X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=deps%2Flightrec%2Flightrec.c;h=1cf755b665167419b81fc5a67055259433d08f1c;hb=f99193c2c2ec0ab56636773949fae34d847d91b0;hp=d5b1de968c57829b43a7abaabf1857fbc0fee90f;hpb=cb72ea130a5ef1b2f47691ed586ad48bb0c39269;p=pcsx_rearmed.git

diff --git a/deps/lightrec/lightrec.c b/deps/lightrec/lightrec.c
index d5b1de96..1cf755b6 100644
--- a/deps/lightrec/lightrec.c
+++ b/deps/lightrec/lightrec.c
@@ -36,20 +36,20 @@ static void lightrec_mtc2(struct lightrec_state *state, u8 reg, u32 data);
 static u32 lightrec_mfc2(struct lightrec_state *state, u8 reg);
 
 static void lightrec_default_sb(struct lightrec_state *state, u32 opcode,
-				void *host, u32 addr, u8 data)
+				void *host, u32 addr, u32 data)
 {
-	*(u8 *)host = data;
+	*(u8 *)host = (u8)data;
 
-	if (!state->invalidate_from_dma_only)
+	if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
 		lightrec_invalidate(state, addr, 1);
 }
 
 static void lightrec_default_sh(struct lightrec_state *state, u32 opcode,
-				void *host, u32 addr, u16 data)
+				void *host, u32 addr, u32 data)
 {
-	*(u16 *)host = HTOLE16(data);
+	*(u16 *)host = HTOLE16((u16)data);
 
-	if (!state->invalidate_from_dma_only)
+	if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
 		lightrec_invalidate(state, addr, 2);
 }
 
@@ -58,7 +58,7 @@ static void lightrec_default_sw(struct lightrec_state *state, u32 opcode,
 {
 	*(u32 *)host = HTOLE32(data);
 
-	if (!state->invalidate_from_dma_only)
+	if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
 		lightrec_invalidate(state, addr, 4);
 }
 
@@ -80,6 +80,27 @@ static u32 lightrec_default_lw(struct lightrec_state *state,
 	return LE32TOH(*(u32 *)host);
 }
 
+static u32 lightrec_default_lwu(struct lightrec_state *state,
+				u32 opcode, void *host, u32 addr)
+{
+	u32 val;
+
+	memcpy(&val, host, 4);
+
+	return LE32TOH(val);
+}
+
+static void lightrec_default_swu(struct lightrec_state *state, u32 opcode,
+				 void *host, u32 addr, u32 data)
+{
+	data = HTOLE32(data);
+
+	memcpy(host, &data, 4);
+
+	if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
+		lightrec_invalidate(state, addr & ~0x3, 8);
+}
+
 static const struct lightrec_mem_map_ops lightrec_default_ops = {
 	.sb = lightrec_default_sb,
 	.sh = lightrec_default_sh,
@@ -87,6 +108,8 @@ static const struct lightrec_mem_map_ops lightrec_default_ops = {
 	.lb = lightrec_default_lb,
 	.lh = lightrec_default_lh,
 	.lw = lightrec_default_lw,
+	.lwu = lightrec_default_lwu,
+	.swu = lightrec_default_swu,
 };
 
 static void __segfault_cb(struct lightrec_state *state, u32 addr,
@@ -94,9 +117,9 @@ static void __segfault_cb(struct lightrec_state *state, u32 addr,
 {
 	lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 	pr_err("Segmentation fault in recompiled code: invalid "
-	       "load/store at address 0x%08x\n", addr);
+	       "load/store at address "PC_FMT"\n", addr);
 	if (block)
-		pr_err("Was executing block PC 0x%08x\n", block->pc);
+		pr_err("Was executing block "PC_FMT"\n", block->pc);
 }
 
 static void lightrec_swl(struct lightrec_state *state,
@@ -290,7 +313,7 @@ u32 lightrec_rw(struct lightrec_state *state, union code op, u32 base,
 		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"
+			pr_debug("Opcode of block at "PC_FMT" has been tagged"
 				 " - flag for recompilation\n", block->pc);
 
 			lut_write(state, lut_offset(block->pc), NULL);
@@ -299,10 +322,10 @@ u32 lightrec_rw(struct lightrec_state *state, union code op, u32 base,
 
 	switch (op.i.op) {
 	case OP_SB:
-		ops->sb(state, opcode, host, addr, (u8) data);
+		ops->sb(state, opcode, host, addr, data);
 		return 0;
 	case OP_SH:
-		ops->sh(state, opcode, host, addr, (u16) data);
+		ops->sh(state, opcode, host, addr, data);
 		return 0;
 	case OP_SWL:
 		lightrec_swl(state, ops, opcode, host, addr, data);
@@ -331,6 +354,11 @@ u32 lightrec_rw(struct lightrec_state *state, union code op, u32 base,
 		return lightrec_lwl(state, ops, opcode, host, addr, data);
 	case OP_LWR:
 		return lightrec_lwr(state, ops, opcode, host, addr, data);
+	case OP_META_LWU:
+		return ops->lwu(state, opcode, host, addr);
+	case OP_META_SWU:
+		ops->swu(state, opcode, host, addr, data);
+		return 0;
 	case OP_LW:
 	default:
 		return ops->lw(state, opcode, host, addr);
@@ -352,6 +380,7 @@ static void lightrec_rw_helper(struct lightrec_state *state,
 	case OP_LWL:
 	case OP_LWR:
 	case OP_LW:
+	case OP_META_LWU:
 		if (OPT_HANDLE_LOAD_DELAYS && unlikely(!state->in_delay_slot_n)) {
 			state->temp_reg = ret;
 			state->in_delay_slot_n = 0xff;
@@ -376,10 +405,10 @@ static void lightrec_rw_generic_cb(struct lightrec_state *state, u32 arg)
 	u16 offset = (u16)arg;
 
 	block = lightrec_find_block_from_lut(state->block_cache,
-					     arg >> 16, state->next_pc);
+					     arg >> 16, state->curr_pc);
 	if (unlikely(!block)) {
-		pr_err("rw_generic: No block found in LUT for PC 0x%x offset 0x%x\n",
-			 state->next_pc, offset);
+		pr_err("rw_generic: No block found in LUT for "PC_FMT" offset 0x%"PRIx16"\n",
+			 state->curr_pc, offset);
 		lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 		return;
 	}
@@ -524,7 +553,7 @@ static void lightrec_mtc0(struct lightrec_state *state, u8 reg, u32 data)
 		status = state->regs.cp0[12];
 
 		/* Handle software interrupts */
-		if (!!(status & cause & 0x300) & status)
+		if ((!!(status & cause & 0x300)) & status)
 			lightrec_set_exit_flags(state, LIGHTREC_EXIT_CHECK_INTERRUPT);
 
 		/* Handle hardware interrupts */
@@ -665,7 +694,7 @@ static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 	u8 old_flags;
 
 	if (block && lightrec_block_is_outdated(state, block)) {
-		pr_debug("Block at PC 0x%08x is outdated!\n", block->pc);
+		pr_debug("Block at "PC_FMT" is outdated!\n", block->pc);
 
 		old_flags = block_set_flags(block, BLOCK_IS_DEAD);
 		if (!(old_flags & BLOCK_IS_DEAD)) {
@@ -685,7 +714,7 @@ static struct block * lightrec_get_block(struct lightrec_state *state, u32 pc)
 	if (!block) {
 		block = lightrec_precompile_block(state, pc);
 		if (!block) {
-			pr_err("Unable to recompile block at PC 0x%x\n", pc);
+			pr_err("Unable to recompile block at "PC_FMT"\n", pc);
 			lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 			return NULL;
 		}
@@ -724,7 +753,7 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 			!block_has_flag(block, BLOCK_IS_DEAD);
 
 		if (unlikely(should_recompile)) {
-			pr_debug("Block at PC 0x%08x should recompile\n", pc);
+			pr_debug("Block at "PC_FMT" should recompile\n", pc);
 
 			if (ENABLE_THREADED_COMPILER) {
 				lightrec_recompiler_add(state->rec, block);
@@ -776,7 +805,7 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 	} while (state->exit_flags == LIGHTREC_EXIT_NORMAL
 		 && state->current_cycle < state->target_cycle);
 
-	state->next_pc = pc;
+	state->curr_pc = pc;
 	return func;
 }
 
@@ -1018,16 +1047,16 @@ static u32 lightrec_memset(struct lightrec_state *state)
 	u32 length = state->regs.gpr[5] * 4;
 
 	if (!map) {
-		pr_err("Unable to find memory map for memset target address "
-		       "0x%x\n", kunseg_pc);
+		pr_err("Unable to find memory map for memset target address "PC_FMT"\n",
+		       kunseg_pc);
 		return 0;
 	}
 
-	pr_debug("Calling host memset, PC 0x%x (host address 0x%" PRIxPTR ") for %u bytes\n",
+	pr_debug("Calling host memset, "PC_FMT" (host address 0x%"PRIxPTR") for %u bytes\n",
 		 kunseg_pc, (uintptr_t)host, length);
 	memset(host, 0, length);
 
-	if (!state->invalidate_from_dma_only)
+	if (!(state->opt_flags & LIGHTREC_OPT_INV_DMA_ONLY))
 		lightrec_invalidate_map(state, map, kunseg_pc, length);
 
 	/* Rough estimation of the number of cycles consumed */
@@ -1046,7 +1075,7 @@ static u32 lightrec_check_load_delay(struct lightrec_state *state, u32 pc, u8 re
 	} else {
 		block = lightrec_get_block(state, pc);
 		if (unlikely(!block)) {
-			pr_err("Unable to get block at PC 0x%08x\n", pc);
+			pr_err("Unable to get block at "PC_FMT"\n", pc);
 			lightrec_set_exit_flags(state, LIGHTREC_EXIT_SEGFAULT);
 			pc = 0;
 		} else {
@@ -1077,6 +1106,14 @@ static void update_cycle_counter_after_c(jit_state_t *_jit)
 	jit_subr(LIGHTREC_REG_CYCLE, JIT_R2, JIT_R1);
 }
 
+static void sync_next_pc(jit_state_t *_jit)
+{
+	if (lightrec_store_next_pc()) {
+		jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
+			    offsetof(struct lightrec_state, next_pc));
+	}
+}
+
 static struct block * generate_dispatcher(struct lightrec_state *state)
 {
 	struct block *block;
@@ -1140,6 +1177,7 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 		 * in JIT_V0 and the address of the block in JIT_V1. */
 		addr4 = jit_indirect();
 
+		sync_next_pc(_jit);
 		update_cycle_counter_before_c(_jit);
 
 		jit_prepare();
@@ -1167,6 +1205,7 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 		 * in question; and if it does, handle it accordingly. */
 		addr5 = jit_indirect();
 
+		sync_next_pc(_jit);
 		update_cycle_counter_before_c(_jit);
 
 		jit_prepare();
@@ -1178,22 +1217,24 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 		jit_retval(JIT_V0);
 
 		update_cycle_counter_after_c(_jit);
-
-		if (OPT_DETECT_IMPOSSIBLE_BRANCHES)
-			jit_patch(jmp2);
 	}
 
+	/* The block will jump here, with the number of cycles remaining in
+	 * LIGHTREC_REG_CYCLE */
+	addr2 = jit_indirect();
+
+	sync_next_pc(_jit);
+
+	if (OPT_HANDLE_LOAD_DELAYS && 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
-	 * LIGHTREC_REG_CYCLE */
-	addr2 = jit_indirect();
-
-	/* Store back the next_pc to the lightrec_state structure */
-	offset = offsetof(struct lightrec_state, next_pc);
+	/* Store back the next PC to the lightrec_state structure */
+	offset = offsetof(struct lightrec_state, curr_pc);
 	jit_stxi_i(offset, LIGHTREC_REG_STATE, JIT_V0);
 
 	/* Jump to end if state->target_cycle < state->current_cycle */
@@ -1254,7 +1295,7 @@ static struct block * generate_dispatcher(struct lightrec_state *state)
 
 	/* Reset JIT_V0 to the next PC */
 	jit_ldxi_ui(JIT_V0, LIGHTREC_REG_STATE,
-		    offsetof(struct lightrec_state, next_pc));
+		    offsetof(struct lightrec_state, curr_pc));
 
 	/* If we get non-NULL, loop */
 	jit_patch_at(jit_bnei(JIT_V1, 0), loop);
@@ -1311,9 +1352,10 @@ union code lightrec_read_opcode(struct lightrec_state *state, u32 pc)
 	return (union code) LE32TOH(*code);
 }
 
-__cnst unsigned int lightrec_cycles_of_opcode(union code code)
+unsigned int lightrec_cycles_of_opcode(const struct lightrec_state *state,
+				       union code code)
 {
-	return 2;
+	return state->cycles_per_op;
 }
 
 void lightrec_free_opcode_list(struct lightrec_state *state, struct opcode *ops)
@@ -1469,6 +1511,8 @@ static bool lightrec_block_is_fully_tagged(const struct block *block)
 		case OP_SWR:
 		case OP_LWC2:
 		case OP_SWC2:
+		case OP_META_LWU:
+		case OP_META_SWU:
 			if (!LIGHTREC_FLAGS_GET_IO_MODE(op->flags))
 				return false;
 			fallthrough;
@@ -1484,7 +1528,7 @@ static void lightrec_reap_block(struct lightrec_state *state, void *data)
 {
 	struct block *block = data;
 
-	pr_debug("Reap dead block at PC 0x%08x\n", block->pc);
+	pr_debug("Reap dead block at "PC_FMT"\n", block->pc);
 	lightrec_unregister_block(state->block_cache, block);
 	lightrec_free_block(state, block);
 }
@@ -1515,6 +1559,7 @@ static void lightrec_reap_opcode_list(struct lightrec_state *state, void *data)
 int lightrec_compile_block(struct lightrec_cstate *cstate,
 			   struct block *block)
 {
+	u32 was_dead[ARRAY_SIZE(cstate->targets) / 8];
 	struct lightrec_state *state = cstate->state;
 	struct lightrec_branch_target *target;
 	bool fully_tagged = false;
@@ -1543,7 +1588,9 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 	block->_jit = _jit;
 
 	lightrec_regcache_reset(cstate->reg_cache);
-	lightrec_preload_pc(cstate->reg_cache);
+
+	if (OPT_PRELOAD_PC && (block->flags & BLOCK_PRELOAD_PC))
+		lightrec_preload_pc(cstate->reg_cache, _jit);
 
 	cstate->cycles = 0;
 	cstate->nb_local_branches = 0;
@@ -1581,7 +1628,7 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 #endif
 		}
 
-		cstate->cycles += lightrec_cycles_of_opcode(elm->c);
+		cstate->cycles += lightrec_cycles_of_opcode(state, elm->c);
 	}
 
 	for (i = 0; i < cstate->nb_local_branches; i++) {
@@ -1631,11 +1678,8 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 	/* Add compiled function to the LUT */
 	lut_write(state, lut_offset(block->pc), block->function);
 
-	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++) {
+	for (i = 0; ENABLE_THREADED_COMPILER && i < cstate->nb_targets; i++) {
 		target = &cstate->targets[i];
 
 		if (!target->offset)
@@ -1643,12 +1687,6 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 
 		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
@@ -1657,17 +1695,24 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 			/* Set the "block dead" flag to prevent the dynarec from
 			 * recompiling this block */
 			old_flags = block_set_flags(block2, BLOCK_IS_DEAD);
+
+			if (old_flags & BLOCK_IS_DEAD)
+				was_dead[i / 32] |= BIT(i % 32);
+			else
+				was_dead[i / 32] &= ~BIT(i % 32);
 		}
 
-		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 (block2)
+			lightrec_recompiler_remove(state->rec, block2);
+	}
 
-			/* If block2 was pending for compilation, cancel it.
-			 * If it's being compiled right now, wait until it
-			 * finishes. */
-			if (block2)
-				lightrec_recompiler_remove(state->rec, block2);
-		}
+	for (i = 0; i < cstate->nb_targets; i++) {
+		target = &cstate->targets[i];
+
+		if (!target->offset)
+			continue;
 
 		/* We know from now on that block2 (if present) isn't going to
 		 * be compiled. We can override the LUT entry with our new
@@ -1675,6 +1720,8 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 		offset = lut_offset(block->pc) + target->offset;
 		lut_write(state, offset, jit_address(target->label));
 
+		offset = block->pc + target->offset * sizeof(u32);
+		block2 = lightrec_find_block(state->block_cache, offset);
 		if (block2) {
 			pr_debug("Reap block 0x%08x as it's covered by block "
 				 "0x%08x\n", block2->pc, block->pc);
@@ -1683,7 +1730,7 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 			if (!ENABLE_THREADED_COMPILER) {
 				lightrec_unregister_block(state->block_cache, block2);
 				lightrec_free_block(state, block2);
-			} else if (!(old_flags & BLOCK_IS_DEAD)) {
+			} else if (!(was_dead[i / 32] & BIT(i % 32))) {
 				lightrec_reaper_add(state->reaper,
 						    lightrec_reap_block,
 						    block2);
@@ -1691,6 +1738,9 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 		}
 	}
 
+	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();
@@ -1702,7 +1752,7 @@ int lightrec_compile_block(struct lightrec_cstate *cstate,
 		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"
+		pr_debug("Block "PC_FMT" is fully tagged"
 			 " - free opcode list\n", block->pc);
 
 		if (ENABLE_THREADED_COMPILER) {
@@ -1763,13 +1813,13 @@ u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
 		target_cycle = UINT_MAX;
 
 	state->target_cycle = target_cycle;
-	state->next_pc = pc;
+	state->curr_pc = pc;
 
 	block_trace = get_next_block_func(state, pc);
 	if (block_trace) {
 		cycles_delta = state->target_cycle - state->current_cycle;
 
-		cycles_delta = (*func)(state, state->next_pc,
+		cycles_delta = (*func)(state, state->curr_pc,
 				       block_trace, cycles_delta);
 
 		state->current_cycle = state->target_cycle - cycles_delta;
@@ -1781,7 +1831,7 @@ u32 lightrec_execute(struct lightrec_state *state, u32 pc, u32 target_cycle)
 	if (LOG_LEVEL >= INFO_L)
 		lightrec_print_info(state);
 
-	return state->next_pc;
+	return state->curr_pc;
 }
 
 u32 lightrec_run_interpreter(struct lightrec_state *state, u32 pc,
@@ -1895,7 +1945,7 @@ struct lightrec_state * lightrec_init(char *argv0,
 	else
 		lut_size = CODE_LUT_SIZE * sizeof(void *);
 
-	init_jit(argv0);
+	init_jit_with_debug(argv0, stdout);
 
 	state = calloc(1, sizeof(*state) + lut_size);
 	if (!state)
@@ -1906,6 +1956,7 @@ struct lightrec_state * lightrec_init(char *argv0,
 	state->tlsf = tlsf;
 	state->with_32bit_lut = with_32bit_lut;
 	state->in_delay_slot_n = 0xff;
+	state->cycles_per_op = 2;
 
 	state->block_cache = lightrec_blockcache_init(state);
 	if (!state->block_cache)
@@ -2052,12 +2103,12 @@ void lightrec_invalidate_all(struct lightrec_state *state)
 	memset(state->code_lut, 0, lut_elm_size(state) * CODE_LUT_SIZE);
 }
 
-void lightrec_set_invalidate_mode(struct lightrec_state *state, bool dma_only)
+void lightrec_set_unsafe_opt_flags(struct lightrec_state *state, u32 flags)
 {
-	if (state->invalidate_from_dma_only != dma_only)
+	if ((flags ^ state->opt_flags) & LIGHTREC_OPT_INV_DMA_ONLY)
 		lightrec_invalidate_all(state);
 
-	state->invalidate_from_dma_only = dma_only;
+	state->opt_flags = flags;
 }
 
 void lightrec_set_exit_flags(struct lightrec_state *state, u32 flags)
@@ -2100,3 +2151,8 @@ struct lightrec_registers * lightrec_get_registers(struct lightrec_state *state)
 {
 	return &state->regs;
 }
+
+void lightrec_set_cycles_per_opcode(struct lightrec_state *state, u32 cycles)
+{
+	state->cycles_per_op = cycles;
+}