1
0
Fork 0
forked from len0rd/rockbox

* ARM asm DSP and codec/plugin functions: Use r12 scratch register properly

* Fix saving another unused reg in dsp code
* Use less regs in the generic ARM mpegplayer adding idct pure DC case
* Fix ARMv6 mpegplayer adding idct using an unsaved register in pure DC case


git-svn-id: svn://svn.rockbox.org/rockbox/trunk@21803 a1c6a512-1295-4272-9138-f99709370657
This commit is contained in:
Jens Arnold 2009-07-12 13:14:35 +00:00
parent e12c1c0a62
commit 02c031709c
6 changed files with 125 additions and 127 deletions

View file

@ -268,7 +268,7 @@
bx lr
mpeg2_idct_copy:
stmfd sp!, { r1-r2, r4-r12, lr }
stmfd sp!, { r1-r2, r4-r11, lr }
bl .idct
ldmfd sp!, { r1-r2 }
mov r11, #0
@ -313,7 +313,7 @@ mpeg2_idct_copy:
add r1, r1, r2
cmp r0, r12
blo 1b
ldmfd sp!, { r4-r12, pc }
ldmfd sp!, { r4-r11, pc }
mpeg2_idct_add:
cmp r0, #129
@ -324,7 +324,7 @@ mpeg2_idct_add:
cmp r1, #0x40
bne 3f
1:
stmfd sp!, { r2-r12, lr }
stmfd sp!, { r2-r11, lr }
bl .idct
ldmfd sp!, { r1-r2 }
mov r11, #0
@ -385,24 +385,20 @@ mpeg2_idct_add:
add r1, r1, r2
cmp r0, r12
blo 2b
ldmfd sp!, { r4-r12, pc }
ldmfd sp!, { r4-r11, pc }
3:
stmfd sp!, { r4-r11 }
stmfd sp!, { r4-r6, lr }
ldrsh r1, [r0, #0] /* r1 = block[0] */
mov r11, #0
strh r11, [r0, #0] /* block[0] = 0 */
strh r11, [r0, #126] /* block[63] = 0 */
mov r4, #0
strh r4, [r0, #0] /* block[0] = 0 */
strh r4, [r0, #126] /* block[63] = 0 */
add r1, r1, #64 /* r1 = DC << 7 */
add r0, r2, r3, asl #3
4:
ldrb r4, [r2, #0]
ldrb r5, [r2, #1]
ldrb r6, [r2, #2]
ldrb r7, [r2, #3]
ldrb r8, [r2, #4]
ldrb r9, [r2, #5]
ldrb r10, [r2, #6]
ldrb r11, [r2, #7]
ldrb lr, [r2, #3]
add r4, r4, r1, asr #7
cmp r4, #255
mvnhi r4, r4, asr #31
@ -415,28 +411,31 @@ mpeg2_idct_add:
cmp r6, #255
mvnhi r6, r6, asr #31
strb r6, [r2, #2]
add r7, r7, r1, asr #7
cmp r7, #255
mvnhi r7, r7, asr #31
strb r7, [r2, #3]
add r8, r8, r1, asr #7
cmp r8, #255
mvnhi r8, r8, asr #31
strb r8, [r2, #4]
add r9, r9, r1, asr #7
cmp r9, #255
mvnhi r9, r9, asr #31
strb r9, [r2, #5]
add r10, r10, r1, asr #7
cmp r10, #255
mvnhi r10, r10, asr #31
strb r10, [r2, #6]
add r11, r11, r1, asr #7
cmp r11, #255
mvnhi r11, r11, asr #31
strb r11, [r2, #7]
add lr, lr, r1, asr #7
cmp lr, #255
mvnhi lr, lr, asr #31
strb lr, [r2, #3]
ldrb r4, [r2, #4]
ldrb r5, [r2, #5]
ldrb r6, [r2, #6]
ldrb lr, [r2, #7]
add r4, r4, r1, asr #7
cmp r4, #255
mvnhi r4, r4, asr #31
strb r4, [r2, #4]
add r5, r5, r1, asr #7
cmp r5, #255
mvnhi r5, r5, asr #31
strb r5, [r2, #5]
add r6, r6, r1, asr #7
cmp r6, #255
mvnhi r6, r6, asr #31
strb r6, [r2, #6]
add lr, lr, r1, asr #7
cmp lr, #255
mvnhi lr, lr, asr #31
strb lr, [r2, #7]
add r2, r2, r3
cmp r2, r0
blo 4b
ldmfd sp!, { r4-r11 }
bx lr
ldmfd sp!, { r4-r6, pc }

View file

@ -196,7 +196,7 @@ L_W0246:
mpeg2_idct_copy:
stmfd sp!, {r1-r2, r4-r12, lr}
stmfd sp!, {r1-r2, r4-r11, lr}
bl .idct
ldmfd sp!, {r1-r2}
@ -230,7 +230,7 @@ mpeg2_idct_copy:
cmp r0, r12
blo 1b
ldmfd sp!, {r4-r12, pc}
ldmfd sp!, {r4-r11, pc}
mpeg2_idct_add:
cmp r0, #129
@ -241,7 +241,7 @@ mpeg2_idct_add:
cmp r1, #0x40
bne 3f
1:
stmfd sp!, {r2-r12, lr}
stmfd sp!, {r2-r11, lr}
bl .idct
ldmfd sp!, {r1-r2}
mov r11, #0
@ -287,21 +287,21 @@ mpeg2_idct_add:
add r1, r1, r2
cmp r0, r12
blo 2b
ldmfd sp!, {r4-r12, pc}
ldmfd sp!, {r4-r11, pc}
3:
stmfd sp!, {r4-r7}
stmfd sp!, {r4-r6, lr}
ldrsh r1, [r0, #0] /* r1 = block[0] */
mov r11, #0
strh r11, [r0, #0] /* block[0] = 0 */
strh r11, [r0, #126] /* block[63] = 0 */
mov r4, #0
strh r4, [r0, #0] /* block[0] = 0 */
strh r4, [r0, #126] /* block[63] = 0 */
add r1, r1, #64 /* r1 = DC << 7 */
add r0, r2, r3, asl #3
4:
ldrb r4, [r2, #0]
ldrb r5, [r2, #1]
ldrb r6, [r2, #2]
ldrb r7, [r2, #3]
ldrb lr, [r2, #3]
add r4, r4, r1, asr #7
usat r4, #8, r4
strb r4, [r2, #0]
@ -311,13 +311,13 @@ mpeg2_idct_add:
add r6, r6, r1, asr #7
usat r6, #8, r6
strb r6, [r2, #2]
add r7, r7, r1, asr #7
usat r7, #8, r7
strb r7, [r2, #3]
add lr, lr, r1, asr #7
usat lr, #8, lr
strb lr, [r2, #3]
ldrb r4, [r2, #4]
ldrb r5, [r2, #5]
ldrb r6, [r2, #6]
ldrb r7, [r2, #7]
ldrb lr, [r2, #7]
add r4, r4, r1, asr #7
usat r4, #8, r4
strb r4, [r2, #4]
@ -327,11 +327,10 @@ mpeg2_idct_add:
add r6, r6, r1, asr #7
usat r6, #8, r6
strb r6, [r2, #6]
add r7, r7, r1, asr #7
usat r7, #8, r7
strb r7, [r2, #7]
add lr, lr, r1, asr #7
usat lr, #8, lr
strb lr, [r2, #7]
add r2, r2, r3
cmp r2, r0
blo 4b
ldmfd sp!, {r4-r7}
bx lr
ldmfd sp!, {r4-r6, pc}