/*
+ * Basic macros to emit x86 instructions and some utils
+ * Copyright (C) 2008,2009,2010 notaz
+ *
+ * This work is licensed under the terms of MAME license.
+ * See COPYING file in the top-level directory.
+ *
+ * note:
+ * temp registers must be eax-edx due to use of SETcc and r/w 8/16.
* note about silly things like emith_eor_r_r_r:
- * these are here because the compiler was designed
- * for ARM as it's primary target.
+ * these are here because the compiler was designed
+ * for ARM as it's primary target.
*/
#include <stdarg.h>
#define CONTEXT_REG xBP
-#define IOP_JMP 0xeb
-#define IOP_JO 0x70
-#define IOP_JNO 0x71
-#define IOP_JB 0x72
-#define IOP_JAE 0x73
-#define IOP_JE 0x74
-#define IOP_JNE 0x75
-#define IOP_JBE 0x76
-#define IOP_JA 0x77
-#define IOP_JS 0x78
-#define IOP_JNS 0x79
-#define IOP_JL 0x7c
-#define IOP_JGE 0x7d
-#define IOP_JLE 0x7e
-#define IOP_JG 0x7f
+#define ICOND_JO 0x00
+#define ICOND_JNO 0x01
+#define ICOND_JB 0x02
+#define ICOND_JAE 0x03
+#define ICOND_JE 0x04
+#define ICOND_JNE 0x05
+#define ICOND_JBE 0x06
+#define ICOND_JA 0x07
+#define ICOND_JS 0x08
+#define ICOND_JNS 0x09
+#define ICOND_JL 0x0c
+#define ICOND_JGE 0x0d
+#define ICOND_JLE 0x0e
+#define ICOND_JG 0x0f
+
+#define IOP_JMP 0xeb
// unified conditions (we just use rel8 jump instructions for x86)
-#define DCOND_EQ IOP_JE
-#define DCOND_NE IOP_JNE
-#define DCOND_MI IOP_JS // MInus
-#define DCOND_PL IOP_JNS // PLus or zero
-#define DCOND_HI IOP_JA // higher (unsigned)
-#define DCOND_HS IOP_JAE // higher || same (unsigned)
-#define DCOND_LO IOP_JB // lower (unsigned)
-#define DCOND_LS IOP_JBE // lower || same (unsigned)
-#define DCOND_GE IOP_JGE // greater || equal (signed)
-#define DCOND_GT IOP_JG // greater (signed)
-#define DCOND_LE IOP_JLE // less || equal (signed)
-#define DCOND_LT IOP_JL // less (signed)
-#define DCOND_VS IOP_JO // oVerflow Set
-#define DCOND_VC IOP_JNO // oVerflow Clear
+#define DCOND_EQ ICOND_JE
+#define DCOND_NE ICOND_JNE
+#define DCOND_MI ICOND_JS // MInus
+#define DCOND_PL ICOND_JNS // PLus or zero
+#define DCOND_HI ICOND_JA // higher (unsigned)
+#define DCOND_HS ICOND_JAE // higher || same (unsigned)
+#define DCOND_LO ICOND_JB // lower (unsigned)
+#define DCOND_LS ICOND_JBE // lower || same (unsigned)
+#define DCOND_GE ICOND_JGE // greater || equal (signed)
+#define DCOND_GT ICOND_JG // greater (signed)
+#define DCOND_LE ICOND_JLE // less || equal (signed)
+#define DCOND_LT ICOND_JL // less (signed)
+#define DCOND_VS ICOND_JO // oVerflow Set
+#define DCOND_VC ICOND_JNO // oVerflow Clear
#define EMIT_PTR(ptr, val, type) \
*(type *)(ptr) = val
#define EMIT_SIB(scale,index,base) \
EMIT(((scale)<<6) | ((index)<<3) | (base), u8)
-#define EMIT_OP_MODRM(op,mod,r,rm) { \
+#define EMIT_OP_MODRM(op,mod,r,rm) do { \
EMIT_OP(op); \
EMIT_MODRM(mod, r, rm); \
-}
+} while (0)
#define JMP8_POS(ptr) \
ptr = tcache_ptr; \
tcache_ptr += 2
#define JMP8_EMIT(op, ptr) \
- EMIT_PTR(ptr, op, u8); \
+ EMIT_PTR(ptr, 0x70|(op), u8); \
+ EMIT_PTR(ptr + 1, (tcache_ptr - (ptr+2)), u8)
+
+#define JMP8_EMIT_NC(ptr) \
+ EMIT_PTR(ptr, IOP_JMP, u8); \
EMIT_PTR(ptr + 1, (tcache_ptr - (ptr+2)), u8)
// _r_r
#define emith_move_r_imm_s8(r, imm) \
emith_move_r_imm(r, (u32)(signed int)(signed char)(imm))
-#define emith_arith_r_imm(op, r, imm) { \
+#define emith_arith_r_imm(op, r, imm) do { \
EMIT_OP_MODRM(0x81, 3, op, r); \
EMIT(imm, u32); \
-}
+} while (0)
-// 2 - adc, 3 - sbb
#define emith_add_r_imm(r, imm) \
emith_arith_r_imm(0, r, imm)
#define emith_or_r_imm(r, imm) \
emith_arith_r_imm(1, r, imm)
+#define emith_adc_r_imm(r, imm) \
+ emith_arith_r_imm(2, r, imm)
+
+#define emith_sbc_r_imm(r, imm) \
+ emith_arith_r_imm(3, r, imm) // sbb
+
#define emith_and_r_imm(r, imm) \
emith_arith_r_imm(4, r, imm)
#define emith_cmp_r_imm(r, imm) \
emith_arith_r_imm(7, r, imm)
-#define emith_tst_r_imm(r, imm) { \
+#define emith_tst_r_imm(r, imm) do { \
EMIT_OP_MODRM(0xf7, 3, 0, r); \
EMIT(imm, u32); \
-}
+} while (0)
// fake
#define emith_bic_r_imm(r, imm) \
emith_arith_r_imm(4, r, ~(imm))
// fake conditionals (using SJMP instead)
-#define emith_add_r_imm_c(cond, r, imm) { \
- (void)(cond); \
- emith_add_r_imm(r, imm); \
-}
-
-#define emith_or_r_imm_c(cond, r, imm) { \
+#define emith_move_r_imm_c(cond, r, imm) { \
(void)(cond); \
- emith_or_r_imm(r, imm); \
+ emith_move_r_imm(r, imm); \
}
-#define emith_eor_r_imm_c(cond, r, imm) { \
+#define emith_add_r_imm_c(cond, r, imm) { \
(void)(cond); \
- emith_eor_r_imm(r, imm); \
+ emith_add_r_imm(r, imm); \
}
#define emith_sub_r_imm_c(cond, r, imm) { \
emith_sub_r_imm(r, imm); \
}
-#define emith_bic_r_imm_c(cond, r, imm) { \
- (void)(cond); \
- emith_bic_r_imm(r, imm); \
-}
+#define emith_or_r_imm_c(cond, r, imm) \
+ emith_or_r_imm(r, imm)
+#define emith_eor_r_imm_c(cond, r, imm) \
+ emith_eor_r_imm(r, imm)
+#define emith_bic_r_imm_c(cond, r, imm) \
+ emith_bic_r_imm(r, imm)
+#define emith_ror_c(cond, d, s, cnt) \
+ emith_ror(d, s, cnt)
+
+#define emith_read_r_r_offs_c(cond, r, rs, offs) \
+ emith_read_r_r_offs(r, rs, offs)
+#define emith_write_r_r_offs_c(cond, r, rs, offs) \
+ emith_write_r_r_offs(r, rs, offs)
+#define emith_read8_r_r_offs_c(cond, r, rs, offs) \
+ emith_read8_r_r_offs(r, rs, offs)
+#define emith_write8_r_r_offs_c(cond, r, rs, offs) \
+ emith_write8_r_r_offs(r, rs, offs)
+#define emith_read16_r_r_offs_c(cond, r, rs, offs) \
+ emith_read16_r_r_offs(r, rs, offs)
+#define emith_write16_r_r_offs_c(cond, r, rs, offs) \
+ emith_write16_r_r_offs(r, rs, offs)
+#define emith_jump_reg_c(cond, r) \
+ emith_jump_reg(r)
+#define emith_jump_ctx_c(cond, offs) \
+ emith_jump_ctx(offs)
+#define emith_ret_c(cond) \
+ emith_ret()
// _r_r_imm
+#define emith_add_r_r_imm(d, s, imm) { \
+ if (d != s) \
+ emith_move_r_r(d, s); \
+ emith_add_r_imm(d, imm); \
+}
+
#define emith_and_r_r_imm(d, s, imm) { \
if (d != s) \
emith_move_r_r(d, s); \
- emith_and_r_imm(d, imm) \
+ emith_and_r_imm(d, imm); \
}
// shift
#define emith_push(r) \
EMIT_OP(0x50 + (r))
+#define emith_push_imm(imm) { \
+ EMIT_OP(0x68); \
+ EMIT(imm, u32); \
+}
+
#define emith_pop(r) \
EMIT_OP(0x58 + (r))
#define emith_setc(r) { \
EMIT_OP(0x0f); \
- EMIT(0x92, u8); \
- EMIT_MODRM(3, 0, r); /* SETC r */ \
-}
-
-// put bit0 of r0 to carry
-#define emith_set_carry(r0) { \
- emith_tst_r_imm(r0, 1); /* clears C */ \
- EMITH_SJMP_START(DCOND_EQ); \
- EMIT_OP(0xf9); /* STC */ \
- EMITH_SJMP_END(DCOND_EQ); \
+ EMIT_OP_MODRM(0x92, 3, 0, r); /* SETC r */ \
}
-// put bit0 of r0 to carry (for subtraction)
-#define emith_set_carry_sub emith_set_carry
-
// XXX: stupid mess
#define emith_mul_(op, dlo, dhi, s1, s2) { \
int rmr; \
#define emith_subf_r_r emith_sub_r_r
#define emith_adcf_r_r emith_adc_r_r
#define emith_sbcf_r_r emith_sbc_r_r
+#define emith_eorf_r_r emith_eor_r_r
#define emith_negcf_r_r emith_negc_r_r
#define emith_lslf emith_lsl
#define emith_rolcf emith_rolc
#define emith_rorcf emith_rorc
-// XXX: offs is 8bit only
-#define emith_ctx_read(r, offs) { \
- EMIT_OP_MODRM(0x8b, 1, r, xBP); \
- EMIT(offs, u8); /* mov tmp, [ebp+#offs] */ \
+#define emith_deref_op(op, r, rs, offs) do { \
+ /* mov r <-> [ebp+#offs] */ \
+ if ((offs) >= 0x80) { \
+ EMIT_OP_MODRM(op, 2, r, rs); \
+ EMIT(offs, u32); \
+ } else { \
+ EMIT_OP_MODRM(op, 1, r, rs); \
+ EMIT(offs, u8); \
+ } \
+} while (0)
+
+#define is_abcdx(r) (xAX <= (r) && (r) <= xDX)
+
+#define emith_read_r_r_offs(r, rs, offs) \
+ emith_deref_op(0x8b, r, rs, offs)
+
+#define emith_write_r_r_offs(r, rs, offs) \
+ emith_deref_op(0x89, r, rs, offs)
+
+// note: don't use prefixes on this
+#define emith_read8_r_r_offs(r, rs, offs) do { \
+ int r_ = r; \
+ if (!is_abcdx(r)) \
+ r_ = rcache_get_tmp(); \
+ emith_deref_op(0x8a, r_, rs, offs); \
+ if ((r) != r_) { \
+ emith_move_r_r(r, r_); \
+ rcache_free_tmp(r_); \
+ } \
+} while (0)
+
+#define emith_write8_r_r_offs(r, rs, offs) do {\
+ int r_ = r; \
+ if (!is_abcdx(r)) { \
+ r_ = rcache_get_tmp(); \
+ emith_move_r_r(r_, r); \
+ } \
+ emith_deref_op(0x88, r_, rs, offs); \
+ if ((r) != r_) \
+ rcache_free_tmp(r_); \
+} while (0)
+
+#define emith_read16_r_r_offs(r, rs, offs) { \
+ EMIT(0x66, u8); /* operand override */ \
+ emith_read_r_r_offs(r, rs, offs); \
}
-#define emith_ctx_write(r, offs) { \
- EMIT_OP_MODRM(0x89, 1, r, xBP); \
- EMIT(offs, u8); /* mov [ebp+#offs], tmp */ \
+#define emith_write16_r_r_offs(r, rs, offs) { \
+ EMIT(0x66, u8); \
+ emith_write_r_r_offs(r, rs, offs); \
+}
+
+#define emith_ctx_read(r, offs) \
+ emith_read_r_r_offs(r, CONTEXT_REG, offs)
+
+#define emith_ctx_write(r, offs) \
+ emith_write_r_r_offs(r, CONTEXT_REG, offs)
+
+#define emith_ctx_read_multiple(r, offs, cnt, tmpr) do { \
+ int r_ = r, offs_ = offs, cnt_ = cnt; \
+ for (; cnt_ > 0; r_++, offs_ += 4, cnt_--) \
+ emith_ctx_read(r_, offs_); \
+} while (0)
+
+#define emith_ctx_write_multiple(r, offs, cnt, tmpr) do { \
+ int r_ = r, offs_ = offs, cnt_ = cnt; \
+ for (; cnt_ > 0; r_++, offs_ += 4, cnt_--) \
+ emith_ctx_write(r_, offs_); \
+} while (0)
+
+// assumes EBX is free
+#define emith_ret_to_ctx(offs) { \
+ emith_pop(xBX); \
+ emith_ctx_write(xBX, offs); \
}
#define emith_jump(ptr) { \
- u32 disp = (u32)ptr - ((u32)tcache_ptr + 5); \
+ u32 disp = (u32)(ptr) - ((u32)tcache_ptr + 5); \
EMIT_OP(0xe9); \
EMIT(disp, u32); \
}
+#define emith_jump_patchable(target) \
+ emith_jump(target)
+
+#define emith_jump_cond(cond, ptr) { \
+ u32 disp = (u32)(ptr) - ((u32)tcache_ptr + 6); \
+ EMIT(0x0f, u8); \
+ EMIT_OP(0x80 | (cond)); \
+ EMIT(disp, u32); \
+}
+
+#define emith_jump_cond_patchable(cond, target) \
+ emith_jump_cond(cond, target)
+
+#define emith_jump_patch(ptr, target) do { \
+ u32 disp_ = (u32)(target) - ((u32)(ptr) + 4); \
+ u32 offs_ = (*(u8 *)(ptr) == 0x0f) ? 2 : 1; \
+ EMIT_PTR((u8 *)(ptr) + offs_, disp_ - offs_, u32); \
+} while (0)
+
+#define emith_jump_at(ptr, target) { \
+ u32 disp_ = (u32)(target) - ((u32)(ptr) + 5); \
+ EMIT_PTR(ptr, 0xe9, u8); \
+ EMIT_PTR((u8 *)(ptr) + 1, disp_, u32); \
+}
+
#define emith_call(ptr) { \
- u32 disp = (u32)ptr - ((u32)tcache_ptr + 5); \
+ u32 disp = (u32)(ptr) - ((u32)tcache_ptr + 5); \
EMIT_OP(0xe8); \
EMIT(disp, u32); \
}
#define emith_call_cond(cond, ptr) \
emith_call(ptr)
-// "simple" or "short" jump
-#define EMITH_SJMP_START(cond) { \
+#define emith_call_reg(r) \
+ EMIT_OP_MODRM(0xff, 3, 2, r)
+
+#define emith_call_ctx(offs) { \
+ EMIT_OP_MODRM(0xff, 2, 2, CONTEXT_REG); \
+ EMIT(offs, u32); \
+}
+
+#define emith_ret() \
+ EMIT_OP(0xc3)
+
+#define emith_jump_reg(r) \
+ EMIT_OP_MODRM(0xff, 3, 4, r)
+
+#define emith_jump_ctx(offs) { \
+ EMIT_OP_MODRM(0xff, 2, 4, CONTEXT_REG); \
+ EMIT(offs, u32); \
+}
+
+#define emith_push_ret()
+
+#define emith_pop_and_ret() \
+ emith_ret()
+
+#define EMITH_JMP_START(cond) { \
u8 *cond_ptr; \
JMP8_POS(cond_ptr)
-#define EMITH_SJMP_END(cond) \
+#define EMITH_JMP_END(cond) \
JMP8_EMIT(cond, cond_ptr); \
}
-#define host_arg2reg(rd, arg) \
- switch (arg) { \
- case 0: rd = xAX; break; \
- case 1: rd = xDX; break; \
- case 2: rd = xCX; break; \
- }
+#define EMITH_JMP3_START(cond) { \
+ u8 *cond_ptr, *else_ptr; \
+ JMP8_POS(cond_ptr)
+
+#define EMITH_JMP3_MID(cond) \
+ JMP8_POS(else_ptr); \
+ JMP8_EMIT(cond, cond_ptr);
+
+#define EMITH_JMP3_END() \
+ JMP8_EMIT_NC(else_ptr); \
+}
+
+// "simple" jump (no more then a few insns)
+// ARM will use conditional instructions here
+#define EMITH_SJMP_START EMITH_JMP_START
+#define EMITH_SJMP_END EMITH_JMP_END
+
+#define EMITH_SJMP3_START EMITH_JMP3_START
+#define EMITH_SJMP3_MID EMITH_JMP3_MID
+#define EMITH_SJMP3_END EMITH_JMP3_END
#define emith_pass_arg_r(arg, reg) { \
int rd = 7; \
emith_move_r_imm(rd, imm); \
}
+#define host_instructions_updated(base, end)
+
+#define host_arg2reg(rd, arg) \
+ switch (arg) { \
+ case 0: rd = xAX; break; \
+ case 1: rd = xDX; break; \
+ case 2: rd = xCX; break; \
+ }
+
/* SH2 drc specific */
-#define emith_sh2_test_t() { \
- int t = rcache_get_reg(SHR_SR, RC_GR_READ); \
- EMIT_OP_MODRM(0xf6, 3, 0, t); \
- EMIT(0x01, u8); /* test <reg>, byte 1 */ \
+#define emith_sh2_drc_entry() { \
+ emith_push(xBX); \
+ emith_push(xBP); \
+ emith_push(xSI); \
+ emith_push(xDI); \
+}
+
+#define emith_sh2_drc_exit() { \
+ emith_pop(xDI); \
+ emith_pop(xSI); \
+ emith_pop(xBP); \
+ emith_pop(xBX); \
+ emith_ret(); \
+}
+
+// assumes EBX is free temporary
+#define emith_sh2_wcall(a, tab, ret_ptr) { \
+ int arg2_; \
+ host_arg2reg(arg2_, 2); \
+ emith_lsr(xBX, a, SH2_WRITE_SHIFT); \
+ EMIT_OP_MODRM(0x8b, 0, xBX, 4); \
+ EMIT_SIB(2, xBX, tab); /* mov ebx, [tab + ebx * 4] */ \
+ emith_ctx_read(arg2_, offsetof(SH2, is_slave)); \
+ emith_push_imm((long)(ret_ptr)); \
+ emith_jump_reg(xBX); \
}
#define emith_sh2_dtbf_loop() { \
emith_asr(tmp_, cr, 2+12); \
JMP8_POS(jmp0); /* no negative cycles */ \
emith_move_r_imm(tmp_, 0); \
- JMP8_EMIT(IOP_JNS, jmp0); \
+ JMP8_EMIT(ICOND_JNS, jmp0); \
emith_and_r_imm(cr, 0xffe); \
emith_subf_r_r(rn, tmp_); \
JMP8_POS(jmp1); /* no overflow */ \
emith_or_r_r(cr, rn); \
emith_or_r_imm(cr, 1); \
emith_move_r_imm(rn, 0); \
- JMP8_EMIT(IOP_JA, jmp1); \
+ JMP8_EMIT(ICOND_JA, jmp1); \
rcache_free_tmp(tmp_); \
}
-#define emith_write_sr(srcr) { \
+#define emith_write_sr(sr, srcr) { \
int tmp_ = rcache_get_tmp(); \
- int srr = rcache_get_reg(SHR_SR, RC_GR_RMW); \
- emith_clear_msb(tmp_, srcr, 20); \
- emith_bic_r_imm(srr, 0xfff); \
- emith_or_r_r(srr, tmp_); \
+ emith_clear_msb(tmp_, srcr, 22); \
+ emith_bic_r_imm(sr, 0x3ff); \
+ emith_or_r_r(sr, tmp_); \
rcache_free_tmp(tmp_); \
}
-#define emith_carry_to_t(srr, is_sub) { \
- int tmp_ = rcache_get_tmp(); \
- emith_setc(tmp_); \
- emith_bic_r_imm(srr, 1); \
- EMIT_OP_MODRM(0x08, 3, tmp_, srr); /* OR srrl, tmpl */ \
- rcache_free_tmp(tmp_); \
-}
+#define emith_tpop_carry(sr, is_sub) \
+ emith_lsr(sr, sr, 1)
+
+#define emith_tpush_carry(sr, is_sub) \
+ emith_adc_r_r(sr, sr)
/*
* if Q
#define emith_sh2_div1_step(rn, rm, sr) { \
u8 *jmp0, *jmp1; \
int tmp_ = rcache_get_tmp(); \
+ emith_eor_r_r(tmp_, tmp_); \
emith_tst_r_imm(sr, Q); /* if (Q ^ M) */ \
JMP8_POS(jmp0); /* je do_sub */ \
emith_add_r_r(rn, rm); \
JMP8_POS(jmp1); /* jmp done */ \
- JMP8_EMIT(IOP_JE, jmp0); /* do_sub: */ \
+ JMP8_EMIT(ICOND_JE, jmp0); /* do_sub: */ \
emith_sub_r_r(rn, rm); \
- JMP8_EMIT(IOP_JMP, jmp1);/* done: */ \
+ JMP8_EMIT_NC(jmp1); /* done: */ \
emith_setc(tmp_); \
- EMIT_OP_MODRM(0x30, 3, tmp_, sr); /* T = Q1 ^ Q2 (byte) */ \
+ EMIT_OP_MODRM(0x31, 3, tmp_, sr); /* T = Q1 ^ Q2 */ \
rcache_free_tmp(tmp_); \
}