d2                    | d2  = p->YcoeffsB[0] - p->buf[YADAPTCOEFFSB]
+    sub.l   %d3, %d2                    | %d2  = p->YcoeffsB[0] - p->buf[YADAPTCOEFFSB]
-    sub.l   %d7, %a0                    | a0  = p->YcoeffsB[1] - p->buf[YADAPTCOEFFSB-1]
+    sub.l   %d7, %a0                    | %a0  = p->YcoeffsB[1] - p->buf[YADAPTCOEFFSB-1]
-    sub.l   %d6, %a1                    | a1  = p->YcoeffsB[2] - p->buf[YADAPTCOEFFSB-2]
+    sub.l   %d6, %a1                    | %a1  = p->YcoeffsB[2] - p->buf[YADAPTCOEFFSB-2]
-    sub.l   %d5, %a2                    | a2  = p->YcoeffsB[3] - p->buf[YADAPTCOEFFSB-3]
+    sub.l   %d5, %a2                    | %a2  = p->YcoeffsB[3] - p->buf[YADAPTCOEFFSB-3]
-    sub.l   %d4, %a3                    | a3  = p->YcoeffsB[4] - p->buf[YADAPTCOEFFSB-4]
+    sub.l   %d4, %a3                    | %a3  = p->YcoeffsB[4] - p->buf[YADAPTCOEFFSB-4]
    
    movem.l %d2/%a0-%a3, (YcoeffsB,%a6) | Save p->YcoeffsB[]
    
-    movem.l (YcoeffsA,%a6), %d4-%d7     | d4  = p->YcoeffsA[0]
+    movem.l (YcoeffsA,%a6), %d4-%d7     | %d4  = p->YcoeffsA[0]
-                                        | d5  = p->YcoeffsA[1]
+                                        | %d5  = p->YcoeffsA[1]
-                                        | d6  = p->YcoeffsA[2]
+                                        | %d6  = p->YcoeffsA[2]
-                                        | d7  = p->YcoeffsA[3]
+                                        | %d7  = p->YcoeffsA[3]
                                        
-    movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | d2  = p->buf[YADAPTCOEFFSA-3]
+    movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | %d2  = p->buf[YADAPTCOEFFSA-3]
-                                        | a0  = p->buf[YADAPTCOEFFSA-2]
+                                        | %a0  = p->buf[YADAPTCOEFFSA-2]
-                                        | a1  = p->buf[YADAPTCOEFFSA-1]
+                                        | %a1  = p->buf[YADAPTCOEFFSA-1]
-                                        | a2  = p->buf[YADAPTCOEFFSA]
+                                        | %a2  = p->buf[YADAPTCOEFFSA]
                                        
-    sub.l   %a2, %d4                    | d4  = p->YcoeffsA[0] - p->buf[YADAPTCOEFFSA]
+    sub.l   %a2, %d4                    | %d4  = p->YcoeffsA[0] - p->buf[YADAPTCOEFFSA]
-    sub.l   %a1, %d5                    | d5  = p->YcoeffsA[1] - p->buf[YADAPTCOEFFSA-1]
+    sub.l   %a1, %d5                    | %d5  = p->YcoeffsA[1] - p->buf[YADAPTCOEFFSA-1]
-    sub.l   %a0, %d6                    | d6  = p->YcoeffsA[2] - p->buf[YADAPTCOEFFSA-2]
+    sub.l   %a0, %d6                    | %d6  = p->YcoeffsA[2] - p->buf[YADAPTCOEFFSA-2]
-    sub.l   %d2, %d7                    | d7  = p->YcoeffsA[3] - p->buf[YADAPTCOEFFSA-3]
+    sub.l   %d2, %d7                    | %d7  = p->YcoeffsA[3] - p->buf[YADAPTCOEFFSA-3]
    
-    movem.l %d4-%d7, (YcoeffsA,%a6)     | Save p->YcoeffsA[]
    bra.s   2f
 1:  | *decoded0 < 0
-    add.l   %d3, %d2                    | d2  = p->YcoeffsB[0] + p->buf[YADAPTCOEFFSB]
+    add.l   %d3, %d2                    | %d2  = p->YcoeffsB[0] + p->buf[YADAPTCOEFFSB]
-    add.l   %d7, %a0                    | a0  = p->YcoeffsB[1] + p->buf[YADAPTCOEFFSB-1]
+    add.l   %d7, %a0                    | %a0  = p->YcoeffsB[1] + p->buf[YADAPTCOEFFSB-1]
-    add.l   %d6, %a1                    | a1  = p->YcoeffsB[2] + p->buf[YADAPTCOEFFSB-2]
+    add.l   %d6, %a1                    | %a1  = p->YcoeffsB[2] + p->buf[YADAPTCOEFFSB-2]
-    add.l   %d5, %a2                    | a2  = p->YcoeffsB[3] + p->buf[YADAPTCOEFFSB-3]
+    add.l   %d5, %a2                    | %a2  = p->YcoeffsB[3] + p->buf[YADAPTCOEFFSB-3]
-    add.l   %d4, %a3                    | a3  = p->YcoeffsB[4] + p->buf[YADAPTCOEFFSB-4]
+    add.l   %d4, %a3                    | %a3  = p->YcoeffsB[4] + p->buf[YADAPTCOEFFSB-4]
    
    movem.l %d2/%a0-%a3, (YcoeffsB,%a6) | Save p->YcoeffsB[]
    
-    movem.l (YcoeffsA,%a6), %d4-%d7     | d4  = p->YcoeffsA[0]
+    movem.l (YcoeffsA,%a6), %d4-%d7     | %d4  = p->YcoeffsA[0]
-                                        | d5  = p->YcoeffsA[1]
+                                        | %d5  = p->YcoeffsA[1]
-                                        | d6  = p->YcoeffsA[2]
+                                        | %d6  = p->YcoeffsA[2]
-                                        | d7  = p->YcoeffsA[3]
+                                        | %d7  = p->YcoeffsA[3]
                                        
-    movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | d2  = p->buf[YADAPTCOEFFSA-3]
+    movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | %d2  = p->buf[YADAPTCOEFFSA-3]
-                                        | a0  = p->buf[YADAPTCOEFFSA-2]
+                                        | %a0  = p->buf[YADAPTCOEFFSA-2]
-                                        | a1  = p->buf[YADAPTCOEFFSA-1]
+                                        | %a1  = p->buf[YADAPTCOEFFSA-1]
-                                        | a2  = p->buf[YADAPTCOEFFSA]
+                                        | %a2  = p->buf[YADAPTCOEFFSA]
                                        
-    add.l   %a2, %d4                    | d4  = p->YcoeffsA[0] + p->buf[YADAPTCOEFFSA]
+    add.l   %a2, %d4                    | %d4  = p->YcoeffsA[0] + p->buf[YADAPTCOEFFSA]
-    add.l   %a1, %d5                    | d5  = p->YcoeffsA[1] + p->buf[YADAPTCOEFFSA-1]
+    add.l   %a1, %d5                    | %d5  = p->YcoeffsA[1] + p->buf[YADAPTCOEFFSA-1]
-    add.l   %a0, %d6                    | d6  = p->YcoeffsA[2] + p->buf[YADAPTCOEFFSA-2]
+    add.l   %a0, %d6                    | %d6  = p->YcoeffsA[2] + p->buf[YADAPTCOEFFSA-2]
-    add.l   %d2, %d7                    | d7  = p->YcoeffsA[3] + p->buf[YADAPTCOEFFSA-3]
+    add.l   %d2, %d7                    | %d7  = p->YcoeffsA[3] + p->buf[YADAPTCOEFFSA-3]
-    
-    movem.l %d4-%d7, (YcoeffsA,%a6)     | Save p->YcoeffsA[]
 2:
+    movem.l %d4-%d7, (YcoeffsA,%a6)     | Save p->YcoeffsA[]
+3:
    | ***** PREDICTOR X *****
    
@@ -293,6 +294,8 @@ predictor_decode_stereo:
    sub.l   %d3, %d2                    
    neg.l   %d2                         | %d2  = %d3 -%d2
+    move.l  %d2, (XDELAYA-4,%a5)        | p->buf[XDELAYA-1]  = %d2
    movem.l (XcoeffsA,%a6), %a0-%a3     | %a0  = p->XcoeffsA[0]
                                        | %a1  = p->XcoeffsA[1]
                                        | %a2  = p->XcoeffsA[2]
@@ -303,27 +306,26 @@ predictor_decode_stereo:
    mac.l   %d1, %a2, %acc0             | %acc0 += p->buf[XDELAYA-2] * p->XcoeffsA[2]
    mac.l   %d0, %a3, %acc0             | %acc0 += p->buf[XDELAYA-3] * p->XcoeffsA[3]
-    move.l  %d2, (XDELAYA-4,%a5)        | p->buf[XDELAYA-1]  = %d2
    move.l  %d3, (XDELAYA,%a5)          | p->buf[XDELAYA]  = %d3
    
-    movclr.l %acc0, %d0
    tst.l   %d2
    beq.s   1f
    spl.b   %d2                         | pos: 0x??????ff, neg: 0x??????00
    extb.l  %d2                         | pos: 0xffffffff, neg: 0x00000000
    or.l    #1, %d2                     | pos: 0xffffffff, neg: 0x00000001
 1:                                      | %d2 = SIGN(%d2)
+    move.l  %d2, (XADAPTCOEFFSA-4,%a5)  | p->buf[XADAPTCOEFFSA-1]  = %d2
    tst.l   %d3
    beq.s   1f
    spl.b   %d3
    extb.l  %d3
    or.l    #1, %d3
 1:                                      | %d3 = SIGN(%d3)
+    move.l  %d3, (XADAPTCOEFFSA,%a5)    | p->buf[XADAPTCOEFFSA]  = %d3
+    movclr.l %acc0, %d0
-    move.l  %d2, (XADAPTCOEFFSA-4,%a5)  | p->buf[XADAPTCOEFFSA-1]  = r2
-    move.l  %d3, (XADAPTCOEFFSA,%a5)    | p->buf[XADAPTCOEFFSA]  = r3
-    
    | NOTE: %d0 now contains predictionA - don't overwrite.
    
    | Predictor X, Filter B
@@ -345,11 +347,13 @@ predictor_decode_stereo:
    sub.l   %d3, %d7
    neg.l   %d7                         | %d7  = %d3 - %d7
-    movem.l (XcoeffsB,%a6), %d2/%a0-%a3     | %d2  = p->XcoeffsB[0]
+    move.l  %d7, (XDELAYB-4,%a5)        | p->buf[XDELAYB-1]  = %d7
-                                            | %a0  = p->XcoeffsB[1]
-                                            | %a1  = p->XcoeffsB[2]
+    movem.l (XcoeffsB,%a6), %d2/%a0-%a3 | %d2  = p->XcoeffsB[0]
-                                            | %a2  = p->XcoeffsB[3]
+                                        | %a0  = p->XcoeffsB[1]
-                                            | %a3  = p->XcoeffsB[4]
+                                        | %a1  = p->XcoeffsB[2]
+                                        | %a2  = p->XcoeffsB[3]
+                                        | %a3  = p->XcoeffsB[4]
    mac.l   %d3, %d2, %acc0             | %acc0  = p->buf[XDELAYB] * p->XcoeffsB[0]
    mac.l   %d7, %a0, %acc0             | %acc0 += p->buf[XDELAYB-1] * p->XcoeffsB[1]
@@ -357,27 +361,26 @@ predictor_decode_stereo:
    mac.l   %d5, %a2, %acc0             | %acc0 += p->buf[XDELAYB-3] * p->XcoeffsB[3]
    mac.l   %d4, %a3, %acc0             | %acc0 += p->buf[XDELAYB-4] * p->XcoeffsB[4]
    
-    move.l  %d7, (XDELAYB-4,%a5)        | p->buf[XDELAYB-1]  = %d7
    move.l  %d3, (XDELAYB, %a5)         | p->buf[XDELAYB]  = %d3
    
-    movclr.l %acc0, %d1
-    
    tst.l   %d7
    beq.s   1f
    spl.b   %d7
    extb.l  %d7
    or.l    #1, %d7
 1:                                      | %d7 = SIGN(%d7)
+    move.l  %d7, (XADAPTCOEFFSB-4,%a5)  | p->buf[XADAPTCOEFFSB-1]  = %d7
    tst.l   %d3
    beq.s   1f
    spl.b   %d3
    extb.l  %d3
    or.l    #1, %d3
 1:                                      | %d3 = SIGN(%d3)
-    move.l  %d7, (XADAPTCOEFFSB-4,%a5)  | p->buf[XADAPTCOEFFSB-1]  = %d7
    move.l  %d3, (XADAPTCOEFFSB, %a5)   | p->buf[XADAPTCOEFFSB]  = %d3
-    
+    movclr.l %acc0, %d1
    | %d0 still contains predictionA
    | %d1 contains predictionB
@@ -405,76 +408,76 @@ predictor_decode_stereo:
    | %a4 contains decoded1
    | %d5 contains *decoded1
-    | %d2, %a0, %a1, %a2, %a31 contain p->XcoeffsB[0..4]
+    | %d2, %a0, %a1, %a2, %a3 contain p->XcoeffsB[0..4]
    | %d7, %d3 contain p->buf[XADAPTCOEFFSB-1] and p->buf[XADAPTCOEFFSB]
    move.l  %d1, (%a4)+                 | *(decoded1++)  = %d1 (p->XfilterA)
    move.l  %a4, (4,%sp)                | save decoded1
    tst.l   %d5
-    beq.s   2f
+    beq.s   3f
    
-    movem.l (XADAPTCOEFFSB-16,%a5), %d4-%d6  | d4  = p->buf[XADAPTCOEFFSB-4]
+    movem.l (XADAPTCOEFFSB-16,%a5), %d4-%d6 | %d4  = p->buf[XADAPTCOEFFSB-4]
-                                            | d5  = p->buf[XADAPTCOEFFSB-3]
+                                            | %d5  = p->buf[XADAPTCOEFFSB-3]
-                                            | d6  = p->buf[XADAPTCOEFFSB-2]
+                                            | %d6  = p->buf[XADAPTCOEFFSB-2]
    bmi.s   1f                          | flags still valid here
    
    | *decoded1 > 0
-    sub.l   %d3, %d2                    | d2  = p->XcoeffsB[0] - p->buf[XADAPTCOEFFSB]
+    sub.l   %d3, %d2                    | %d2  = p->XcoeffsB[0] - p->buf[XADAPTCOEFFSB]
-    sub.l   %d7, %a0                    | a0  = p->XcoeffsB[1] - p->buf[XADAPTCOEFFSB-1]
+    sub.l   %d7, %a0                    | %a0  = p->XcoeffsB[1] - p->buf[XADAPTCOEFFSB-1]
-    sub.l   %d6, %a1                    | a1  = p->XcoeffsB[2] - p->buf[XADAPTCOEFFSB-2]
+    sub.l   %d6, %a1                    | %a1  = p->XcoeffsB[2] - p->buf[XADAPTCOEFFSB-2]
-    sub.l   %d5, %a2                    | a2  = p->XcoeffsB[3] - p->buf[XADAPTCOEFFSB-3]
+    sub.l   %d5, %a2                    | %a2  = p->XcoeffsB[3] - p->buf[XADAPTCOEFFSB-3]
-    sub.l   %d4, %a3                    | a3  = p->XcoeffsB[4] - p->buf[XADAPTCOEFFSB-4]
+    sub.l   %d4, %a3                    | %a3  = p->XcoeffsB[4] - p->buf[XADAPTCOEFFSB-4]
    
    movem.l %d2/%a0-%a3, (XcoeffsB,%a6) | Save p->XcoeffsB[]
    
-    movem.l (XcoeffsA,%a6), %d4-%d7     | d4  = p->XcoeffsA[0]
+    movem.l (XcoeffsA,%a6), %d4-%d7     | %d4  = p->XcoeffsA[0]
-                                        | d5  = p->XcoeffsA[1]
+                                        | %d5  = p->XcoeffsA[1]
-                                        | d6  = p->XcoeffsA[2]
+                                        | %d6  = p->XcoeffsA[2]
-                                        | d7  = p->XcoeffsA[3]
+                                        | %d7  = p->XcoeffsA[3]
                                        
-    movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | d2  = p->buf[XADAPTCOEFFSA-3]
+    movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | %d2  = p->buf[XADAPTCOEFFSA-3]
-                                        | a0  = p->buf[XADAPTCOEFFSA-2]
+                                        | %a0  = p->buf[XADAPTCOEFFSA-2]
-                                        | a1  = p->buf[XADAPTCOEFFSA-1]
+                                        | %a1  = p->buf[XADAPTCOEFFSA-1]
-                                        | a2  = p->buf[XADAPTCOEFFSA]
+                                        | %a2  = p->buf[XADAPTCOEFFSA]
                                        
-    sub.l   %a2, %d4                    | d4  = p->XcoeffsA[0] - p->buf[XADAPTCOEFFSA]
+    sub.l   %a2, %d4                    | %d4  = p->XcoeffsA[0] - p->buf[XADAPTCOEFFSA]
-    sub.l   %a1, %d5                    | d5  = p->XcoeffsA[1] - p->buf[XADAPTCOEFFSA-1]
+    sub.l   %a1, %d5                    | %d5  = p->XcoeffsA[1] - p->buf[XADAPTCOEFFSA-1]
-    sub.l   %a0, %d6                    | d6  = p->XcoeffsA[2] - p->buf[XADAPTCOEFFSA-2]
+    sub.l   %a0, %d6                    | %d6  = p->XcoeffsA[2] - p->buf[XADAPTCOEFFSA-2]
-    sub.l   %d2, %d7                    | d7  = p->XcoeffsA[3] - p->buf[XADAPTCOEFFSA-3]
+    sub.l   %d2, %d7                    | %d7  = p->XcoeffsA[3] - p->buf[XADAPTCOEFFSA-3]
    
-    movem.l %d4-%d7, (XcoeffsA,%a6)     | Save p->XcoeffsA[]
    bra.s   2f
 1:  | *decoded1 < 0
-    add.l   %d3, %d2                    | d2  = p->XcoeffsB[0] + p->buf[XADAPTCOEFFSB]
+    add.l   %d3, %d2                    | %d2  = p->XcoeffsB[0] + p->buf[XADAPTCOEFFSB]
-    add.l   %d7, %a0                    | a0  = p->XcoeffsB[1] + p->buf[XADAPTCOEFFSB-1]
+    add.l   %d7, %a0                    | %a0  = p->XcoeffsB[1] + p->buf[XADAPTCOEFFSB-1]
-    add.l   %d6, %a1                    | a1  = p->XcoeffsB[2] + p->buf[XADAPTCOEFFSB-2]
+    add.l   %d6, %a1                    | %a1  = p->XcoeffsB[2] + p->buf[XADAPTCOEFFSB-2]
-    add.l   %d5, %a2                    | a2  = p->XcoeffsB[3] + p->buf[XADAPTCOEFFSB-3]
+    add.l   %d5, %a2                    | %a2  = p->XcoeffsB[3] + p->buf[XADAPTCOEFFSB-3]
-    add.l   %d4, %a3                    | a3  = p->XcoeffsB[4] + p->buf[XADAPTCOEFFSB-4]
+    add.l   %d4, %a3                    | %a3  = p->XcoeffsB[4] + p->buf[XADAPTCOEFFSB-4]
    
    movem.l %d2/%a0-%a3, (XcoeffsB,%a6) | Save p->XcoeffsB[]
    
-    movem.l (XcoeffsA,%a6), %d4-%d7     | d4  = p->XcoeffsA[0]
+    movem.l (XcoeffsA,%a6), %d4-%d7     | %d4  = p->XcoeffsA[0]
-                                        | d5  = p->XcoeffsA[1]
+                                        | %d5  = p->XcoeffsA[1]
-                                        | d6  = p->XcoeffsA[2]
+                                        | %d6  = p->XcoeffsA[2]
-                                        | d7  = p->XcoeffsA[3]
+                                        | %d7  = p->XcoeffsA[3]
-    movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | d2  = p->buf[XADAPTCOEFFSA-3]
+    movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 | %d2  = p->buf[XADAPTCOEFFSA-3]
-                                        | a0  = p->buf[XADAPTCOEFFSA-2]
+                                        | %a0  = p->buf[XADAPTCOEFFSA-2]
-                                        | a1  = p->buf[XADAPTCOEFFSA-1]
+                                        | %a1  = p->buf[XADAPTCOEFFSA-1]
-                                        | a2  = p->buf[XADAPTCOEFFSA]
+                                        | %a2  = p->buf[XADAPTCOEFFSA]
                                        
-    add.l   %a2, %d4                    | d4  = p->XcoeffsA[0] + p->buf[XADAPTCOEFFSA]
+    add.l   %a2, %d4                    | %d4  = p->XcoeffsA[0] + p->buf[XADAPTCOEFFSA]
-    add.l   %a1, %d5                    | d5  = p->XcoeffsA[1] + p->buf[XADAPTCOEFFSA-1]
+    add.l   %a1, %d5                    | %d5  = p->XcoeffsA[1] + p->buf[XADAPTCOEFFSA-1]
-    add.l   %a0, %d6                    | d6  = p->XcoeffsA[2] + p->buf[XADAPTCOEFFSA-2]
+    add.l   %a0, %d6                    | %d6  = p->XcoeffsA[2] + p->buf[XADAPTCOEFFSA-2]
-    add.l   %d2, %d7                    | d7  = p->XcoeffsA[3] + p->buf[XADAPTCOEFFSA-3]
+    add.l   %d2, %d7                    | %d7  = p->XcoeffsA[3] + p->buf[XADAPTCOEFFSA-3]
-    
-    movem.l %d4-%d7, (XcoeffsA,%a6)     | Save p->XcoeffsA[]
 2:
+    movem.l %d4-%d7, (XcoeffsA,%a6)     | Save p->XcoeffsA[]
+3:
    | ***** COMMON *****
    
@@ -484,10 +487,10 @@ predictor_decode_stereo:
    
    cmp.l   %a3, %a5                    
    bne.s   .endofloop
+                                              
    | The history buffer is full, we need to do a memmove:
-    
-    lea.l   (historybuffer,%a6), %a3
+    lea.l   (historybuffer,%a6), %a3    
    | dest = %a3 (p->historybuffer)
    | src = %a5 (p->buf)
@@ -515,7 +518,7 @@ predictor_decode_stereo:
    lea.l   (40,%a3), %a3
    lea.l   (historybuffer,%a6), %a5    | p->buf = &p->historybuffer[0]
-    
 .endofloop:
    subq.l  #1, (8,%sp)                 | decrease loop count
    bne.w   .loop
author	Jens Arnold <amiconn@rockbox.org>	2007-10-19 22:57:19 +0000
committer	Jens Arnold <amiconn@rockbox.org>	2007-10-19 22:57:19 +0000
commit	87f5359d604a3d51526965a97c74896f392ed444 (patch)
tree	b24323088f7b2b7e470d5b61feca19794f7b22a8
parent	152f405cce90a285b9c4f1bcaba27dbccc51fdf2 (diff)
download	rockbox-87f5359d604a3d51526965a97c74896f392ed444.tar.gz rockbox-87f5359d604a3d51526965a97c74896f392ed444.zip

diff --git a/apps/codecs/demac/libdemac/predictor-cf.S b/apps/codecs/demac/libdemac/predictor-cf.S index 19873420c3..3b9489e791 100644 --- a/apps/codecs/demac/libdemac/predictor-cf.S +++ b/apps/codecs/demac/libdemac/predictor-cf.S
@@ -97,6 +97,8 @@ predictor_decode_stereo:
97	sub.l %d3, %d2	97	sub.l %d3, %d2
98	neg.l %d2 \| %d2 = %d3 - %d2	98	neg.l %d2 \| %d2 = %d3 - %d2
99		99
		100	move.l %d2, (YDELAYA-4,%a5) \| p->buf[YDELAYA-1] = %d2
		101
100	movem.l (YcoeffsA,%a6), %a0-%a3 \| %a0 = p->YcoeffsA[0]	102	movem.l (YcoeffsA,%a6), %a0-%a3 \| %a0 = p->YcoeffsA[0]
101	\| %a1 = p->YcoeffsA[1]	103	\| %a1 = p->YcoeffsA[1]
102	\| %a2 = p->YcoeffsA[2]	104	\| %a2 = p->YcoeffsA[2]
@@ -107,10 +109,7 @@ predictor_decode_stereo:
107	mac.l %d1, %a2, %acc0 \| %acc0 += p->buf[YDELAYA-2] * p->YcoeffsA[2]	109	mac.l %d1, %a2, %acc0 \| %acc0 += p->buf[YDELAYA-2] * p->YcoeffsA[2]
108	mac.l %d0, %a3, %acc0 \| %acc0 += p->buf[YDELAYA-3] * p->YcoeffsA[3]	110	mac.l %d0, %a3, %acc0 \| %acc0 += p->buf[YDELAYA-3] * p->YcoeffsA[3]
109		111
110	move.l %d2, (YDELAYA-4,%a5) \| p->buf[YDELAYA-1] = %d2
111	move.l %d3, (YDELAYA,%a5) \| p->buf[YDELAYA] = %d3	112	move.l %d3, (YDELAYA,%a5) \| p->buf[YDELAYA] = %d3
112
113	movclr.l %acc0, %d0
114		113
115	tst.l %d2	114	tst.l %d2
116	beq.s 1f	115	beq.s 1f
@@ -118,16 +117,18 @@ predictor_decode_stereo:
118	extb.l %d2 \| pos: 0xffffffff, neg: 0x00000000	117	extb.l %d2 \| pos: 0xffffffff, neg: 0x00000000
119	or.l #1, %d2 \| pos: 0xffffffff, neg: 0x00000001	118	or.l #1, %d2 \| pos: 0xffffffff, neg: 0x00000001
120	1: \| %d2 = SIGN(%d2)	119	1: \| %d2 = SIGN(%d2)
		120	move.l %d2, (YADAPTCOEFFSA-4,%a5) \| p->buf[YADAPTCOEFFSA-1] = %d2
		121
121	tst.l %d3	122	tst.l %d3
122	beq.s 1f	123	beq.s 1f
123	spl.b %d3	124	spl.b %d3
124	extb.l %d3	125	extb.l %d3
125	or.l #1, %d3	126	or.l #1, %d3
126	1: \| %d3 = SIGN(%d3)	127	1: \| %d3 = SIGN(%d3)
127
128	move.l %d2, (YADAPTCOEFFSA-4,%a5) \| p->buf[YADAPTCOEFFSA-1] = %d2
129	move.l %d3, (YADAPTCOEFFSA,%a5) \| p->buf[YADAPTCOEFFSA] = %d3	128	move.l %d3, (YADAPTCOEFFSA,%a5) \| p->buf[YADAPTCOEFFSA] = %d3
130		129
		130	movclr.l %acc0, %d0
		131
131	\| NOTE: %d0 now contains predictionA - don't overwrite.	132	\| NOTE: %d0 now contains predictionA - don't overwrite.
132		133
133	\| Predictor Y, Filter B	134	\| Predictor Y, Filter B
@@ -149,11 +150,13 @@ predictor_decode_stereo:
149	sub.l %d3, %d7	150	sub.l %d3, %d7
150	neg.l %d7 \| %d7 = %d3 - %d7	151	neg.l %d7 \| %d7 = %d3 - %d7
151		152
152	movem.l (YcoeffsB,%a6), %d2/%a0-%a3 \| %d2 = p->YcoeffsB[0]	153	move.l %d7, (YDELAYB-4,%a5) \| p->buf[YDELAYB-1] = %d7
153	\| %a0 = p->YcoeffsB[1]	154
154	\| %a1 = p->YcoeffsB[2]	155	movem.l (YcoeffsB,%a6), %d2/%a0-%a3 \| %d2 = p->YcoeffsB[0]
155	\| %a2 = p->YcoeffsB[3]	156	\| %a0 = p->YcoeffsB[1]
156	\| %a3 = p->YcoeffsB[4]	157	\| %a1 = p->YcoeffsB[2]
		158	\| %a2 = p->YcoeffsB[3]
		159	\| %a3 = p->YcoeffsB[4]
157		160
158	mac.l %d3, %d2, %acc0 \| %acc0 = p->buf[YDELAYB] * p->YcoeffsB[0]	161	mac.l %d3, %d2, %acc0 \| %acc0 = p->buf[YDELAYB] * p->YcoeffsB[0]
159	mac.l %d7, %a0, %acc0 \| %acc0 += p->buf[YDELAYB-1] * p->YcoeffsB[1]	162	mac.l %d7, %a0, %acc0 \| %acc0 += p->buf[YDELAYB-1] * p->YcoeffsB[1]
@@ -161,26 +164,24 @@ predictor_decode_stereo:
161	mac.l %d5, %a2, %acc0 \| %acc0 += p->buf[YDELAYB-3] * p->YcoeffsB[3]	164	mac.l %d5, %a2, %acc0 \| %acc0 += p->buf[YDELAYB-3] * p->YcoeffsB[3]
162	mac.l %d4, %a3, %acc0 \| %acc0 += p->buf[YDELAYB-4] * p->YcoeffsB[4]	165	mac.l %d4, %a3, %acc0 \| %acc0 += p->buf[YDELAYB-4] * p->YcoeffsB[4]
163		166
164	move.l %d7, (YDELAYB-4,%a5) \| p->buf[YDELAYB-1] = %d7
165	move.l %d3, (YDELAYB, %a5) \| p->buf[YDELAYB] = %d3	167	move.l %d3, (YDELAYB, %a5) \| p->buf[YDELAYB] = %d3
166		168
167	movclr.l %acc0, %d1
168
169	tst.l %d7	169	tst.l %d7
170	beq.s 1f	170	beq.s 1f
171	spl.b %d7	171	spl.b %d7
172	extb.l %d7	172	extb.l %d7
173	or.l #1, %d7	173	or.l #1, %d7
174	1: \| %d7 = SIGN(%d7)	174	1: \| %d7 = SIGN(%d7)
		175	move.l %d7, (YADAPTCOEFFSB-4,%a5) \| p->buf[YADAPTCOEFFSB-1] = %d7
175	tst.l %d3	176	tst.l %d3
176	beq.s 1f	177	beq.s 1f
177	spl.b %d3	178	spl.b %d3
178	extb.l %d3	179	extb.l %d3
179	or.l #1, %d3	180	or.l #1, %d3
180	1: \| %d3 = SIGN(%d3)	181	1: \| %d3 = SIGN(%d3)
181
182	move.l %d7, (YADAPTCOEFFSB-4,%a5) \| p->buf[YADAPTCOEFFSB-1] = %d7
183	move.l %d3, (YADAPTCOEFFSB, %a5) \| p->buf[YADAPTCOEFFSB] = %d3	182	move.l %d3, (YADAPTCOEFFSB, %a5) \| p->buf[YADAPTCOEFFSB] = %d3
		183
		184	movclr.l %acc0, %d1
184		185
185	\| %d0 still contains predictionA	186	\| %d0 still contains predictionA
186	\| %d1 contains predictionB	187	\| %d1 contains predictionB
@@ -215,70 +216,70 @@ predictor_decode_stereo:
215	move.l %d1, (%a4)+ \| *(decoded0++) = %d1 (p->YfilterA)	216	move.l %d1, (%a4)+ \| *(decoded0++) = %d1 (p->YfilterA)
216	move.l %a4, (%sp) \| save decoded0	217	move.l %a4, (%sp) \| save decoded0
217	tst.l %d5	218	tst.l %d5
218	beq.s 2f	219	beq.s 3f
219		220
220	movem.l (YADAPTCOEFFSB-16,%a5), %d4-%d6 \| d4 = p->buf[YADAPTCOEFFSB-4]	221	movem.l (YADAPTCOEFFSB-16,%a5), %d4-%d6 \| %d4 = p->buf[YADAPTCOEFFSB-4]
221	\| d5 = p->buf[YADAPTCOEFFSB-3]	222	\| %d5 = p->buf[YADAPTCOEFFSB-3]
222	\| d6 = p->buf[YADAPTCOEFFSB-2]	223	\| %d6 = p->buf[YADAPTCOEFFSB-2]
223		224
224	bmi.s 1f \| flags still valid here	225	bmi.s 1f \| flags still valid here
225		226
226	\| *decoded0 > 0	227	\| *decoded0 > 0
227		228
228	sub.l %d3, %d2 \| d2 = p->YcoeffsB[0] - p->buf[YADAPTCOEFFSB]	229	sub.l %d3, %d2 \| %d2 = p->YcoeffsB[0] - p->buf[YADAPTCOEFFSB]
229	sub.l %d7, %a0 \| a0 = p->YcoeffsB[1] - p->buf[YADAPTCOEFFSB-1]	230	sub.l %d7, %a0 \| %a0 = p->YcoeffsB[1] - p->buf[YADAPTCOEFFSB-1]
230	sub.l %d6, %a1 \| a1 = p->YcoeffsB[2] - p->buf[YADAPTCOEFFSB-2]	231	sub.l %d6, %a1 \| %a1 = p->YcoeffsB[2] - p->buf[YADAPTCOEFFSB-2]
231	sub.l %d5, %a2 \| a2 = p->YcoeffsB[3] - p->buf[YADAPTCOEFFSB-3]	232	sub.l %d5, %a2 \| %a2 = p->YcoeffsB[3] - p->buf[YADAPTCOEFFSB-3]
232	sub.l %d4, %a3 \| a3 = p->YcoeffsB[4] - p->buf[YADAPTCOEFFSB-4]	233	sub.l %d4, %a3 \| %a3 = p->YcoeffsB[4] - p->buf[YADAPTCOEFFSB-4]
233		234
234	movem.l %d2/%a0-%a3, (YcoeffsB,%a6) \| Save p->YcoeffsB[]	235	movem.l %d2/%a0-%a3, (YcoeffsB,%a6) \| Save p->YcoeffsB[]
235		236
236	movem.l (YcoeffsA,%a6), %d4-%d7 \| d4 = p->YcoeffsA[0]	237	movem.l (YcoeffsA,%a6), %d4-%d7 \| %d4 = p->YcoeffsA[0]
237	\| d5 = p->YcoeffsA[1]	238	\| %d5 = p->YcoeffsA[1]
238	\| d6 = p->YcoeffsA[2]	239	\| %d6 = p->YcoeffsA[2]
239	\| d7 = p->YcoeffsA[3]	240	\| %d7 = p->YcoeffsA[3]
240		241
241	movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| d2 = p->buf[YADAPTCOEFFSA-3]	242	movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| %d2 = p->buf[YADAPTCOEFFSA-3]
242	\| a0 = p->buf[YADAPTCOEFFSA-2]	243	\| %a0 = p->buf[YADAPTCOEFFSA-2]
243	\| a1 = p->buf[YADAPTCOEFFSA-1]	244	\| %a1 = p->buf[YADAPTCOEFFSA-1]
244	\| a2 = p->buf[YADAPTCOEFFSA]	245	\| %a2 = p->buf[YADAPTCOEFFSA]
245		246
246	sub.l %a2, %d4 \| d4 = p->YcoeffsA[0] - p->buf[YADAPTCOEFFSA]	247	sub.l %a2, %d4 \| %d4 = p->YcoeffsA[0] - p->buf[YADAPTCOEFFSA]
247	sub.l %a1, %d5 \| d5 = p->YcoeffsA[1] - p->buf[YADAPTCOEFFSA-1]	248	sub.l %a1, %d5 \| %d5 = p->YcoeffsA[1] - p->buf[YADAPTCOEFFSA-1]
248	sub.l %a0, %d6 \| d6 = p->YcoeffsA[2] - p->buf[YADAPTCOEFFSA-2]	249	sub.l %a0, %d6 \| %d6 = p->YcoeffsA[2] - p->buf[YADAPTCOEFFSA-2]
249	sub.l %d2, %d7 \| d7 = p->YcoeffsA[3] - p->buf[YADAPTCOEFFSA-3]	250	sub.l %d2, %d7 \| %d7 = p->YcoeffsA[3] - p->buf[YADAPTCOEFFSA-3]
250		251
251	movem.l %d4-%d7, (YcoeffsA,%a6) \| Save p->YcoeffsA[]
252	bra.s 2f	252	bra.s 2f
253		253
254	1: \| *decoded0 < 0	254	1: \| *decoded0 < 0
255		255
256	add.l %d3, %d2 \| d2 = p->YcoeffsB[0] + p->buf[YADAPTCOEFFSB]	256	add.l %d3, %d2 \| %d2 = p->YcoeffsB[0] + p->buf[YADAPTCOEFFSB]
257	add.l %d7, %a0 \| a0 = p->YcoeffsB[1] + p->buf[YADAPTCOEFFSB-1]	257	add.l %d7, %a0 \| %a0 = p->YcoeffsB[1] + p->buf[YADAPTCOEFFSB-1]
258	add.l %d6, %a1 \| a1 = p->YcoeffsB[2] + p->buf[YADAPTCOEFFSB-2]	258	add.l %d6, %a1 \| %a1 = p->YcoeffsB[2] + p->buf[YADAPTCOEFFSB-2]
259	add.l %d5, %a2 \| a2 = p->YcoeffsB[3] + p->buf[YADAPTCOEFFSB-3]	259	add.l %d5, %a2 \| %a2 = p->YcoeffsB[3] + p->buf[YADAPTCOEFFSB-3]
260	add.l %d4, %a3 \| a3 = p->YcoeffsB[4] + p->buf[YADAPTCOEFFSB-4]	260	add.l %d4, %a3 \| %a3 = p->YcoeffsB[4] + p->buf[YADAPTCOEFFSB-4]
261		261
262	movem.l %d2/%a0-%a3, (YcoeffsB,%a6) \| Save p->YcoeffsB[]	262	movem.l %d2/%a0-%a3, (YcoeffsB,%a6) \| Save p->YcoeffsB[]
263		263
264	movem.l (YcoeffsA,%a6), %d4-%d7 \| d4 = p->YcoeffsA[0]	264	movem.l (YcoeffsA,%a6), %d4-%d7 \| %d4 = p->YcoeffsA[0]
265	\| d5 = p->YcoeffsA[1]	265	\| %d5 = p->YcoeffsA[1]
266	\| d6 = p->YcoeffsA[2]	266	\| %d6 = p->YcoeffsA[2]
267	\| d7 = p->YcoeffsA[3]	267	\| %d7 = p->YcoeffsA[3]
268		268
269	movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| d2 = p->buf[YADAPTCOEFFSA-3]	269	movem.l (YADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| %d2 = p->buf[YADAPTCOEFFSA-3]
270	\| a0 = p->buf[YADAPTCOEFFSA-2]	270	\| %a0 = p->buf[YADAPTCOEFFSA-2]
271	\| a1 = p->buf[YADAPTCOEFFSA-1]	271	\| %a1 = p->buf[YADAPTCOEFFSA-1]
272	\| a2 = p->buf[YADAPTCOEFFSA]	272	\| %a2 = p->buf[YADAPTCOEFFSA]
273		273
274	add.l %a2, %d4 \| d4 = p->YcoeffsA[0] + p->buf[YADAPTCOEFFSA]	274	add.l %a2, %d4 \| %d4 = p->YcoeffsA[0] + p->buf[YADAPTCOEFFSA]
275	add.l %a1, %d5 \| d5 = p->YcoeffsA[1] + p->buf[YADAPTCOEFFSA-1]	275	add.l %a1, %d5 \| %d5 = p->YcoeffsA[1] + p->buf[YADAPTCOEFFSA-1]
276	add.l %a0, %d6 \| d6 = p->YcoeffsA[2] + p->buf[YADAPTCOEFFSA-2]	276	add.l %a0, %d6 \| %d6 = p->YcoeffsA[2] + p->buf[YADAPTCOEFFSA-2]
277	add.l %d2, %d7 \| d7 = p->YcoeffsA[3] + p->buf[YADAPTCOEFFSA-3]	277	add.l %d2, %d7 \| %d7 = p->YcoeffsA[3] + p->buf[YADAPTCOEFFSA-3]
278
279	movem.l %d4-%d7, (YcoeffsA,%a6) \| Save p->YcoeffsA[]
280		278
281	2:	279	2:
		280	movem.l %d4-%d7, (YcoeffsA,%a6) \| Save p->YcoeffsA[]
		281
		282	3:
282		283
283	\| *** PREDICTOR X ***	284	\| *** PREDICTOR X ***
284		285
@@ -293,6 +294,8 @@ predictor_decode_stereo:
293	sub.l %d3, %d2	294	sub.l %d3, %d2
294	neg.l %d2 \| %d2 = %d3 -%d2	295	neg.l %d2 \| %d2 = %d3 -%d2
295		296
		297	move.l %d2, (XDELAYA-4,%a5) \| p->buf[XDELAYA-1] = %d2
		298
296	movem.l (XcoeffsA,%a6), %a0-%a3 \| %a0 = p->XcoeffsA[0]	299	movem.l (XcoeffsA,%a6), %a0-%a3 \| %a0 = p->XcoeffsA[0]
297	\| %a1 = p->XcoeffsA[1]	300	\| %a1 = p->XcoeffsA[1]
298	\| %a2 = p->XcoeffsA[2]	301	\| %a2 = p->XcoeffsA[2]
@@ -303,27 +306,26 @@ predictor_decode_stereo:
303	mac.l %d1, %a2, %acc0 \| %acc0 += p->buf[XDELAYA-2] * p->XcoeffsA[2]	306	mac.l %d1, %a2, %acc0 \| %acc0 += p->buf[XDELAYA-2] * p->XcoeffsA[2]
304	mac.l %d0, %a3, %acc0 \| %acc0 += p->buf[XDELAYA-3] * p->XcoeffsA[3]	307	mac.l %d0, %a3, %acc0 \| %acc0 += p->buf[XDELAYA-3] * p->XcoeffsA[3]
305		308
306	move.l %d2, (XDELAYA-4,%a5) \| p->buf[XDELAYA-1] = %d2
307	move.l %d3, (XDELAYA,%a5) \| p->buf[XDELAYA] = %d3	309	move.l %d3, (XDELAYA,%a5) \| p->buf[XDELAYA] = %d3
308		310
309	movclr.l %acc0, %d0
310
311	tst.l %d2	311	tst.l %d2
312	beq.s 1f	312	beq.s 1f
313	spl.b %d2 \| pos: 0x??????ff, neg: 0x??????00	313	spl.b %d2 \| pos: 0x??????ff, neg: 0x??????00
314	extb.l %d2 \| pos: 0xffffffff, neg: 0x00000000	314	extb.l %d2 \| pos: 0xffffffff, neg: 0x00000000
315	or.l #1, %d2 \| pos: 0xffffffff, neg: 0x00000001	315	or.l #1, %d2 \| pos: 0xffffffff, neg: 0x00000001
316	1: \| %d2 = SIGN(%d2)	316	1: \| %d2 = SIGN(%d2)
		317	move.l %d2, (XADAPTCOEFFSA-4,%a5) \| p->buf[XADAPTCOEFFSA-1] = %d2
		318
317	tst.l %d3	319	tst.l %d3
318	beq.s 1f	320	beq.s 1f
319	spl.b %d3	321	spl.b %d3
320	extb.l %d3	322	extb.l %d3
321	or.l #1, %d3	323	or.l #1, %d3
322	1: \| %d3 = SIGN(%d3)	324	1: \| %d3 = SIGN(%d3)
		325	move.l %d3, (XADAPTCOEFFSA,%a5) \| p->buf[XADAPTCOEFFSA] = %d3
		326
		327	movclr.l %acc0, %d0
323		328
324	move.l %d2, (XADAPTCOEFFSA-4,%a5) \| p->buf[XADAPTCOEFFSA-1] = r2
325	move.l %d3, (XADAPTCOEFFSA,%a5) \| p->buf[XADAPTCOEFFSA] = r3
326
327	\| NOTE: %d0 now contains predictionA - don't overwrite.	329	\| NOTE: %d0 now contains predictionA - don't overwrite.
328		330
329	\| Predictor X, Filter B	331	\| Predictor X, Filter B
@@ -345,11 +347,13 @@ predictor_decode_stereo:
345	sub.l %d3, %d7	347	sub.l %d3, %d7
346	neg.l %d7 \| %d7 = %d3 - %d7	348	neg.l %d7 \| %d7 = %d3 - %d7
347		349
348	movem.l (XcoeffsB,%a6), %d2/%a0-%a3 \| %d2 = p->XcoeffsB[0]	350	move.l %d7, (XDELAYB-4,%a5) \| p->buf[XDELAYB-1] = %d7
349	\| %a0 = p->XcoeffsB[1]	351
350	\| %a1 = p->XcoeffsB[2]	352	movem.l (XcoeffsB,%a6), %d2/%a0-%a3 \| %d2 = p->XcoeffsB[0]
351	\| %a2 = p->XcoeffsB[3]	353	\| %a0 = p->XcoeffsB[1]
352	\| %a3 = p->XcoeffsB[4]	354	\| %a1 = p->XcoeffsB[2]
		355	\| %a2 = p->XcoeffsB[3]
		356	\| %a3 = p->XcoeffsB[4]
353		357
354	mac.l %d3, %d2, %acc0 \| %acc0 = p->buf[XDELAYB] * p->XcoeffsB[0]	358	mac.l %d3, %d2, %acc0 \| %acc0 = p->buf[XDELAYB] * p->XcoeffsB[0]
355	mac.l %d7, %a0, %acc0 \| %acc0 += p->buf[XDELAYB-1] * p->XcoeffsB[1]	359	mac.l %d7, %a0, %acc0 \| %acc0 += p->buf[XDELAYB-1] * p->XcoeffsB[1]
@@ -357,27 +361,26 @@ predictor_decode_stereo:
357	mac.l %d5, %a2, %acc0 \| %acc0 += p->buf[XDELAYB-3] * p->XcoeffsB[3]	361	mac.l %d5, %a2, %acc0 \| %acc0 += p->buf[XDELAYB-3] * p->XcoeffsB[3]
358	mac.l %d4, %a3, %acc0 \| %acc0 += p->buf[XDELAYB-4] * p->XcoeffsB[4]	362	mac.l %d4, %a3, %acc0 \| %acc0 += p->buf[XDELAYB-4] * p->XcoeffsB[4]
359		363
360	move.l %d7, (XDELAYB-4,%a5) \| p->buf[XDELAYB-1] = %d7
361	move.l %d3, (XDELAYB, %a5) \| p->buf[XDELAYB] = %d3	364	move.l %d3, (XDELAYB, %a5) \| p->buf[XDELAYB] = %d3
362		365
363	movclr.l %acc0, %d1
364
365	tst.l %d7	366	tst.l %d7
366	beq.s 1f	367	beq.s 1f
367	spl.b %d7	368	spl.b %d7
368	extb.l %d7	369	extb.l %d7
369	or.l #1, %d7	370	or.l #1, %d7
370	1: \| %d7 = SIGN(%d7)	371	1: \| %d7 = SIGN(%d7)
		372	move.l %d7, (XADAPTCOEFFSB-4,%a5) \| p->buf[XADAPTCOEFFSB-1] = %d7
		373
371	tst.l %d3	374	tst.l %d3
372	beq.s 1f	375	beq.s 1f
373	spl.b %d3	376	spl.b %d3
374	extb.l %d3	377	extb.l %d3
375	or.l #1, %d3	378	or.l #1, %d3
376	1: \| %d3 = SIGN(%d3)	379	1: \| %d3 = SIGN(%d3)
377
378	move.l %d7, (XADAPTCOEFFSB-4,%a5) \| p->buf[XADAPTCOEFFSB-1] = %d7
379	move.l %d3, (XADAPTCOEFFSB, %a5) \| p->buf[XADAPTCOEFFSB] = %d3	380	move.l %d3, (XADAPTCOEFFSB, %a5) \| p->buf[XADAPTCOEFFSB] = %d3
380		381
		382	movclr.l %acc0, %d1
		383
381	\| %d0 still contains predictionA	384	\| %d0 still contains predictionA
382	\| %d1 contains predictionB	385	\| %d1 contains predictionB
383		386
@@ -405,76 +408,76 @@ predictor_decode_stereo:
405	\| %a4 contains decoded1	408	\| %a4 contains decoded1
406	\| %d5 contains *decoded1	409	\| %d5 contains *decoded1
407		410
408	\| %d2, %a0, %a1, %a2, %a31 contain p->XcoeffsB[0..4]	411	\| %d2, %a0, %a1, %a2, %a3 contain p->XcoeffsB[0..4]
409	\| %d7, %d3 contain p->buf[XADAPTCOEFFSB-1] and p->buf[XADAPTCOEFFSB]	412	\| %d7, %d3 contain p->buf[XADAPTCOEFFSB-1] and p->buf[XADAPTCOEFFSB]
410		413
411	move.l %d1, (%a4)+ \| *(decoded1++) = %d1 (p->XfilterA)	414	move.l %d1, (%a4)+ \| *(decoded1++) = %d1 (p->XfilterA)
412	move.l %a4, (4,%sp) \| save decoded1	415	move.l %a4, (4,%sp) \| save decoded1
413	tst.l %d5	416	tst.l %d5
414	beq.s 2f	417	beq.s 3f
415		418
416	movem.l (XADAPTCOEFFSB-16,%a5), %d4-%d6 \| d4 = p->buf[XADAPTCOEFFSB-4]	419	movem.l (XADAPTCOEFFSB-16,%a5), %d4-%d6 \| %d4 = p->buf[XADAPTCOEFFSB-4]
417	\| d5 = p->buf[XADAPTCOEFFSB-3]	420	\| %d5 = p->buf[XADAPTCOEFFSB-3]
418	\| d6 = p->buf[XADAPTCOEFFSB-2]	421	\| %d6 = p->buf[XADAPTCOEFFSB-2]
419		422
420	bmi.s 1f \| flags still valid here	423	bmi.s 1f \| flags still valid here
421		424
422	\| *decoded1 > 0	425	\| *decoded1 > 0
423		426
424	sub.l %d3, %d2 \| d2 = p->XcoeffsB[0] - p->buf[XADAPTCOEFFSB]	427	sub.l %d3, %d2 \| %d2 = p->XcoeffsB[0] - p->buf[XADAPTCOEFFSB]
425	sub.l %d7, %a0 \| a0 = p->XcoeffsB[1] - p->buf[XADAPTCOEFFSB-1]	428	sub.l %d7, %a0 \| %a0 = p->XcoeffsB[1] - p->buf[XADAPTCOEFFSB-1]
426	sub.l %d6, %a1 \| a1 = p->XcoeffsB[2] - p->buf[XADAPTCOEFFSB-2]	429	sub.l %d6, %a1 \| %a1 = p->XcoeffsB[2] - p->buf[XADAPTCOEFFSB-2]
427	sub.l %d5, %a2 \| a2 = p->XcoeffsB[3] - p->buf[XADAPTCOEFFSB-3]	430	sub.l %d5, %a2 \| %a2 = p->XcoeffsB[3] - p->buf[XADAPTCOEFFSB-3]
428	sub.l %d4, %a3 \| a3 = p->XcoeffsB[4] - p->buf[XADAPTCOEFFSB-4]	431	sub.l %d4, %a3 \| %a3 = p->XcoeffsB[4] - p->buf[XADAPTCOEFFSB-4]
429		432
430	movem.l %d2/%a0-%a3, (XcoeffsB,%a6) \| Save p->XcoeffsB[]	433	movem.l %d2/%a0-%a3, (XcoeffsB,%a6) \| Save p->XcoeffsB[]
431		434
432	movem.l (XcoeffsA,%a6), %d4-%d7 \| d4 = p->XcoeffsA[0]	435	movem.l (XcoeffsA,%a6), %d4-%d7 \| %d4 = p->XcoeffsA[0]
433	\| d5 = p->XcoeffsA[1]	436	\| %d5 = p->XcoeffsA[1]
434	\| d6 = p->XcoeffsA[2]	437	\| %d6 = p->XcoeffsA[2]
435	\| d7 = p->XcoeffsA[3]	438	\| %d7 = p->XcoeffsA[3]
436		439
437	movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| d2 = p->buf[XADAPTCOEFFSA-3]	440	movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| %d2 = p->buf[XADAPTCOEFFSA-3]
438	\| a0 = p->buf[XADAPTCOEFFSA-2]	441	\| %a0 = p->buf[XADAPTCOEFFSA-2]
439	\| a1 = p->buf[XADAPTCOEFFSA-1]	442	\| %a1 = p->buf[XADAPTCOEFFSA-1]
440	\| a2 = p->buf[XADAPTCOEFFSA]	443	\| %a2 = p->buf[XADAPTCOEFFSA]
441		444
442	sub.l %a2, %d4 \| d4 = p->XcoeffsA[0] - p->buf[XADAPTCOEFFSA]	445	sub.l %a2, %d4 \| %d4 = p->XcoeffsA[0] - p->buf[XADAPTCOEFFSA]
443	sub.l %a1, %d5 \| d5 = p->XcoeffsA[1] - p->buf[XADAPTCOEFFSA-1]	446	sub.l %a1, %d5 \| %d5 = p->XcoeffsA[1] - p->buf[XADAPTCOEFFSA-1]
444	sub.l %a0, %d6 \| d6 = p->XcoeffsA[2] - p->buf[XADAPTCOEFFSA-2]	447	sub.l %a0, %d6 \| %d6 = p->XcoeffsA[2] - p->buf[XADAPTCOEFFSA-2]
445	sub.l %d2, %d7 \| d7 = p->XcoeffsA[3] - p->buf[XADAPTCOEFFSA-3]	448	sub.l %d2, %d7 \| %d7 = p->XcoeffsA[3] - p->buf[XADAPTCOEFFSA-3]
446		449
447	movem.l %d4-%d7, (XcoeffsA,%a6) \| Save p->XcoeffsA[]
448	bra.s 2f	450	bra.s 2f
449		451
450	1: \| *decoded1 < 0	452	1: \| *decoded1 < 0
451		453
452	add.l %d3, %d2 \| d2 = p->XcoeffsB[0] + p->buf[XADAPTCOEFFSB]	454	add.l %d3, %d2 \| %d2 = p->XcoeffsB[0] + p->buf[XADAPTCOEFFSB]
453	add.l %d7, %a0 \| a0 = p->XcoeffsB[1] + p->buf[XADAPTCOEFFSB-1]	455	add.l %d7, %a0 \| %a0 = p->XcoeffsB[1] + p->buf[XADAPTCOEFFSB-1]
454	add.l %d6, %a1 \| a1 = p->XcoeffsB[2] + p->buf[XADAPTCOEFFSB-2]	456	add.l %d6, %a1 \| %a1 = p->XcoeffsB[2] + p->buf[XADAPTCOEFFSB-2]
455	add.l %d5, %a2 \| a2 = p->XcoeffsB[3] + p->buf[XADAPTCOEFFSB-3]	457	add.l %d5, %a2 \| %a2 = p->XcoeffsB[3] + p->buf[XADAPTCOEFFSB-3]
456	add.l %d4, %a3 \| a3 = p->XcoeffsB[4] + p->buf[XADAPTCOEFFSB-4]	458	add.l %d4, %a3 \| %a3 = p->XcoeffsB[4] + p->buf[XADAPTCOEFFSB-4]
457		459
458	movem.l %d2/%a0-%a3, (XcoeffsB,%a6) \| Save p->XcoeffsB[]	460	movem.l %d2/%a0-%a3, (XcoeffsB,%a6) \| Save p->XcoeffsB[]
459		461
460	movem.l (XcoeffsA,%a6), %d4-%d7 \| d4 = p->XcoeffsA[0]	462	movem.l (XcoeffsA,%a6), %d4-%d7 \| %d4 = p->XcoeffsA[0]
461	\| d5 = p->XcoeffsA[1]	463	\| %d5 = p->XcoeffsA[1]
462	\| d6 = p->XcoeffsA[2]	464	\| %d6 = p->XcoeffsA[2]
463	\| d7 = p->XcoeffsA[3]	465	\| %d7 = p->XcoeffsA[3]
464		466
465	movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| d2 = p->buf[XADAPTCOEFFSA-3]	467	movem.l (XADAPTCOEFFSA-12,%a5), %d2/%a0-%a2 \| %d2 = p->buf[XADAPTCOEFFSA-3]
466	\| a0 = p->buf[XADAPTCOEFFSA-2]	468	\| %a0 = p->buf[XADAPTCOEFFSA-2]
467	\| a1 = p->buf[XADAPTCOEFFSA-1]	469	\| %a1 = p->buf[XADAPTCOEFFSA-1]
468	\| a2 = p->buf[XADAPTCOEFFSA]	470	\| %a2 = p->buf[XADAPTCOEFFSA]
469		471
470	add.l %a2, %d4 \| d4 = p->XcoeffsA[0] + p->buf[XADAPTCOEFFSA]	472	add.l %a2, %d4 \| %d4 = p->XcoeffsA[0] + p->buf[XADAPTCOEFFSA]
471	add.l %a1, %d5 \| d5 = p->XcoeffsA[1] + p->buf[XADAPTCOEFFSA-1]	473	add.l %a1, %d5 \| %d5 = p->XcoeffsA[1] + p->buf[XADAPTCOEFFSA-1]
472	add.l %a0, %d6 \| d6 = p->XcoeffsA[2] + p->buf[XADAPTCOEFFSA-2]	474	add.l %a0, %d6 \| %d6 = p->XcoeffsA[2] + p->buf[XADAPTCOEFFSA-2]
473	add.l %d2, %d7 \| d7 = p->XcoeffsA[3] + p->buf[XADAPTCOEFFSA-3]	475	add.l %d2, %d7 \| %d7 = p->XcoeffsA[3] + p->buf[XADAPTCOEFFSA-3]
474
475	movem.l %d4-%d7, (XcoeffsA,%a6) \| Save p->XcoeffsA[]
476		476
477	2:	477	2:
		478	movem.l %d4-%d7, (XcoeffsA,%a6) \| Save p->XcoeffsA[]
		479
		480	3:
478		481
479	\| *** COMMON ***	482	\| *** COMMON ***
480		483
@@ -484,10 +487,10 @@ predictor_decode_stereo:
484		487
485	cmp.l %a3, %a5	488	cmp.l %a3, %a5
486	bne.s .endofloop	489	bne.s .endofloop
487		490
488	\| The history buffer is full, we need to do a memmove:	491	\| The history buffer is full, we need to do a memmove:
489		492
490	lea.l (historybuffer,%a6), %a3	493	lea.l (historybuffer,%a6), %a3
491		494
492	\| dest = %a3 (p->historybuffer)	495	\| dest = %a3 (p->historybuffer)
493	\| src = %a5 (p->buf)	496	\| src = %a5 (p->buf)
@@ -515,7 +518,7 @@ predictor_decode_stereo:
515	lea.l (40,%a3), %a3	518	lea.l (40,%a3), %a3
516		519
517	lea.l (historybuffer,%a6), %a5 \| p->buf = &p->historybuffer[0]	520	lea.l (historybuffer,%a6), %a5 \| p->buf = &p->historybuffer[0]
518		521
519	.endofloop:	522	.endofloop:
520	subq.l #1, (8,%sp) \| decrease loop count	523	subq.l #1, (8,%sp) \| decrease loop count
521	bne.w .loop	524	bne.w .loop