Refactor arm9 exception handling code

exceptions/arm9/source/handlers.s

@@ -5,12 +5,24 @@
 @   This is part of Luma3DS, see LICENSE.txt for details
 @
 
-.macro GEN_HANDLER name, addr_offset
+.macro GEN_HANDLER name
     .global \name
     .type   \name, %function
     \name:
         stmfd sp!, {r0-r7}      @ FIQ has its own r8-r14 regs
-    ldr r0, =\addr_offset
+        ldr r1, =\@             @ macro expansion counter
+        b _commonHandler
+
+    .size   \name, . - \name
+.endm
+
+.text
+.arm
+.align 4
+
+.global _commonHandler
+.type   _commonHandler, %function
+_commonHandler:
     sub r0, lr, r0          @ address of instruction that triggered the exception; we will handle the undef+Thumb case later
     mrs r2, spsr
     
@@ -25,23 +37,15 @@
     msr cpsr_c, r3          @ restore processor mode
     mov sp, r6
     
-    stmfd sp!, {r0,r2}      @ it's a bit of a mess, but we will fix that later
+    stmfd sp!, {r2,lr}      @ it's a bit of a mess, but we will fix that later
-                            @ order of regs now: pc, spsr, r8-r14, r0-r7
+                            @ order of saved regs now: cpsr, pc + (2/4/8), r8-r14, r0-r7
     mov r0, sp
-    ldr r1, =\@             @ macro expansion counter
     b mainHandler
 
-    .size   \name, . - \name
+GEN_HANDLER FIQHandler
-.endm
+GEN_HANDLER undefinedInstructionHandler
-
+GEN_HANDLER prefetchAbortHandler
-.text
+GEN_HANDLER dataAbortHandler
-.arm
-.align 4
-
-    GEN_HANDLER FIQHandler, 4
-    GEN_HANDLER undefinedInstructionHandler, 4
-    GEN_HANDLER prefetchAbortHandler, 4
-    GEN_HANDLER dataAbortHandler, 8
 
 .global setupStack
 .type   setupStack, %function

exceptions/arm9/source/main.c

@@ -9,32 +9,37 @@
 #include "i2c.h"
 #include "handlers.h"
 
-#define TEMP_BUFFER     0x1FF80000 //We choose AXIWRAM as tmp buffer since it's usually ARM11 payloads running there
 #define FINAL_BUFFER    0x25000000
+#define SP              ((void *)0x02000000)    //We make the (full descending) stack point to the end of ITCM for our exception handlers. 
+                                                //It doesn't matter if we're overwriting stuff, since we're going to reboot.
+
+#define REG_DUMP_SIZE   (4*17)
+#define CODE_DUMP_SIZE  48
 #define STACK_DUMP_SIZE 0x2000
-#define SP              ((void *)(TEMP_BUFFER + 4 * STACK_DUMP_SIZE))
+#define OTHER_DATA_SIZE 0
 
 void __attribute__((noreturn)) mainHandler(u32 regs[17], u32 type)
 {
-    vu32 *dump = (u32 *)TEMP_BUFFER;
+    //vu32 *dump = (u32 *)TEMP_BUFFER;
+    u32 dump[(40 + REG_DUMP_SIZE + CODE_DUMP_SIZE + STACK_DUMP_SIZE + OTHER_DATA_SIZE) / 4];
     
     dump[0] = 0xDEADC0DE;               //Magic
     dump[1] = 0xDEADCAFE;               //Magic
     dump[2] = (1 << 16) | 0;            //Dump format version number
     dump[3] = 9;                        //Processor
     dump[4] = type;                     //Exception type
-    dump[6] = 4 * 17;                   //Register dump size (r0-r12, sp, lr, pc, cpsr)
+    dump[6] = REG_DUMP_SIZE;            //Register dump size (r0-r12, sp, lr, pc, cpsr)
-    dump[7] = 40;                       //Code dump size (10 ARM instructions, up to 20 Thumb instructions).
+    dump[7] = CODE_DUMP_SIZE;           //Code dump size (10 ARM instructions, up to 20 Thumb instructions).
     dump[8] = STACK_DUMP_SIZE;          //Stack dump size
-    dump[9] = 0;                        //Other data size
+    dump[9] = OTHER_DATA_SIZE;          //Other data size
-    dump[5] = 40 + dump[6] + dump[7] + dump[8] + dump[9]; //Total size
+    dump[5] = sizeof(dump);             //Total size
 
     //Dump registers
-    //Current order of regs: pc, spsr, r8-r12, sp, lr, r0-r7
+    //Current order of saved regs: cpsr, pc, r8-r12, sp, lr, r0-r7
     vu32 *regdump = dump + 10;
 
-    u32 cpsr = regs[1];
+    u32 cpsr = regs[0];
-    u32 pc   = regs[0] + (((cpsr & 0x20) != 0 && type == 1) ? 2 : 0);
+    u32 pc   = regs[1] - ((type < 3) ? (((cpsr & 0x20) != 0 && type == 1) ? 2 : 4) : 8);
 
     regdump[15] = pc;
     regdump[16] = cpsr;
@@ -60,7 +65,7 @@ void __attribute__((noreturn)) mainHandler(u32 regs[17], u32 type)
         stackdump[i] = sp[i];
 
     vu32 *final = (u32 *)FINAL_BUFFER;
-    for(u32 i = 0; i < dump[5] / 4; i++)
+    for(u32 i = 0; i < sizeof(dump) / 4; i++)
         final[i] = dump[i];
 
     i2cWriteRegister(I2C_DEV_MCU, 0x20, 1 << 2); //Reboot