hint-less mode updated for SVP, state load callback, etc

[picodrive.git] / Pico / carthw / svp / stub_arm.S

diff --git a/Pico/carthw/svp/stub_arm.S b/Pico/carthw/svp/stub_arm.S
index b812b61..d215388 100644 (file)
--- a/Pico/carthw/svp/stub_arm.S
+++ b/Pico/carthw/svp/stub_arm.S
@@ -7,8 +7,11 @@
 .global tcache
 
 .global flush_inval_caches
-.global regfile_load
-.global regfile_store
+.global ssp_regfile_load
+.global ssp_regfile_store
+.global ssp_drc_entry
+.global ssp_drc_next
+.global ssp_hle_800
 
 @ translation cache buffer
 .text
@@ -45,7 +48,7 @@ flush_inval_caches:
 
 @ trashes r2,r3
 
-regfile_load:
+ssp_regfile_load:
     ldr     r7, =ssp
     ldr     r7, [r7]
     add     r2, r7, #0x400
@@ -62,7 +65,7 @@ regfile_load:
     orrne   r8, r8, #0x8
     tst     r6,     #0x20000000
     orrne   r8, r8, #0x4                @ sss0 *           NZ..
-    orr     r6, r8, r9, lsr #12         @ sss0 * .uu. .lll ....
+    orr     r6, r8, r9, lsr #12         @ sss0 * .uu. .lll NZ..
 
     ldr     r8, [r7, #0x440]            @ r0-r2
     ldr     r9, [r7, #0x444]            @ r4-r6
@@ -70,7 +73,7 @@ regfile_load:
     bx      lr
 
 
-regfile_store:
+ssp_regfile_store:
     str     r10,[r7, #(0x400+7*4)]      @ P
     str     r8, [r7, #0x440]            @ r0-r2
     str     r9, [r7, #0x444]            @ r4-r6
@@ -93,4 +96,100 @@ regfile_store:
     bx      lr
 
 
+#define SSP_OFFS_GR         0x400
+#define SSP_PM0                 8
+#define SSP_PC                  6
+#define SSP_OFFS_EMUSTAT    0x484 // emu_status
+#define SSP_OFFS_IRAM_DIRTY 0x494
+#define SSP_OFFS_IRAM_CTX   0x498 // iram_context
+#define SSP_OFFS_BLTAB      0x49c // block_table
+#define SSP_OFFS_BLTAB_IRAM 0x4a0
+#define SSP_OFFS_TMP0       0x4a4
+#define SSP_WAIT_PM0       0x2000
+
+
+ssp_drc_entry:
+    stmfd   sp!, {r4-r11, lr}
+    mov     r11, r0
+    bl      ssp_regfile_load
+
+ssp_drc_next:
+    cmp     r11, #0
+    bmi     ssp_drc_end
+
+    ldr     r0, [r7, #(SSP_OFFS_GR+SSP_PC*4)]
+    mov     r0, r0, lsr #16
+    str     r0, [r7, #SSP_OFFS_TMP0]
+    cmp     r0, #0x400
+    blt     ssp_de_iram
+
+    ldr     r1, [r7, #SSP_OFFS_BLTAB]
+    ldr     r1, [r1, r0, lsl #2]
+    tst     r1, r1
+    bxne    r1
+    bl      ssp_translate_block
+    ldr     r2, [r7, #SSP_OFFS_TMP0]           @ entry PC
+    ldr     r1, [r7, #SSP_OFFS_BLTAB]
+    str     r0, [r1, r2, lsl #2]
+    bx      r0
+
+ssp_de_iram:
+    ldr     r1, [r7, #SSP_OFFS_IRAM_DIRTY]
+    tst     r1, r1
+    ldreq   r1, [r7, #SSP_OFFS_IRAM_CTX]
+    beq     ssp_de_iram_ctx
+
+    bl      ssp_get_iram_context
+    mov     r1, #0
+    str     r1, [r7, #SSP_OFFS_IRAM_DIRTY]
+    mov     r1, r0
+    str     r1, [r7, #SSP_OFFS_IRAM_CTX]
+    ldr     r0, [r7, #SSP_OFFS_TMP0]           @ entry PC
+    
+ssp_de_iram_ctx:
+    ldr     r2, [r7, #SSP_OFFS_BLTAB_IRAM]
+    add     r2, r2, r1, lsl #12                        @ block_tab_iram + iram_context * 0x800/2*4
+    add     r2, r2, r0, lsl #2
+    ldr     r1, [r2]
+    tst     r1, r1
+    bxne    r1
+    str     r2, [r7, #SSP_OFFS_TMP0]
+    bl      ssp_translate_block
+    ldr     r2, [r7, #SSP_OFFS_TMP0]           @ &block_table_iram[iram_context][rPC]
+    str     r0, [r2]
+    bx      r0
+
+ssp_drc_end:
+    bl      ssp_regfile_store
+    mov     r0, r11
+    ldmfd   sp!, {r4-r11, lr}
+    bx      lr
+
+
+
+@ ld      A, PM0
+@ andi    2
+@ bra     z=1, gloc_0800
+ssp_hle_800:
+    @ block prologue
+@    stmfd   sp!, {r4-r11, lr}
+@    bl      regfile_load
+@    mov     r11, #0
+    
+    ldr     r0, [r7, #(SSP_OFFS_GR+SSP_PM0*4)]
+    ldr     r1, [r7, #SSP_OFFS_EMUSTAT]
+    tst     r0, #0x20000
+    orreq   r1, r1, #SSP_WAIT_PM0
+    addeq   r11,r11, #1024
+    streq   r1, [r7, #SSP_OFFS_EMUSTAT]
+    movne   r0,     #0x04000000
+    orrne   r0, r0, #0x00040000
+    strne   r0, [r7, #(SSP_OFFS_GR+SSP_PC*4)]
+
+    bl      ssp_drc_next
+@    bl      regfile_store
+@    add     r0, r11, #3
+@    ldmfd   sp!, {r4-r11, lr}
+@    bx      lr
+