X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=deps%2Flightrec%2Flightrec.c;h=d5b1de968c57829b43a7abaabf1857fbc0fee90f;hb=2ec79b77bf0925c69ece6e51fc219a1e783aa432;hp=ffa40f0938af4634c1607ad014cac6e269ceaa3f;hpb=03535202b4b624c534340322646fb7f4062e3f53;p=pcsx_rearmed.git

diff --git a/deps/lightrec/lightrec.c b/deps/lightrec/lightrec.c
index ffa40f09..d5b1de96 100644
--- 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 */
@@ -201,7 +198,7 @@ static void lightrec_invalidate_map(struct lightrec_state *state,
 	}
 }
 
-enum psx_map
+static enum psx_map
 lightrec_get_map_idx(struct lightrec_state *state, u32 kaddr)
 {
 	const struct lightrec_mem_map *map;
@@ -240,32 +237,64 @@ lightrec_get_map(struct lightrec_state *state, void **host, u32 kaddr)
 	return map;
 }
 
-u32 lightrec_rw(struct lightrec_state *state, union code op,
-		u32 addr, u32 data, u32 *flags, struct block *block)
+u32 lightrec_rw(struct lightrec_state *state, union code op, u32 base,
+		u32 data, u32 *flags, struct block *block, u16 offset)
 {
 	const struct lightrec_mem_map *map;
 	const struct lightrec_mem_map_ops *ops;
 	u32 opcode = op.opcode;
+	bool was_tagged = true;
+	u16 old_flags;
+	u32 addr;
 	void *host;
 
-	addr += (s16) op.i.imm;
+	addr = kunseg(base + (s16) op.i.imm);
 
-	map = lightrec_get_map(state, &host, kunseg(addr));
+	map = lightrec_get_map(state, &host, addr);
 	if (!map) {
 		__segfault_cb(state, addr, block);
 		return 0;
 	}
 
-	if (unlikely(map->ops)) {
+	if (flags)
+		was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(*flags);
+
+	if (likely(!map->ops)) {
+		if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags)) {
+			/* Force parallel port accesses as HW accesses, because
+			 * the direct-I/O emitters can't differenciate it. */
+			if (unlikely(map == &state->maps[PSX_MAP_PARALLEL_PORT]))
+				*flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
+			/* If the base register is 0x0, be extra suspicious.
+			 * Some games (e.g. Sled Storm) actually do segmentation
+			 * faults by using uninitialized pointers, which are
+			 * later initialized to point to hardware registers. */
+			else if (op.i.rs && base == 0x0)
+				*flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
+			else
+				*flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
+		}
+
+		ops = &lightrec_default_ops;
+	} else if (flags &&
+		   LIGHTREC_FLAGS_GET_IO_MODE(*flags) == LIGHTREC_IO_DIRECT_HW) {
+		ops = &lightrec_default_ops;
+	} else {
 		if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
 			*flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_HW);
 
 		ops = map->ops;
-	} else {
-		if (flags && !LIGHTREC_FLAGS_GET_IO_MODE(*flags))
-			*flags |= LIGHTREC_IO_MODE(LIGHTREC_IO_DIRECT);
+	}
 
-		ops = &lightrec_default_ops;
+	if (!was_tagged) {
+		old_flags = block_set_flags(block, BLOCK_SHOULD_RECOMPILE);
+
+		if (!(old_flags & BLOCK_SHOULD_RECOMPILE)) {
+			pr_debug("Opcode of block at PC 0x%08x has been tagged"
+				 " - flag for recompilation\n", block->pc);
+
+			lut_write(state, lut_offset(block->pc), NULL);
+		}
 	}
 
 	switch (op.i.op) {
@@ -310,10 +339,10 @@ u32 lightrec_rw(struct lightrec_state *state, union code op,
 
 static void lightrec_rw_helper(struct lightrec_state *state,
 			       union code op, u32 *flags,
-			       struct block *block)
+			       struct block *block, u16 offset)
 {
 	u32 ret = lightrec_rw(state, op, state->regs.gpr[op.i.rs],
-			      state->regs.gpr[op.i.rt], flags, block);
+			      state->regs.gpr[op.i.rt], flags, block, offset);
 
 	switch (op.i.op) {
 	case OP_LB:
@@ -323,8 +352,12 @@ static void lightrec_rw_helper(struct lightrec_state *state,
 	case OP_LWL:
 	case OP_LWR:
 	case OP_LW:
-		if (op.i.rt)
+		if (OPT_HANDLE_LOAD_DELAYS && unlikely(!state->in_delay_slot_n)) {
+			state->temp_reg = ret;
+			state->in_delay_slot_n = 0xff;
+		} else if (op.i.rt) {
 			state->regs.gpr[op.i.rt] = ret;
+		}
 		fallthrough;
 	default:
 		break;
@@ -333,14 +366,13 @@ static void lightrec_rw_helper(struct lightrec_state *state,
 
 static void lightrec_rw_cb(struct lightrec_state *state, u32 arg)
 {
-	lightrec_rw_helper(state, (union code) arg, NULL, NULL);
+	lightrec_rw_helper(state, (union code) arg, NULL, NULL, 0);
 }
 
 static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 {
 	struct block *block;
 	struct opcode *op;
-	bool was_tagged;
 	u16 offset = (u16)arg;
 
 	block = lightrec_find_block_from_lut(state->block_cache,
@@ -348,21 +380,12 @@ static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 	if (unlikely(!block)) {
 		pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
 			 state->next_pc, offset);
+		lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 		return;
 	}
 
 	op = &block->opcode_list[offset];
-	was_tagged = LIGHTREC_FLAGS_GET_IO_MODE(op->flags);
-
-	lightrec_rw_helper(state, op->c, &op->flags, block);
-
-	if (!was_tagged) {
-		pr_debug("Opcode of block at PC 0x%08x has been tagged - flag "
-			 "for recompilation\n", block->pc);
-
-		block->flags |= BLOCK_SHOULD_RECOMPILE;
-		lut_write(state, lut_offset(block->pc), NULL);
-	}
+	lightrec_rw_helper(state, op->c, &op->flags, block, offset);
 }
 
 static u32 clamp_s32(s32 val, s32 min, s32 max)
@@ -418,12 +441,47 @@ static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg)
 
 u32 lightrec_mfc(struct lightrec_state *state, union code op)
 {
+	u32 val;
+
 	if (op.i.op == OP_CP0)
 		return state->regs.cp0[op.r.rd];
-	else if (op.r.rs == OP_CP2_BASIC_MFC2)
-		return lightrec_mfc2(state, op.r.rd);
-	else
-		return state->regs.cp2c[op.r.rd];
+
+	if (op.i.op == OP_SWC2) {
+		val = lightrec_mfc2(state, op.i.rt);
+	} else if (op.r.rs == OP_CP2_BASIC_MFC2)
+		val = lightrec_mfc2(state, op.r.rd);
+	else {
+		val = state->regs.cp2c[op.r.rd];
+
+		switch (op.r.rd) {
+		case 4:
+		case 12:
+		case 20:
+		case 26:
+		case 27:
+		case 29:
+		case 30:
+			val = (u32)(s16)val;
+			fallthrough;
+		default:
+			break;
+		}
+	}
+
+	if (state->ops.cop2_notify)
+		(*state->ops.cop2_notify)(state, op.opcode, val);
+
+	return val;
+}
+
+static void lightrec_mfc_cb(struct lightrec_state *state, union code op)
+{
+	u32 rt = lightrec_mfc(state, op);
+
+	if (op.i.op == OP_SWC2)
+		state->temp_reg = rt;
+	else if (op.r.rt)
+		state->regs.gpr[op.r.rt] = rt;
 }
 
 static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
@@ -540,21 +598,36 @@ static void lightrec_ctc2(struct lightrec_state *state, u8 reg, u32 data)
 	}
 }
 
-void lightrec_mtc(struct lightrec_state *state, union code op, u32 data)
+void lightrec_mtc(struct lightrec_state *state, union code op, u8 reg, u32 data)
 {
-	if (op.i.op == OP_CP0)
-		lightrec_mtc0(state, op.r.rd, data);
-	else if (op.r.rs == OP_CP2_BASIC_CTC2)
-		lightrec_ctc2(state, op.r.rd, data);
-	else
-		lightrec_mtc2(state, op.r.rd, data);
+	if (op.i.op == OP_CP0) {
+		lightrec_mtc0(state, reg, data);
+	} else {
+		if (op.i.op == OP_LWC2 || op.r.rs != OP_CP2_BASIC_CTC2)
+			lightrec_mtc2(state, reg, data);
+		else
+			lightrec_ctc2(state, reg, data);
+
+		if (state->ops.cop2_notify)
+			(*state->ops.cop2_notify)(state, op.opcode, data);
+	}
 }
 
 static void lightrec_mtc_cb(struct lightrec_state *state, u32 arg)
 {
 	union code op = (union code) arg;
+	u32 data;
+	u8 reg;
 
-	lightrec_mtc(state, op, state->regs.gpr[op.r.rt]);
+	if (op.i.op == OP_LWC2) {
+		data = state->temp_reg;
+		reg = op.i.rt;
+	} else {
+		data = state->regs.gpr[op.r.rt];
+		reg = op.r.rd;
+	}
+
+	lightrec_mtc(state, op, reg, data);
 }
 
 void lightrec_rfe(struct lightrec_state *state)
@@ -586,31 +659,26 @@ static void lightrec_cp_cb(struct lightrec_state *state, u32 arg)
 	lightrec_cp(state, (union code) arg);
 }
 
-static void lightrec_syscall_cb(struct lightrec_state *state)
-{
-	lightrec_set_exit_flags(state, LIGHTREC_EXIT_SYSCALL);
-}
-
-static void lightrec_break_cb(struct lightrec_state *state)
-{
-	lightrec_set_exit_flags(state, LIGHTREC_EXIT_BREAK);
-}
-
 static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 {
 	struct block *block = lightrec_find_block(state->block_cache, pc);
+	u8 old_flags;
 
 	if (block && lightrec_block_is_outdated(state, block)) {
 		pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
 
-		/* Make sure the recompiler isn't processing the block we'll
-		 * destroy */
-		if (ENABLE_THREADED_COMPILER)
-			lightrec_recompiler_remove(state->rec, block);
+		old_flags = block_set_flags(block, BLOCK_IS_DEAD);
+		if (!(old_flags & BLOCK_IS_DEAD)) {
+			/* Make sure the recompiler isn't processing the block
+			 * we'll destroy */
+			if (ENABLE_THREADED_COMPILER)
+				lightrec_recompiler_remove(state->rec, block);
+
+			lightrec_unregister_block(state->block_cache, block);
+			remove_from_code_lut(state->block_cache, block);
+			lightrec_free_block(state, block);
+		}
 
-		lightrec_unregister_block(state->block_cache, block);
-		remove_from_code_lut(state->block_cache, block);
-		lightrec_free_block(state, block);
 		block = NULL;
 	}
 
@@ -635,7 +703,7 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 	void *func;
 	int err;
 
-	for (;;) {
+	do {
 		func = lut_read(state, lut_offset(pc));
 		if (func && func != state->get_next_block)
 			break;
@@ -645,19 +713,19 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 		if (unlikely(!block))
 			break;
 
-		if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET)) {
+		if (OPT_REPLACE_MEMSET &&
+		    block_has_flag(block, BLOCK_IS_MEMSET)) {
 			func = state->memset_func;
 			break;
 		}
 
-		should_recompile = block->flags & BLOCK_SHOULD_RECOMPILE &&
-			!(block->flags & BLOCK_IS_DEAD);
+		should_recompile = block_has_flag(block, BLOCK_SHOULD_RECOMPILE) &&
+			!block_has_flag(block, BLOCK_NEVER_COMPILE) &&
+			!block_has_flag(block, BLOCK_IS_DEAD);
 
 		if (unlikely(should_recompile)) {
 			pr_debug("Block at PC 0x%08x should recompile\n", pc);
 
-			lightrec_unregister(MEM_FOR_CODE, block->code_size);
-
 			if (ENABLE_THREADED_COMPILER) {
 				lightrec_recompiler_add(state->rec, block);
 			} else {
@@ -677,12 +745,12 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 		if (likely(func))
 			break;
 
-		if (unlikely(block->flags & BLOCK_NEVER_COMPILE)) {
+		if (unlikely(block_has_flag(block, BLOCK_NEVER_COMPILE))) {
 			pc = lightrec_emulate_block(state, block, pc);
 
 		} else if (!ENABLE_THREADED_COMPILER) {
 			/* Block wasn't compiled yet - run the interpreter */
-			if (block->flags & BLOCK_FULLY_TAGGED)
+			if (block_has_flag(block, BLOCK_FULLY_TAGGED))
 				pr_debug("Block fully tagged, skipping first pass\n");
 			else if (ENABLE_FIRST_PASS && likely(!should_recompile))
 				pc = lightrec_emulate_block(state, block, pc);
@@ -693,29 +761,25 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 				state->exit_flags = LIGHTREC_EXIT_NOMEM;
 				return NULL;
 			}
+		} else if (unlikely(block_has_flag(block, BLOCK_IS_DEAD))) {
+			/*
+			 * If the block is dead but has never been compiled,
+			 * then its function pointer is NULL and we cannot
+			 * execute the block. In that case, reap all the dead
+			 * blocks now, and in the next loop we will create a
+			 * new block.
+			 */
+			lightrec_reaper_reap(state->reaper);
 		} else {
 			lightrec_recompiler_add(state->rec, block);
 		}
-
-		if (state->exit_flags != LIGHTREC_EXIT_NORMAL ||
-		    state->current_cycle >= state->target_cycle)
-			break;
-	}
+	} while (state->exit_flags == LIGHTREC_EXIT_NORMAL
+		 && state->current_cycle < state->target_cycle);
 
 	state->next_pc = pc;
 	return func;
 }
 
-static s32 c_function_wrapper(struct lightrec_state *state, s32 cycles_delta,
-			      void (*f)(struct lightrec_state *, u32), u32 arg)
-{
-	state->current_cycle = state->target_cycle - cycles_delta;
-
-	(*f)(state, arg);
-
-	return state->target_cycle - state->current_cycle;
-}
-
 static void * lightrec_alloc_code(struct lightrec_state *state, size_t size)
 {
 	void *code;
@@ -757,6 +821,8 @@ static void lightrec_free_code(struct lightrec_state *state, void *ptr)
 		lightrec_code_alloc_unlock(state);
 }
 
+static char lightning_code_data[0x80000];
+
 static void * lightrec_emit_code(struct lightrec_state *state,
 				 const struct block *block,
 				 jit_state_t *_jit, unsigned int *size)
@@ -767,7 +833,9 @@ static void * lightrec_emit_code(struct lightrec_state *state,
 
 	jit_realize();
 
-	if (!ENABLE_DISASSEMBLER)
+	if (ENABLE_DISASSEMBLER)
+		jit_set_data(lightning_code_data, sizeof(lightning_code_data), 0);
+	else
 		jit_set_data(NULL, 0, JIT_DISABLE_DATA | JIT_DISABLE_NOTE);
 
 	if (has_code_buffer) {
@@ -813,6 +881,9 @@ static void * lightrec_emit_code(struct lightrec_state *state,
 
 	*size = (unsigned int) new_code_size;
 
+	if (state->ops.code_inv)
+		state->ops.code_inv(code, new_code_size);
+
 	return code;
 }
 
@@ -821,9 +892,17 @@ static struct block * generate_wrapper(struct lightrec_state *state)
 	struct block *block;
 	jit_state_t *_jit;
 	unsigned int i;
-	int stack_ptr;
-	jit_node_t *to_tramp, *to_fn_epilog;
 	jit_node_t *addr[C_WRAPPERS_COUNT - 1];
+	jit_node_t *to_end[C_WRAPPERS_COUNT - 1];
+	u8 tmp = JIT_R1;
+
+#ifdef __sh__
+	/* On SH, GBR-relative loads target the r0 register.
+	 * Use it as the temporary register to factorize the move to
+	 * JIT_R1. */
+	if (LIGHTREC_REG_STATE == _GBR)
+		tmp = _R0;
+#endif
 
 	block = lightrec_malloc(state, MEM_FOR_IR, sizeof(*block));
 	if (!block)
@@ -840,67 +919,70 @@ static struct block * generate_wrapper(struct lightrec_state *state)
 	jit_prolog();
 	jit_tramp(256);
 
-	/* Add entry points; separate them by opcodes that increment
-	 * LIGHTREC_REG_STATE (since we cannot touch other registers).
-	 * The difference will then tell us which C function to call. */
+	/* Add entry points */
 	for (i = C_WRAPPERS_COUNT - 1; i > 0; i--) {
-		jit_addi(LIGHTREC_REG_STATE, LIGHTREC_REG_STATE, __WORDSIZE / 8);
+		jit_ldxi(tmp, LIGHTREC_REG_STATE,
+			 offsetof(struct lightrec_state, c_wrappers[i]));
+		to_end[i - 1] = jit_b();
 		addr[i - 1] = jit_indirect();
 	}
 
+	jit_ldxi(tmp, LIGHTREC_REG_STATE,
+		 offsetof(struct lightrec_state, c_wrappers[0]));
+
+	for (i = 0; i < C_WRAPPERS_COUNT - 1; i++)
+		jit_patch(to_end[i]);
+	jit_movr(JIT_R1, tmp);
+
 	jit_epilog();
 	jit_prolog();
 
-	stack_ptr = jit_allocai(sizeof(uintptr_t) * NUM_TEMPS);
-
 	/* Save all temporaries on stack */
-	for (i = 0; i < NUM_TEMPS; i++)
-		jit_stxi(stack_ptr + i * sizeof(uintptr_t), JIT_FP, JIT_R(i));
-
-	jit_getarg(JIT_R1, jit_arg());
+	for (i = 0; i < NUM_TEMPS; i++) {
+		if (i + FIRST_TEMP != 1) {
+			jit_stxi(offsetof(struct lightrec_state, wrapper_regs[i]),
+				 LIGHTREC_REG_STATE, JIT_R(i + FIRST_TEMP));
+		}
+	}
 
-	/* Jump to the trampoline */
-	to_tramp = jit_jmpi();
+	jit_getarg(JIT_R2, jit_arg());
 
-	/* The trampoline will jump back here */
-	to_fn_epilog = jit_label();
+	jit_prepare();
+	jit_pushargr(LIGHTREC_REG_STATE);
+	jit_pushargr(JIT_R2);
 
-	/* Restore temporaries from stack */
-	for (i = 0; i < NUM_TEMPS; i++)
-		jit_ldxi(JIT_R(i), JIT_FP, stack_ptr + i * sizeof(uintptr_t));
+	jit_ldxi_ui(JIT_R2, LIGHTREC_REG_STATE,
+		    offsetof(struct lightrec_state, target_cycle));
 
-	jit_ret();
-	jit_epilog();
+	/* state->current_cycle = state->target_cycle - delta; */
+	jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, LIGHTREC_REG_CYCLE);
+	jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
+		   LIGHTREC_REG_STATE, LIGHTREC_REG_CYCLE);
 
-	/* Trampoline entry point.
-	 * The sole purpose of the trampoline is to cheese Lightning not to
-	 * save/restore the callee-saved register LIGHTREC_REG_CYCLE, since we
-	 * do want to return to the caller with this register modified. */
-	jit_prolog();
-	jit_tramp(256);
-	jit_patch(to_tramp);
+	/* Call the wrapper function */
+	jit_finishr(JIT_R1);
 
-	/* Retrieve the wrapper function */
-	jit_ldxi(JIT_R0, LIGHTREC_REG_STATE,
-		 offsetof(struct lightrec_state, c_wrappers));
+	/* delta = state->target_cycle - state->current_cycle */;
+	jit_ldxi_ui(LIGHTREC_REG_CYCLE, LIGHTREC_REG_STATE,
+		    offsetof(struct lightrec_state, current_cycle));
+	jit_ldxi_ui(JIT_R1, LIGHTREC_REG_STATE,
+		    offsetof(struct lightrec_state, target_cycle));
+	jit_subr(LIGHTREC_REG_CYCLE, JIT_R1, LIGHTREC_REG_CYCLE);
 
-	/* Restore LIGHTREC_REG_STATE to its correct value */
-	jit_movi(LIGHTREC_REG_STATE, (uintptr_t) state);
-
-	jit_prepare();
-	jit_pushargr(LIGHTREC_REG_STATE);
-	jit_pushargr(LIGHTREC_REG_CYCLE);
-	jit_pushargr(JIT_R0);
-	jit_pushargr(JIT_R1);
-	jit_finishi(c_function_wrapper);
-	jit_retval_i(LIGHTREC_REG_CYCLE);
+	/* Restore temporaries from stack */
+	for (i = 0; i < NUM_TEMPS; i++) {
+		if (i + FIRST_TEMP != 1) {
+			jit_ldxi(JIT_R(i + FIRST_TEMP), LIGHTREC_REG_STATE,
+				 offsetof(struct lightrec_state, wrapper_regs[i]));
+		}
+	}
 
-	jit_patch_at(jit_jmpi(), to_fn_epilog);
+	jit_ret();
 	jit_epilog();
 
 	block->_jit = _jit;
 	block->opcode_list = NULL;
-	block->flags = 0;
+	block->flags = BLOCK_NO_OPCODE_LIST;
 	block->nb_ops = 0;
 
 	block->function = lightrec_emit_code(state, block, _jit,
@@ -952,11 +1034,54 @@ static u32 lightrec_memset(struct lightrec_state *state)
 	return 8 + 5 * (length  + 3 / 4);
 }
 
+static u32 lightrec_check_load_delay(struct lightrec_state *state, u32 pc, u8 reg)
+{
+	struct block *block;
+	union code first_op;
+
+	first_op = lightrec_read_opcode(state, pc);
+
+	if (likely(!opcode_reads_register(first_op, reg))) {
+		state->regs.gpr[reg] = state->temp_reg;
+	} else {
+		block = lightrec_get_block(state, pc);
+		if (unlikely(!block)) {
+			pr_err("Unable to get block at PC 0x%08x\n", pc);
+			lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
+			pc = 0;
+		} else {
+			pc = lightrec_handle_load_delay(state, block, pc, reg);
+		}
+	}
+
+	return pc;
+}
+
+static void update_cycle_counter_before_c(jit_state_t *_jit)
+{
+	/* update state->current_cycle */
+	jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
+		   offsetof(struct lightrec_state, target_cycle));
+	jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
+	jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
+		   LIGHTREC_REG_STATE, JIT_R1);
+}
+
+static void update_cycle_counter_after_c(jit_state_t *_jit)
+{
+	/* Recalc the delta */
+	jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
+		   offsetof(struct lightrec_state, current_cycle));
+	jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
+		   offsetof(struct lightrec_state, target_cycle));
+	jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
+}
+
 static struct block * generate_dispatcher(struct lightrec_state *state)
 {
 	struct block *block;
 	jit_state_t *_jit;
-	jit_node_t *to_end, *loop, *addr, *addr2, *addr3;
+	jit_node_t *to_end, *loop, *addr, *addr2, *addr3, *addr4, *addr5, *jmp, *jmp2;
 	unsigned int i;
 	u32 offset;
 
@@ -974,36 +1099,93 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 	jit_prolog();
 	jit_frame(256);
 
-	jit_getarg(JIT_R0, jit_arg());
+	jit_getarg(LIGHTREC_REG_STATE, jit_arg());
+	jit_getarg(JIT_V0, jit_arg());
+	jit_getarg(JIT_V1, jit_arg());
 	jit_getarg_i(LIGHTREC_REG_CYCLE, jit_arg());
 
 	/* Force all callee-saved registers to be pushed on the stack */
 	for (i = 0; i < NUM_REGS; i++)
-		jit_movr(JIT_V(i), JIT_V(i));
-
-	/* Pass lightrec_state structure to blocks, using the last callee-saved
-	 * register that Lightning provides */
-	jit_movi(LIGHTREC_REG_STATE, (intptr_t) state);
+		jit_movr(JIT_V(i + FIRST_REG), JIT_V(i + FIRST_REG));
 
 	loop = jit_label();
 
 	/* Call the block's code */
-	jit_jmpr(JIT_R0);
+	jit_jmpr(JIT_V1);
 
 	if (OPT_REPLACE_MEMSET) {
 		/* Blocks will jump here when they need to call
 		 * lightrec_memset() */
 		addr3 = jit_indirect();
 
+		jit_movr(JIT_V1, LIGHTREC_REG_CYCLE);
+
 		jit_prepare();
 		jit_pushargr(LIGHTREC_REG_STATE);
+
 		jit_finishi(lightrec_memset);
+		jit_retval(LIGHTREC_REG_CYCLE);
 
 		jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
 			    offsetof(struct lightrec_state, regs.gpr[31]));
+		jit_subr(LIGHTREC_REG_CYCLE, JIT_V1, LIGHTREC_REG_CYCLE);
+
+		if (OPT_DETECT_IMPOSSIBLE_BRANCHES || OPT_HANDLE_LOAD_DELAYS)
+			jmp = jit_b();
+	}
+
+	if (OPT_DETECT_IMPOSSIBLE_BRANCHES) {
+		/* Blocks will jump here when they reach a branch that should
+		 * be executed with the interpreter, passing the branch's PC
+		 * in JIT_V0 and the address of the block in JIT_V1. */
+		addr4 = jit_indirect();
+
+		update_cycle_counter_before_c(_jit);
+
+		jit_prepare();
+		jit_pushargr(LIGHTREC_REG_STATE);
+		jit_pushargr(JIT_V1);
+		jit_pushargr(JIT_V0);
+		jit_finishi(lightrec_emulate_block);
+
+		jit_retval(JIT_V0);
+
+		update_cycle_counter_after_c(_jit);
+
+		if (OPT_HANDLE_LOAD_DELAYS)
+			jmp2 = jit_b();
 
-		jit_retval(JIT_R0);
-		jit_subr(LIGHTREC_REG_CYCLE, LIGHTREC_REG_CYCLE, JIT_R0);
+	}
+
+	if (OPT_HANDLE_LOAD_DELAYS) {
+		/* Blocks will jump here when they reach a branch with a load
+		 * opcode in its delay slot. The delay slot has already been
+		 * executed; the load value is in (state->temp_reg), and the
+		 * register number is in JIT_V1.
+		 * Jump to a C function which will evaluate the branch target's
+		 * first opcode, to make sure that it does not read the register
+		 * in question; and if it does, handle it accordingly. */
+		addr5 = jit_indirect();
+
+		update_cycle_counter_before_c(_jit);
+
+		jit_prepare();
+		jit_pushargr(LIGHTREC_REG_STATE);
+		jit_pushargr(JIT_V0);
+		jit_pushargr(JIT_V1);
+		jit_finishi(lightrec_check_load_delay);
+
+		jit_retval(JIT_V0);
+
+		update_cycle_counter_after_c(_jit);
+
+		if (OPT_DETECT_IMPOSSIBLE_BRANCHES)
+			jit_patch(jmp2);
+	}
+
+	if (OPT_REPLACE_MEMSET
+	    && (OPT_DETECT_IMPOSSIBLE_BRANCHES || OPT_HANDLE_LOAD_DELAYS)) {
+		jit_patch(jmp);
 	}
 
 	/* The block will jump here, with the number of cycles remaining in
@@ -1018,61 +1200,64 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 	to_end = jit_blei(LIGHTREC_REG_CYCLE, 0);
 
 	/* Convert next PC to KUNSEG and avoid mirrors */
-	jit_andi(JIT_R0, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
-	jit_rshi_u(JIT_R1, JIT_R0, 28);
+	jit_andi(JIT_V1, JIT_V0, 0x10000000 | (RAM_SIZE - 1));
+	jit_rshi_u(JIT_R1, JIT_V1, 28);
 	jit_andi(JIT_R2, JIT_V0, BIOS_SIZE - 1);
 	jit_addi(JIT_R2, JIT_R2, RAM_SIZE);
-	jit_movnr(JIT_R0, JIT_R2, JIT_R1);
+	jit_movnr(JIT_V1, JIT_R2, JIT_R1);
 
 	/* If possible, use the code LUT */
 	if (!lut_is_32bit(state))
-		jit_lshi(JIT_R0, JIT_R0, 1);
-	jit_addr(JIT_R0, JIT_R0, LIGHTREC_REG_STATE);
+		jit_lshi(JIT_V1, JIT_V1, 1);
+	jit_add_state(JIT_V1, JIT_V1);
 
 	offset = offsetof(struct lightrec_state, code_lut);
 	if (lut_is_32bit(state))
-		jit_ldxi_ui(JIT_R0, JIT_R0, offset);
+		jit_ldxi_ui(JIT_V1, JIT_V1, offset);
 	else
-		jit_ldxi(JIT_R0, JIT_R0, offset);
+		jit_ldxi(JIT_V1, JIT_V1, offset);
 
 	/* If we get non-NULL, loop */
-	jit_patch_at(jit_bnei(JIT_R0, 0), loop);
+	jit_patch_at(jit_bnei(JIT_V1, 0), loop);
+
+	/* The code LUT will be set to this address when the block at the target
+	 * PC has been preprocessed but not yet compiled by the threaded
+	 * recompiler */
+	addr = jit_indirect();
 
 	/* Slow path: call C function get_next_block_func() */
 
 	if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 		/* We may call the interpreter - update state->current_cycle */
-		jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
-			   offsetof(struct lightrec_state, target_cycle));
-		jit_subr(JIT_R1, JIT_R2, LIGHTREC_REG_CYCLE);
-		jit_stxi_i(offsetof(struct lightrec_state, current_cycle),
-			   LIGHTREC_REG_STATE, JIT_R1);
+		update_cycle_counter_before_c(_jit);
 	}
 
-	/* The code LUT will be set to this address when the block at the target
-	 * PC has been preprocessed but not yet compiled by the threaded
-	 * recompiler */
-	addr = jit_indirect();
-
-	/* Get the next block */
 	jit_prepare();
 	jit_pushargr(LIGHTREC_REG_STATE);
 	jit_pushargr(JIT_V0);
+
+	/* Save the cycles register if needed */
+	if (!(ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES))
+		jit_movr(JIT_V0, LIGHTREC_REG_CYCLE);
+
+	/* Get the next block */
 	jit_finishi(&get_next_block_func);
-	jit_retval(JIT_R0);
+	jit_retval(JIT_V1);
 
 	if (ENABLE_FIRST_PASS || OPT_DETECT_IMPOSSIBLE_BRANCHES) {
 		/* The interpreter may have updated state->current_cycle and
 		 * state->target_cycle - recalc the delta */
-		jit_ldxi_i(JIT_R1, LIGHTREC_REG_STATE,
-			   offsetof(struct lightrec_state, current_cycle));
-		jit_ldxi_i(JIT_R2, LIGHTREC_REG_STATE,
-			   offsetof(struct lightrec_state, target_cycle));
-		jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
+		update_cycle_counter_after_c(_jit);
+	} else {
+		jit_movr(LIGHTREC_REG_CYCLE, JIT_V0);
 	}
 
+	/* Reset JIT_V0 to the next PC */
+	jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
+		    offsetof(struct lightrec_state, next_pc));
+
 	/* If we get non-NULL, loop */
-	jit_patch_at(jit_bnei(JIT_R0, 0), loop);
+	jit_patch_at(jit_bnei(JIT_V1, 0), loop);
 
 	/* When exiting, the recompiled code will jump to that address */
 	jit_note(__FILE__, __LINE__);
@@ -1083,7 +1268,7 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 
 	block->_jit = _jit;
 	block->opcode_list = NULL;
-	block->flags = 0;
+	block->flags = BLOCK_NO_OPCODE_LIST;
 	block->nb_ops = 0;
 
 	block->function = lightrec_emit_code(state, block, _jit,
@@ -1092,6 +1277,10 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 		goto err_free_block;
 
 	state->eob_wrapper_func = jit_address(addr2);
+	if (OPT_DETECT_IMPOSSIBLE_BRANCHES)
+		state->interpreter_func = jit_address(addr4);
+	if (OPT_HANDLE_LOAD_DELAYS)
+		state->ds_check_func = jit_address(addr5);
 	if (OPT_REPLACE_MEMSET)
 		state->memset_func = jit_address(addr3);
 	state->get_next_block = jit_address(addr);
@@ -1122,16 +1311,18 @@ union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
 	return (union code) LE32TOH(*code);
 }
 
-unsigned int lightrec_cycles_of_opcode(union code code)
+__cnst unsigned int lightrec_cycles_of_opcode(union code code)
 {
 	return 2;
 }
 
-void lightrec_free_opcode_list(struct lightrec_state *state, struct block *block)
+void lightrec_free_opcode_list(struct lightrec_state *state, struct opcode *ops)
 {
+	struct opcode_list *list = container_of(ops, struct opcode_list, ops);
+
 	lightrec_free(state, MEM_FOR_IR,
-		      sizeof(*block->opcode_list) * block->nb_ops,
-		      block->opcode_list);
+		      sizeof(*list) + list->nb_ops * sizeof(struct opcode),
+		      list);
 }
 
 static unsigned int lightrec_get_mips_block_len(const u32 *src)
@@ -1153,25 +1344,28 @@ static unsigned int lightrec_get_mips_block_len(const u32 *src)
 static struct opcode * lightrec_disassemble(struct lightrec_state *state,
 					    const u32 *src, unsigned int *len)
 {
-	struct opcode *list;
+	struct opcode_list *list;
 	unsigned int i, length;
 
 	length = lightrec_get_mips_block_len(src);
 
-	list = lightrec_malloc(state, MEM_FOR_IR, sizeof(*list) * length);
+	list = lightrec_malloc(state, MEM_FOR_IR,
+			       sizeof(*list) + sizeof(struct opcode) * length);
 	if (!list) {
 		pr_err("Unable to allocate memory\n");
 		return NULL;
 	}
 
+	list->nb_ops = (u16) length;
+
 	for (i = 0; i < length; i++) {
-		list[i].opcode = LE32TOH(src[i]);
-		list[i].flags = 0;
+		list->ops[i].opcode = LE32TOH(src[i]);
+		list->ops[i].flags = 0;
 	}
 
 	*len = length * sizeof(u32);
 
-	return list;
+	return list->ops;
 }
 
 static struct block * lightrec_precompile_block(struct lightrec_state *state,
@@ -1179,11 +1373,12 @@ static struct block * lightrec_precompile_block(struct lightrec_state *state,
 {
 	struct opcode *list;
 	struct block *block;
-	void *host;
+	void *host, *addr;
 	const struct lightrec_mem_map *map = lightrec_get_map(state, &host, kunseg(pc));
 	const u32 *code = (u32 *) host;
 	unsigned int length;
 	bool fully_tagged;
+	u8 block_flags = 0;
 
 	if (!map)
 		return NULL;
@@ -1209,9 +1404,6 @@ static struct block * lightrec_precompile_block(struct lightrec_state *state,
 	block->flags = 0;
 	block->code_size = 0;
 	block->precompile_date = state->current_cycle;
-#if ENABLE_THREADED_COMPILER
-	block->op_list_freed = (atomic_flag)ATOMIC_FLAG_INIT;
-#endif
 	block->nb_ops = length / sizeof(u32);
 
 	lightrec_optimize(state, block);
@@ -1227,21 +1419,22 @@ static struct block * lightrec_precompile_block(struct lightrec_state *state,
 
 	pr_debug("Block size: %hu opcodes\n", block->nb_ops);
 
-	/* If the first opcode is an 'impossible' branch, never compile the
-	 * block */
-	if (should_emulate(block->opcode_list))
-		block->flags |= BLOCK_NEVER_COMPILE;
-
 	fully_tagged = lightrec_block_is_fully_tagged(block);
 	if (fully_tagged)
-		block->flags |= BLOCK_FULLY_TAGGED;
+		block_flags |= BLOCK_FULLY_TAGGED;
 
-	if (OPT_REPLACE_MEMSET && (block->flags & BLOCK_IS_MEMSET))
-		lut_write(state, lut_offset(pc), state->memset_func);
+	if (block_flags)
+		block_set_flags(block, block_flags);
 
 	block->hash = lightrec_calculate_block_hash(block);
 
-	pr_debug("Recompile count: %u\n", state->nb_precompile++);
+	if (OPT_REPLACE_MEMSET && block_has_flag(block, BLOCK_IS_MEMSET))
+		addr = state->memset_func;
+	else
+		addr = state->get_next_block;
+	lut_write(state, lut_offset(pc), addr);
+
+	pr_debug("Blocks created: %u\n", ++state->nb_precompile);
 
 	return block;
 }
@@ -1254,8 +1447,12 @@ static bool lightrec_block_is_fully_tagged(const struct block *block)
 	for (i = 0; i < block->nb_ops; i++) {
 		op = &block->opcode_list[i];
 
-		/* Verify that all load/stores of the opcode list
-		 * Check all loads/stores of the opcode list and mark the
+		/* If we have one branch that must be emulated, we cannot trash
+		 * the opcode list. */
+		if (should_emulate(op))
+			return false;
+
+		/* Check all loads/stores of the opcode list and mark the
 		 * block as fully compiled if they all have been tagged. */
 		switch (op->c.i.op) {
 		case OP_LB:
@@ -1310,24 +1507,31 @@ static void lightrec_reap_function(struct lightrec_state *state, void *data)
 	lightrec_free_function(state, data);
 }
 
+static void lightrec_reap_opcode_list(struct lightrec_state *state, void *data)
+{
+	lightrec_free_opcode_list(state, data);
+}
+
 int lightrec_compile_block(struct lightrec_cstate *cstate,
 			   struct block *block)
 {
 	struct lightrec_state *state = cstate->state;
 	struct lightrec_branch_target *target;
-	bool op_list_freed = false, fully_tagged = false;
+	bool fully_tagged = false;
 	struct block *block2;
 	struct opcode *elm;
 	jit_state_t *_jit, *oldjit;
 	jit_node_t *start_of_block;
 	bool skip_next = false;
 	void *old_fn, *new_fn;
+	size_t old_code_size;
 	unsigned int i, j;
+	u8 old_flags;
 	u32 offset;
 
 	fully_tagged = lightrec_block_is_fully_tagged(block);
 	if (fully_tagged)
-		block->flags |= BLOCK_FULLY_TAGGED;
+		block_set_flags(block, BLOCK_FULLY_TAGGED);
 
 	_jit = jit_new_state();
 	if (!_jit)
@@ -1335,13 +1539,16 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 
 	oldjit = block->_jit;
 	old_fn = block->function;
+	old_code_size = block->code_size;
 	block->_jit = _jit;
 
 	lightrec_regcache_reset(cstate->reg_cache);
+	lightrec_preload_pc(cstate->reg_cache);
+
 	cstate->cycles = 0;
-	cstate->nb_branches = 0;
 	cstate->nb_local_branches = 0;
 	cstate->nb_targets = 0;
+	cstate->no_load_delay = false;
 
 	jit_prolog();
 	jit_tramp(256);
@@ -1360,7 +1567,7 @@ 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_jump_to_interpreter(cstate, block, i);
 			skip_next = !op_flag_no_ds(elm->flags);
 		} else {
 			lightrec_rec_opcode(cstate, block, i);
@@ -1377,9 +1584,6 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 		cstate->cycles += lightrec_cycles_of_opcode(elm->c);
 	}
 
-	for (i = 0; i < cstate->nb_branches; i++)
-		jit_patch(cstate->branches[i]);
-
 	for (i = 0; i < cstate->nb_local_branches; i++) {
 		struct lightrec_branch *branch = &cstate->local_branches[i];
 
@@ -1403,7 +1607,6 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 			pr_err("Unable to find branch target\n");
 	}
 
-	jit_patch_abs(jit_jmpi(), state->eob_wrapper_func);
 	jit_ret();
 	jit_epilog();
 
@@ -1412,22 +1615,24 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 		if (!ENABLE_THREADED_COMPILER)
 			pr_err("Unable to compile block!\n");
 		block->_jit = oldjit;
+		jit_clear_state();
 		_jit_destroy_state(_jit);
 		return -ENOMEM;
 	}
 
+	/* Pause the reaper, because lightrec_reset_lut_offset() may try to set
+	 * the old block->function pointer to the code LUT. */
+	if (ENABLE_THREADED_COMPILER)
+		lightrec_reaper_pause(state->reaper);
+
 	block->function = new_fn;
-	block->flags &= ~BLOCK_SHOULD_RECOMPILE;
+	block_clear_flags(block, BLOCK_SHOULD_RECOMPILE);
 
 	/* Add compiled function to the LUT */
 	lut_write(state, lut_offset(block->pc), block->function);
 
-	if (ENABLE_THREADED_COMPILER) {
-		/* Since we might try to reap the same block multiple times,
-		 * we need the reaper to wait until everything has been
-		 * submitted, so that the duplicate entries can be dropped. */
-		lightrec_reaper_pause(state->reaper);
-	}
+	if (ENABLE_THREADED_COMPILER)
+		lightrec_reaper_continue(state->reaper);
 
 	/* Detect old blocks that have been covered by the new one */
 	for (i = 0; i < cstate->nb_targets; i++) {
@@ -1437,6 +1642,13 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 			continue;
 
 		offset = block->pc + target->offset * sizeof(u32);
+
+		/* Pause the reaper while we search for the block until we set
+		 * the BLOCK_IS_DEAD flag, otherwise the block may be removed
+		 * under our feet. */
+		if (ENABLE_THREADED_COMPILER)
+			lightrec_reaper_pause(state->reaper);
+
 		block2 = lightrec_find_block(state->block_cache, offset);
 		if (block2) {
 			/* No need to check if block2 is compilable - it must
@@ -1444,12 +1656,16 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 
 			/* Set the "block dead" flag to prevent the dynarec from
 			 * recompiling this block */
-			block2->flags |= BLOCK_IS_DEAD;
+			old_flags = block_set_flags(block2, BLOCK_IS_DEAD);
+		}
+
+		if (ENABLE_THREADED_COMPILER) {
+			lightrec_reaper_continue(state->reaper);
 
 			/* If block2 was pending for compilation, cancel it.
 			 * If it's being compiled right now, wait until it
 			 * finishes. */
-			if (ENABLE_THREADED_COMPILER)
+			if (block2)
 				lightrec_recompiler_remove(state->rec, block2);
 		}
 
@@ -1464,20 +1680,17 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 				 "0x%08x\n", block2->pc, block->pc);
 
 			/* Finally, reap the block. */
-			if (ENABLE_THREADED_COMPILER) {
+			if (!ENABLE_THREADED_COMPILER) {
+				lightrec_unregister_block(state->block_cache, block2);
+				lightrec_free_block(state, block2);
+			} else if (!(old_flags & BLOCK_IS_DEAD)) {
 				lightrec_reaper_add(state->reaper,
 						    lightrec_reap_block,
 						    block2);
-			} else {
-				lightrec_unregister_block(state->block_cache, block2);
-				lightrec_free_block(state, block2);
 			}
 		}
 	}
 
-	if (ENABLE_THREADED_COMPILER)
-		lightrec_reaper_continue(state->reaper);
-
 	if (ENABLE_DISASSEMBLER) {
 		pr_debug("Compiling block at PC: 0x%08x\n", block->pc);
 		jit_disassemble();
@@ -1485,15 +1698,20 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 
 	jit_clear_state();
 
-#if ENABLE_THREADED_COMPILER
 	if (fully_tagged)
-		op_list_freed = atomic_flag_test_and_set(&block->op_list_freed);
-#endif
-	if (fully_tagged && !op_list_freed) {
+		old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
+
+	if (fully_tagged && !(old_flags & BLOCK_NO_OPCODE_LIST)) {
 		pr_debug("Block PC 0x%08x is fully tagged"
 			 " - free opcode list\n", block->pc);
-		lightrec_free_opcode_list(state, block);
-		block->opcode_list = NULL;
+
+		if (ENABLE_THREADED_COMPILER) {
+			lightrec_reaper_add(state->reaper,
+					    lightrec_reap_opcode_list,
+					    block->opcode_list);
+		} else {
+			lightrec_free_opcode_list(state, block->opcode_list);
+		}
 	}
 
 	if (oldjit) {
@@ -1509,8 +1727,12 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 			_jit_destroy_state(oldjit);
 			lightrec_free_function(state, old_fn);
 		}
+
+		lightrec_unregister(MEM_FOR_CODE, old_code_size);
 	}
 
+	pr_debug("Blocks compiled: %u\n", ++state->nb_compile);
+
 	return 0;
 }
 
@@ -1530,7 +1752,7 @@ static void lightrec_print_info(struct lightrec_state *state)
 
 u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
 {
-	s32 (*func)(void *, s32) = (void *)state->dispatcher->function;
+	s32 (*func)(struct lightrec_state *, u32, void *, s32) = (void *)state->dispatcher->function;
 	void *block_trace;
 	s32 cycles_delta;
 
@@ -1547,7 +1769,8 @@ u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
 	if (block_trace) {
 		cycles_delta = state->target_cycle - state->current_cycle;
 
-		cycles_delta = (*func)(block_trace, cycles_delta);
+		cycles_delta = (*func)(state, state->next_pc,
+				       block_trace, cycles_delta);
 
 		state->current_cycle = state->target_cycle - cycles_delta;
 	}
@@ -1561,20 +1784,24 @@ u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
 	return state->next_pc;
 }
 
-u32 lightrec_execute_one(struct lightrec_state *state, u32 pc)
+u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc,
+			     u32 target_cycle)
 {
-	return lightrec_execute(state, pc, state->current_cycle);
-}
-
-u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
-{
-	struct block *block = lightrec_get_block(state, pc);
-	if (!block)
-		return 0;
+	struct block *block;
 
 	state->exit_flags = LIGHTREC_EXIT_NORMAL;
+	state->target_cycle = target_cycle;
 
-	pc = lightrec_emulate_block(state, block, pc);
+	do {
+		block = lightrec_get_block(state, pc);
+		if (!block)
+			break;
+
+		pc = lightrec_emulate_block(state, block, pc);
+
+		if (ENABLE_THREADED_COMPILER)
+			lightrec_reaper_reap(state->reaper);
+	} while (state->current_cycle < state->target_cycle);
 
 	if (LOG_LEVEL >= INFO_L)
 		lightrec_print_info(state);
@@ -1584,9 +1811,13 @@ u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc)
 
 void lightrec_free_block(struct lightrec_state *state, struct block *block)
 {
+	u8 old_flags;
+
 	lightrec_unregister(MEM_FOR_MIPS_CODE, block->nb_ops * sizeof(u32));
-	if (block->opcode_list)
-		lightrec_free_opcode_list(state, block);
+	old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST);
+
+	if (!(old_flags & BLOCK_NO_OPCODE_LIST))
+		lightrec_free_opcode_list(state, block->opcode_list);
 	if (block->_jit)
 		_jit_destroy_state(block->_jit);
 	if (block->function) {
@@ -1639,6 +1870,11 @@ struct lightrec_state * lightrec_init(char *argv0,
 		return NULL;
 	}
 
+	if (ops->cop2_notify)
+		pr_debug("Optional cop2_notify callback in lightrec_ops\n");
+	else
+		pr_debug("No optional cop2_notify callback in lightrec_ops\n");
+
 	if (ENABLE_CODE_BUFFER && nb > PSX_MAP_CODE_BUFFER
 	    && codebuf_map->address) {
 		tlsf = tlsf_create_with_pool(codebuf_map->address,
@@ -1669,6 +1905,7 @@ struct lightrec_state * lightrec_init(char *argv0,
 
 	state->tlsf = tlsf;
 	state->with_32bit_lut = with_32bit_lut;
+	state->in_delay_slot_n = 0xff;
 
 	state->block_cache = lightrec_blockcache_init(state);
 	if (!state->block_cache)
@@ -1703,10 +1940,9 @@ struct lightrec_state * lightrec_init(char *argv0,
 
 	state->c_wrappers[C_WRAPPER_RW] = lightrec_rw_cb;
 	state->c_wrappers[C_WRAPPER_RW_GENERIC] = lightrec_rw_generic_cb;
+	state->c_wrappers[C_WRAPPER_MFC] = lightrec_mfc_cb;
 	state->c_wrappers[C_WRAPPER_MTC] = lightrec_mtc_cb;
 	state->c_wrappers[C_WRAPPER_CP] = lightrec_cp_cb;
-	state->c_wrappers[C_WRAPPER_SYSCALL] = lightrec_syscall_cb;
-	state->c_wrappers[C_WRAPPER_BREAK] = lightrec_break_cb;
 
 	map = &state->maps[PSX_MAP_BIOS];
 	state->offset_bios = (uintptr_t)map->address - map->pc;
@@ -1714,6 +1950,9 @@ struct lightrec_state * lightrec_init(char *argv0,
 	map = &state->maps[PSX_MAP_SCRATCH_PAD];
 	state->offset_scratch = (uintptr_t)map->address - map->pc;
 
+	map = &state->maps[PSX_MAP_HW_REGISTERS];
+	state->offset_io = (uintptr_t)map->address - map->pc;
+
 	map = &state->maps[PSX_MAP_KERNEL_USER_RAM];
 	state->offset_ram = (uintptr_t)map->address - map->pc;
 
@@ -1725,6 +1964,7 @@ struct lightrec_state * lightrec_init(char *argv0,
 	if (state->offset_bios == 0 &&
 	    state->offset_scratch == 0 &&
 	    state->offset_ram == 0 &&
+	    state->offset_io == 0 &&
 	    state->mirrors_mapped) {
 		pr_info("Memory map is perfect. Emitted code will be best.\n");
 	} else {