* notes:
* - tcache, block descriptor, link buffer overflows result in sh2_translate()
* failure, followed by full tcache invalidation for that region
- * - jumps between blocks are tracked for SMC handling (in block_links[])
+ * - jumps between blocks are tracked for SMC handling (in block_links[]),
+ * except jumps between different tcaches
+ * - non-main block entries are called subblocks, as they have same tracking
+ * structures that main blocks have.
+ *
+ * implemented:
+ * - static register allocation
+ * - remaining register caching and tracking in temporaries
+ * - block-local branch linking
+ * - block linking (except between tcaches)
+ * - some constant propagation
+ *
+ * TODO:
+ * - proper SMC handling
+ * - better constant propagation
+ * - stack caching?
+ * - bug fixing
*/
#include <stddef.h>
#include <stdio.h>
#include "sh2.h"
#include "compiler.h"
#include "../drc/cmn.h"
+#include "../debug.h"
+
+// features
+#define PROPAGATE_CONSTANTS 1
+#define LINK_BRANCHES 1
+
+// limits (per block)
+#define BLOCK_CYCLE_LIMIT 100
+#define MAX_BLOCK_SIZE (BLOCK_CYCLE_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_LOCAL_BRANCHES 32
// debug stuff {
#ifndef DRC_DEBUG
#define do_host_disasm(x)
#endif
-#if (DRC_DEBUG & 4)
-static void REGPARM(3) *sh2_drc_announce_entry(void *block, SH2 *sh2, u32 sr)
+#if (DRC_DEBUG & 4) || defined(PDB)
+static void REGPARM(3) *sh2_drc_log_entry(void *block, SH2 *sh2, u32 sr)
{
- if (block != NULL)
+ if (block != NULL) {
dbg(4, "= %csh2 enter %08x %p, c=%d", sh2->is_slave ? 's' : 'm',
sh2->pc, block, (signed int)sr >> 12);
+ pdb_step(sh2, sh2->pc);
+ }
return block;
}
#endif
// } debug
-#define BLOCK_CYCLE_LIMIT 100
-#define MAX_BLOCK_SIZE (BLOCK_CYCLE_LIMIT * 6 * 6)
#define TCACHE_BUFFERS 3
// we have 3 translation cache buffers, split from one drc/cmn buffer.
typedef struct block_desc_ {
u32 addr; // SH2 PC address
- u32 end_addr; // TODO rm?
void *tcache_ptr; // translated block for above PC
struct block_desc_ *next; // next block with the same PC hash
#if (DRC_DEBUG & 1)
typedef struct block_link_ {
u32 target_pc;
- void *jump;
+ void *jump; // insn address
// struct block_link_ *next;
} block_link;
enum {
HR_FREE,
HR_CACHED, // 'val' has sh2_reg_e
- HR_CACHED_DIRTY,
- HR_CONST, // 'val' has constant
+// HR_CONST, // 'val' has a constant
HR_TEMP, // reg used for temp storage
};
+enum {
+ HRF_DIRTY = 1 << 0, // reg has "dirty" value to be written to ctx
+ HRF_LOCKED = 1 << 1, // HR_CACHED can't be evicted
+};
+
typedef struct {
- u8 reg;
- u8 type;
- u16 stamp; // kind of a timestamp
- u32 val;
+ u32 hreg:5; // "host" reg
+ u32 greg:5; // "guest" reg
+ u32 type:3;
+ u32 flags:3;
+ u32 stamp:16; // kind of a timestamp
} temp_reg_t;
// note: reg_temp[] must have at least the amount of
static void (*sh2_drc_dispatcher)(void);
static void (*sh2_drc_exit)(void);
static void (*sh2_drc_test_irq)(void);
+
+static u32 REGPARM(2) (*sh2_drc_read8)(u32 a, SH2 *sh2);
+static u32 REGPARM(2) (*sh2_drc_read16)(u32 a, SH2 *sh2);
+static u32 REGPARM(2) (*sh2_drc_read32)(u32 a, SH2 *sh2);
static void REGPARM(2) (*sh2_drc_write8)(u32 a, u32 d);
static void REGPARM(2) (*sh2_drc_write8_slot)(u32 a, u32 d);
static void REGPARM(2) (*sh2_drc_write16)(u32 a, u32 d);
static void REGPARM(2) (*sh2_drc_write16_slot)(u32 a, u32 d);
+static int REGPARM(3) (*sh2_drc_write32)(u32 a, u32 d, SH2 *sh2);
extern void REGPARM(2) sh2_do_op(SH2 *sh2, int opcode);
-static void flush_tcache(int tcid)
+// address space stuff
+static void *dr_get_pc_base(u32 pc, int is_slave)
+{
+ void *ret = NULL;
+ u32 mask = 0;
+
+ if ((pc & ~0x7ff) == 0) {
+ // BIOS
+ ret = is_slave ? Pico32xMem->sh2_rom_s : Pico32xMem->sh2_rom_m;
+ mask = 0x7ff;
+ }
+ else if ((pc & 0xfffff000) == 0xc0000000) {
+ // data array
+ ret = Pico32xMem->data_array[is_slave];
+ mask = 0xfff;
+ }
+ else if ((pc & 0xc6000000) == 0x06000000) {
+ // SDRAM
+ ret = Pico32xMem->sdram;
+ mask = 0x03ffff;
+ }
+ else if ((pc & 0xc6000000) == 0x02000000) {
+ // ROM
+ ret = Pico.rom;
+ mask = 0x3fffff;
+ }
+
+ if (ret == NULL)
+ return (void *)-1; // NULL is valid value
+
+ return (char *)ret - (pc & ~mask);
+}
+
+static int dr_ctx_get_mem_ptr(u32 a, u32 *mask)
+{
+ int poffs = -1;
+
+ if ((a & ~0x7ff) == 0) {
+ // BIOS
+ poffs = offsetof(SH2, p_bios);
+ *mask = 0x7ff;
+ }
+ else if ((a & 0xfffff000) == 0xc0000000) {
+ // data array
+ poffs = offsetof(SH2, p_da);
+ *mask = 0xfff;
+ }
+ else if ((a & 0xc6000000) == 0x06000000) {
+ // SDRAM
+ poffs = offsetof(SH2, p_sdram);
+ *mask = 0x03ffff;
+ }
+ else if ((a & 0xc6000000) == 0x02000000) {
+ // ROM
+ poffs = offsetof(SH2, p_rom);
+ *mask = 0x3fffff;
+ }
+
+ return poffs;
+}
+
+static block_desc *dr_get_bd(u32 pc, int is_slave, int *tcache_id)
+{
+ *tcache_id = 0;
+
+ // we have full block id tables for data_array and RAM
+ // BIOS goes to data_array table too
+ if ((pc & 0xe0000000) == 0xc0000000 || (pc & ~0xfff) == 0) {
+ int blkid = Pico32xMem->drcblk_da[is_slave][(pc & 0xfff) >> SH2_DRCBLK_DA_SHIFT];
+ *tcache_id = 1 + is_slave;
+ if (blkid & 1)
+ return &block_tables[*tcache_id][blkid >> 1];
+ }
+ // RAM
+ else if ((pc & 0xc6000000) == 0x06000000) {
+ int blkid = Pico32xMem->drcblk_ram[(pc & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT];
+ if (blkid & 1)
+ return &block_tables[0][blkid >> 1];
+ }
+ // ROM
+ else if ((pc & 0xc6000000) == 0x02000000) {
+ block_desc *bd = HASH_FUNC(hash_table, pc);
+
+ for (; bd != NULL; bd = bd->next)
+ if (bd->addr == pc)
+ return bd;
+ }
+
+ return NULL;
+}
+
+// ---------------------------------------------------------------
+
+// block management
+static void REGPARM(1) flush_tcache(int tcid)
{
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]);
block_counts[tcid] = 0;
+ block_link_counts[tcid] = 0;
tcache_ptrs[tcid] = tcache_bases[tcid];
if (tcid == 0) { // ROM, RAM
memset(hash_table, 0, sizeof(hash_table[0]) * MAX_HASH_ENTRIES);
#endif
}
+#if LINK_BRANCHES
// add block links (tracked branches)
static int dr_add_block_link(u32 target_pc, void *jump, int tcache_id)
{
return 0;
}
+#endif
-static void *dr_find_block(block_desc *tab, u32 addr)
-{
- for (tab = tab->next; tab != NULL; tab = tab->next)
- if (tab->addr == addr)
- break;
-
- if (tab != NULL)
- return tab->tcache_ptr;
-
- printf("block miss for %08x\n", addr);
- return NULL;
-}
-
-static block_desc *dr_add_block(u32 addr, int tcache_id, int *blk_id)
+static block_desc *dr_add_block(u32 addr, int is_slave, int *blk_id)
{
- int *bcount = &block_counts[tcache_id];
block_desc *bd;
+ int tcache_id;
+ int *bcount;
+
+ bd = dr_get_bd(addr, is_slave, &tcache_id);
+ if (bd != NULL) {
+ dbg(1, "block override for %08x", addr);
+ bd->tcache_ptr = tcache_ptr;
+ *blk_id = bd - block_tables[tcache_id];
+ return bd;
+ }
+ bcount = &block_counts[tcache_id];
if (*bcount >= block_max_counts[tcache_id]) {
printf("bd overflow for tcache %d\n", tcache_id);
return NULL;
}
+ if (*bcount == 0)
+ (*bcount)++; // not using descriptor 0
bd = &block_tables[tcache_id][*bcount];
bd->addr = addr;
return bd;
}
+static void REGPARM(3) *dr_lookup_block(u32 pc, int is_slave, int *tcache_id)
+{
+ block_desc *bd = NULL;
+ void *block = NULL;
+
+ bd = dr_get_bd(pc, is_slave, tcache_id);
+ if (bd != NULL)
+ block = bd->tcache_ptr;
+
+#if (DRC_DEBUG & 1)
+ if (bd != NULL)
+ bd->refcount++;
+#endif
+ return block;
+}
+
+static void *dr_prepare_ext_branch(u32 pc, SH2 *sh2, int tcache_id)
+{
+#if LINK_BRANCHES
+ int target_tcache_id;
+ void *target;
+ int ret;
+
+ target = dr_lookup_block(pc, sh2->is_slave, &target_tcache_id);
+ if (target_tcache_id == tcache_id) {
+ // allow linking blocks only from local cache
+ ret = dr_add_block_link(pc, tcache_ptr, tcache_id);
+ if (ret < 0)
+ return NULL;
+ }
+ if (target == NULL || target_tcache_id != tcache_id)
+ target = sh2_drc_dispatcher;
+
+ return target;
+#else
+ return sh2_drc_dispatcher;
+#endif
+}
+
+static void dr_link_blocks(void *target, u32 pc, int tcache_id)
+{
+#if LINK_BRANCHES
+ block_link *bl = block_links[tcache_id];
+ int cnt = block_link_counts[tcache_id];
+ int i;
+
+ for (i = 0; i < cnt; i++) {
+ if (bl[i].target_pc == pc) {
+ dbg(1, "- link from %p", bl[i].jump);
+ emith_jump_patch(bl[i].jump, target);
+ // XXX: sync ARM caches (old jump should be fine)?
+ }
+ }
+#endif
+}
+
#define ADD_TO_ARRAY(array, count, item, failcode) \
array[count++] = item; \
if (count >= ARRAY_SIZE(array)) { \
failcode; \
}
-int find_in_array(u32 *array, size_t size, u32 what)
+static int find_in_array(u32 *array, size_t size, u32 what)
{
size_t i;
for (i = 0; i < size; i++)
// ---------------------------------------------------------------
-// register chache
+// register cache / constant propagation stuff
+typedef enum {
+ RC_GR_READ,
+ RC_GR_WRITE,
+ RC_GR_RMW,
+} rc_gr_mode;
+
+static int rcache_get_reg_(sh2_reg_e r, rc_gr_mode mode, int do_locking);
+
+// guest regs with constants
+static u32 dr_gcregs[24];
+// a mask of constant/dirty regs
+static u32 dr_gcregs_mask;
+static u32 dr_gcregs_dirty;
+
+#if PROPAGATE_CONSTANTS
+static void gconst_new(sh2_reg_e r, u32 val)
+{
+ int i;
+
+ dr_gcregs_mask |= 1 << r;
+ dr_gcregs_dirty |= 1 << r;
+ dr_gcregs[r] = val;
+
+ // throw away old r that we might have cached
+ for (i = ARRAY_SIZE(reg_temp) - 1; i >= 0; i--) {
+ if ((reg_temp[i].type == HR_CACHED) &&
+ reg_temp[i].greg == r) {
+ reg_temp[i].type = HR_FREE;
+ reg_temp[i].flags = 0;
+ }
+ }
+}
+#endif
+
+static int gconst_get(sh2_reg_e r, u32 *val)
+{
+ if (dr_gcregs_mask & (1 << r)) {
+ *val = dr_gcregs[r];
+ return 1;
+ }
+ return 0;
+}
+
+static int gconst_check(sh2_reg_e r)
+{
+ if ((dr_gcregs_mask | dr_gcregs_dirty) & (1 << r))
+ return 1;
+ return 0;
+}
+
+// update hr if dirty, else do nothing
+static int gconst_try_read(int hr, sh2_reg_e r)
+{
+ if (dr_gcregs_dirty & (1 << r)) {
+ emith_move_r_imm(hr, dr_gcregs[r]);
+ dr_gcregs_dirty &= ~(1 << r);
+ return 1;
+ }
+ return 0;
+}
+
+static void gconst_check_evict(sh2_reg_e r)
+{
+ if (dr_gcregs_mask & (1 << r))
+ // no longer cached in reg, make dirty again
+ dr_gcregs_dirty |= 1 << r;
+}
+
+static void gconst_kill(sh2_reg_e r)
+{
+ dr_gcregs_mask &= ~(1 << r);
+ dr_gcregs_dirty &= ~(1 << r);
+}
+
+static void gconst_clean(void)
+{
+ int i;
+
+ for (i = 0; i < ARRAY_SIZE(dr_gcregs); i++)
+ if (dr_gcregs_dirty & (1 << i)) {
+ // using RC_GR_READ here: it will call gconst_try_read,
+ // cache the reg and mark it dirty.
+ rcache_get_reg_(i, RC_GR_READ, 0);
+ }
+}
+
+static void gconst_invalidate(void)
+{
+ dr_gcregs_mask = dr_gcregs_dirty = 0;
+}
+
static u16 rcache_counter;
static temp_reg_t *rcache_evict(void)
u16 min_stamp = (u16)-1;
for (i = 0; i < ARRAY_SIZE(reg_temp); i++) {
- if (reg_temp[i].type == HR_CACHED || reg_temp[i].type == HR_CACHED_DIRTY)
- if (reg_temp[i].stamp <= min_stamp) {
- min_stamp = reg_temp[i].stamp;
- oldest = i;
- }
+ if (reg_temp[i].type == HR_CACHED && !(reg_temp[i].flags & HRF_LOCKED) &&
+ reg_temp[i].stamp <= min_stamp) {
+ min_stamp = reg_temp[i].stamp;
+ oldest = i;
+ }
}
if (oldest == -1) {
}
i = oldest;
- if (reg_temp[i].type == HR_CACHED_DIRTY) {
- // writeback
- emith_ctx_write(reg_temp[i].reg, reg_temp[i].val * 4);
+ if (reg_temp[i].type == HR_CACHED) {
+ if (reg_temp[i].flags & HRF_DIRTY)
+ // writeback
+ emith_ctx_write(reg_temp[i].hreg, reg_temp[i].greg * 4);
+ gconst_check_evict(reg_temp[i].greg);
}
+ reg_temp[i].type = HR_FREE;
+ reg_temp[i].flags = 0;
return ®_temp[i];
}
-typedef enum {
- RC_GR_READ,
- RC_GR_WRITE,
- RC_GR_RMW,
-} rc_gr_mode;
+static int get_reg_static(sh2_reg_e r, rc_gr_mode mode)
+{
+ int i = reg_map_g2h[r];
+ if (i != -1) {
+ if (mode != RC_GR_WRITE)
+ gconst_try_read(i, r);
+ }
+ return i;
+}
// note: must not be called when doing conditional code
-static int rcache_get_reg(sh2_reg_e r, rc_gr_mode mode)
+static int rcache_get_reg_(sh2_reg_e r, rc_gr_mode mode, int do_locking)
{
temp_reg_t *tr;
- int i;
+ int i, ret;
- // maybe already statically mapped?
- i = reg_map_g2h[r];
- if (i != -1)
- return i;
+ // maybe statically mapped?
+ ret = get_reg_static(r, mode);
+ if (ret != -1)
+ goto end;
rcache_counter++;
// maybe already cached?
+ // if so, prefer against gconst (they must be in sync)
for (i = ARRAY_SIZE(reg_temp) - 1; i >= 0; i--) {
- if ((reg_temp[i].type == HR_CACHED || reg_temp[i].type == HR_CACHED_DIRTY) &&
- reg_temp[i].val == r)
- {
+ if (reg_temp[i].type == HR_CACHED && reg_temp[i].greg == r) {
reg_temp[i].stamp = rcache_counter;
if (mode != RC_GR_READ)
- reg_temp[i].type = HR_CACHED_DIRTY;
- return reg_temp[i].reg;
+ reg_temp[i].flags |= HRF_DIRTY;
+ ret = reg_temp[i].hreg;
+ goto end;
}
}
// use any free reg
for (i = ARRAY_SIZE(reg_temp) - 1; i >= 0; i--) {
- if (reg_temp[i].type == HR_FREE || reg_temp[i].type == HR_CONST) {
+ if (reg_temp[i].type == HR_FREE) {
tr = ®_temp[i];
goto do_alloc;
}
tr = rcache_evict();
do_alloc:
- if (mode != RC_GR_WRITE)
- emith_ctx_read(tr->reg, r * 4);
-
- tr->type = mode != RC_GR_READ ? HR_CACHED_DIRTY : HR_CACHED;
- tr->val = r;
+ tr->type = HR_CACHED;
+ if (do_locking)
+ tr->flags |= HRF_LOCKED;
+ if (mode != RC_GR_READ)
+ tr->flags |= HRF_DIRTY;
+ tr->greg = r;
tr->stamp = rcache_counter;
- return tr->reg;
+ ret = tr->hreg;
+
+ if (mode != RC_GR_WRITE) {
+ if (gconst_check(r)) {
+ if (gconst_try_read(ret, r))
+ tr->flags |= HRF_DIRTY;
+ }
+ else
+ emith_ctx_read(tr->hreg, r * 4);
+ }
+
+end:
+ if (mode != RC_GR_READ)
+ gconst_kill(r);
+
+ return ret;
+}
+
+static int rcache_get_reg(sh2_reg_e r, rc_gr_mode mode)
+{
+ return rcache_get_reg_(r, mode, 1);
}
static int rcache_get_tmp(void)
int i;
for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
- if (reg_temp[i].type == HR_FREE || reg_temp[i].type == HR_CONST) {
+ if (reg_temp[i].type == HR_FREE) {
tr = ®_temp[i];
goto do_alloc;
}
do_alloc:
tr->type = HR_TEMP;
- return tr->reg;
+ return tr->hreg;
}
static int rcache_get_arg_id(int arg)
host_arg2reg(r, arg);
for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
- if (reg_temp[i].reg == r)
+ if (reg_temp[i].hreg == r)
break;
- if (i == ARRAY_SIZE(reg_temp))
- // let's just say it's untracked arg reg
- return r;
+ if (i == ARRAY_SIZE(reg_temp)) // can't happen
+ exit(1);
- if (reg_temp[i].type == HR_CACHED_DIRTY) {
+ if (reg_temp[i].type == HR_CACHED) {
// writeback
- emith_ctx_write(reg_temp[i].reg, reg_temp[i].val * 4);
+ if (reg_temp[i].flags & HRF_DIRTY)
+ emith_ctx_write(reg_temp[i].hreg, reg_temp[i].greg * 4);
+ gconst_check_evict(reg_temp[i].greg);
}
else if (reg_temp[i].type == HR_TEMP) {
printf("arg %d reg %d already used, aborting\n", arg, r);
exit(1);
}
+ reg_temp[i].type = HR_FREE;
+ reg_temp[i].flags = 0;
+
return i;
}
// get a reg to be used as function arg
-// it's assumed that regs are cleaned before call
static int rcache_get_tmp_arg(int arg)
{
int id = rcache_get_arg_id(arg);
reg_temp[id].type = HR_TEMP;
- return reg_temp[id].reg;
+ return reg_temp[id].hreg;
}
-// same but caches reg. RC_GR_READ only.
+// same but caches a reg. RC_GR_READ only.
static int rcache_get_reg_arg(int arg, sh2_reg_e r)
{
int i, srcr, dstr, dstid;
+ int dirty = 0, src_dirty = 0;
dstid = rcache_get_arg_id(arg);
- dstr = reg_temp[dstid].reg;
+ dstr = reg_temp[dstid].hreg;
// maybe already statically mapped?
- srcr = reg_map_g2h[r];
+ srcr = get_reg_static(r, RC_GR_READ);
if (srcr != -1)
goto do_cache;
// maybe already cached?
for (i = ARRAY_SIZE(reg_temp) - 1; i >= 0; i--) {
- if ((reg_temp[i].type == HR_CACHED || reg_temp[i].type == HR_CACHED_DIRTY) &&
- reg_temp[i].val == r)
+ if ((reg_temp[i].type == HR_CACHED) &&
+ reg_temp[i].greg == r)
{
- srcr = reg_temp[i].reg;
+ srcr = reg_temp[i].hreg;
+ if (reg_temp[i].flags & HRF_DIRTY)
+ src_dirty = 1;
goto do_cache;
}
}
// must read
srcr = dstr;
- emith_ctx_read(srcr, r * 4);
+ if (gconst_check(r)) {
+ if (gconst_try_read(srcr, r))
+ dirty = 1;
+ }
+ else
+ emith_ctx_read(srcr, r * 4);
do_cache:
- if (srcr != dstr)
+ if (dstr != srcr)
emith_move_r_r(dstr, srcr);
+#if 1
+ else
+ dirty |= src_dirty;
+
+ if (dirty)
+ // must clean, callers might want to modify the arg before call
+ emith_ctx_write(dstr, r * 4);
+#else
+ if (dirty)
+ reg_temp[dstid].flags |= HRF_DIRTY;
+#endif
reg_temp[dstid].stamp = ++rcache_counter;
reg_temp[dstid].type = HR_CACHED;
- reg_temp[dstid].val = r;
+ reg_temp[dstid].greg = r;
+ reg_temp[dstid].flags |= HRF_LOCKED;
return dstr;
}
{
int i;
for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
- if (reg_temp[i].reg == hr)
+ if (reg_temp[i].hreg == hr)
break;
if (i == ARRAY_SIZE(reg_temp) || reg_temp[i].type != HR_TEMP) {
}
reg_temp[i].type = HR_FREE;
+ reg_temp[i].flags = 0;
+}
+
+static void rcache_unlock(int hr)
+{
+ int i;
+ for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
+ if (reg_temp[i].type == HR_CACHED && reg_temp[i].hreg == hr)
+ reg_temp[i].flags &= ~HRF_LOCKED;
+}
+
+static void rcache_unlock_all(void)
+{
+ int i;
+ for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
+ reg_temp[i].flags &= ~HRF_LOCKED;
}
static void rcache_clean(void)
{
int i;
+ gconst_clean();
+
for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
- if (reg_temp[i].type == HR_CACHED_DIRTY) {
+ if (reg_temp[i].type == HR_CACHED && (reg_temp[i].flags & HRF_DIRTY)) {
// writeback
- emith_ctx_write(reg_temp[i].reg, reg_temp[i].val * 4);
- reg_temp[i].type = HR_CACHED;
+ emith_ctx_write(reg_temp[i].hreg, reg_temp[i].greg * 4);
+ reg_temp[i].flags &= ~HRF_DIRTY;
}
}
static void rcache_invalidate(void)
{
int i;
- for (i = 0; i < ARRAY_SIZE(reg_temp); i++)
+ for (i = 0; i < ARRAY_SIZE(reg_temp); i++) {
reg_temp[i].type = HR_FREE;
+ reg_temp[i].flags = 0;
+ }
rcache_counter = 0;
+
+ gconst_invalidate();
}
static void rcache_flush(void)
// ---------------------------------------------------------------
+static int emit_get_rbase_and_offs(u32 a, u32 *offs)
+{
+ u32 mask = 0;
+ int poffs;
+ int hr;
+
+ poffs = dr_ctx_get_mem_ptr(a, &mask);
+ if (poffs == -1)
+ return -1;
+
+ // XXX: could use some related reg
+ hr = rcache_get_tmp();
+ emith_ctx_read(hr, poffs);
+ emith_add_r_imm(hr, a & mask & ~0xff);
+ *offs = a & 0xff; // XXX: ARM oriented..
+ return hr;
+}
+
static void emit_move_r_imm32(sh2_reg_e dst, u32 imm)
{
- // TODO: propagate this constant
+#if PROPAGATE_CONSTANTS
+ gconst_new(dst, imm);
+#else
int hr = rcache_get_reg(dst, RC_GR_WRITE);
emith_move_r_imm(hr, imm);
+#endif
}
static void emit_move_r_r(sh2_reg_e dst, sh2_reg_e src)
// arguments must be ready
// reg cache must be clean before call
-static int emit_memhandler_read(int size)
+static int emit_memhandler_read_(int size, int ram_check)
{
int arg0, arg1;
host_arg2reg(arg0, 0);
+ rcache_clean();
+
// must writeback cycles for poll detection stuff
+ // FIXME: rm
if (reg_map_g2h[SHR_SR] != -1)
emith_ctx_write(reg_map_g2h[SHR_SR], SHR_SR * 4);
+
arg1 = rcache_get_tmp_arg(1);
emith_move_r_r(arg1, CONTEXT_REG);
-#if 1
- if (Pico.rom == (void *)0x02000000 && Pico32xMem->sdram == (void *)0x06000000) {
+#ifndef PDB_NET
+ if (ram_check && Pico.rom == (void *)0x02000000 && Pico32xMem->sdram == (void *)0x06000000) {
int tmp = rcache_get_tmp();
emith_and_r_r_imm(tmp, arg0, 0xfb000000);
emith_cmp_r_imm(tmp, 0x02000000);
emith_eor_r_imm_c(DCOND_EQ, arg0, 1);
emith_read8_r_r_offs_c(DCOND_EQ, arg0, arg0, 0);
EMITH_SJMP3_MID(DCOND_NE);
- emith_call_cond(DCOND_NE, p32x_sh2_read8);
+ emith_call_cond(DCOND_NE, sh2_drc_read8);
EMITH_SJMP3_END();
break;
case 1: // 16
EMITH_SJMP3_START(DCOND_NE);
emith_read16_r_r_offs_c(DCOND_EQ, arg0, arg0, 0);
EMITH_SJMP3_MID(DCOND_NE);
- emith_call_cond(DCOND_NE, p32x_sh2_read16);
+ emith_call_cond(DCOND_NE, sh2_drc_read16);
EMITH_SJMP3_END();
break;
case 2: // 32
emith_read_r_r_offs_c(DCOND_EQ, arg0, arg0, 0);
emith_ror_c(DCOND_EQ, arg0, arg0, 16);
EMITH_SJMP3_MID(DCOND_NE);
- emith_call_cond(DCOND_NE, p32x_sh2_read32);
+ emith_call_cond(DCOND_NE, sh2_drc_read32);
EMITH_SJMP3_END();
break;
}
{
switch (size) {
case 0: // 8
- emith_call(p32x_sh2_read8);
+ emith_call(sh2_drc_read8);
break;
case 1: // 16
- emith_call(p32x_sh2_read16);
+ emith_call(sh2_drc_read16);
break;
case 2: // 32
- emith_call(p32x_sh2_read32);
+ emith_call(sh2_drc_read32);
break;
}
}
return rcache_get_tmp_arg(0);
}
+static int emit_memhandler_read(int size)
+{
+ return emit_memhandler_read_(size, 1);
+}
+
+static int emit_memhandler_read_rr(sh2_reg_e rd, sh2_reg_e rs, u32 offs, int size)
+{
+ int hr, hr2, ram_check = 1;
+ u32 val, offs2;
+
+ if (gconst_get(rs, &val)) {
+ hr = emit_get_rbase_and_offs(val + offs, &offs2);
+ if (hr != -1) {
+ hr2 = rcache_get_reg(rd, RC_GR_WRITE);
+ switch (size) {
+ case 0: // 8
+ emith_read8_r_r_offs(hr2, hr, offs2 ^ 1);
+ emith_sext(hr2, hr2, 8);
+ break;
+ case 1: // 16
+ emith_read16_r_r_offs(hr2, hr, offs2);
+ emith_sext(hr2, hr2, 16);
+ break;
+ case 2: // 32
+ emith_read_r_r_offs(hr2, hr, offs2);
+ emith_ror(hr2, hr2, 16);
+ break;
+ }
+ rcache_free_tmp(hr);
+ return hr2;
+ }
+
+ ram_check = 0;
+ }
+
+ hr = rcache_get_reg_arg(0, rs);
+ if (offs != 0)
+ emith_add_r_imm(hr, offs);
+ hr = emit_memhandler_read_(size, ram_check);
+ hr2 = rcache_get_reg(rd, RC_GR_WRITE);
+ if (size != 2) {
+ emith_sext(hr2, hr, (size == 1) ? 16 : 8);
+ } else
+ emith_move_r_r(hr2, hr);
+ rcache_free_tmp(hr);
+
+ return hr2;
+}
+
static void emit_memhandler_write(int size, u32 pc, int delay)
{
int ctxr;
break;
case 2: // 32
emith_move_r_r(ctxr, CONTEXT_REG);
- emith_call(p32x_sh2_write32);
+ emith_call(sh2_drc_write32);
break;
}
rcache_invalidate();
static int emit_indirect_indexed_read(int rx, int ry, int size)
{
int a0, t;
- rcache_clean();
a0 = rcache_get_reg_arg(0, rx);
t = rcache_get_reg(ry, RC_GR_READ);
emith_add_r_r(a0, t);
{
int tmp;
- rcache_clean();
rcache_get_reg_arg(0, rn);
tmp = emit_memhandler_read(size);
emith_ctx_write(tmp, offsetof(SH2, drc_tmp));
rcache_free_tmp(tmp);
tmp = rcache_get_reg(rn, RC_GR_RMW);
emith_add_r_imm(tmp, 1 << size);
+ rcache_unlock(tmp);
- rcache_clean();
rcache_get_reg_arg(0, rm);
*rmr = emit_memhandler_read(size);
*rnr = rcache_get_tmp();
emith_ctx_read(*rnr, offsetof(SH2, drc_tmp));
tmp = rcache_get_reg(rm, RC_GR_RMW);
emith_add_r_imm(tmp, 1 << size);
+ rcache_unlock(tmp);
}
static void emit_do_static_regs(int is_write, int tmpr)
host_arg2reg(arg1, 1);
host_arg2reg(arg2, 2);
-#if (DRC_DEBUG & 4)
+#if (DRC_DEBUG & 4) || defined(PDB)
+ emit_do_static_regs(1, arg2);
emith_move_r_r(arg1, CONTEXT_REG);
emith_move_r_r(arg2, rcache_get_reg(SHR_SR, RC_GR_READ));
- emith_call(sh2_drc_announce_entry);
+ emith_call(sh2_drc_log_entry);
rcache_invalidate();
#endif
emith_tst_r_r(arg0, arg0);
EMITH_SJMP_END(DCOND_EQ);
}
-static void REGPARM(3) *lookup_block(u32 pc, int is_slave, int *tcache_id)
-{
- block_desc *bd = NULL;
- void *block = NULL;
- *tcache_id = 0;
-
- // we have full block id tables for data_array and RAM
- // BIOS goes to data_array table too
- if ((pc & 0xe0000000) == 0xc0000000 || (pc & ~0xfff) == 0) {
- int blkid = Pico32xMem->drcblk_da[is_slave][(pc & 0xfff) >> SH2_DRCBLK_DA_SHIFT];
- *tcache_id = 1 + is_slave;
- if (blkid & 1) {
- bd = &block_tables[*tcache_id][blkid >> 1];
- block = bd->tcache_ptr;
- }
- }
- // RAM
- else if ((pc & 0xc6000000) == 0x06000000) {
- int blkid = Pico32xMem->drcblk_ram[(pc & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT];
- if (blkid & 1) {
- bd = &block_tables[0][blkid >> 1];
- block = bd->tcache_ptr;
- }
- }
- // ROM
- else if ((pc & 0xc6000000) == 0x02000000) {
- bd = HASH_FUNC(hash_table, pc);
-
- if (bd != NULL) {
- if (bd->addr == pc)
- block = bd->tcache_ptr;
- else
- block = dr_find_block(bd, pc);
- }
- }
-
-#if (DRC_DEBUG & 1)
- if (bd != NULL)
- bd->refcount++;
-#endif
- return block;
-}
-
-void dr_link_blocks(void *target, u32 pc, int tcache_id)
-{
- block_link *bl = block_links[tcache_id];
- int cnt = block_link_counts[tcache_id];
- int i;
-
- for (i = 0; i < cnt; i++) {
- if (bl[i].target_pc == pc) {
- dbg(1, "- link from %p", bl[i].jump);
- emith_jump_patch(bl[i].jump, target);
- // XXX: sync ARM caches (old jump should be fine)?
- }
- }
-}
-
-void *dr_prepare_ext_branch(u32 pc, SH2 *sh2, int tcache_id)
-{
- int target_tcache_id;
- void *target;
- int ret;
-
- target = lookup_block(pc, sh2->is_slave, &target_tcache_id);
- if (target_tcache_id == tcache_id) {
- // allow linking blocks only from local cache
- ret = dr_add_block_link(pc, tcache_ptr, tcache_id);
- if (ret < 0)
- return NULL;
- }
- if (target == NULL || target_tcache_id != tcache_id)
- target = sh2_drc_dispatcher;
-
- return target;
-}
-
#define DELAYED_OP \
drcf.delayed_op = 2
goto default_; \
}
+#define FETCH_OP(pc) \
+ dr_pc_base[(pc) / 2]
+
+#define FETCH32(a) \
+ ((dr_pc_base[(a) / 2] << 16) | dr_pc_base[(a) / 2 + 1])
+
#define GET_Fx() \
((op >> 4) & 0x0f)
if (GET_Fx() >= n) \
goto default_
-#define MAX_LOCAL_BRANCHES 32
-
// op_flags: data from 1st pass
#define OP_FLAGS(pc) op_flags[((pc) - base_pc) / 2]
#define OF_DELAY_OP (1 << 0)
static void REGPARM(2) *sh2_translate(SH2 *sh2, int tcache_id)
{
// XXX: maybe use structs instead?
- void *branch_target_ptr[MAX_LOCAL_BRANCHES];
u32 branch_target_pc[MAX_LOCAL_BRANCHES];
+ void *branch_target_ptr[MAX_LOCAL_BRANCHES];
+ int branch_target_blkid[MAX_LOCAL_BRANCHES];
int branch_target_count = 0;
void *branch_patch_ptr[MAX_LOCAL_BRANCHES];
u32 branch_patch_pc[MAX_LOCAL_BRANCHES];
int branch_patch_count = 0;
+ u32 literal_addr[MAX_LITERALS];
+ int literal_addr_count = 0;
int pending_branch_cond = -1;
int pending_branch_pc = 0;
u8 op_flags[BLOCK_CYCLE_LIMIT + 1];
u32 pc, base_pc, end_pc, out_pc;
void *block_entry;
block_desc *this_block;
+ u16 *dr_pc_base;
int blkid_main = 0;
+ int skip_op = 0;
u32 tmp, tmp2;
int cycles;
int op;
base_pc = sh2->pc;
- // validate PC
- tmp = base_pc >> 29;
- if ((tmp != 0 && tmp != 1 && tmp != 6) || base_pc == 0) {
+ // get base/validate PC
+ dr_pc_base = dr_get_pc_base(base_pc, sh2->is_slave);
+ if (dr_pc_base == (void *)-1) {
printf("invalid PC, aborting: %08x\n", base_pc);
// FIXME: be less destructive
exit(1);
}
tcache_ptr = tcache_ptrs[tcache_id];
- this_block = dr_add_block(base_pc, tcache_id, &blkid_main);
+ this_block = dr_add_block(base_pc, sh2->is_slave, &blkid_main);
if (this_block == NULL)
return NULL;
// 1st pass: scan forward for local branches
memset(op_flags, 0, sizeof(op_flags));
for (cycles = 0, pc = base_pc; cycles < BLOCK_CYCLE_LIMIT; cycles++, pc += 2) {
- op = p32x_sh2_read16(pc, sh2);
+ op = FETCH_OP(pc);
if ((op & 0xf000) == 0xa000 || (op & 0xf000) == 0xb000) { // BRA, BSR
signed int offs = ((signed int)(op << 20) >> 19);
pc += 2;
}
branch_target_count = tmp;
memset(branch_target_ptr, 0, sizeof(branch_target_ptr[0]) * branch_target_count);
+ memset(branch_target_blkid, 0, sizeof(branch_target_blkid[0]) * branch_target_count);
// -------------------------------------------------
// 2nd pass: actual compilation
if (drcf.delayed_op > 0)
drcf.delayed_op--;
+ op = FETCH_OP(pc);
+
i = find_in_array(branch_target_pc, branch_target_count, pc);
- if (i >= 0)
+ if (i >= 0 || pc == base_pc)
{
- if (pc != sh2->pc)
+ if (pc != base_pc)
{
/* make "subblock" - just a mid-block entry */
block_desc *subblock;
- u16 *drcblk;
- int blkid;
sr = rcache_get_reg(SHR_SR, RC_GR_RMW);
FLUSH_CYCLES(sr);
- rcache_flush();
+ // decide if to flush rcache
+ if ((op & 0xf0ff) == 0x4010 && FETCH_OP(pc + 2) == 0x8bfd) // DT; BF #-2
+ rcache_clean();
+ else
+ rcache_flush();
do_host_disasm(tcache_id);
- subblock = dr_add_block(pc, tcache_id, &blkid);
+ dbg(1, "-- %csh2 subblock #%d,%d %08x -> %p", sh2->is_slave ? 's' : 'm',
+ tcache_id, branch_target_blkid[i], pc, tcache_ptr);
+
+ subblock = dr_add_block(pc, sh2->is_slave, &branch_target_blkid[i]);
if (subblock == NULL)
return NULL;
- subblock->end_addr = pc;
-
- if (tcache_id != 0) { // data array, BIOS
- drcblk = Pico32xMem->drcblk_da[sh2->is_slave];
- drcblk += (pc & 0x00fff) >> SH2_DRCBLK_DA_SHIFT;
- *drcblk = (blkid << 1) | 1;
- } else if ((this_block->addr & 0xc7fc0000) == 0x06000000) { // DRAM
- drcblk = Pico32xMem->drcblk_ram;
- drcblk += (pc & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT;
- *drcblk = (blkid << 1) | 1;
- }
-
- dbg(1, "-- %csh2 subblock #%d,%d %08x -> %p", sh2->is_slave ? 's' : 'm',
- tcache_id, blkid, pc, tcache_ptr);
// since we made a block entry, link any other blocks that jump to current pc
dr_link_blocks(tcache_ptr, pc, tcache_id);
}
- branch_target_ptr[i] = tcache_ptr;
+ if (i >= 0)
+ branch_target_ptr[i] = tcache_ptr;
// must update PC
emit_move_r_imm32(SHR_PC, pc);
sr = rcache_get_reg(SHR_SR, RC_GR_READ);
emith_cmp_r_imm(sr, 0);
emith_jump_cond(DCOND_LE, sh2_drc_exit);
+ do_host_disasm(tcache_id);
+ rcache_unlock_all();
}
- op = p32x_sh2_read16(pc, sh2);
-
#if (DRC_DEBUG & 3)
insns_compiled++;
#if (DRC_DEBUG & 2)
pc += 2;
cycles++;
+ if (skip_op > 0) {
+ skip_op--;
+ continue;
+ }
+
switch ((op >> 12) & 0x0f)
{
/////////////////////////////////////////////
emith_or_r_imm(sr, T);
break;
case 2: // CLRMAC 0000000000101000
- tmp = rcache_get_reg(SHR_MACL, RC_GR_WRITE);
- emith_move_r_imm(tmp, 0);
- tmp = rcache_get_reg(SHR_MACH, RC_GR_WRITE);
- emith_move_r_imm(tmp, 0);
+ emit_move_r_imm32(SHR_MACL, 0);
+ emit_move_r_imm32(SHR_MACH, 0);
break;
default:
goto default_;
goto end_op;
case 2: // RTE 0000000000101011
DELAYED_OP;
- rcache_clean();
// pop PC
- rcache_get_reg_arg(0, SHR_SP);
- tmp = emit_memhandler_read(2);
- tmp2 = rcache_get_reg(SHR_PC, RC_GR_WRITE);
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
- rcache_clean();
+ emit_memhandler_read_rr(SHR_PC, SHR_SP, 0, 2);
// pop SR
tmp = rcache_get_reg_arg(0, SHR_SP);
emith_add_r_imm(tmp, 4);
goto end_op;
case 0x0f: // MAC.L @Rm+,@Rn+ 0000nnnnmmmm1111
emit_indirect_read_double(&tmp, &tmp2, GET_Rn(), GET_Rm(), 2);
- sr = rcache_get_reg(SHR_SR, RC_GR_READ);
tmp4 = rcache_get_reg(SHR_MACH, RC_GR_RMW);
/* MS 16 MAC bits unused if saturated */
+ sr = rcache_get_reg(SHR_SR, RC_GR_READ);
emith_tst_r_imm(sr, S);
EMITH_SJMP_START(DCOND_EQ);
emith_clear_msb_c(DCOND_NE, tmp4, tmp4, 16);
EMITH_SJMP_END(DCOND_EQ);
+ rcache_unlock(sr);
tmp3 = rcache_get_reg(SHR_MACL, RC_GR_RMW); // might evict SR
emith_mula_s64(tmp3, tmp4, tmp, tmp2);
rcache_free_tmp(tmp2);
rcache_clean();
tmp = rcache_get_reg_arg(0, GET_Rn());
tmp2 = rcache_get_reg_arg(1, GET_Rm());
- emith_add_r_imm(tmp, (op & 0x0f) * 4);
+ if (op & 0x0f)
+ emith_add_r_imm(tmp, (op & 0x0f) * 4);
emit_memhandler_write(2, pc, drcf.delayed_op);
goto end_op;
emith_tpush_carry(sr, 0);
goto end_op;
case 1: // DT Rn 0100nnnn00010000
- if (p32x_sh2_read16(pc, sh2) == 0x8bfd) { // BF #-2
- emith_sh2_dtbf_loop();
- goto end_op;
- }
- tmp = rcache_get_reg(GET_Rn(), RC_GR_RMW);
sr = rcache_get_reg(SHR_SR, RC_GR_RMW);
if (drcf.delayed_op)
DELAY_SAVE_T(sr);
+ if (FETCH_OP(pc) == 0x8bfd) { // BF #-2
+ if (gconst_get(GET_Rn(), &tmp)) {
+ // XXX: limit burned cycles
+ emit_move_r_imm32(GET_Rn(), 0);
+ emith_or_r_imm(sr, T);
+ cycles += tmp * 4 + 1; // +1 syncs with noconst version, not sure why
+ skip_op = 1;
+ }
+ else
+ emith_sh2_dtbf_loop();
+ goto end_op;
+ }
+ tmp = rcache_get_reg(GET_Rn(), RC_GR_RMW);
emith_bic_r_imm(sr, T);
emith_subf_r_imm(tmp, 1);
emit_or_t_if_eq(sr);
default:
goto default_;
}
- rcache_clean();
rcache_get_reg_arg(0, GET_Rn());
tmp2 = emit_memhandler_read(2);
if (tmp == SHR_SR) {
break;
case 1: // TAS.B @Rn 0100nnnn00011011
// XXX: is TAS working on 32X?
- rcache_clean();
rcache_get_reg_arg(0, GET_Rn());
tmp = emit_memhandler_read(0);
sr = rcache_get_reg(SHR_SR, RC_GR_RMW);
}
goto end_op;
case 0x0f:
- // MAC @Rm+,@Rn+ 0100nnnnmmmm1111
+ // MAC.W @Rm+,@Rn+ 0100nnnnmmmm1111
emit_indirect_read_double(&tmp, &tmp2, GET_Rn(), GET_Rm(), 1);
emith_sext(tmp, tmp, 16);
emith_sext(tmp2, tmp2, 16);
/////////////////////////////////////////////
case 0x05:
// MOV.L @(disp,Rm),Rn 0101nnnnmmmmdddd
- rcache_clean();
- tmp = rcache_get_reg_arg(0, GET_Rm());
- emith_add_r_imm(tmp, (op & 0x0f) * 4);
- tmp = emit_memhandler_read(2);
- tmp2 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
+ emit_memhandler_read_rr(GET_Rn(), GET_Rm(), (op & 0x0f) * 4, 2);
goto end_op;
/////////////////////////////////////////////
case 0x04: // MOV.B @Rm+,Rn 0110nnnnmmmm0100
case 0x05: // MOV.W @Rm+,Rn 0110nnnnmmmm0101
case 0x06: // MOV.L @Rm+,Rn 0110nnnnmmmm0110
- rcache_clean();
- rcache_get_reg_arg(0, GET_Rm());
- tmp = emit_memhandler_read(op & 3);
- tmp2 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
- if ((op & 3) != 2) {
- emith_sext(tmp2, tmp, (op & 1) ? 16 : 8);
- } else
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
+ emit_memhandler_read_rr(GET_Rn(), GET_Rm(), 0, op & 3);
if ((op & 7) >= 4 && GET_Rn() != GET_Rm()) {
tmp = rcache_get_reg(GET_Rm(), RC_GR_RMW);
emith_add_r_imm(tmp, (1 << (op & 3)));
tmp = rcache_get_reg_arg(0, GET_Rm());
tmp2 = rcache_get_reg_arg(1, SHR_R0);
tmp3 = (op & 0x100) >> 8;
- emith_add_r_imm(tmp, (op & 0x0f) << tmp3);
+ if (op & 0x0f)
+ emith_add_r_imm(tmp, (op & 0x0f) << tmp3);
emit_memhandler_write(tmp3, pc, drcf.delayed_op);
goto end_op;
case 0x0400: // MOV.B @(disp,Rm),R0 10000100mmmmdddd
case 0x0500: // MOV.W @(disp,Rm),R0 10000101mmmmdddd
- rcache_clean();
- tmp = rcache_get_reg_arg(0, GET_Rm());
- tmp3 = (op & 0x100) >> 8;
- emith_add_r_imm(tmp, (op & 0x0f) << tmp3);
- tmp = emit_memhandler_read(tmp3);
- tmp2 = rcache_get_reg(0, RC_GR_WRITE);
- emith_sext(tmp2, tmp, 8 << tmp3);
- rcache_free_tmp(tmp);
+ tmp = (op & 0x100) >> 8;
+ emit_memhandler_read_rr(SHR_R0, GET_Rm(), (op & 0x0f) << tmp, tmp);
goto end_op;
case 0x0800: // CMP/EQ #imm,R0 10001000iiiiiiii
// XXX: could use cmn
/////////////////////////////////////////////
case 0x09:
// MOV.W @(disp,PC),Rn 1001nnnndddddddd
- rcache_clean();
- tmp = rcache_get_tmp_arg(0);
- emith_move_r_imm(tmp, pc + (op & 0xff) * 2 + 2);
- tmp = emit_memhandler_read(1);
- tmp2 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
- emith_sext(tmp2, tmp, 16);
- rcache_free_tmp(tmp);
+ tmp = pc + (op & 0xff) * 2 + 2;
+#if PROPAGATE_CONSTANTS
+ if (tmp < end_pc + MAX_LITERAL_OFFSET && literal_addr_count < MAX_LITERALS) {
+ ADD_TO_ARRAY(literal_addr, literal_addr_count, tmp,);
+ gconst_new(GET_Rn(), (u32)(int)(signed short)FETCH_OP(tmp));
+ }
+ else
+#endif
+ {
+ tmp2 = rcache_get_tmp_arg(0);
+ emith_move_r_imm(tmp2, tmp);
+ tmp2 = emit_memhandler_read(1);
+ tmp3 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
+ emith_sext(tmp3, tmp2, 16);
+ rcache_free_tmp(tmp2);
+ }
goto end_op;
/////////////////////////////////////////////
case 0x0400: // MOV.B @(disp,GBR),R0 11000100dddddddd
case 0x0500: // MOV.W @(disp,GBR),R0 11000101dddddddd
case 0x0600: // MOV.L @(disp,GBR),R0 11000110dddddddd
- rcache_clean();
- tmp = rcache_get_reg_arg(0, SHR_GBR);
- tmp3 = (op & 0x300) >> 8;
- emith_add_r_imm(tmp, (op & 0xff) << tmp3);
- tmp = emit_memhandler_read(tmp3);
- tmp2 = rcache_get_reg(0, RC_GR_WRITE);
- if (tmp3 != 2) {
- emith_sext(tmp2, tmp, 8 << tmp3);
- } else
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
+ tmp = (op & 0x300) >> 8;
+ emit_memhandler_read_rr(SHR_R0, SHR_GBR, (op & 0xff) << tmp, tmp);
goto end_op;
case 0x0300: // TRAPA #imm 11000011iiiiiiii
tmp = rcache_get_reg(SHR_SP, RC_GR_RMW);
emith_sub_r_imm(tmp, 4*2);
- rcache_clean();
// push SR
tmp = rcache_get_reg_arg(0, SHR_SP);
emith_add_r_imm(tmp, 4);
emith_move_r_imm(tmp, pc);
emit_memhandler_write(2, pc, drcf.delayed_op);
// obtain new PC
- tmp = rcache_get_reg_arg(0, SHR_VBR);
- emith_add_r_imm(tmp, (op & 0xff) * 4);
- tmp = emit_memhandler_read(2);
- tmp2 = rcache_get_reg(SHR_PC, RC_GR_WRITE);
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
+ emit_memhandler_read_rr(SHR_PC, SHR_VBR, (op & 0xff) * 4, 2);
out_pc = (u32)-1;
cycles += 7;
goto end_op;
/////////////////////////////////////////////
case 0x0d:
// MOV.L @(disp,PC),Rn 1101nnnndddddddd
- rcache_clean();
- tmp = rcache_get_tmp_arg(0);
- emith_move_r_imm(tmp, (pc + (op & 0xff) * 4 + 2) & ~3);
- tmp = emit_memhandler_read(2);
- tmp2 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
- emith_move_r_r(tmp2, tmp);
- rcache_free_tmp(tmp);
+ tmp = (pc + (op & 0xff) * 4 + 2) & ~3;
+#if PROPAGATE_CONSTANTS
+ if (tmp < end_pc + MAX_LITERAL_OFFSET && literal_addr_count < MAX_LITERALS) {
+ ADD_TO_ARRAY(literal_addr, literal_addr_count, tmp,);
+ gconst_new(GET_Rn(), FETCH32(tmp));
+ }
+ else
+#endif
+ {
+ tmp2 = rcache_get_tmp_arg(0);
+ emith_move_r_imm(tmp2, tmp);
+ tmp2 = emit_memhandler_read(2);
+ tmp3 = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
+ emith_move_r_r(tmp3, tmp2);
+ rcache_free_tmp(tmp2);
+ }
goto end_op;
/////////////////////////////////////////////
case 0x0e:
// MOV #imm,Rn 1110nnnniiiiiiii
- tmp = rcache_get_reg(GET_Rn(), RC_GR_WRITE);
- emith_move_r_imm_s8(tmp, op & 0xff);
+ emit_move_r_imm32(GET_Rn(), (u32)(signed int)(signed char)op);
goto end_op;
default:
}
end_op:
+ rcache_unlock_all();
+
// conditional branch handling (with/without delay)
if (pending_branch_cond != -1 && drcf.delayed_op != 2)
{
else
emith_tst_r_imm(sr, T);
+#if LINK_BRANCHES
if (find_in_array(branch_target_pc, branch_target_count, target_pc) >= 0) {
// local branch
// XXX: jumps back can be linked already
break;
}
}
- else {
+ else
+#endif
+ {
// can't resolve branch locally, make a block exit
emit_move_r_imm32(SHR_PC, target_pc);
rcache_clean();
emith_jump_patch(branch_patch_ptr[i], target);
}
- this_block->end_addr = pc;
+ end_pc = pc;
// mark memory blocks as containing compiled code
- if (tcache_id != 0) {
- // data array, BIOS
- u16 *drcblk = Pico32xMem->drcblk_da[sh2->is_slave];
- tmp = (this_block->addr & 0xfff) >> SH2_DRCBLK_DA_SHIFT;
- tmp2 = (this_block->end_addr & 0xfff) >> SH2_DRCBLK_DA_SHIFT;
- drcblk[tmp] = (blkid_main << 1) | 1;
- for (++tmp; tmp < tmp2; tmp++) {
- if (drcblk[tmp])
- continue; // dont overwrite overlay block(s)
- drcblk[tmp] = blkid_main << 1;
+ // override any overlay blocks as they become unreachable anyway
+ if (tcache_id != 0 || (this_block->addr & 0xc7fc0000) == 0x06000000)
+ {
+ u16 *drc_ram_blk = NULL;
+ u32 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;
}
- }
- else if ((this_block->addr & 0xc7fc0000) == 0x06000000) { // DRAM
- tmp = (this_block->addr & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT;
- tmp2 = (this_block->end_addr & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT;
- Pico32xMem->drcblk_ram[tmp] = (blkid_main << 1) | 1;
- for (++tmp; tmp < tmp2; tmp++) {
- if (Pico32xMem->drcblk_ram[tmp])
- continue;
- Pico32xMem->drcblk_ram[tmp] = blkid_main << 1;
+ else if ((this_block->addr & 0xc7fc0000) == 0x06000000) {
+ // SDRAM
+ drc_ram_blk = Pico32xMem->drcblk_ram;
+ shift = SH2_DRCBLK_RAM_SHIFT;
+ mask = 0x3ffff;
+ }
+
+ drc_ram_blk[(base_pc >> shift) & mask] = (blkid_main << 1) | 1;
+ for (pc = base_pc + 2; pc < end_pc; pc += 2)
+ drc_ram_blk[(pc >> shift) & mask] = blkid_main << 1;
+
+ // mark subblocks
+ for (i = 0; i < branch_target_count; i++)
+ if (branch_target_blkid[i] != 0)
+ drc_ram_blk[(branch_target_pc[i] >> shift) & mask] =
+ (branch_target_blkid[i] << 1) | 1;
+
+ // mark literals
+ for (i = 0; i < literal_addr_count; i++) {
+ tmp = literal_addr[i];
+ //printf("marking literal %08x\n", tmp);
+ drc_ram_blk[(tmp >> shift) & mask] = blkid_main << 1;
+ if (!(tmp & 3)) // assume long
+ drc_ram_blk[((tmp + 2) >> shift) & mask] = blkid_main << 1;
}
}
tcache_ptrs[tcache_id] = tcache_ptr;
-#ifdef ARM
- cache_flush_d_inval_i(block_entry, tcache_ptr);
-#endif
+ host_instructions_updated(block_entry, tcache_ptr);
do_host_disasm(tcache_id);
dbg(1, " block #%d,%d tcache %d/%d, insns %d -> %d %.3f",
int arg0, arg1, arg2, sr, tmp;
void *sh2_drc_write_end, *sh2_drc_write_slot_end;
+ sh2_drc_write32 = p32x_sh2_write32;
+ sh2_drc_read8 = p32x_sh2_read8;
+ sh2_drc_read16 = p32x_sh2_read16;
+ sh2_drc_read32 = p32x_sh2_read32;
+
host_arg2reg(arg0, 0);
host_arg2reg(arg1, 1);
host_arg2reg(arg2, 2);
emith_ctx_read(arg0, SHR_PC * 4);
emith_ctx_read(arg1, offsetof(SH2, is_slave));
emith_add_r_r_imm(arg2, CONTEXT_REG, offsetof(SH2, drc_tmp));
- emith_call(lookup_block);
+ emith_call(dr_lookup_block);
emit_block_entry();
// lookup failed, call sh2_translate()
emith_move_r_r(arg0, CONTEXT_REG);
tmp = rcache_get_reg_arg(1, SHR_SR);
emith_clear_msb(tmp, tmp, 22);
emith_move_r_r(arg2, CONTEXT_REG);
- emith_call(p32x_sh2_write32);
+ emith_call(p32x_sh2_write32); // XXX: use sh2_drc_write32?
rcache_invalidate();
// push PC
rcache_get_reg_arg(0, SHR_SP);
EMITH_SJMP_START(DCOND_NE);
emith_jump_ctx_c(DCOND_EQ, offsetof(SH2, drc_tmp)); // return
EMITH_SJMP_END(DCOND_NE);
- // since PC is up to date, jump to it's block instead of returning
emith_call(sh2_drc_test_irq);
emith_jump_ctx(offsetof(SH2, drc_tmp));
emith_ctx_read(arg2, offsetof(SH2, write16_tab));
emith_sh2_wcall(arg0, arg2, sh2_drc_write_slot_end);
+#ifdef PDB_NET
+ // debug
+ #define MAKE_READ_WRAPPER(func) { \
+ void *tmp = (void *)tcache_ptr; \
+ emith_push_ret(); \
+ emith_call(func); \
+ emith_ctx_read(arg2, offsetof(SH2, pdb_io_csum[0])); \
+ emith_addf_r_r(arg2, arg0); \
+ emith_ctx_write(arg2, offsetof(SH2, pdb_io_csum[0])); \
+ emith_ctx_read(arg2, offsetof(SH2, pdb_io_csum[1])); \
+ emith_adc_r_imm(arg2, 0x01000000); \
+ emith_ctx_write(arg2, offsetof(SH2, pdb_io_csum[1])); \
+ emith_pop_and_ret(); \
+ func = tmp; \
+ }
+ #define MAKE_WRITE_WRAPPER(func) { \
+ void *tmp = (void *)tcache_ptr; \
+ emith_ctx_read(arg2, offsetof(SH2, pdb_io_csum[0])); \
+ emith_addf_r_r(arg2, arg1); \
+ emith_ctx_write(arg2, offsetof(SH2, pdb_io_csum[0])); \
+ emith_ctx_read(arg2, offsetof(SH2, pdb_io_csum[1])); \
+ emith_adc_r_imm(arg2, 0x01000000); \
+ emith_ctx_write(arg2, offsetof(SH2, pdb_io_csum[1])); \
+ emith_move_r_r(arg2, CONTEXT_REG); \
+ emith_jump(func); \
+ func = tmp; \
+ }
+
+ MAKE_READ_WRAPPER(sh2_drc_read8);
+ MAKE_READ_WRAPPER(sh2_drc_read16);
+ MAKE_READ_WRAPPER(sh2_drc_read32);
+ MAKE_WRITE_WRAPPER(sh2_drc_write8);
+ MAKE_WRITE_WRAPPER(sh2_drc_write8_slot);
+ MAKE_WRITE_WRAPPER(sh2_drc_write16);
+ MAKE_WRITE_WRAPPER(sh2_drc_write16_slot);
+ MAKE_WRITE_WRAPPER(sh2_drc_write32);
+#if (DRC_DEBUG & 2)
+ host_dasm_new_symbol(sh2_drc_read8);
+ host_dasm_new_symbol(sh2_drc_read16);
+ host_dasm_new_symbol(sh2_drc_read32);
+ host_dasm_new_symbol(sh2_drc_write32);
+#endif
+#endif
+
rcache_invalidate();
#if (DRC_DEBUG & 2)
host_dasm_new_symbol(sh2_drc_entry);
#endif
}
-static void sh2_smc_rm_block(u16 *drcblk, u16 *p, block_desc *btab, u32 a)
+static void *sh2_smc_rm_block_entry(block_desc *bd, int tcache_id)
+{
+ void *tmp;
+
+ // XXX: kill links somehow?
+ dbg(1, " killing entry %08x, blkid %d", bd->addr, bd - block_tables[tcache_id]);
+ if (bd->addr == 0 || bd->tcache_ptr == NULL) {
+ printf(" killing dead block!? %08x\n", bd->addr);
+ return bd->tcache_ptr;
+ }
+
+ // 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);
+
+ // attempt to handle self-modifying blocks by exiting at nearest known PC
+ tmp = tcache_ptr;
+ tcache_ptr = bd->tcache_ptr;
+ emit_move_r_imm32(SHR_PC, bd->addr);
+ rcache_flush();
+ emith_jump(sh2_drc_dispatcher);
+ tcache_ptr = tmp;
+
+ bd->addr = 0;
+ return bd->tcache_ptr;
+}
+
+static void sh2_smc_rm_block(u32 a, u16 *drc_ram_blk, int tcache_id, u32 shift, u32 mask)
{
- u16 id = *p >> 1;
- block_desc *bd = btab + id;
-
- // FIXME: skip subblocks; do both directions
- // FIXME: collect all branches
- dbg(1, " killing block %08x", bd->addr);
- bd->addr = bd->end_addr = 0;
-
- while (p > drcblk && (p[-1] >> 1) == id)
- p--;
-
- // check for possible overlay block
- if (p > 0 && p[-1] != 0) {
- bd = btab + (p[-1] >> 1);
- if (bd->addr <= a && a < bd->end_addr)
- sh2_smc_rm_block(drcblk, p - 1, btab, a);
+ //block_link *bl = block_links[tcache_id];
+ //int bl_count = block_link_counts[tcache_id];
+ block_desc *btab = block_tables[tcache_id];
+ u16 *p = drc_ram_blk + ((a & mask) >> shift);
+ u16 *pmax = drc_ram_blk + (mask >> shift);
+ void *tcache_min, *tcache_max;
+ int zeros;
+ u16 *pt;
+
+ // Figure out what the main block is, as subblocks also have the flag set.
+ // This relies on sub having single entry. It's possible that innocent
+ // block might be hit, but that's not such a big deal.
+ if ((p[0] >> 1) != (p[1] >> 1)) {
+ for (; p > drc_ram_blk; p--)
+ if (p[-1] == 0 || (p[-1] >> 1) == (*p >> 1))
+ break;
}
+ pt = p;
+
+ for (; p > drc_ram_blk; p--)
+ if ((*p & 1))
+ break;
+
+ if (!(*p & 1)) {
+ printf("smc rm: missing block start for %08x?\n", a);
+ p = pt;
+ }
+
+ if (*p == 0)
+ return;
- do {
- *p++ = 0;
+ tcache_min = tcache_max = sh2_smc_rm_block_entry(&btab[*p >> 1], tcache_id);
+ *p = 0;
+
+ for (p++, zeros = 0; p < pmax && zeros < MAX_LITERAL_OFFSET / 2; p++) {
+ int id = *p >> 1;
+ if (id == 0) {
+ // 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)
+ zeros++;
+ continue;
+ }
+ if (*p & 1) {
+ if (id == (p[1] >> 1))
+ // hit other block
+ break;
+ tcache_max = sh2_smc_rm_block_entry(&btab[id], tcache_id);
+ }
+ *p = 0;
}
- while ((*p >> 1) == id);
+
+ host_instructions_updated(tcache_min, (void *)((char *)tcache_max + 4*4 + 4));
}
void sh2_drc_wcheck_ram(unsigned int a, int val, int cpuid)
{
- u16 *drcblk = Pico32xMem->drcblk_ram;
- u16 *p = drcblk + ((a & 0x3ffff) >> SH2_DRCBLK_RAM_SHIFT);
-
dbg(1, "%csh2 smc check @%08x", cpuid ? 's' : 'm', a);
- sh2_smc_rm_block(drcblk, p, block_tables[0], a);
+ sh2_smc_rm_block(a, Pico32xMem->drcblk_ram, 0, SH2_DRCBLK_RAM_SHIFT, 0x3ffff);
}
void sh2_drc_wcheck_da(unsigned int a, int val, int cpuid)
{
- u16 *drcblk = Pico32xMem->drcblk_da[cpuid];
- u16 *p = drcblk + ((a & 0xfff) >> SH2_DRCBLK_DA_SHIFT);
-
dbg(1, "%csh2 smc check @%08x", cpuid ? 's' : 'm', a);
- sh2_smc_rm_block(drcblk, p, block_tables[1 + cpuid], a);
+ sh2_smc_rm_block(a, Pico32xMem->drcblk_da[cpuid],
+ 1 + cpuid, SH2_DRCBLK_DA_SHIFT, 0xfff);
}
void sh2_execute(SH2 *sh2c, int cycles)
}
#if (DRC_DEBUG & 1)
-static void block_stats(void)
+void block_stats(void)
{
int c, b, i, total = 0;
+ printf("block stats:\n");
for (b = 0; b < ARRAY_SIZE(block_tables); b++)
for (i = 0; i < block_counts[b]; i++)
if (block_tables[b][i].addr != 0)
flush_tcache(2);
}
+void sh2_drc_mem_setup(SH2 *sh2)
+{
+ // fill the convenience pointers
+ sh2->p_bios = sh2->is_slave ? Pico32xMem->sh2_rom_s : Pico32xMem->sh2_rom_m;
+ sh2->p_da = Pico32xMem->data_array[sh2->is_slave];
+ sh2->p_sdram = Pico32xMem->sdram;
+ sh2->p_rom = Pico.rom;
+}
+
int sh2_drc_init(SH2 *sh2)
{
int i;
drc_cmn_init();
tcache_ptr = tcache;
sh2_generate_utils();
-#ifdef ARM
- cache_flush_d_inval_i(tcache, tcache_ptr);
-#endif
+ host_instructions_updated(tcache, tcache_ptr);
tcache_bases[0] = tcache_ptrs[0] = tcache_ptr;
for (i = 1; i < ARRAY_SIZE(tcache_bases); i++)