fixed mp3 buffer overlapping of GP2X sound buffer

[picodrive.git] / platform / gp2x / code940 / 940init.s

@ vim:filetype=armasm:\r
\r
@ ARM940 initialization.\r
@ Based on ogg940 code by Dzz.\r
@ (c) Copyright 2007, Grazvydas "notaz" Ignotas\r
\r
.equ mmsp2_regs, (0xc0000000-0x02000000) @ assume we live @ 0x2000000 bank\r
.equ shared_ctl,  0x00200000             @ this is where shared_ctl struncture is located\r
\r
\r
@ exception table:\r
.global code940\r
code940:\r
    b .b_reset                    @ reset\r
    b .b_undef                    @ undefined instructions\r
    b .b_swi                      @ software interrupt\r
    b .b_pabort                   @ prefetch abort\r
    b .b_dabort                   @ data abort\r
    b .b_reserved                 @ reserved\r
    b .b_irq                      @ IRQ\r
    b .b_fiq                      @ FIQ\r
\r
@ test\r
.b_reset:\r
    mov     r12, #0\r
    b       .Begin\r
.b_undef:\r
    mov     r12, #1\r
    b       .Begin\r
.b_swi:\r
    mov     r12, #2\r
    b       .Begin\r
.b_pabort:\r
    mov     r12, #3\r
    b       .Begin\r
.b_dabort:\r
    mov     r12, #4\r
    b       .Begin\r
.b_reserved:\r
    mov     r12, #5\r
    b       .Begin\r
.b_irq:\r
    mov     sp, #0x100000       @ reset stack\r
    sub     sp, sp, #4\r
    mov     r0, #shared_ctl     @ remember where we were when interrupt happened\r
    add     r0, r0, #0x20\r
    str     lr, [r0]\r
    mov     r0, #shared_ctl     @ increment exception counter (for debug)\r
    add     r0, r0, #(6*4)\r
    ldr     r1, [r0]\r
    add     r1, r1, #1\r
    str     r1, [r0]\r
\r
    bl Main940\r
\r
    @ we should never get here\r
    b .b_reserved\r
\r
\r
.b_fiq:\r
    mov     r12, #7\r
    b       .Begin\r
\r
.Begin:\r
    mov sp, #0x100000           @ set the stack top (1M)\r
    sub sp, sp, #4              @ minus 4\r
\r
    @ set up memory region 0 -- the whole 4GB address space\r
    mov r0, #(0x1f<<1)|1        @ region data\r
    mcr p15, 0, r0, c6, c0, 0   @ opcode2 ~ data/instr\r
    mcr p15, 0, r0, c6, c0, 1\r
\r
    @ set up region 1 which is the first 2 megabytes.\r
    mov r0, #(0x14<<1)|1        @ region data\r
    mcr p15, 0, r0, c6, c1, 0\r
    mcr p15, 0, r0, c6, c1, 1\r
\r
    @ set up region 2: 64k 0x200000-0x210000\r
    mov r0, #(0x0f<<1)|1\r
    orr r0, r0, #0x200000\r
    mcr p15, 0, r0, c6, c2, 0\r
    mcr p15, 0, r0, c6, c2, 1\r
\r
    @ set up region 3: 64k 0xbe000000-0xbe010000 (hw control registers)\r
    mov r0, #(0x0f<<1)|1\r
    orr r0, r0, #mmsp2_regs\r
    mcr p15, 0, r0, c6, c3, 0\r
    mcr p15, 0, r0, c6, c3, 1\r
\r
    @ region 4: 4K 0x00000000-0x00001000 (boot code protection region)\r
    mov r0, #(0x0b<<1)|1\r
    mcr p15, 0, r0, c6, c4, 0\r
    mcr p15, 0, r0, c6, c4, 1\r
\r
    @ region 5: 4M 0x00400000-0x00800000 (mp3 area part1)\r
    mov r0, #(0x15<<1)|1\r
    orr r0, r0, #0x00400000\r
    mcr p15, 0, r0, c6, c5, 0\r
    mcr p15, 0, r0, c6, c5, 1\r
\r
    @ region 6: 8M 0x00800000-0x01000000 (mp3 area part2)\r
    mov r0, #(0x16<<1)|1\r
    orr r0, r0, #0x00800000\r
    mcr p15, 0, r0, c6, c6, 0\r
    mcr p15, 0, r0, c6, c6, 1\r
\r
    @ set regions 1, 4, 5 and 6 to be cacheable (so the first 2M and mp3 area will be cacheable)\r
    mov r0, #(1<<1)|(1<<4)|(1<<5)|(1<<6)\r
    mcr p15, 0, r0, c2, c0, 0\r
    mcr p15, 0, r0, c2, c0, 1\r
\r
    @ set region 1 to be bufferable too (only data)\r
    mov r0, #(1<<1)\r
    mcr p15, 0, r0, c3, c0, 0\r
\r
    @ set access protection\r
    @ data: [full, full, no, full, full, full, no access] for regions [6 5 4 3 2 1 0]\r
    mov r0,     #       (3<<12)|(3<<10)|(0<<8)\r
    orr r0, r0, #(3<<6)|(3<< 4)|(3<< 2)|(0<<0)\r
    mcr p15, 0, r0, c5, c0, 0\r
    @ instructions: [no, no, full, no, no, full, no]\r
    mov r0,     #       (0<<12)|(0<<10)|(3<<8)\r
    orr r0, r0, #(0<<6)|(0<< 4)|(3<< 2)|(0<<0)\r
    mcr p15, 0, r0, c5, c0, 1\r
\r
    mrc p15, 0, r0, c1, c0, 0   @ fetch current control reg\r
    orr r0, r0, #1              @ 0x00000001: enable protection unit\r
    orr r0, r0, #4              @ 0x00000004: enable D cache\r
    orr r0, r0, #0x1000         @ 0x00001000: enable I cache\r
@    bic r0, r0, #0xC0000000\r
@    orr r0, r0, #0x40000000     @ 0x40000000: synchronous, faster?\r
    orr r0, r0, #0xC0000000     @ 0xC0000000: async\r
    mcr p15, 0, r0, c1, c0, 0   @ set control reg\r
\r
    @ flush (invalidate) the cache (just in case)\r
    mov r0, #0\r
    mcr p15, 0, r0, c7, c6, 0\r
\r
    @ remember which exception vector we came from (increment counter for debug)\r
    mov     r0, #shared_ctl\r
    add     r0, r0, r12, lsl #2\r
    ldr     r1, [r0]\r
    add     r1, r1, #1\r
    str     r1, [r0]\r
    \r
    @ remember last lr (for debug)\r
    mov     r0, #shared_ctl\r
    add     r0, r0, #0x20\r
    str     lr, [r0]\r
\r
    @ ready to take first job-interrupt\r
wait_for_irq:\r
    mrs     r0, cpsr\r
    bic     r0, r0, #0x80\r
    msr     cpsr_c, r0              @ enable interrupts\r
\r
    mov     r0, #0\r
    mcr     p15, 0, r0, c7, c0, 4   @ wait for IRQ\r
@    mcr     p15, 0, r0, c15, c8, 2\r
    nop\r
    nop\r
    b       .b_reserved\r
\r
\r
\r
@ next job getter\r
.global wait_get_job @ int oldjob\r
\r
wait_get_job:\r
    mov     r3, #mmsp2_regs\r
    orr     r2, r3, #0x3B00\r
    orr     r2, r2, #0x0046         @ DUALPEND940 register\r
    ldrh    r12,[r2]\r
\r
    tst     r0, r0\r
    beq     wgj_no_old\r
    sub     r0, r0, #1\r
    mov     r1, #1\r
    mov     r1, r1, lsl r0\r
    strh    r1, [r2]                @ clear finished job's pending bit\r
    bic     r12,r12,r1\r
\r
wgj_no_old:\r
    tst     r12,r12\r
    beq     wgj_no_jobs\r
    mov     r0, #0\r
wgj_loop:\r
    add     r0, r0, #1\r
    movs    r12,r12,lsr #1\r
    bxcs    lr\r
    b       wgj_loop\r
\r
wgj_no_jobs:\r
    mvn     r0, #0\r
    orr     r2, r3, #0x4500\r
    str     r0, [r2]            @ clear all pending interrupts in irq controller's SRCPND register\r
    orr     r2, r2, #0x0010\r
    str     r0, [r2]            @ clear all pending interrupts in irq controller's INTPND register\r
    b       wait_for_irq\r
\r
.pool\r
\r
\r
\r
\r
@ some asm utils are also defined here:\r
.global spend_cycles @ c\r
\r
spend_cycles:\r
    mov     r0, r0, lsr #2  @ 4 cycles/iteration\r
    sub     r0, r0, #2      @ entry/exit/init\r
.sc_loop:\r
    subs    r0, r0, #1\r
    bpl     .sc_loop\r
\r
    bx      lr\r
\r
\r
@ clean-flush function from ARM940T technical reference manual\r
.global dcache_clean_flush\r
\r
dcache_clean_flush:\r
    mov     r1, #0                  @ init line counter\r
ccf_outer_loop:\r
    mov     r0, #0                  @ segment counter\r
ccf_inner_loop:\r
    orr     r2, r1, r0              @ make segment and line address\r
    mcr     p15, 0, r2, c7, c14, 2  @ clean and flush that line\r
    add     r0, r0, #0x10           @ incremet segment counter\r
    cmp     r0, #0x40               @ complete all 4 segments?\r
    bne     ccf_inner_loop\r
    add     r1, r1, #0x04000000     @ increment line counter\r
    cmp     r1, #0                  @ complete all lines?\r
    bne     ccf_outer_loop\r
    bx      lr\r
\r
\r
\r
@ clean-only version\r
.global dcache_clean\r
\r
dcache_clean:\r
    mov     r1, #0                  @ init line counter\r
cf_outer_loop:\r
    mov     r0, #0                  @ segment counter\r
cf_inner_loop:\r
    orr     r2, r1, r0              @ make segment and line address\r
    mcr     p15, 0, r2, c7, c10, 2  @ clean that line\r
    add     r0, r0, #0x10           @ incremet segment counter\r
    cmp     r0, #0x40               @ complete all 4 segments?\r
    bne     cf_inner_loop\r
    add     r1, r1, #0x04000000     @ increment line counter\r
    cmp     r1, #0                  @ complete all lines?\r
    bne     cf_outer_loop\r
    bx      lr\r
\r
\r
@ drain write buffer\r
.global drain_wb\r
\r
drain_wb:\r
    mov     r0, #0\r
    mcr     p15, 0, r0, c7, c10, 4\r
    bx      lr\r
\r
\r
.global set_if_not_changed @ int *val, int oldval, int newval\r
\r
set_if_not_changed:\r
    swp    r3, r2, [r0]\r
    cmp    r1, r3\r
    bxeq   lr\r
    strne  r3, [r0] @ restore value which was changed there by other core\r
    bx     lr\r
\r
\r
\r
@ pad the protected region.\r
.rept 1024\r
.long 0\r
.endr\r
\r