improve sh/hi levels

[picodrive.git] / pico / draw_arm.s

diff --git a/pico/draw_arm.s b/pico/draw_arm.s
index de40eaa..82bb603 100644 (file)
--- a/pico/draw_arm.s
+++ b/pico/draw_arm.s
@@ -1595,53 +1595,26 @@ FinalizeLineBGR444:
 @ @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@\r
 \r
 \r
-@ hilights 2 pixels in RGB555/BGR555 format\r
-.macro TileDoShHi2Pixels555 reg\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #27\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #26\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #27\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #27\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #26\r
-    adds    \reg, \reg, #0x40000000\r
-    orrcs   \reg, \reg, #0xf0000000\r
-    mov     \reg, \reg, ror #27\r
-.endm\r
-\r
-\r
 @ Convert 0000bbb0 ggg0rrr0\r
 @ to      rrrrrggg gggbbbbb\r
 \r
-@ r2,r3,r9 - scratch, lr = 0x001c001c, r8 = 0x00030003\r
+@ r2,r3 - scratch, lr = 0x001c001c, r8 = 0x08610861\r
 .macro convRGB565 reg\r
-    and     r2,   lr,   \reg,lsl #1\r
-    and     r9,   r8,   \reg,lsr #2\r
-    orr     r2,   r2,   r9           @ r2=red\r
-    and     r3,   lr,   \reg,lsr #7\r
-    and     r9,   r8,   \reg,lsr #10\r
-    orr     r3,   r3,   r9           @ r3=blue\r
-    and     \reg, \reg, lr,  lsl #3\r
-    orr     \reg, \reg, \reg,lsl #3  @ green\r
-    orr     \reg, \reg, r2,  lsl #11 @ add red back\r
-    orr     \reg, \reg, r3           @ add blue back\r
+    and     r2,   lr,   \reg,lsr #7  @ b\r
+    and     r3,   lr,   \reg,lsr #3  @ g\r
+    and     \reg, lr,   \reg,lsl #1  @ r\r
+    orr     r2,   r2,   r3,  lsl #6\r
+    orr     \reg, r2,   \reg,lsl #11\r
+\r
+    and     r2,   r8,   \reg,lsr #4\r
+    orr     \reg, \reg, r2\r
 .endm\r
 \r
-@ trashes: r2-r9,r12,lr; r0,r1 are advanced\r
+@ trashes: r2-r8,r12,lr; r8 = 0x08610861; r0,r1 are advanced\r
 .macro vidConvCpyRGB565_local\r
     mov     r12, r2, lsr #3  @ repeats\r
     mov     lr, #0x001c0000\r
     orr     lr, lr,  #0x01c  @ lr == pattern 0x001c001c\r
-    mov     r8, #0x00030000\r
-    orr     r8, r8,  #0x003\r
 \r
 0:\r
     ldmia   r1!, {r4-r7}\r
@@ -1663,6 +1636,9 @@ FinalizeLineBGR444:
 \r
 vidConvCpyRGB565: @ void *to, void *from, int pixels\r
     stmfd   sp!, {r4-r9,lr}\r
+    mov     r8,     #0x0061\r
+    orr     r8, r8, #0x0800\r
+    orr     r8, r8, lsl #16\r
     vidConvCpyRGB565_local\r
     ldmfd   sp!, {r4-r9,lr}\r
     bx      lr\r
@@ -1676,9 +1652,8 @@ PicoDoHighPal555:
     ldr     r8, =(Pico+0x22228)  @ Pico.video\r
 \r
 PicoDoHighPal555_nopush:\r
-    str     r1, [sp, #-8]        @ is called from FinalizeLine555?\r
+    orr     r9, r1, r0, lsl #31  @ 0:called from FinalizeLine555, 31: s/h\r
 \r
-    str     r0, [sp, #-4]\r
     ldr     r0, =HighPal\r
 \r
     mov     r1, #0\r
@@ -1686,10 +1661,13 @@ PicoDoHighPal555_nopush:
 \r
     sub     r1, r8, #0x128       @ r1=Pico.cram\r
     mov     r2, #0x40\r
+    mov     r8,     #0x0061\r
+    orr     r8, r8, #0x0800\r
+    orr     r8, r8, lsl #16\r
+\r
     vidConvCpyRGB565_local\r
 \r
-    ldr     r0, [sp, #-4]\r
-    tst     r0, r0\r
+    tst     r9, #(1<<31)\r
     beq     PicoDoHighPal555_end\r
 \r
     ldr     r3, =HighPal\r
@@ -1711,19 +1689,28 @@ PicoDoHighPal555_nopush:
     bne     .fl_loopcpRGB555_sh\r
 \r
     @ hilighted pixels:\r
+    @  t = ((dpal[i] >> 1) & 0x738e738e) + 0x738e738e;\r
+    @  t |= (t >> 4) & 0x08610861;\r
+    @ r8=0x08610861\r
     sub     r3, r3, #0x40*2\r
-    mov     lr, #0x40/2\r
+    mov     lr, #0x40/4\r
 .fl_loopcpRGB555_hi:\r
-    ldr     r1, [r3], #4\r
-    TileDoShHi2Pixels555 r1\r
-    str     r1, [r4], #4\r
+    ldmia   r3!, {r1,r6}\r
+    and     r1, r12, r1, lsr #1\r
+    and     r6, r12, r6, lsr #1\r
+    add     r1, r12, r1\r
+    add     r6, r12, r6\r
+    and     r5, r8, r1, lsr #4\r
+    and     r7, r8, r6, lsr #4\r
+    orr     r1, r1, r5\r
+    orr     r6, r6, r7\r
+    stmia   r4!, {r1,r6}\r
     subs    lr, lr, #1\r
     bne     .fl_loopcpRGB555_hi\r
     mov     r0, #1\r
 \r
 PicoDoHighPal555_end:\r
-    ldr     r1, [sp, #-8]\r
-    tst     r1, r1\r
+    tst     r9, #1\r
     ldmeqfd sp!, {r4-r9,pc}\r
 \r
     ldr     r8, =(Pico+0x22228)  @ Pico.video\r