svp compiler direct calls

[picodrive.git] / Pico / carthw / svp / gen_arm.c

#define EMIT(x) *tcache_ptr++ = x

#define A_R4M  (1 << 4)
#define A_R5M  (1 << 5)
#define A_R6M  (1 << 6)
#define A_R7M  (1 << 7)
#define A_R8M  (1 << 8)
#define A_R9M  (1 << 9)
#define A_R10M (1 << 10)
#define A_R11M (1 << 11)
#define A_R14M (1 << 14)

#define A_COND_AL 0xe

/* addressing mode 1 */
#define A_AM1_LSL 0
#define A_AM1_LSR 1
#define A_AM1_ASR 2
#define A_AM1_ROR 3

#define A_AM1_IMM(ror2,imm8)                  (((ror2)<<8) | (imm8) | 0x02000000)
#define A_AM1_REG_XIMM(shift_imm,shift_op,rm) (((shift_imm)<<7) | ((shift_op)<<5) | (rm))

/* data processing op */
#define A_OP_ORR 0xc
#define A_OP_MOV 0xd

#define EOP_C_DOP_X(cond,op,s,rn,rd,shifter_op) \
	EMIT(((cond)<<28) | ((op)<< 21) | ((s)<<20) | ((rn)<<16) | ((rd)<<12) | (shifter_op))

#define EOP_C_DOP_IMM(cond,op,s,rn,rd,ror2,imm8)             EOP_C_DOP_X(cond,op,s,rn,rd,A_AM1_IMM(ror2,imm8))
#define EOP_C_DOP_REG(cond,op,s,rn,rd,shift_imm,shift_op,rm) EOP_C_DOP_X(cond,op,s,rn,rd,A_AM1_REG_XIMM(shift_imm,shift_op,rm))

#define EOP_MOV_IMM(s,   rd,ror2,imm8) EOP_C_DOP_IMM(A_COND_AL,A_OP_MOV,s, 0,rd,ror2,imm8)
#define EOP_ORR_IMM(s,rn,rd,ror2,imm8) EOP_C_DOP_IMM(A_COND_AL,A_OP_ORR,s,rn,rd,ror2,imm8)

#define EOP_MOV_REG(s,   rd,shift_imm,shift_op,rm) EOP_C_DOP_REG(A_COND_AL,A_OP_MOV,s, 0,rd,shift_imm,shift_op,rm)

#define EOP_MOV_REG_SIMPLE(rd,rm) EOP_MOV_REG(0,rd,0,A_AM1_LSL,rm)

/* ldr and str */
#define EOP_C_XXR_IMM(cond,u,b,l,rn,rd,offset_12) \
	EMIT(((cond)<<28) | 0x05000000 | ((u)<<23) | ((b)<<22) | ((l)<<20) | ((rn)<<16) | ((rd)<<12) | (offset_12))

#define EOP_LDR_IMM(   rd,rn,offset_12) EOP_C_XXR_IMM(A_COND_AL,1,0,1,rn,rd,offset_12)
#define EOP_LDR_NEGIMM(rd,rn,offset_12) EOP_C_XXR_IMM(A_COND_AL,0,0,1,rn,rd,offset_12)
#define EOP_LDR_SIMPLE(rd,rn)           EOP_C_XXR_IMM(A_COND_AL,1,0,1,rn,rd,0)
#define EOP_STR_IMM(   rd,rn,offset_12) EOP_C_XXR_IMM(A_COND_AL,1,0,0,rn,rd,offset_12)
#define EOP_STR_SIMPLE(rd,rn)           EOP_C_XXR_IMM(A_COND_AL,1,0,0,rn,rd,0)

/* ldm and stm */
#define EOP_XXM(cond,p,u,s,w,l,rn,list) \
	EMIT(((cond)<<28) | (1<<27) | ((p)<<24) | ((u)<<23) | ((s)<<22) | ((w)<<21) | ((l)<<20) | ((rn)<<16) | (list))

#define EOP_STMFD_ST(list) EOP_XXM(A_COND_AL,1,0,0,1,0,13,list)
#define EOP_LDMFD_ST(list) EOP_XXM(A_COND_AL,0,1,0,1,1,13,list)

/* branches */
#define EOP_C_BX(cond,rm) \
	EMIT(((cond)<<28) | 0x012fff10 | (rm))

#define EOP_BX(rm) EOP_C_BX(A_COND_AL,rm)

#define EOP_C_B(cond,l,signed_immed_24) \
	EMIT(((cond)<<28) | 0x0a000000 | ((l)<<24) | (signed_immed_24))

#define EOP_B( signed_immed_24) EOP_C_B(A_COND_AL,0,signed_immed_24)
#define EOP_BL(signed_immed_24) EOP_C_B(A_COND_AL,1,signed_immed_24)


static void emit_mov_const(int d, unsigned int val)
{
	int need_or = 0;
	if (val & 0xff000000) {
		EOP_MOV_IMM(0, d,  8/2, (val>>24)&0xff);
		need_or = 1;
	}
	if (val & 0x00ff0000) {
		EOP_C_DOP_IMM(A_COND_AL,need_or ? A_OP_ORR : A_OP_MOV, 0, need_or ? d : 0, d, 16/2, (val>>16)&0xff);
		need_or = 1;
	}
	if (val & 0x0000ff00) {
		EOP_C_DOP_IMM(A_COND_AL,need_or ? A_OP_ORR : A_OP_MOV, 0, need_or ? d : 0, d, 24/2, (val>>8)&0xff);
		need_or = 1;
	}
	if ((val &0x000000ff) || !need_or)
		EOP_C_DOP_IMM(A_COND_AL,need_or ? A_OP_ORR : A_OP_MOV, 0, need_or ? d : 0, d, 0, val&0xff);
}

/*
static void check_offset_12(unsigned int val)
{
	if (!(val & ~0xfff)) return;
	printf("offset_12 overflow %04x\n", val);
	exit(1);
}
*/

static void check_offset_24(int val)
{
	if (val >= (int)0xff000000 && val <= 0x00ffffff) return;
	printf("offset_24 overflow %08x\n", val);
	exit(1);
}

static void emit_call(void *target)
{
	int val = (unsigned int *)target - tcache_ptr - 2;
	check_offset_24(val);

	EOP_BL(val & 0xffffff);			// bl target
}

static void emit_block_prologue(void)
{
	// stack regs
	EOP_STMFD_ST(A_R4M|A_R5M|A_R6M|A_R7M|A_R8M|A_R9M|A_R10M|A_R11M|A_R14M);	// stmfd r13!, {r4-r11,lr}
	emit_call(regfile_load);
}

static void emit_block_epilogue(int icount)
{
	emit_call(regfile_store);
	EOP_LDMFD_ST(A_R4M|A_R5M|A_R6M|A_R7M|A_R8M|A_R9M|A_R10M|A_R11M|A_R14M);	// ldmfd r13!, {r4-r11,lr}
	emit_mov_const(0, icount);
	EOP_BX(14);				// bx r14
}

static void emit_pc_dump(int pc)
{
	emit_mov_const(3, pc<<16);
	EOP_STR_IMM(3,7,0x400+6*4);		// str r3, [r7, #(0x400+6*8)]
}

static void emit_interpreter_call(void *target)
{
	emit_call(regfile_store);
	emit_call(target);
	emit_call(regfile_load);
}

static void handle_caches()
{
#ifdef ARM
	extern void flush_inval_caches(const void *start_addr, const void *end_addr);
	flush_inval_caches(tcache, tcache_ptr);
#else
#error wth
#endif
}