Another small optimisation for greyscale blitting on SH1 (making better use of page mode by reordering accesses).

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@16053 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 28 insertions, 27 deletions

diff --git a/firmware/target/sh/archos/lcd-as-archos-bitmap.S b/firmware/target/sh/archos/lcd-as-archos-bitmap.S
index a60ea9aad1..ae7b7f85cc 100644
--- a/firmware/target/sh/archos/lcd-as-archos-bitmap.S
+++ b/firmware/target/sh/archos/lcd-as-archos-bitmap.S
@@ -223,22 +223,22 @@ _lcd_write_data:
  *   r4 - current value address
  *   r5 - current phase address
  *   r6 - end address
- *   r7/r8 - current/next pixel phase
- *   r9 - current pixel value
- *   r10  - 0x00000080 \
- *   r11 - 0x00008000  > for phase sign check
- *   r12 - 0x00800000 /
+ *   r7/r8  - current/next block of phases (alternating)
+ *   r9/r10 - current blocks of values
+ *   r11 - 0x00000080 \
+ *   r12 - 0x00008000  > for phase sign check
+ *   r13 - 0x00800000 /
  */
 
 _lcd_grey_data:
     mov.l   r8, @-r15       /* save r8 */
-    shll2   r6              /*    v */
     mov.l   r9, @-r15       /* save r9 */
-    shll    r6              /* r6 *= 8; (8 pixels per block) */
     mov.l   r10, @-r15      /* save r10 */
-    add     r4, r6          /* end address */
+    shll2   r6              /*    v */
+    mov.l   r11, @-r15      /* save r11 */
+    shll    r6              /* r6 *= 8; (8 pixels per block) */
     mov.l   .lcdr, r3       /* put lcd data port address in r3 */
-    nop                     /* keep alignment */
+    add     r4, r6          /* end address */
 
     /* This code will fail if an interrupt changes the contents of PBDRL.
      * If so, we must disable the interrupt here. If disabling interrupts
@@ -248,17 +248,17 @@ _lcd_grey_data:
 
     mov.b   @r3, r0         /* r0 = PBDRL */
     or      #(LCD_DS|LCD_SD), r0    /* r0 |= LCD_DS|LCD_SD */
-    mov.l   r11, @-r15      /* save r11 */
-    and     #(~(LCD_CS|LCD_SC)), r0 /* r0 &= ~(LCD_CS|LCD_SC) */
     mov.l   r12, @-r15      /* save r12 */
+    and     #(~(LCD_CS|LCD_SC)), r0 /* r0 &= ~(LCD_CS|LCD_SC) */
+    mov.l   r13, @-r15      /* save r13 */
     neg     r0, r2          /* r2 = 0 - r0 */
 
     /* loop exploits that SD is on bit 0 for recorders and Ondios */
 
-    mov.w   .ptest, r10
-    swap.b  r10, r11
+    mov.w   .ptest, r11
+    swap.b  r11, r12
     mov.l   @r5, r7
-    swap.w  r10, r12
+    swap.w  r11, r13
     mov.l   .pmask, r0
 
 .greyloop:
@@ -269,66 +269,67 @@ _lcd_grey_data:
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    tst     r12, r7
-    mov.l   @(4,r5), r8
+    tst     r13, r7
+    mov.l   @r4+, r10
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    tst     r11, r7
+    tst     r12, r7
+    mov.l   @(4,r5), r8
     negc    r2, r1
-    tst     r10, r7
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
+    tst     r11, r7
     or      r0, r7
-    sub     r9, r7
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
     cmp/pz  r8
-    mov.l   r7, @r5
+    sub     r9, r7
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    tst     r12, r8
-    mov.l   @r4+, r9
+    tst     r13, r8
+    mov.l   r7, @r5
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    tst     r11, r8
+    tst     r12, r8
     mov.l   @(8,r5), r7
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    tst     r10, r8
+    tst     r11, r8
     or      r0, r8
     negc    r2, r1
     mov.b   r1, @r3
     add     #(LCD_SC), r1
     mov.b   r1, @r3
 
-    sub     r9, r8
+    sub     r10, r8
     mov.l   r8, @(4,r5)
     
     add     #8, r5
     cmp/hi  r4, r6
     bt      .greyloop
 
-    mov.l   @r15+, r12      /* restore r12 */
+    mov.l   @r15+, r13      /* restore r13 */
     mov     #(LCD_CS|LCD_DS|LCD_SD|LCD_SC), r0
-    mov.l   @r15+, r11      /* restore r11 */
+    mov.l   @r15+, r12      /* restore r12 */
     or      r0, r1          /* restore port */
+    mov.l   @r15+, r11      /* restore r11 */
     mov.l   @r15+, r10      /* restore r10 */
     mov.l   @r15+, r9       /* restore r9 */
     mov.l   @r15+, r8       /* restore r8 */