git subrepo pull --force deps/lightrec

subrepo:
  subdir:   "deps/lightrec"
  merged:   "782c220e5d"
upstream:
  origin:   "https://github.com/pcercuei/lightrec.git"
  branch:   "master"
  commit:   "782c220e5d"
git-subrepo:
  version:  "0.4.3"
  origin:   "https://github.com/ingydotnet/git-subrepo.git"
  commit:   "2f68596"

diff --git a/deps/lightrec/.gitrepo b/deps/lightrec/.gitrepo
index 3db9689..ceb8643 100644 (file)
--- a/deps/lightrec/.gitrepo
+++ b/deps/lightrec/.gitrepo
@@ -6,7 +6,7 @@
 [subrepo]
        remote = https://github.com/pcercuei/lightrec.git
        branch = master
-       commit = 75374fa374a367c85f5cac6c1b9fc2c583298d42
-       parent = 22a3b99d24fada1c7a5e7ab4458c132a0c387d50
+       commit = 782c220e5dd9bdf179415a4d058823ff439254b8
+       parent = 5ced3945423cda0010597b27b7da6bce77b12baa
        method = merge
        cmdver = 0.4.3

diff --git a/deps/lightrec/emitter.c b/deps/lightrec/emitter.c
index 270fd48..3f7da58 100644 (file)
--- a/deps/lightrec/emitter.c
+++ b/deps/lightrec/emitter.c
@@ -87,21 +87,16 @@ 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);
+       jit_subi(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, state->cycles);
 
        lightrec_jump_to_eob(state, _jit);
 }
@@ -250,7 +245,7 @@ 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) {
@@ -1127,6 +1122,13 @@ 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,
@@ -1258,7 +1260,7 @@ 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,
@@ -1582,7 +1584,7 @@ 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,
@@ -1966,8 +1968,10 @@ rec_mtc0(struct lightrec_cstate *state, const struct block *block, u16 offset)
        lightrec_free_reg(reg_cache, rt);
 
        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,

diff --git a/deps/lightrec/emitter.h b/deps/lightrec/emitter.h
index b7f54fd..4cbe8da 100644 (file)
--- a/deps/lightrec/emitter.h
+++ b/deps/lightrec/emitter.h
@@ -13,7 +13,7 @@ struct lightrec_cstate;
 struct opcode;
 
 void lightrec_rec_opcode(struct lightrec_cstate *state, const struct block *block, u16 offset);
-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);
 
 #endif /* __EMITTER_H__ */

diff --git a/deps/lightrec/lightrec-private.h b/deps/lightrec/lightrec-private.h
index 6b6f721..56032f5 100644 (file)
--- a/deps/lightrec/lightrec-private.h
+++ b/deps/lightrec/lightrec-private.h
@@ -16,8 +16,15 @@
 #include <stdatomic.h>
 #endif
 
+#ifdef _MSC_BUILD
+#include <immintrin.h>
+#endif
+
 #define ARRAY_SIZE(x) (sizeof(x) ? sizeof(x) / sizeof((x)[0]) : 0)
 
+#define GENMASK(h, l) \
+       (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
+
 #ifdef __GNUC__
 #      define likely(x)       __builtin_expect(!!(x),1)
 #      define unlikely(x)     __builtin_expect(!!(x),0)
@@ -51,10 +58,12 @@
 
 #ifdef _MSC_BUILD
 #      define popcount32(x)    __popcnt(x)
-#      define ffs32(x)         (31 - __lzcnt(x))
+#      define clz32(x)         _lzcnt_u32(x)
+#      define ctz32(x)         _tzcnt_u32(x)
 #else
 #      define popcount32(x)    __builtin_popcount(x)
-#      define ffs32(x)         (__builtin_ffs(x) - 1)
+#      define clz32(x)         __builtin_clz(x)
+#      define ctz32(x)         __builtin_ctz(x)
 #endif
 
 /* Flags for (struct block *)->flags */

diff --git a/deps/lightrec/lightrec.c b/deps/lightrec/lightrec.c
index 43636b2..be4da10 100644 (file)
--- a/deps/lightrec/lightrec.c
+++ b/deps/lightrec/lightrec.c
@@ -28,9 +28,6 @@
 #include <stddef.h>
 #include <string.h>
 
-#define GENMASK(h, l) \
-       (((uintptr_t)-1 << (l)) & ((uintptr_t)-1 >> (__WORDSIZE - 1 - (h))))
-
 static struct block * lightrec_precompile_block(struct lightrec_state *state,
                                                u32 pc);
 static bool lightrec_block_is_fully_tagged(const struct block *block);
@@ -107,7 +104,7 @@ static void lightrec_swl(struct lightrec_state *state,
                         u32 opcode, void *host, u32 addr, u32 data)
 {
        unsigned int shift = addr & 0x3;
-       unsigned int mask = GENMASK(31, (shift + 1) * 8);
+       unsigned int mask = shift < 3 ? GENMASK(31, (shift + 1) * 8) : 0;
        u32 old_data;
 
        /* Align to 32 bits */
@@ -171,7 +168,7 @@ static u32 lightrec_lwr(struct lightrec_state *state,
                        u32 opcode, void *host, u32 addr, u32 data)
 {
        unsigned int shift = addr & 0x3;
-       unsigned int mask = GENMASK(31, 32 - shift * 8);
+       unsigned int mask = shift ? GENMASK(31, 32 - shift * 8) : 0;
        u32 old_data;
 
        /* Align to 32 bits */
@@ -1423,7 +1420,7 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
                        pr_debug("Branch at offset 0x%x will be emulated\n",
                                 i << 2);
 
-                       lightrec_emit_eob(cstate, block, i, false);
+                       lightrec_emit_eob(cstate, block, i);
                        skip_next = !op_flag_no_ds(elm->flags);
                } else {
                        lightrec_rec_opcode(cstate, block, i);

diff --git a/deps/lightrec/optimizer.c b/deps/lightrec/optimizer.c
index 3e94cff..10067a7 100644 (file)
--- a/deps/lightrec/optimizer.c
+++ b/deps/lightrec/optimizer.c
@@ -1112,7 +1112,7 @@ static int lightrec_transform_ops(struct lightrec_state *state, struct block *bl
                                else
                                        op->i.op = OP_META_MULTU2;
 
-                               op->r.op = ffs32(values[op->r.rt]);
+                               op->r.op = ctz32(values[op->r.rt]);
                                break;
                        case OP_SPECIAL_OR:
                        case OP_SPECIAL_ADD: