H300: lcd_write_data() in pure asm, essentially the same speed as before, but working correctly for all allowed data lengths. Removed RAM waste by declaring inline functions static.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@8575 a1c6a512-1295-4272-9138-f99709370657

2 files changed, 122 insertions, 86 deletions

diff --git a/firmware/drivers/lcd-h300.c b/firmware/drivers/lcd-h300.c
index d84edd6b2e..8ee6b1c47e 100644
--- a/firmware/drivers/lcd-h300.c
+++ b/firmware/drivers/lcd-h300.c
@@ -77,55 +77,18 @@ static bool display_on=false; /* is the display turned on? */
 
 
 /* called very frequently - inline! */
-inline void lcd_write_reg(int reg, int val)
+static inline void lcd_write_reg(int reg, int val)
 {
     *(volatile unsigned short *)0xf0000000 = reg;
     *(volatile unsigned short *)0xf0000002 = val;
 }
 
 /* called very frequently - inline! */
-inline void lcd_begin_write_gram(void)
+static inline void lcd_begin_write_gram(void)
 {
     *(volatile unsigned short *)0xf0000000 = R_WRITE_DATA_2_GRAM;
 }
 
-/* called very frequently - inline! */
-inline void lcd_write_data(const unsigned short* p_bytes, int count) ICODE_ATTR;
-inline void lcd_write_data(const unsigned short* p_bytes, int count)
-{
-    int precount = ((-(size_t)p_bytes) & 0xf) / 2;
-    count -= precount;
-    while(precount--)
-        *(volatile unsigned short *)0xf0000002 = *p_bytes++;
-    while((count -= 8) >= 0) asm (
-            "\n\tmovem.l (%[p_bytes]),%%d1-%%d4\
-             \n\tswap %%d1\
-             \n\tmove.w %%d1,(%[dest])\
-             \n\tswap %%d1\
-             \n\tmove.w %%d1,(%[dest])\
-             \n\tswap %%d2\
-             \n\tmove.w %%d2,(%[dest])\
-             \n\tswap %%d2\
-             \n\tmove.w %%d2,(%[dest])\
-             \n\tswap %%d3\
-             \n\tmove.w %%d3,(%[dest])\
-             \n\tswap %%d3\
-             \n\tmove.w %%d3,(%[dest])\
-             \n\tswap %%d4\
-             \n\tmove.w %%d4,(%[dest])\
-             \n\tswap %%d4\
-             \n\tmove.w %%d4,(%[dest])\
-             \n\tlea.l (16,%[p_bytes]),%[p_bytes]"
-             : [p_bytes] "+a" (p_bytes)
-             : [dest] "a" ((volatile unsigned short *)0xf0000002)
-             : "d1", "d2", "d3", "d4", "memory");
-    if (count != 0) {
-        count += 8;
-        while(count--)
-            *(volatile unsigned short *)0xf0000002 = *p_bytes++;
-    }
-}
-
 /*** hardware configuration ***/
 
 void lcd_set_contrast(int val)
@@ -406,8 +369,8 @@ void lcd_update_rect(int x, int y, int width, int height)
         lcd_write_reg(R_HORIZ_RAM_ADDR_POS, (ymax<<8) | y); 
         
         /* vert ram addr */ 
-        lcd_write_reg(R_VERT_RAM_ADDR_POS,((x+width-1)<<8) | x); 
-        lcd_write_reg(R_RAM_ADDR_SET, (x<<8) | y); 
+        lcd_write_reg(R_VERT_RAM_ADDR_POS,((x+width-1)<<8) | x);
+        lcd_write_reg(R_RAM_ADDR_SET, (x<<8) | y);
         lcd_begin_write_gram(); 
 
         /* Copy specified rectangle bitmap to hardware */ 
@@ -421,6 +384,6 @@ void lcd_update_rect(int x, int y, int width, int height)
         lcd_write_reg(R_HORIZ_RAM_ADDR_POS, 0xaf00); 
 
         /* vert ram addr: 0 - 219 */  
-        lcd_write_reg(R_VERT_RAM_ADDR_POS, 0xdb00); 
+        lcd_write_reg(R_VERT_RAM_ADDR_POS, 0xdb00);
     }
 }
diff --git a/firmware/drivers/lcd.S b/firmware/drivers/lcd.S
index 79c6e32f33..2b50de5190 100755
--- a/firmware/drivers/lcd.S
+++ b/firmware/drivers/lcd.S
@@ -313,62 +313,135 @@ _lcd_write_data:
     .type       lcd_write_command,@function
 
 lcd_write_command:
-        move.l  (4,%sp),%d0
-        lea MBAR2,%a1
-        move.l   #~8,%d1
-        and.l   %d1,(0xb4,%a1)
-        move.w  %d0,0xf0000000
-        rts
+    move.l  (4,%sp),%d0
+    lea MBAR2,%a1
+    move.l   #~8,%d1
+    and.l   %d1,(0xb4,%a1)
+    move.w  %d0,0xf0000000
+    rts
 
     .align      2
     .global     lcd_write_command_ex
     .type       lcd_write_command_ex,@function
 
 lcd_write_command_ex:
-        lea MBAR2,%a1
-        
-        move.l  (4,%sp),%d0     /* Command */
-
-        move.l   #~8,%d1        /* Set A0 = 0 */
-        and.l   %d1,(0xb4,%a1)
-        move.w  %d0,0xf0000000  /* Write to LCD */
-
-        not.l   %d1             /* Set A0 = 1 */
-        or.l    %d1,(0xb4,%a1)
-
-        move.l  (8,%sp),%d0     /* Data */
-        cmp.l   #0xffffffff,%d0 /* -1? */
-        beq.b   .last
-        move.w  %d0,0xf0000000  /* Write to LCD */
-        
-        move.l  (12,%sp),%d0     /* Data */
-        cmp.l   #0xffffffff,%d0 /* -1? */
-        beq.b   .last
-        move.w  %d0,0xf0000000  /* Write to LCD */
-        
-.last:  
-        rts
+    lea MBAR2,%a1
+
+    move.l  (4,%sp),%d0     /* Command */
+
+    move.l   #~8,%d1        /* Set A0 = 0 */
+    and.l   %d1,(0xb4,%a1)
+    move.w  %d0,0xf0000000  /* Write to LCD */
+
+    not.l   %d1             /* Set A0 = 1 */
+    or.l    %d1,(0xb4,%a1)
+
+    move.l  (8,%sp),%d0     /* Data */
+    cmp.l   #0xffffffff,%d0 /* -1? */
+    beq.b   .last
+    move.w  %d0,0xf0000000  /* Write to LCD */
+
+    move.l  (12,%sp),%d0    /* Data */
+    cmp.l   #0xffffffff,%d0 /* -1? */
+    beq.b   .last
+    move.w  %d0,0xf0000000  /* Write to LCD */
+
+.last:
+    rts
 
     .align      2
     .global     lcd_write_data
     .type       lcd_write_data,@function
 
 lcd_write_data:
-        move.l  (4,%sp),%a0       /* Data pointer */
-        move.l  (8,%sp),%d0     /* Length */
-        lea MBAR2,%a1
-        moveq   #8,%d1
-        or.l    %d1,(0xb4,%a1)
+    move.l  (4,%sp),%a0     /* Data pointer */
+    move.l  (8,%sp),%d0     /* Length */
+    lea MBAR2,%a1
+    moveq   #8,%d1
+    or.l    %d1,(0xb4,%a1)
 
-        lea     0xf0000000,%a1
+    lea     0xf0000000,%a1
 .loop:
-        /* When running in IRAM, this loop takes 7 cycles plus the LCD write.
-           The 7 cycles are necessary to follow the LCD timing specs
-           at 140MHz */
-        move.b  (%a0)+,%d1  /* 3(1/0) */
-        move.w  %d1,(%a1)   /* 1(0/1) */
-        subq.l  #1,%d0      /* 1(0/0) */
-        nop                 /* 1(0/0) */
-        bne     .loop       /* 2(0/0) */
-        rts
+    /* When running in IRAM, this loop takes 7 cycles plus the LCD write.
+       The 7 cycles are necessary to follow the LCD timing specs
+       at 140MHz */
+    move.b  (%a0)+,%d1      /* 3(1/0) */
+    move.w  %d1,(%a1)       /* 1(0/1) */
+    subq.l  #1,%d0          /* 1(0/0) */
+    nop                     /* 1(0/0) */
+    bne     .loop           /* 2(0/0) */
+    rts
+#elif defined(IRIVER_H300_SERIES)
+    .section    .icode,"ax",@progbits
+
+    .align      2
+    .global     lcd_write_data
+    .type       lcd_write_data,@function
+
+lcd_write_data:
+    move.l  (4,%sp),%a0     /* data pointer */
+    move.l  (8,%sp),%d0     /* length in words */
+    add.l   %d0,%d0         /* words -> bytes */
+    add.l   %a0,%d0         /* -> end address */
+    lea.l   0xf0000002,%a1  /* LCD data port */
+    
+    moveq.l #31,%d1
+    add.l   %a0,%d1
+    and.l   #0xFFFFFFF0,%d1 /* %d1 = first line bound + 16 */
+    cmp.l   %d1,%d0         /* at least one full line to send? */
+    blo.b   .words2_loop    /* no: skip to word loop */
+    
+    subq.l  #8,%d1
+    subq.l  #8,%d1          /* %d1 = first line bound */
+
+    cmp.l   %a0,%d1         /* any leading words? */
+    bls.b   .words1_end     /* no: skip leading word loop */
+    
+.words1_loop:
+    move.w  (%a0)+,(%a1)    /* transfer word */
+    cmp.l   %a0,%d1         /* run %a0 up to first line bound */
+    bhi.b   .words1_loop
+    
+.words1_end:
+    lea.l   (-16,%sp),%sp   /* free up some registers */
+    movem.l %d2-%d4/%a2,(%sp)
+    move.l  %d0,%a2
+    lea.l   (-15,%a2),%a2   /* %a2 = end address - 15 (one line/pass) */
+    
+    /* burst-optimised line transfers */
+.line_loop:
+    movem.l (%a0),%d1-%d4   /* burst-read line */
+    lea.l   (16,%a0),%a0    /* increment address */
+    swap    %d1             /* send data to LCD in correct order... */
+    move.w  %d1,(%a1)
+    swap    %d1
+    move.w  %d1,(%a1)
+    swap    %d2
+    move.w  %d2,(%a1)
+    swap    %d2
+    move.w  %d2,(%a1)
+    swap    %d3
+    move.w  %d3,(%a1)
+    swap    %d3
+    move.w  %d3,(%a1)
+    swap    %d4
+    move.w  %d4,(%a1)
+    swap    %d4
+    move.w  %d4,(%a1)
+    cmp.l   %a0,%a2         /* run %a0 up to last line bound */
+    bhi.b   .line_loop
+    
+    movem.l (%sp),%d2-%d4/%a2
+    lea.l   (16,%sp),%sp    /* restore registers */
+    
+    cmp.l   %a0,%d0         /* any trailing words? */
+    bls.b   .words2_end     /* no: get otta here */
+
+.words2_loop:
+    move.w  (%a0)+,(%a1)    /* transfer word */
+    cmp.l   %a0,%d0         /* run %a0 up to end address */
+    bhi.b   .words2_loop
+
+.words2_end:
+    rts
 #endif