X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=deps%2Flightrec%2Foptimizer.c;h=562f7e009ac98692209000c2a6ae3ca7375c7f15;hb=02487de7ff9fcbb6d7d692a6b3ae6e6539708abc;hp=f719192b6997cab3a2dc9b47a73dd9d86994aa33;hpb=fd58fa3240e46c4d6027c28f3882a4c584af7123;p=pcsx_rearmed.git diff --git a/deps/lightrec/optimizer.c b/deps/lightrec/optimizer.c index f719192b..562f7e00 100644 --- a/deps/lightrec/optimizer.c +++ b/deps/lightrec/optimizer.c @@ -277,7 +277,8 @@ static bool reg_is_dead(const struct opcode *list, unsigned int offset, u8 reg) return true; if (has_delay_slot(list[i].c)) { - if (list[i].flags & LIGHTREC_NO_DS) + if (list[i].flags & LIGHTREC_NO_DS || + opcode_reads_register(list[i + 1].c, reg)) return false; return opcode_writes_register(list[i + 1].c, reg); @@ -459,16 +460,18 @@ bool load_in_delay_slot(union code op) return false; } -static u32 lightrec_propagate_consts(const struct opcode *op, u32 known, u32 *v) +static u32 lightrec_propagate_consts(const struct opcode *op, + const struct opcode *prev, + u32 known, u32 *v) { - union code c = op->c; + union code c = prev->c; /* Register $zero is always, well, zero */ known |= BIT(0); v[0] = 0; if (op->flags & LIGHTREC_SYNC) - return 0; + return BIT(0); switch (c.i.op) { case OP_SPECIAL: @@ -814,18 +817,81 @@ static void lightrec_optimize_sll_sra(struct opcode *list, unsigned int offset) to_nop->opcode = 0; } +static void lightrec_remove_useless_lui(struct block *block, unsigned int offset, + u32 known, u32 *values) +{ + struct opcode *list = block->opcode_list, + *op = &block->opcode_list[offset]; + int reader; + + if (!(op->flags & LIGHTREC_SYNC) && (known & BIT(op->i.rt)) && + values[op->i.rt] == op->i.imm << 16) { + pr_debug("Converting duplicated LUI to NOP\n"); + op->opcode = 0x0; + return; + } + + if (op->i.imm != 0 || op->i.rt == 0) + return; + + reader = find_next_reader(list, offset + 1, op->i.rt); + if (reader <= 0) + return; + + if (opcode_writes_register(list[reader].c, op->i.rt) || + reg_is_dead(list, reader, op->i.rt)) { + pr_debug("Removing useless LUI 0x0\n"); + + if (list[reader].i.rs == op->i.rt) + list[reader].i.rs = 0; + if (list[reader].i.op == OP_SPECIAL && + list[reader].i.rt == op->i.rt) + list[reader].i.rt = 0; + op->opcode = 0x0; + } +} + +static void lightrec_modify_lui(struct block *block, unsigned int offset) +{ + union code c, *lui = &block->opcode_list[offset].c; + bool stop = false, stop_next = false; + unsigned int i; + + for (i = offset + 1; !stop && i < block->nb_ops; i++) { + c = block->opcode_list[i].c; + stop = stop_next; + + if ((opcode_is_store(c) && c.i.rt == lui->i.rt) + || (!opcode_is_load(c) && opcode_reads_register(c, lui->i.rt))) + break; + + if (opcode_writes_register(c, lui->i.rt)) { + pr_debug("Convert LUI at offset 0x%x to kuseg\n", + i - 1 << 2); + lui->i.imm = kunseg(lui->i.imm << 16) >> 16; + break; + } + + if (has_delay_slot(c)) + stop_next = true; + } +} + static int lightrec_transform_ops(struct lightrec_state *state, struct block *block) { struct opcode *list = block->opcode_list; - struct opcode *op; + struct opcode *prev, *op = NULL; u32 known = BIT(0); u32 values[32] = { 0 }; unsigned int i; - int reader; for (i = 0; i < block->nb_ops; i++) { + prev = op; op = &list[i]; + if (prev) + known = lightrec_propagate_consts(op, prev, known, values); + /* Transform all opcodes detected as useless to real NOPs * (0x0: SLL r0, r0, #0) */ if (op->opcode != 0 && is_nop(op->c)) { @@ -856,30 +922,8 @@ static int lightrec_transform_ops(struct lightrec_state *state, struct block *bl break; case OP_LUI: - if (!(op->flags & LIGHTREC_SYNC) && - (known & BIT(op->i.rt)) && - values[op->i.rt] == op->i.imm << 16) { - pr_debug("Converting duplicated LUI to NOP\n"); - op->opcode = 0x0; - } - - if (op->i.imm != 0 || op->i.rt == 0) - break; - - reader = find_next_reader(list, i + 1, op->i.rt); - if (reader > 0 && - (opcode_writes_register(list[reader].c, op->i.rt) || - reg_is_dead(list, reader, op->i.rt))) { - - pr_debug("Removing useless LUI 0x0\n"); - - if (list[reader].i.rs == op->i.rt) - list[reader].i.rs = 0; - if (list[reader].i.op == OP_SPECIAL && - list[reader].i.rt == op->i.rt) - list[reader].i.rt = 0; - op->opcode = 0x0; - } + lightrec_modify_lui(block, i); + lightrec_remove_useless_lui(block, i, known, values); break; /* Transform ORI/ADDI/ADDIU with imm #0 or ORR/ADD/ADDU/SUB/SUBU @@ -933,8 +977,6 @@ static int lightrec_transform_ops(struct lightrec_state *state, struct block *bl default: /* fall-through */ break; } - - known = lightrec_propagate_consts(op, known, values); } return 0; @@ -1228,16 +1270,20 @@ static int lightrec_early_unload(struct lightrec_state *state, struct block *blo static int lightrec_flag_io(struct lightrec_state *state, struct block *block) { - const struct lightrec_mem_map *map; - struct opcode *list; + struct opcode *prev = NULL, *list = NULL; + enum psx_map psx_map; u32 known = BIT(0); u32 values[32] = { 0 }; unsigned int i; - u32 val; + u32 val, kunseg_val; for (i = 0; i < block->nb_ops; i++) { + prev = list; list = &block->opcode_list[i]; + if (prev) + known = lightrec_propagate_consts(list, prev, known, values); + switch (list->i.op) { case OP_SB: case OP_SH: @@ -1279,24 +1325,43 @@ static int lightrec_flag_io(struct lightrec_state *state, struct block *block) case OP_LWR: case OP_LWC2: if (OPT_FLAG_IO && (known & BIT(list->i.rs))) { - val = kunseg(values[list->i.rs] + (s16) list->i.imm); - map = lightrec_get_map(state, NULL, val); + val = values[list->i.rs] + (s16) list->i.imm; + kunseg_val = kunseg(val); + psx_map = lightrec_get_map_idx(state, kunseg_val); + + switch (psx_map) { + case PSX_MAP_KERNEL_USER_RAM: + if (val == kunseg_val) + list->flags |= LIGHTREC_NO_MASK; + /* fall-through */ + case PSX_MAP_MIRROR1: + case PSX_MAP_MIRROR2: + case PSX_MAP_MIRROR3: + pr_debug("Flaging opcode %u as RAM access\n", i); + list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_RAM); + break; + case PSX_MAP_BIOS: + pr_debug("Flaging opcode %u as BIOS access\n", i); + list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_BIOS); + break; + case PSX_MAP_SCRATCH_PAD: + pr_debug("Flaging opcode %u as scratchpad access\n", i); + list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_SCRATCH); - if (!map || map->ops || - map == &state->maps[PSX_MAP_PARALLEL_PORT]) { - pr_debug("Flagging opcode %u as accessing I/O registers\n", + /* Consider that we're never going to run code from + * the scratchpad. */ + list->flags |= LIGHTREC_NO_INVALIDATE; + break; + default: + pr_debug("Flagging opcode %u as I/O access\n", i); - list->flags |= LIGHTREC_HW_IO; - } else { - pr_debug("Flaging opcode %u as direct memory access\n", i); - list->flags |= LIGHTREC_DIRECT_IO; + list->flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW); + break; } } default: /* fall-through */ break; } - - known = lightrec_propagate_consts(list, known, values); } return 0; @@ -1483,15 +1548,19 @@ static bool lightrec_always_skip_div_check(void) static int lightrec_flag_mults_divs(struct lightrec_state *state, struct block *block) { - struct opcode *list; + struct opcode *prev, *list = NULL; u8 reg_hi, reg_lo; unsigned int i; u32 known = BIT(0); u32 values[32] = { 0 }; for (i = 0; i < block->nb_ops - 1; i++) { + prev = list; list = &block->opcode_list[i]; + if (prev) + known = lightrec_propagate_consts(list, prev, known, values); + if (list->i.op != OP_SPECIAL) continue; @@ -1507,14 +1576,12 @@ static int lightrec_flag_mults_divs(struct lightrec_state *state, struct block * case OP_SPECIAL_MULTU: break; default: - known = lightrec_propagate_consts(list, known, values); continue; } /* Don't support opcodes in delay slots */ if ((i && has_delay_slot(block->opcode_list[i - 1].c)) || (list->flags & LIGHTREC_NO_DS)) { - known = lightrec_propagate_consts(list, known, values); continue; } @@ -1558,8 +1625,6 @@ static int lightrec_flag_mults_divs(struct lightrec_state *state, struct block * } else { list->r.imm = 0; } - - known = lightrec_propagate_consts(list, known, values); } return 0;