X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=deps%2Flightrec%2Femitter.c;h=0eff0ce3d893c4fd0b6b5a0d071b1880c57b9671;hb=54604e030df86ba542633e543c5bd269b2e391bc;hp=523daf7716815551121e5668d70d4aee964bb9ae;hpb=6ce0b00a79f4fa7c4a1735be9f48bfbc644b080c;p=pcsx_rearmed.git diff --git a/deps/lightrec/emitter.c b/deps/lightrec/emitter.c index 523daf77..0eff0ce3 100644 --- a/deps/lightrec/emitter.c +++ b/deps/lightrec/emitter.c @@ -21,6 +21,11 @@ static void rec_SPECIAL(struct lightrec_cstate *state, const struct block *block static void rec_REGIMM(struct lightrec_cstate *state, const struct block *block, u16 offset); static void rec_CP0(struct lightrec_cstate *state, const struct block *block, u16 offset); static void rec_CP2(struct lightrec_cstate *state, const struct block *block, u16 offset); +static void rec_cp2_do_mtc2(struct lightrec_cstate *state, + const struct block *block, u16 offset, u8 reg, u8 in_reg); +static void rec_cp2_do_mfc2(struct lightrec_cstate *state, + const struct block *block, u16 offset, + u8 reg, u8 out_reg); static void unknown_opcode(struct lightrec_cstate *state, const struct block *block, u16 offset) { @@ -38,6 +43,16 @@ lightrec_jump_to_eob(struct lightrec_cstate *state, jit_state_t *_jit) jit_patch_abs(jit_jmpi(), state->state->eob_wrapper_func); } +static void update_ra_register(struct regcache *reg_cache, jit_state_t *_jit, + u8 ra_reg, u32 pc, u32 link) +{ + u8 link_reg; + + link_reg = lightrec_alloc_reg_out(reg_cache, _jit, ra_reg, 0); + lightrec_load_imm(reg_cache, _jit, link_reg, pc, link); + lightrec_free_reg(reg_cache, link_reg); +} + static void lightrec_emit_end_of_block(struct lightrec_cstate *state, const struct block *block, u16 offset, s8 reg_new_pc, u32 imm, u8 ra_reg, @@ -51,18 +66,19 @@ static void lightrec_emit_end_of_block(struct lightrec_cstate *state, jit_note(__FILE__, __LINE__); - if (link) { - /* Update the $ra register */ - u8 link_reg = lightrec_alloc_reg_out(reg_cache, _jit, ra_reg, 0); - jit_movi(link_reg, link); - lightrec_free_reg(reg_cache, link_reg); - } + if (link && ra_reg != reg_new_pc) + update_ra_register(reg_cache, _jit, ra_reg, block->pc, link); - if (reg_new_pc < 0) { - reg_new_pc = lightrec_alloc_reg(reg_cache, _jit, JIT_V0); - lightrec_lock_reg(reg_cache, _jit, reg_new_pc); + if (reg_new_pc < 0) + lightrec_load_next_pc_imm(reg_cache, _jit, block->pc, imm); + else + lightrec_load_next_pc(reg_cache, _jit, reg_new_pc); - jit_movi(reg_new_pc, imm); + if (link && ra_reg == reg_new_pc) { + /* Handle the special case: JALR $r0, $r0 + * In that case the target PC should be the old value of the + * register. */ + update_ra_register(reg_cache, _jit, ra_reg, block->pc, link); } if (has_delay_slot(op->c) && @@ -77,8 +93,6 @@ static void lightrec_emit_end_of_block(struct lightrec_cstate *state, /* Clean the remaining registers */ lightrec_clean_regs(reg_cache, _jit); - jit_movr(JIT_V0, reg_new_pc); - if (cycles && update_cycles) { jit_subi(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, cycles); pr_debug("EOB: %u cycles\n", cycles); @@ -87,53 +101,35 @@ static void lightrec_emit_end_of_block(struct lightrec_cstate *state, lightrec_jump_to_eob(state, _jit); } -void lightrec_emit_eob(struct lightrec_cstate *state, const struct block *block, - u16 offset, bool after_op) +void lightrec_emit_eob(struct lightrec_cstate *state, + const struct block *block, u16 offset) { struct regcache *reg_cache = state->reg_cache; jit_state_t *_jit = block->_jit; - union code c = block->opcode_list[offset].c; - u32 cycles = state->cycles; - - if (after_op) - cycles += lightrec_cycles_of_opcode(c); lightrec_clean_regs(reg_cache, _jit); - jit_movi(JIT_V0, block->pc + (offset << 2)); - jit_subi(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, cycles); + lightrec_load_imm(reg_cache, _jit, JIT_V0, block->pc, + block->pc + (offset << 2)); + jit_subi(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, state->cycles); lightrec_jump_to_eob(state, _jit); } -static u8 get_jr_jalr_reg(struct lightrec_cstate *state, const struct block *block, u16 offset) -{ - struct regcache *reg_cache = state->reg_cache; - jit_state_t *_jit = block->_jit; - const struct opcode *op = &block->opcode_list[offset]; - u8 rs; - - rs = lightrec_request_reg_in(reg_cache, _jit, op->r.rs, JIT_V0); - lightrec_lock_reg(reg_cache, _jit, rs); - - return rs; -} - static void rec_special_JR(struct lightrec_cstate *state, const struct block *block, u16 offset) { - u8 rs = get_jr_jalr_reg(state, block, offset); + union code c = block->opcode_list[offset].c; _jit_name(block->_jit, __func__); - lightrec_emit_end_of_block(state, block, offset, rs, 0, 31, 0, true); + lightrec_emit_end_of_block(state, block, offset, c.r.rs, 0, 31, 0, true); } static void rec_special_JALR(struct lightrec_cstate *state, const struct block *block, u16 offset) { - u8 rs = get_jr_jalr_reg(state, block, offset); union code c = block->opcode_list[offset].c; _jit_name(block->_jit, __func__); - lightrec_emit_end_of_block(state, block, offset, rs, 0, c.r.rd, + lightrec_emit_end_of_block(state, block, offset, c.r.rs, 0, c.r.rd, get_branch_pc(block, offset, 2), true); } @@ -204,12 +200,12 @@ static void rec_b(struct lightrec_cstate *state, const struct block *block, u16 const struct opcode *op = &block->opcode_list[offset], *next = &block->opcode_list[offset + 1]; jit_node_t *addr; - u8 link_reg, rs, rt; bool is_forward = (s16)op->i.imm >= -1; int op_cycles = lightrec_cycles_of_opcode(op->c); u32 target_offset, cycles = state->cycles + op_cycles; bool no_indirection = false; u32 next_pc; + u8 rs, rt; jit_note(__FILE__, __LINE__); @@ -250,15 +246,11 @@ static void rec_b(struct lightrec_cstate *state, const struct block *block, u16 if (op_flag_local_branch(op->flags)) { /* Recompile the delay slot */ - if (next && next->opcode && !op_flag_no_ds(op->flags)) + if (!op_flag_no_ds(op->flags) && next->opcode) lightrec_rec_opcode(state, block, offset + 1); - if (link) { - /* Update the $ra register */ - link_reg = lightrec_alloc_reg_out(reg_cache, _jit, 31, 0); - jit_movi(link_reg, link); - lightrec_free_reg(reg_cache, link_reg); - } + if (link) + update_ra_register(reg_cache, _jit, 31, block->pc, link); /* Clean remaining registers */ lightrec_clean_regs(reg_cache, _jit); @@ -292,13 +284,8 @@ static void rec_b(struct lightrec_cstate *state, const struct block *block, u16 lightrec_regcache_leave_branch(reg_cache, regs_backup); - if (bz && link) { - /* Update the $ra register */ - link_reg = lightrec_alloc_reg_out(reg_cache, _jit, - 31, REG_EXT); - jit_movi(link_reg, (s32)link); - lightrec_free_reg(reg_cache, link_reg); - } + if (bz && link) + update_ra_register(reg_cache, _jit, 31, block->pc, link); if (!op_flag_no_ds(op->flags) && next->opcode) lightrec_rec_opcode(state, block, offset + 1); @@ -947,11 +934,11 @@ static void rec_alu_div(struct lightrec_cstate *state, if (!op_flag_no_lo(flags)) { if (is_signed) { - jit_lti(lo, rs, 0); + jit_ltr(lo, rs, rt); jit_lshi(lo, lo, 1); jit_subi(lo, lo, 1); } else { - jit_movi(lo, 0xffffffff); + jit_subi(lo, rt, 1); } } @@ -1127,6 +1114,13 @@ static u32 rec_ram_mask(struct lightrec_state *state) return (RAM_SIZE << (state->mirrors_mapped * 2)) - 1; } +static u32 rec_io_mask(const struct lightrec_state *state) +{ + u32 length = state->maps[PSX_MAP_HW_REGISTERS].length; + + return GENMASK(31 - clz32(length - 1), 0); +} + static void rec_store_memory(struct lightrec_cstate *cstate, const struct block *block, u16 offset, jit_code_t code, @@ -1148,8 +1142,10 @@ static void rec_store_memory(struct lightrec_cstate *cstate, ((!state->mirrors_mapped && !no_mask) || (invalidate && ((imm & 0x3) || simm + lut_offt != (s16)(simm + lut_offt)))); bool need_tmp = !no_mask || addr_offset || add_imm || invalidate; + bool swc2 = c.i.op == OP_SWC2; + u8 in_reg = swc2 ? REG_CP2_TEMP : c.i.rt; - rt = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rt, 0); + rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0); rs = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rs, 0); if (need_tmp) tmp = lightrec_alloc_reg_temp(reg_cache, _jit); @@ -1177,7 +1173,7 @@ static void rec_store_memory(struct lightrec_cstate *cstate, addr_reg2 = addr_reg; } - if (is_big_endian() && swap_code && c.i.rt) { + if (is_big_endian() && swap_code && in_reg) { tmp3 = lightrec_alloc_reg_temp(reg_cache, _jit); jit_new_node_ww(swap_code, tmp3, rt); @@ -1258,7 +1254,7 @@ static void rec_store_io(struct lightrec_cstate *cstate, return rec_store_memory(cstate, block, offset, code, swap_code, cstate->state->offset_io, - 0x1fffffff, false); + rec_io_mask(cstate->state), false); } static void rec_store_direct_no_invalidate(struct lightrec_cstate *cstate, @@ -1271,7 +1267,8 @@ static void rec_store_direct_no_invalidate(struct lightrec_cstate *cstate, union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; jit_node_t *to_not_ram, *to_end; - u8 tmp, tmp2, rs, rt; + bool swc2 = c.i.op == OP_SWC2; + u8 tmp, tmp2, rs, rt, in_reg = swc2 ? REG_CP2_TEMP : c.i.rt; s16 imm; jit_note(__FILE__, __LINE__); @@ -1315,9 +1312,9 @@ static void rec_store_direct_no_invalidate(struct lightrec_cstate *cstate, lightrec_free_reg(reg_cache, tmp2); } - rt = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rt, 0); + rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0); - if (is_big_endian() && swap_code && c.i.rt) { + if (is_big_endian() && swap_code && in_reg) { tmp2 = lightrec_alloc_reg_temp(reg_cache, _jit); jit_new_node_ww(swap_code, tmp2, rt); @@ -1341,7 +1338,9 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; jit_node_t *to_not_ram, *to_end; - u8 tmp, tmp2, tmp3, rs, rt; + bool swc2 = c.i.op == OP_SWC2; + u8 tmp, tmp2, tmp3, masked_reg, rs, rt; + u8 in_reg = swc2 ? REG_CP2_TEMP : c.i.rt; jit_note(__FILE__, __LINE__); @@ -1360,10 +1359,21 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block lightrec_free_reg(reg_cache, rs); tmp = lightrec_alloc_reg_temp(reg_cache, _jit); - to_not_ram = jit_bgti(tmp2, ram_size); + if (state->offset_ram != state->offset_scratch) { + to_not_ram = jit_bgti(tmp2, ram_size); + masked_reg = tmp2; + } else { + jit_lti_u(tmp, tmp2, ram_size); + jit_movnr(tmp, tmp2, tmp); + masked_reg = tmp; + } /* Compute the offset to the code LUT */ - jit_andi(tmp, tmp2, (RAM_SIZE - 1) & ~3); + if (c.i.op == OP_SW) + jit_andi(tmp, masked_reg, RAM_SIZE - 1); + else + jit_andi(tmp, masked_reg, (RAM_SIZE - 1) & ~3); + if (!lut_is_32bit(state)) jit_lshi(tmp, tmp, 1); jit_addr(tmp, LIGHTREC_REG_STATE, tmp); @@ -1378,10 +1388,9 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block jit_movi(tmp, state->offset_ram); to_end = jit_b(); + jit_patch(to_not_ram); } - jit_patch(to_not_ram); - if (state->offset_ram || state->offset_scratch) jit_movi(tmp, state->offset_scratch); @@ -1394,9 +1403,9 @@ static void rec_store_direct(struct lightrec_cstate *cstate, const struct block lightrec_free_reg(reg_cache, tmp); lightrec_free_reg(reg_cache, tmp3); - rt = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rt, 0); + rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, 0); - if (is_big_endian() && swap_code && c.i.rt) { + if (is_big_endian() && swap_code && in_reg) { tmp = lightrec_alloc_reg_temp(reg_cache, _jit); jit_new_node_ww(swap_code, tmp, rt); @@ -1416,10 +1425,26 @@ static void rec_store(struct lightrec_cstate *state, jit_code_t code, jit_code_t swap_code) { u32 flags = block->opcode_list[offset].flags; + u32 mode = LIGHTREC_FLAGS_GET_IO_MODE(flags); bool no_invalidate = op_flag_no_invalidate(flags) || state->state->invalidate_from_dma_only; + union code c = block->opcode_list[offset].c; + bool is_swc2 = c.i.op == OP_SWC2; + + if (is_swc2) { + switch (mode) { + case LIGHTREC_IO_RAM: + case LIGHTREC_IO_SCRATCH: + case LIGHTREC_IO_DIRECT: + case LIGHTREC_IO_DIRECT_HW: + rec_cp2_do_mfc2(state, block, offset, c.i.rt, REG_CP2_TEMP); + break; + default: + break; + } + } - switch (LIGHTREC_FLAGS_GET_IO_MODE(flags)) { + switch (mode) { case LIGHTREC_IO_RAM: rec_store_ram(state, block, offset, code, swap_code, !no_invalidate); @@ -1440,8 +1465,11 @@ static void rec_store(struct lightrec_cstate *state, break; default: rec_io(state, block, offset, true, false); - break; + return; } + + if (is_swc2) + lightrec_discard_reg_if_loaded(state->reg_cache, REG_CP2_TEMP); } static void rec_SB(struct lightrec_cstate *state, @@ -1463,7 +1491,9 @@ static void rec_SW(struct lightrec_cstate *state, const struct block *block, u16 offset) { - _jit_name(block->_jit, __func__); + union code c = block->opcode_list[offset].c; + + _jit_name(block->_jit, c.i.op == OP_SWC2 ? "rec_SWC2" : "rec_SW"); rec_store(state, block, offset, jit_code_stxi_i, jit_code_bswapr_ui); } @@ -1482,13 +1512,6 @@ static void rec_SWR(struct lightrec_cstate *state, rec_io(state, block, offset, true, false); } -static void rec_SWC2(struct lightrec_cstate *state, - const struct block *block, u16 offset) -{ - _jit_name(block->_jit, __func__); - rec_io(state, block, offset, false, false); -} - static void rec_load_memory(struct lightrec_cstate *cstate, const struct block *block, u16 offset, jit_code_t code, jit_code_t swap_code, bool is_unsigned, @@ -1497,19 +1520,23 @@ static void rec_load_memory(struct lightrec_cstate *cstate, struct regcache *reg_cache = cstate->reg_cache; struct opcode *op = &block->opcode_list[offset]; jit_state_t *_jit = block->_jit; - u8 rs, rt, addr_reg, flags = REG_EXT; + u8 rs, rt, out_reg, addr_reg, flags = REG_EXT; bool no_mask = op_flag_no_mask(op->flags); union code c = op->c; s16 imm; - if (!c.i.rt) + if (c.i.op == OP_LWC2) + out_reg = REG_CP2_TEMP; + else if (c.i.rt) + out_reg = c.i.rt; + else return; if (is_unsigned) flags |= REG_ZEXT; rs = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rs, 0); - rt = lightrec_alloc_reg_out(reg_cache, _jit, c.i.rt, flags); + rt = lightrec_alloc_reg_out(reg_cache, _jit, out_reg, flags); if (!cstate->state->mirrors_mapped && c.i.imm && !no_mask) { jit_addi(rt, rs, (s16)c.i.imm); @@ -1582,7 +1609,7 @@ static void rec_load_io(struct lightrec_cstate *cstate, _jit_note(block->_jit, __FILE__, __LINE__); rec_load_memory(cstate, block, offset, code, swap_code, is_unsigned, - cstate->state->offset_io, 0x1fffffff); + cstate->state->offset_io, rec_io_mask(cstate->state)); } static void rec_load_direct(struct lightrec_cstate *cstate, @@ -1595,10 +1622,14 @@ static void rec_load_direct(struct lightrec_cstate *cstate, union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; jit_node_t *to_not_ram, *to_not_bios, *to_end, *to_end2; - u8 tmp, rs, rt, addr_reg, flags = REG_EXT; + u8 tmp, rs, rt, out_reg, addr_reg, flags = REG_EXT; s16 imm; - if (!c.i.rt) + if (c.i.op == OP_LWC2) + out_reg = REG_CP2_TEMP; + else if (c.i.rt) + out_reg = c.i.rt; + else return; if (is_unsigned) @@ -1606,7 +1637,7 @@ static void rec_load_direct(struct lightrec_cstate *cstate, jit_note(__FILE__, __LINE__); rs = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rs, 0); - rt = lightrec_alloc_reg_out(reg_cache, _jit, c.i.rt, flags); + rt = lightrec_alloc_reg_out(reg_cache, _jit, out_reg, flags); if ((state->offset_ram == state->offset_bios && state->offset_ram == state->offset_scratch && @@ -1698,7 +1729,8 @@ static void rec_load(struct lightrec_cstate *state, const struct block *block, u16 offset, jit_code_t code, jit_code_t swap_code, bool is_unsigned) { - u32 flags = block->opcode_list[offset].flags; + const struct opcode *op = &block->opcode_list[offset]; + u32 flags = op->flags; switch (LIGHTREC_FLAGS_GET_IO_MODE(flags)) { case LIGHTREC_IO_RAM: @@ -1718,7 +1750,12 @@ static void rec_load(struct lightrec_cstate *state, const struct block *block, break; default: rec_io(state, block, offset, false, true); - break; + return; + } + + if (op->i.op == OP_LWC2) { + rec_cp2_do_mtc2(state, block, offset, op->i.rt, REG_CP2_TEMP); + lightrec_discard_reg_if_loaded(state->reg_cache, REG_CP2_TEMP); } } @@ -1762,6 +1799,7 @@ static void rec_LWR(struct lightrec_cstate *state, const struct block *block, u1 static void rec_LW(struct lightrec_cstate *state, const struct block *block, u16 offset) { + union code c = block->opcode_list[offset].c; jit_code_t code; if (is_big_endian() && __WORDSIZE == 64) @@ -1769,16 +1807,10 @@ static void rec_LW(struct lightrec_cstate *state, const struct block *block, u16 else code = jit_code_ldxi_i; - _jit_name(block->_jit, __func__); + _jit_name(block->_jit, c.i.op == OP_LWC2 ? "rec_LWC2" : "rec_LW"); rec_load(state, block, offset, code, jit_code_bswapr_ui, false); } -static void rec_LWC2(struct lightrec_cstate *state, const struct block *block, u16 offset) -{ - _jit_name(block->_jit, __func__); - rec_io(state, block, offset, false, false); -} - static void rec_break_syscall(struct lightrec_cstate *state, const struct block *block, u16 offset, u32 exit_code) @@ -1817,6 +1849,20 @@ static void rec_special_BREAK(struct lightrec_cstate *state, rec_break_syscall(state, block, offset, LIGHTREC_EXIT_BREAK); } +static void rec_mfc(struct lightrec_cstate *state, const struct block *block, u16 offset) +{ + struct regcache *reg_cache = state->reg_cache; + union code c = block->opcode_list[offset].c; + jit_state_t *_jit = block->_jit; + + jit_note(__FILE__, __LINE__); + + if (c.i.op != OP_SWC2) + lightrec_clean_reg_if_loaded(reg_cache, _jit, c.i.rt, true); + + call_to_c_wrapper(state, block, c.opcode, C_WRAPPER_MFC); +} + static void rec_mtc(struct lightrec_cstate *state, const struct block *block, u16 offset) { struct regcache *reg_cache = state->reg_cache; @@ -1871,6 +1917,7 @@ rec_mtc0(struct lightrec_cstate *state, const struct block *block, u16 offset) const union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; u8 rt, tmp = 0, tmp2, status; + jit_node_t *to_end; jit_note(__FILE__, __LINE__); @@ -1943,19 +1990,29 @@ rec_mtc0(struct lightrec_cstate *state, const struct block *block, u16 offset) jit_orr(tmp, tmp, tmp2); } + lightrec_free_reg(reg_cache, rt); + if (c.r.rd == 12 || c.r.rd == 13) { - jit_stxi_i(offsetof(struct lightrec_state, exit_flags), - LIGHTREC_REG_STATE, tmp); + to_end = jit_beqi(tmp, 0); - lightrec_free_reg(reg_cache, tmp); - lightrec_free_reg(reg_cache, tmp2); - } + jit_ldxi_i(tmp2, LIGHTREC_REG_STATE, + offsetof(struct lightrec_state, target_cycle)); + jit_subr(tmp2, tmp2, LIGHTREC_REG_CYCLE); + jit_movi(LIGHTREC_REG_CYCLE, 0); + jit_stxi_i(offsetof(struct lightrec_state, target_cycle), + LIGHTREC_REG_STATE, tmp2); + jit_stxi_i(offsetof(struct lightrec_state, current_cycle), + LIGHTREC_REG_STATE, tmp2); - lightrec_free_reg(reg_cache, rt); + + jit_patch(to_end); + } if (!op_flag_no_ds(block->opcode_list[offset].flags) && - (c.r.rd == 12 || c.r.rd == 13)) - lightrec_emit_eob(state, block, offset + 1, true); + (c.r.rd == 12 || c.r.rd == 13)) { + state->cycles += lightrec_cycles_of_opcode(c); + lightrec_emit_eob(state, block, offset + 1); + } } static void rec_cp0_MFC0(struct lightrec_cstate *state, @@ -2006,21 +2063,29 @@ static unsigned int cp2c_s_offset(u8 reg) return cp2c_i_offset(reg) + is_big_endian() * 2; } -static void rec_cp2_basic_MFC2(struct lightrec_cstate *state, - const struct block *block, u16 offset) +static void rec_cp2_do_mfc2(struct lightrec_cstate *state, + const struct block *block, u16 offset, + u8 reg, u8 out_reg) { struct regcache *reg_cache = state->reg_cache; - const union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; const u32 zext_regs = 0x300f0080; u8 rt, tmp, tmp2, tmp3, out, flags; - u8 reg = c.r.rd == 15 ? 14 : c.r.rd; unsigned int i; _jit_name(block->_jit, __func__); + if (state->state->ops.cop2_notify) { + /* We must call cop2_notify, handle that in C. */ + rec_mfc(state, block, offset); + return; + } + flags = (zext_regs & BIT(reg)) ? REG_ZEXT : REG_EXT; - rt = lightrec_alloc_reg_out(reg_cache, _jit, c.r.rt, flags); + rt = lightrec_alloc_reg_out(reg_cache, _jit, out_reg, flags); + + if (reg == 15) + reg = 14; switch (reg) { case 1: @@ -2077,6 +2142,14 @@ static void rec_cp2_basic_MFC2(struct lightrec_cstate *state, lightrec_free_reg(reg_cache, rt); } +static void rec_cp2_basic_MFC2(struct lightrec_cstate *state, + const struct block *block, u16 offset) +{ + const union code c = block->opcode_list[offset].c; + + rec_cp2_do_mfc2(state, block, offset, c.r.rd, c.r.rt); +} + static void rec_cp2_basic_CFC2(struct lightrec_cstate *state, const struct block *block, u16 offset) { @@ -2087,6 +2160,12 @@ static void rec_cp2_basic_CFC2(struct lightrec_cstate *state, _jit_name(block->_jit, __func__); + if (state->state->ops.cop2_notify) { + /* We must call cop2_notify, handle that in C. */ + rec_mfc(state, block, offset); + return; + } + switch (c.r.rd) { case 4: case 12: @@ -2100,33 +2179,39 @@ static void rec_cp2_basic_CFC2(struct lightrec_cstate *state, break; default: rt = lightrec_alloc_reg_out(reg_cache, _jit, c.r.rt, REG_ZEXT); - jit_ldxi_i(rt, LIGHTREC_REG_STATE, cp2c_i_offset(c.r.rd)); + jit_ldxi_ui(rt, LIGHTREC_REG_STATE, cp2c_i_offset(c.r.rd)); break; } lightrec_free_reg(reg_cache, rt); } -static void rec_cp2_basic_MTC2(struct lightrec_cstate *state, - const struct block *block, u16 offset) +static void rec_cp2_do_mtc2(struct lightrec_cstate *state, + const struct block *block, u16 offset, + u8 reg, u8 in_reg) { struct regcache *reg_cache = state->reg_cache; - const union code c = block->opcode_list[offset].c; jit_state_t *_jit = block->_jit; jit_node_t *loop, *to_loop; u8 rt, tmp, tmp2, flags = 0; _jit_name(block->_jit, __func__); - if (c.r.rd == 31) + if (state->state->ops.cop2_notify) { + /* We must call cop2_notify, handle that in C. */ + rec_mtc(state, block, offset); + return; + } + + if (reg == 31) return; - if (c.r.rd == 30) + if (reg == 30) flags |= REG_EXT; - rt = lightrec_alloc_reg_in(reg_cache, _jit, c.r.rt, flags); + rt = lightrec_alloc_reg_in(reg_cache, _jit, in_reg, flags); - switch (c.r.rd) { + switch (reg) { case 15: tmp = lightrec_alloc_reg_temp(reg_cache, _jit); jit_ldxi_i(tmp, LIGHTREC_REG_STATE, cp2d_i_offset(13)); @@ -2185,13 +2270,21 @@ static void rec_cp2_basic_MTC2(struct lightrec_cstate *state, lightrec_free_reg(reg_cache, tmp2); break; default: - jit_stxi_i(cp2d_i_offset(c.r.rd), LIGHTREC_REG_STATE, rt); + jit_stxi_i(cp2d_i_offset(reg), LIGHTREC_REG_STATE, rt); break; } lightrec_free_reg(reg_cache, rt); } +static void rec_cp2_basic_MTC2(struct lightrec_cstate *state, + const struct block *block, u16 offset) +{ + const union code c = block->opcode_list[offset].c; + + rec_cp2_do_mtc2(state, block, offset, c.r.rd, c.r.rt); +} + static void rec_cp2_basic_CTC2(struct lightrec_cstate *state, const struct block *block, u16 offset) { @@ -2202,6 +2295,12 @@ static void rec_cp2_basic_CTC2(struct lightrec_cstate *state, _jit_name(block->_jit, __func__); + if (state->state->ops.cop2_notify) { + /* We must call cop2_notify, handle that in C. */ + rec_mtc(state, block, offset); + return; + } + rt = lightrec_alloc_reg_in(reg_cache, _jit, c.r.rt, 0); switch (c.r.rd) { @@ -2294,24 +2393,46 @@ static void rec_meta_MOV(struct lightrec_cstate *state, const struct block *block, u16 offset) { struct regcache *reg_cache = state->reg_cache; - union code c = block->opcode_list[offset].c; + const struct opcode *op = &block->opcode_list[offset]; + union code c = op->c; jit_state_t *_jit = block->_jit; + bool unload_rd; u8 rs, rd; _jit_name(block->_jit, __func__); jit_note(__FILE__, __LINE__); - if (c.r.rs) + + unload_rd = OPT_EARLY_UNLOAD + && LIGHTREC_FLAGS_GET_RD(op->flags) == LIGHTREC_REG_UNLOAD; + + if (c.r.rs || unload_rd) rs = lightrec_alloc_reg_in(reg_cache, _jit, c.r.rs, 0); - rd = lightrec_alloc_reg_out(reg_cache, _jit, c.r.rd, REG_EXT); - if (c.r.rs == 0) - jit_movi(rd, 0); - else - jit_extr_i(rd, rs); + if (unload_rd) { + /* If the destination register will be unloaded right after the + * MOV meta-opcode, we don't actually need to write any host + * register - we can just store the source register directly to + * the register cache, at the offset corresponding to the + * destination register. */ + lightrec_discard_reg_if_loaded(reg_cache, c.r.rd); + + jit_stxi_i(offsetof(struct lightrec_state, regs.gpr) + + c.r.rd << 2, LIGHTREC_REG_STATE, rs); - if (c.r.rs) lightrec_free_reg(reg_cache, rs); - lightrec_free_reg(reg_cache, rd); + } else { + rd = lightrec_alloc_reg_out(reg_cache, _jit, c.r.rd, REG_EXT); + + if (c.r.rs == 0) + jit_movi(rd, 0); + else + jit_extr_i(rd, rs); + + lightrec_free_reg(reg_cache, rd); + } + + if (c.r.rs || unload_rd) + lightrec_free_reg(reg_cache, rs); } static void rec_meta_EXTC_EXTS(struct lightrec_cstate *state, @@ -2350,6 +2471,7 @@ static void rec_meta_MULT2(struct lightrec_cstate *state, u32 flags = block->opcode_list[offset].flags; bool is_signed = c.i.op == OP_META_MULT2; u8 rs, lo, hi, rflags = 0, hiflags = 0; + unsigned int i; if (!op_flag_no_hi(flags) && c.r.op < 32) { rflags = is_signed ? REG_EXT : REG_ZEXT; @@ -2361,28 +2483,38 @@ static void rec_meta_MULT2(struct lightrec_cstate *state, rs = lightrec_alloc_reg_in(reg_cache, _jit, c.i.rs, rflags); - if (!op_flag_no_lo(flags)) { - lo = lightrec_alloc_reg_out(reg_cache, _jit, reg_lo, 0); - - if (c.r.op < 32) - jit_lshi(lo, rs, c.r.op); - else - jit_movi(lo, 0); + /* + * We must handle the case where one of the output registers is our rs + * input register. Thanksfully, computing LO/HI can be done in any + * order. Here, we make sure that the computation that overwrites the + * input register is always performed last. + */ + for (i = 0; i < 2; i++) { + if ((!i ^ (reg_lo == c.i.rs)) && !op_flag_no_lo(flags)) { + lo = lightrec_alloc_reg_out(reg_cache, _jit, reg_lo, 0); + + if (c.r.op < 32) + jit_lshi(lo, rs, c.r.op); + else + jit_movi(lo, 0); - lightrec_free_reg(reg_cache, lo); - } + lightrec_free_reg(reg_cache, lo); + continue; + } - if (!op_flag_no_hi(flags)) { - hi = lightrec_alloc_reg_out(reg_cache, _jit, reg_hi, hiflags); + if ((!!i ^ (reg_lo == c.i.rs)) && !op_flag_no_hi(flags)) { + hi = lightrec_alloc_reg_out(reg_cache, _jit, + reg_hi, hiflags); - if (c.r.op >= 32) - jit_lshi(hi, rs, c.r.op - 32); - else if (is_signed) - jit_rshi(hi, rs, 32 - c.r.op); - else - jit_rshi_u(hi, rs, 32 - c.r.op); + if (c.r.op >= 32) + jit_lshi(hi, rs, c.r.op - 32); + else if (is_signed) + jit_rshi(hi, rs, 32 - c.r.op); + else + jit_rshi_u(hi, rs, 32 - c.r.op); - lightrec_free_reg(reg_cache, hi); + lightrec_free_reg(reg_cache, hi); + } } lightrec_free_reg(reg_cache, rs); @@ -2423,8 +2555,8 @@ static const lightrec_rec_func_t rec_standard[64] = { [OP_SWL] = rec_SWL, [OP_SW] = rec_SW, [OP_SWR] = rec_SWR, - [OP_LWC2] = rec_LWC2, - [OP_SWC2] = rec_SWC2, + [OP_LWC2] = rec_LW, + [OP_SWC2] = rec_SW, [OP_META_MOV] = rec_meta_MOV, [OP_META_EXTC] = rec_meta_EXTC_EXTS,