X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=cpu%2Fsh2%2Fcompiler.c;h=df2f0d390a1bb16f3167f5eab3c65a2889ba66b8;hb=e15536775962fef2058ffa43cbb290271f649c17;hp=644714e81f641f01ce97d10f102a2030c362a94e;hpb=569420b0d64d27db5f55de50b26d70f130c37cf7;p=picodrive.git diff --git a/cpu/sh2/compiler.c b/cpu/sh2/compiler.c index 644714e8..df2f0d39 100644 --- a/cpu/sh2/compiler.c +++ b/cpu/sh2/compiler.c @@ -41,11 +41,11 @@ #define LINK_BRANCHES 1 // limits (per block) -#define MAX_BLOCK_SIZE (BLOCK_CYCLE_LIMIT * 6 * 6) +#define MAX_BLOCK_SIZE (BLOCK_INSN_LIMIT * 6 * 6) // max literal offset from the block end #define MAX_LITERAL_OFFSET 32*2 -#define MAX_LITERALS (BLOCK_CYCLE_LIMIT / 4) +#define MAX_LITERALS (BLOCK_INSN_LIMIT / 4) #define MAX_LOCAL_BRANCHES 32 /// @@ -58,9 +58,10 @@ #ifdef DRC_SH2 // debug stuff -// 1 - ? -// 2 - ? -// 4 - log asm +// 1 - warnings/errors +// 2 - block info/smc +// 4 - asm +// 8 - runtime entries // { #ifndef DRC_DEBUG #define DRC_DEBUG 0 @@ -112,7 +113,7 @@ static void REGPARM(3) *sh2_drc_log_entry(void *block, SH2 *sh2, u32 sr) // and can be discarded early // XXX: need to tune sizes static const int tcache_sizes[TCACHE_BUFFERS] = { - DRC_TCACHE_SIZE * 6 / 8, // ROM, DRAM + DRC_TCACHE_SIZE * 6 / 8, // ROM (rarely used), DRAM DRC_TCACHE_SIZE / 8, // BIOS, data array in master sh2 DRC_TCACHE_SIZE / 8, // ... slave }; @@ -123,34 +124,48 @@ static u8 *tcache_ptrs[TCACHE_BUFFERS]; // ptr for code emiters static u8 *tcache_ptr; -typedef struct block_desc_ { +struct block_desc { u32 addr; // SH2 PC address u32 end_addr; // address after last op void *tcache_ptr; // translated block for above PC - struct block_desc_ *next; // next block with the same PC hash + struct block_desc *next; // next block with the same PC hash #if (DRC_DEBUG & 2) int refcount; #endif -} block_desc; +}; -typedef struct block_link_ { +struct block_link { u32 target_pc; void *jump; // insn address // struct block_link_ *next; -} block_link; +}; static const int block_max_counts[TCACHE_BUFFERS] = { 4*1024, 256, 256, }; -static block_desc *block_tables[TCACHE_BUFFERS]; -static block_link *block_links[TCACHE_BUFFERS]; +static struct block_desc *block_tables[TCACHE_BUFFERS]; +static struct block_link *block_links[TCACHE_BUFFERS]; static int block_counts[TCACHE_BUFFERS]; static int block_link_counts[TCACHE_BUFFERS]; -#define BLOCKID_OVERLAP 0xfffe -#define BLOCKID_MAX block_max_counts[0] +// used for invalidation +static const int ram_sizes[TCACHE_BUFFERS] = { + 0x40000, + 0x1000, + 0x1000, +}; +#define ADDR_TO_BLOCK_PAGE 0x100 + +struct block_list { + struct block_desc *block; + struct block_list *next; +}; + +// array of pointers to block_lists for RAM and 2 data arrays +// each array has len: sizeof(mem) / ADDR_TO_BLOCK_PAGE +static struct block_list **inval_lookup[TCACHE_BUFFERS]; // host register tracking enum { @@ -234,10 +249,10 @@ static temp_reg_t reg_temp[] = { // ROM hash table #define MAX_HASH_ENTRIES 1024 #define HASH_MASK (MAX_HASH_ENTRIES - 1) -static void **hash_table; +static struct block_desc **hash_table; #define HASH_FUNC(hash_tab, addr) \ - ((block_desc **)(hash_tab))[(addr) & HASH_MASK] + (hash_tab)[(addr) & HASH_MASK] static void REGPARM(1) (*sh2_drc_entry)(SH2 *sh2); static void (*sh2_drc_dispatcher)(void); @@ -282,7 +297,7 @@ static int dr_ctx_get_mem_ptr(u32 a, u32 *mask) return poffs; } -static block_desc *dr_get_bd(u32 pc, int is_slave, int *tcache_id) +static struct block_desc *dr_get_bd(u32 pc, int is_slave, int *tcache_id) { *tcache_id = 0; @@ -302,7 +317,7 @@ static block_desc *dr_get_bd(u32 pc, int is_slave, int *tcache_id) } // ROM else if ((pc & 0xc6000000) == 0x02000000) { - block_desc *bd = HASH_FUNC(hash_table, pc); + struct block_desc *bd = HASH_FUNC(hash_table, pc); for (; bd != NULL; bd = bd->next) if (bd->addr == pc) @@ -315,8 +330,50 @@ static block_desc *dr_get_bd(u32 pc, int is_slave, int *tcache_id) // --------------------------------------------------------------- // block management +static void add_to_block_list(struct block_list **blist, struct block_desc *block) +{ + struct block_list *added = malloc(sizeof(*added)); + if (!added) { + elprintf(EL_ANOMALY, "drc OOM (1)"); + return; + } + added->block = block; + added->next = *blist; + *blist = added; +} + +static void rm_from_block_list(struct block_list **blist, struct block_desc *block) +{ + struct block_list *prev = NULL, *current = *blist; + for (; current != NULL; prev = current, current = current->next) { + if (current->block == block) { + if (prev == NULL) + *blist = current->next; + else + prev->next = current->next; + free(current); + return; + } + } + dbg(1, "can't rm block %p (%08x-%08x)", + block, block->addr, block->end_addr); +} + +static void rm_block_list(struct block_list **blist) +{ + struct block_list *tmp, *current = *blist; + while (current != NULL) { + tmp = current; + current = current->next; + free(tmp); + } + *blist = NULL; +} + static void REGPARM(1) flush_tcache(int tcid) { + int i; + dbg(1, "tcache #%d flush! (%d/%d, bds %d/%d)", tcid, tcache_ptrs[tcid] - tcache_bases[tcid], tcache_sizes[tcid], block_counts[tcid], block_max_counts[tcid]); @@ -333,13 +390,16 @@ static void REGPARM(1) flush_tcache(int tcid) #if (DRC_DEBUG & 4) tcache_dsm_ptrs[tcid] = tcache_bases[tcid]; #endif + + for (i = 0; i < ram_sizes[tcid] / ADDR_TO_BLOCK_PAGE; i++) + rm_block_list(&inval_lookup[tcid][i]); } #if LINK_BRANCHES // add block links (tracked branches) static int dr_add_block_link(u32 target_pc, void *jump, int tcache_id) { - block_link *bl = block_links[tcache_id]; + struct block_link *bl = block_links[tcache_id]; int cnt = block_link_counts[tcache_id]; if (cnt >= block_max_counts[tcache_id] * 2) { @@ -355,9 +415,9 @@ static int dr_add_block_link(u32 target_pc, void *jump, int tcache_id) } #endif -static block_desc *dr_add_block(u32 addr, u32 end_addr, int is_slave, int *blk_id) +static struct block_desc *dr_add_block(u32 addr, u32 end_addr, int is_slave, int *blk_id) { - block_desc *bd; + struct block_desc *bd; int tcache_id; int *bcount; @@ -401,7 +461,7 @@ static block_desc *dr_add_block(u32 addr, u32 end_addr, int is_slave, int *blk_i static void REGPARM(3) *dr_lookup_block(u32 pc, int is_slave, int *tcache_id) { - block_desc *bd = NULL; + struct block_desc *bd = NULL; void *block = NULL; bd = dr_get_bd(pc, is_slave, tcache_id); @@ -447,7 +507,7 @@ static void *dr_prepare_ext_branch(u32 pc, SH2 *sh2, int tcache_id) static void dr_link_blocks(void *target, u32 pc, int tcache_id) { #if LINK_BRANCHES - block_link *bl = block_links[tcache_id]; + struct block_link *bl = block_links[tcache_id]; int cnt = block_link_counts[tcache_id]; int i; @@ -1198,7 +1258,7 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id) int literal_addr_count = 0; int pending_branch_cond = -1; int pending_branch_pc = 0; - u8 op_flags[BLOCK_CYCLE_LIMIT]; + u8 op_flags[BLOCK_INSN_LIMIT]; struct { u32 delayed_op:2; u32 test_irq:1; @@ -1208,7 +1268,7 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id) // PC of current, first, last, last_target_blk SH2 insn u32 pc, base_pc, end_pc, out_pc; void *block_entry; - block_desc *this_block; + struct block_desc *this_block; u16 *dr_pc_base; int blkid_main = 0; int skip_op = 0; @@ -1245,17 +1305,17 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id) return NULL; block_entry = tcache_ptr; - dbg(2, "== %csh2 block #%d,%d %08x -> %p", sh2->is_slave ? 's' : 'm', - tcache_id, blkid_main, base_pc, block_entry); + dbg(2, "== %csh2 block #%d,%d %08x-%08x -> %p", sh2->is_slave ? 's' : 'm', + tcache_id, blkid_main, base_pc, end_pc, block_entry); dr_link_blocks(tcache_ptr, base_pc, tcache_id); // collect branch_targets that don't land on delay slots - for (pc = base_pc; pc <= end_pc; pc += 2) { - if (!(OP_FLAGS(pc) & OF_TARGET)) + for (pc = base_pc; pc < end_pc; pc += 2) { + if (!(OP_FLAGS(pc) & OF_BTARGET)) continue; if (OP_FLAGS(pc) & OF_DELAY_OP) { - OP_FLAGS(pc) &= ~OF_TARGET; + OP_FLAGS(pc) &= ~OF_BTARGET; continue; } ADD_TO_ARRAY(branch_target_pc, branch_target_count, pc, break); @@ -1279,13 +1339,13 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id) op = FETCH_OP(pc); - if ((OP_FLAGS(pc) & OF_TARGET) || pc == base_pc) + if ((OP_FLAGS(pc) & OF_BTARGET) || pc == base_pc) { i = find_in_array(branch_target_pc, branch_target_count, pc); if (pc != base_pc) { /* make "subblock" - just a mid-block entry */ - block_desc *subblock; + struct block_desc *subblock; sr = rcache_get_reg(SHR_SR, RC_GR_RMW); FLUSH_CYCLES(sr); @@ -1327,7 +1387,8 @@ static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id) #endif #if (DRC_DEBUG & 4) DasmSH2(sh2dasm_buff, pc, op); - printf("%08x %04x %s\n", pc, op, sh2dasm_buff); + printf("%c%08x %04x %s\n", (OP_FLAGS(pc) & OF_BTARGET) ? '*' : ' ', + pc, op, sh2dasm_buff); #endif #ifdef DRC_CMP //if (out_pc != 0 && out_pc != (u32)-1) @@ -2582,47 +2643,48 @@ end_op: if (tcache_id != 0 || (this_block->addr & 0xc7fc0000) == 0x06000000) { u16 *p, *drc_ram_blk = NULL; - u32 mask = 0, shift = 0; + u32 addr, mask = 0, shift = 0; if (tcache_id != 0) { // data array, BIOS drc_ram_blk = Pico32xMem->drcblk_da[sh2->is_slave]; shift = SH2_DRCBLK_DA_SHIFT; - mask = 0xfff/2; + mask = 0xfff; } else if ((this_block->addr & 0xc7fc0000) == 0x06000000) { // SDRAM drc_ram_blk = Pico32xMem->drcblk_ram; shift = SH2_DRCBLK_RAM_SHIFT; - mask = 0x3ffff/2; + mask = 0x3ffff; } - drc_ram_blk[(base_pc >> shift) & mask] = (blkid_main << 1) | 1; + drc_ram_blk[(base_pc & mask) >> shift] = (blkid_main << 1) | 1; for (pc = base_pc + 2; pc < end_pc; pc += 2) { - p = &drc_ram_blk[(pc >> shift) & mask]; - if (*p && *p != (blkid_main << 1)) - *p = BLOCKID_OVERLAP; // block intersection.. - else - *p = blkid_main << 1; + p = &drc_ram_blk[(pc & mask) >> shift]; + *p = blkid_main << 1; } // mark block entries (used by dr_get_bd()) for (i = 0; i < branch_target_count; i++) if (branch_target_blkid[i] != 0) - drc_ram_blk[(branch_target_pc[i] >> shift) & mask] = + drc_ram_blk[(branch_target_pc[i] & mask) >> shift] = (branch_target_blkid[i] << 1) | 1; // mark literals for (i = 0; i < literal_addr_count; i++) { tmp = literal_addr[i]; - p = &drc_ram_blk[(tmp >> shift) & mask]; - if (*p && *p != (blkid_main << 1)) - *p = BLOCKID_OVERLAP; - else - *p = blkid_main << 1; + p = &drc_ram_blk[(tmp & mask) >> shift]; + *p = blkid_main << 1; if (!(tmp & 3) && shift == 1) p[1] = p[0]; // assume long } + + // add to invalidation lookup lists + addr = base_pc & ~(ADDR_TO_BLOCK_PAGE - 1); + for (; addr < end_pc + MAX_LITERAL_OFFSET; addr += ADDR_TO_BLOCK_PAGE) { + i = (addr & mask) / ADDR_TO_BLOCK_PAGE; + add_to_block_list(&inval_lookup[tcache_id][i], this_block); + } } tcache_ptrs[tcache_id] = tcache_ptr; @@ -2631,9 +2693,9 @@ end_op: do_host_disasm(tcache_id); dbg(2, " block #%d,%d tcache %d/%d, insns %d -> %d %.3f", - tcache_id, block_counts[tcache_id], + tcache_id, blkid_main, tcache_ptr - tcache_bases[tcache_id], tcache_sizes[tcache_id], - insns_compiled, host_insn_count, (double)host_insn_count / insns_compiled); + insns_compiled, host_insn_count, (float)host_insn_count / insns_compiled); if ((sh2->pc & 0xc6000000) == 0x02000000) // ROM dbg(2, " hash collisions %d/%d", hash_collisions, block_counts[tcache_id]); /* @@ -2859,9 +2921,10 @@ static void sh2_generate_utils(void) #endif } -static void sh2_smc_rm_block_entry(block_desc *bd, int tcache_id) +static void sh2_smc_rm_block_entry(struct block_desc *bd, int tcache_id, u32 ram_mask) { void *tmp; + u32 i, addr; // XXX: kill links somehow? dbg(2, " killing entry %08x-%08x, blkid %d,%d", @@ -2871,6 +2934,13 @@ static void sh2_smc_rm_block_entry(block_desc *bd, int tcache_id) return; } + // remove from inval_lookup + addr = bd->addr & ~(ADDR_TO_BLOCK_PAGE - 1); + for (; addr < bd->end_addr; addr += ADDR_TO_BLOCK_PAGE) { + i = (addr & ram_mask) / ADDR_TO_BLOCK_PAGE; + rm_from_block_list(&inval_lookup[tcache_id][i], bd); + } + // since we never reuse space of dead blocks, // insert jump to dispatcher for blocks that are linked to this point //emith_jump_at(bd->tcache_ptr, sh2_drc_dispatcher); @@ -2890,72 +2960,33 @@ static void sh2_smc_rm_block_entry(block_desc *bd, int tcache_id) static void sh2_smc_rm_block(u32 a, u16 *drc_ram_blk, int tcache_id, u32 shift, u32 mask) { - block_desc *btab = block_tables[tcache_id]; - u16 *p = drc_ram_blk + ((a & mask) >> shift); - u16 *pmax = drc_ram_blk + (mask >> shift); - u32 id = ~0, end_addr; - int max_zeros = MAX_LITERAL_OFFSET >> shift; - int i, zeros; - - if (*p == 0 || (*p >> 1) >= BLOCKID_MAX) { - u32 from = ~0, to = 0; - dbg(1, "slow-remove blocks at @%08x", a); - for (i = 0; i < block_counts[tcache_id]; i++) { - if (btab[i].addr <= a && a < btab[i].end_addr) { - if (btab[i].addr < from) - from = btab[i].addr; - if (btab[i].end_addr > to) - to = btab[i].end_addr; - sh2_smc_rm_block_entry(&btab[i], tcache_id); - } - } - if (from < to) { - p = drc_ram_blk + ((from & mask) >> shift); - memset(p, 0, (to - from) >> (shift - 1)); - } - return; - } - - // use end_addr to distinguish the same block - end_addr = btab[*p >> 1].end_addr; - - // go up to the start - for (zeros = 0; p > drc_ram_blk && zeros < max_zeros; p--) { - // there can be holes because games sometimes keep variables - // directly in literal pool and we don't inline them - // to avoid recompile (Star Wars Arcade) - if (p[-1] == 0) { - zeros++; + struct block_list **blist = NULL, *entry; + u32 from = ~0, to = 0; + struct block_desc *block; + + blist = &inval_lookup[tcache_id][(a & mask) / ADDR_TO_BLOCK_PAGE]; + entry = *blist; + while (entry != NULL) { + block = entry->block; + if (block->addr <= a && a < block->end_addr) { + if (block->addr < from) + from = block->addr; + if (block->end_addr > to) + to = block->end_addr; + + sh2_smc_rm_block_entry(block, tcache_id, mask); + + // entry lost, restart search + entry = *blist; continue; } - zeros = 0; - if ((p[-1] >> 1) >= BLOCKID_MAX) - break; - if (btab[p[-1] >> 1].end_addr != end_addr) - break; + entry = entry->next; } - if (!(*p & 1)) - dbg(1, "smc rm: missing block start for %08x?", a); - - // now go down and kill everything - for (zeros = 0; p < pmax && zeros < max_zeros; p++) { - if (*p == 0) { - zeros++; - continue; - } - zeros = 0; - if ((*p >> 1) >= BLOCKID_MAX) - break; - if ((*p >> 1) == id) { - *p = 0; - continue; - } - id = *p >> 1; - if (btab[id].end_addr != end_addr) - break; - *p = 0; - sh2_smc_rm_block_entry(&btab[id], tcache_id); + // clear entry points + if (from < to) { + u16 *p = drc_ram_blk + ((from & mask) >> shift); + memset(p, 0, (to - from) >> (shift - 1)); } } @@ -3005,7 +3036,7 @@ void block_stats(void) total += block_tables[b][i].refcount; for (c = 0; c < 10; c++) { - block_desc *blk, *maxb = NULL; + struct block_desc *blk, *maxb = NULL; int max = 0; for (b = 0; b < ARRAY_SIZE(block_tables); b++) { for (i = 0; i < block_counts[b]; i++) { @@ -3062,6 +3093,11 @@ int sh2_drc_init(SH2 *sh2) block_links[i] = calloc(block_max_counts[i] * 2, sizeof(*block_links[0])); if (block_links[i] == NULL) goto fail; + + inval_lookup[i] = calloc(ram_sizes[i] / ADDR_TO_BLOCK_PAGE, + sizeof(inval_lookup[0])); + if (inval_lookup[i] == NULL) + goto fail; } memset(block_counts, 0, sizeof(block_counts)); memset(block_link_counts, 0, sizeof(block_link_counts)); @@ -3107,6 +3143,8 @@ void sh2_drc_finish(SH2 *sh2) { int i; + sh2_drc_flush_all(); + if (block_tables[0] != NULL) { block_stats(); @@ -3124,6 +3162,10 @@ void sh2_drc_finish(SH2 *sh2) if (block_links[i] == NULL) free(block_links[i]); block_links[i] = NULL; + + if (inval_lookup[i] == NULL) + free(inval_lookup[i]); + inval_lookup[i] = NULL; } drc_cmn_cleanup(); @@ -3175,19 +3217,19 @@ void scan_block(u32 base_pc, int is_slave, u8 *op_flags, u32 *end_pc) u32 pc, target, op; int cycles; - memset(op_flags, 0, BLOCK_CYCLE_LIMIT); + memset(op_flags, 0, BLOCK_INSN_LIMIT); dr_pc_base = dr_get_pc_base(base_pc, is_slave); - for (cycles = 0, pc = base_pc; cycles < BLOCK_CYCLE_LIMIT-1; cycles++, pc += 2) { + for (cycles = 0, pc = base_pc; cycles < BLOCK_INSN_LIMIT-1; cycles++, pc += 2) { op = FETCH_OP(pc); if ((op & 0xf000) == 0xa000 || (op & 0xf000) == 0xb000) { // BRA, BSR signed int offs = ((signed int)(op << 20) >> 19); pc += 2; OP_FLAGS(pc) |= OF_DELAY_OP; target = pc + offs + 2; - if (base_pc <= target && target < base_pc + BLOCK_CYCLE_LIMIT * 2) - OP_FLAGS(target) |= OF_TARGET; + if (base_pc <= target && target < base_pc + BLOCK_INSN_LIMIT * 2) + OP_FLAGS(target) |= OF_BTARGET; break; } if ((op & 0xf000) == 0) { @@ -3212,8 +3254,8 @@ void scan_block(u32 base_pc, int is_slave, u8 *op_flags, u32 *end_pc) if (op & 0x0400) OP_FLAGS(pc + 2) |= OF_DELAY_OP; target = pc + offs + 4; - if (base_pc <= target && target < base_pc + BLOCK_CYCLE_LIMIT * 2) - OP_FLAGS(target) |= OF_TARGET; + if (base_pc <= target && target < base_pc + BLOCK_INSN_LIMIT * 2) + OP_FLAGS(target) |= OF_BTARGET; } if ((op & 0xff00) == 0xc300) // TRAPA break;