[arm] Improve external stores and make them faster
While at it, speed up palette writes too.
This commit is contained in:
David Guillen Fandos
parent
fda42c959f
commit
6c4ffc4db2
|
|
@ -1754,9 +1754,6 @@ static void trace_instruction(u32 pc, u32 mode)
|
||||||
|
|
||||||
/* TODO: Make these use cached registers. Implement iwram_stack_optimize. */
|
/* TODO: Make these use cached registers. Implement iwram_stack_optimize. */
|
||||||
|
|
||||||
#define thumb_block_address_preadjust_up() \
|
|
||||||
generate_add_imm(reg_s0, (bit_count[reg_list] * 4), 0) \
|
|
||||||
|
|
||||||
#define thumb_block_address_preadjust_down() \
|
#define thumb_block_address_preadjust_down() \
|
||||||
generate_sub_imm(reg_s0, (bit_count[reg_list] * 4), 0) \
|
generate_sub_imm(reg_s0, (bit_count[reg_list] * 4), 0) \
|
||||||
|
|
||||||
|
|
@ -1772,11 +1769,6 @@ static void trace_instruction(u32 pc, u32 mode)
|
||||||
generate_add_reg_reg_imm(reg_a0, reg_s0, (bit_count[reg_list] * 4), 0); \
|
generate_add_reg_reg_imm(reg_a0, reg_s0, (bit_count[reg_list] * 4), 0); \
|
||||||
thumb_generate_store_reg(reg_a0, base_reg) \
|
thumb_generate_store_reg(reg_a0, base_reg) \
|
||||||
|
|
||||||
#define thumb_block_address_postadjust_down(base_reg) \
|
|
||||||
generate_mov(reg_a0, reg_s0); \
|
|
||||||
generate_sub_imm(reg_a0, (bit_count[reg_list] * 4), 0); \
|
|
||||||
thumb_generate_store_reg(reg_a0, base_reg) \
|
|
||||||
|
|
||||||
#define thumb_block_address_postadjust_pop_pc(base_reg) \
|
#define thumb_block_address_postadjust_pop_pc(base_reg) \
|
||||||
generate_add_reg_reg_imm(reg_a0, reg_s0, \
|
generate_add_reg_reg_imm(reg_a0, reg_s0, \
|
||||||
((bit_count[reg_list] + 1) * 4), 0); \
|
((bit_count[reg_list] + 1) * 4), 0); \
|
||||||
|
|
|
||||||
|
|
@ -77,6 +77,7 @@ _##symbol:
|
||||||
#define PAL_RAM_OFF 0x900
|
#define PAL_RAM_OFF 0x900
|
||||||
#define RDMAP_OFF 0xD00
|
#define RDMAP_OFF 0xD00
|
||||||
#define IOREG_OFF 0x8D00
|
#define IOREG_OFF 0x8D00
|
||||||
|
#define PAL_CONV_OFF 0x9100
|
||||||
|
|
||||||
|
|
||||||
#if __ARM_ARCH >= 6
|
#if __ARM_ARCH >= 6
|
||||||
|
|
@ -490,6 +491,15 @@ return_to_main:
|
||||||
#define execute_store_builder(store_type, str_op, str_op16, load_op, tnum) ;\
|
#define execute_store_builder(store_type, str_op, str_op16, load_op, tnum) ;\
|
||||||
;\
|
;\
|
||||||
ext_store_u##store_type: ;\
|
ext_store_u##store_type: ;\
|
||||||
|
save_flags() ;\
|
||||||
|
ldr r2, [lr] /* load PC */;\
|
||||||
|
str r2, [reg_base, #REG_PC] /* write out PC */;\
|
||||||
|
store_align_##store_type() ;\
|
||||||
|
call_c_function(write_memory##store_type) ;\
|
||||||
|
restore_flags() ;\
|
||||||
|
add pc, lr, #4 /* return */;\
|
||||||
|
;\
|
||||||
|
ext_io_store_u##store_type: ;\
|
||||||
save_flags() ;\
|
save_flags() ;\
|
||||||
ldr r2, [lr] /* load PC */;\
|
ldr r2, [lr] /* load PC */;\
|
||||||
str r2, [reg_base, #REG_PC] /* write out PC */;\
|
str r2, [reg_base, #REG_PC] /* write out PC */;\
|
||||||
|
|
@ -548,20 +558,20 @@ ext_store_ignore:
|
||||||
add pc, lr, #4 @ return
|
add pc, lr, #4 @ return
|
||||||
|
|
||||||
#define store_lookup_table(store_type) ;\
|
#define store_lookup_table(store_type) ;\
|
||||||
.word ext_store_u##store_type /* -1: ignore, for > 0x0F */;\
|
.word ext_store_ignore /* -1: ignore, for > 0x0F */;\
|
||||||
.word ext_store_ignore /* 0x00: BIOS, ignore */;\
|
.word ext_store_ignore /* 0x00: BIOS, ignore */;\
|
||||||
.word ext_store_ignore /* 0x01: ignore */;\
|
.word ext_store_ignore /* 0x01: ignore */;\
|
||||||
.word ext_store_ewram_u##store_type /* 0x02: ewram */;\
|
.word ext_store_ewram_u##store_type /* 0x02: ewram */;\
|
||||||
.word ext_store_iwram_u##store_type /* 0x03: iwram */;\
|
.word ext_store_iwram_u##store_type /* 0x03: iwram */;\
|
||||||
.word ext_store_u##store_type /* 0x04: I/O regs */;\
|
.word ext_io_store_u##store_type /* 0x04: I/O regs */;\
|
||||||
.word ext_store_u##store_type /* 0x05: palette RAM */;\
|
.word ext_store_palette_u##store_type /* 0x05: palette RAM */;\
|
||||||
.word ext_store_vram_u##store_type /* 0x06: vram */;\
|
.word ext_store_vram_u##store_type /* 0x06: vram */;\
|
||||||
.word ext_store_oam_ram_u##store_type /* 0x07: oam ram */;\
|
.word ext_store_oam_ram_u##store_type /* 0x07: oam ram */;\
|
||||||
.word ext_store_u##store_type /* 0x08: gamepak: ignore */;\
|
.word ext_store_u##store_type /* 0x08: gamepak: ignore */;\
|
||||||
.word ext_store_u##store_type /* 0x09: gamepak: ignore */;\
|
.word ext_store_ignore /* 0x09: gamepak: ignore */;\
|
||||||
.word ext_store_u##store_type /* 0x0A: gamepak: ignore */;\
|
.word ext_store_ignore /* 0x0A: gamepak: ignore */;\
|
||||||
.word ext_store_u##store_type /* 0x0B: gamepak: ignore */;\
|
.word ext_store_ignore /* 0x0B: gamepak: ignore */;\
|
||||||
.word ext_store_u##store_type /* 0x0C: gamepak: ignore */;\
|
.word ext_store_ignore /* 0x0C: gamepak: ignore */;\
|
||||||
.word ext_store_u##store_type /* 0x0D: EEPROM */;\
|
.word ext_store_u##store_type /* 0x0D: EEPROM */;\
|
||||||
.word ext_store_u##store_type /* 0x0E: backup */;\
|
.word ext_store_u##store_type /* 0x0E: backup */;\
|
||||||
.word ext_store_ignore /* 0x0F: ignore */;\
|
.word ext_store_ignore /* 0x0F: ignore */;\
|
||||||
|
|
@ -570,14 +580,62 @@ execute_store_builder(8, strb, strh, ldrb, 0)
|
||||||
execute_store_builder(16, strh, strh, ldrh, 1)
|
execute_store_builder(16, strh, strh, ldrh, 1)
|
||||||
execute_store_builder(32, str, str, ldr, 2)
|
execute_store_builder(32, str, str, ldr, 2)
|
||||||
|
|
||||||
|
@ Palette writes are special since they are converted on the fly for speed
|
||||||
|
|
||||||
|
ext_store_palette_u8:
|
||||||
|
bic r1, r1, #0xff00 // Duplicate the byte
|
||||||
|
orr r1, r1, lsl #8
|
||||||
|
ext_store_palette_u16:
|
||||||
|
mask_addr_16(10) // Accesses are always 16 bit
|
||||||
|
add r2, reg_base, #PAL_RAM_OFF // r2 = palette base
|
||||||
|
strh r1, [r0, r2] // store data
|
||||||
|
|
||||||
|
and r2, r1, #0x3E0 // Convert color point
|
||||||
|
lsl r2, r2, #1
|
||||||
|
orr r2, r1, lsl #11
|
||||||
|
and r1, r1, #0x7C00
|
||||||
|
orr r2, r1, lsr #10
|
||||||
|
|
||||||
|
add r1, reg_base, #PAL_CONV_OFF // r1 = converted palette ram
|
||||||
|
strh r2, [r0, r1] // Converted value write (r2)
|
||||||
|
add pc, lr, #4 // return
|
||||||
|
|
||||||
|
ext_store_palette_u32_safe:
|
||||||
|
sub lr, lr, #4
|
||||||
|
ext_store_palette_u32:
|
||||||
|
mask_addr_32(10) // Accesses are always 16 bit
|
||||||
|
add r2, reg_base, #PAL_RAM_OFF // r2 = palette base
|
||||||
|
str r1, [r0, r2] // store data
|
||||||
|
add r2, reg_base, #PAL_CONV_OFF // r2 = converted palette ram
|
||||||
|
|
||||||
|
lsr r9, r1, #10
|
||||||
|
and r9, r9, #0x1F
|
||||||
|
orr r9, r1, lsl #11
|
||||||
|
bic r1, r1, #0x1F
|
||||||
|
bic r1, r1, #0xFC00
|
||||||
|
orr r9, r1, lsl #1
|
||||||
|
strh r9, [r0, r2] // Write first halfword
|
||||||
|
add r0, r0, #2
|
||||||
|
|
||||||
|
lsr r1, r1, #16
|
||||||
|
and r9, r1, #0x3E0
|
||||||
|
lsl r9, r9, #1
|
||||||
|
orr r9, r1, lsl #11
|
||||||
|
and r1, r1, #0x7C00
|
||||||
|
orr r9, r1, lsr #10
|
||||||
|
strh r9, [r0, r2] // Write second halfword
|
||||||
|
|
||||||
|
add pc, lr, #4 // return
|
||||||
|
|
||||||
|
|
||||||
@ This is a store that is executed in a strm case (so no SMC checks in-between)
|
@ This is a store that is executed in a strm case (so no SMC checks in-between)
|
||||||
|
|
||||||
ext_store_u32_safe:
|
ext_store_u32_safe:
|
||||||
str lr, [reg_base, #REG_SAVE3] @ Restore lr
|
ext_io_store_u32_safe:
|
||||||
save_flags()
|
save_flags()
|
||||||
call_c_function(write_memory32) @ Perform 32bit store
|
call_c_function(write_memory32) @ Perform 32bit store
|
||||||
restore_flags()
|
restore_flags()
|
||||||
ldr pc, [reg_base, #REG_SAVE3] @ return
|
bx lr @ Return
|
||||||
|
|
||||||
ext_store_iwram_u32_safe:
|
ext_store_iwram_u32_safe:
|
||||||
mask_addr_8(15) @ Mask to mirror memory (no need to align!)
|
mask_addr_8(15) @ Mask to mirror memory (no need to align!)
|
||||||
|
|
@ -834,7 +892,6 @@ defsymbl(memory_map_read)
|
||||||
.space 0x8000
|
.space 0x8000
|
||||||
defsymbl(io_registers)
|
defsymbl(io_registers)
|
||||||
.space 0x400
|
.space 0x400
|
||||||
|
|
||||||
defsymbl(palette_ram_converted)
|
defsymbl(palette_ram_converted)
|
||||||
.space 0x400
|
.space 0x400
|
||||||
|
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue