git subrepo pull (merge) deps/lightrec

subrepo:
  subdir:   "deps/lightrec"
  merged:   "b9cf86d3"
upstream:
  origin:   "https://github.com/pcercuei/lightrec.git"
  branch:   "master"
  commit:   "9a14de7d"
git-subrepo:
  version:  "0.4.1"
  origin:   "https://github.com/ingydotnet/git-subrepo"
  commit:   "a04d8c2"

deps/lightrec/.gitrepo

@@ -6,7 +6,7 @@
 [subrepo]
 	remote = https://github.com/pcercuei/lightrec.git
 	branch = master
-	commit = 96b4f0314f1906d78069ec45a18d7cc660e73aeb
+	commit = 9a14de7dca83b4e52de043b45e7288c87e74a8f2
 	parent = 9f797430963d9cf0fcef7d963466f9cac7026de2
 	method = merge
 	cmdver = 0.4.1

deps/lightrec/CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.0)
-project(lightrec LANGUAGES C VERSION 0.2)
+project(lightrec LANGUAGES C VERSION 0.3)
 
 set(BUILD_SHARED_LIBS ON CACHE BOOL "Build shared libraries")
 if (NOT BUILD_SHARED_LIBS)

deps/lightrec/blockcache.c

@@ -21,12 +21,10 @@
 #include <stdlib.h>
 
 /* Must be power of two */
-#define TINY_LUT_SIZE 0x100
 #define LUT_SIZE 0x4000
 
 struct blockcache {
 	struct lightrec_state *state;
-	struct block * tiny_lut[TINY_LUT_SIZE];
 	struct block * lut[LUT_SIZE];
 };
 
@@ -36,39 +34,36 @@ struct block * lightrec_find_block(struct blockcache *cache, u32 pc)
 
 	pc = kunseg(pc);
 
-	block = cache->tiny_lut[(pc >> 2) & (TINY_LUT_SIZE - 1)];
-	if (likely(block && kunseg(block->pc) == pc))
-		return block;
-
-	block = cache->lut[(pc >> 2) & (LUT_SIZE - 1)];
 	for (block = cache->lut[(pc >> 2) & (LUT_SIZE - 1)];
-	     block; block = block->next) {
-		if (kunseg(block->pc) == pc) {
-			cache->tiny_lut[(pc >> 2) & (TINY_LUT_SIZE - 1)] = block;
+	     block; block = block->next)
+		if (kunseg(block->pc) == pc)
 			return block;
-		}
-	}
 
 	return NULL;
 }
 
-static void remove_from_code_lut(struct blockcache *cache, struct block *block)
+void remove_from_code_lut(struct blockcache *cache, struct block *block)
 {
 	struct lightrec_state *state = block->state;
+	const struct opcode *op;
+	u32 offset = lut_offset(block->pc);
 
 	/* Use state->get_next_block in the code LUT, which basically
 	 * calls back get_next_block_func(), until the compiler
 	 * overrides this. This is required, as a NULL value in the code
 	 * LUT means an outdated block. */
-	state->code_lut[lut_offset(block->pc)] = state->get_next_block;
+	state->code_lut[offset] = state->get_next_block;
+
+	for (op = block->opcode_list; op; op = op->next)
+		if (op->c.i.op == OP_META_SYNC)
+			state->code_lut[offset + op->offset] = NULL;
+
 }
 
 void lightrec_mark_for_recompilation(struct blockcache *cache,
 				     struct block *block)
 {
 	block->flags |= BLOCK_SHOULD_RECOMPILE;
-
-	remove_from_code_lut(cache, block);
 }
 
 void lightrec_register_block(struct blockcache *cache, struct block *block)
@@ -81,7 +76,6 @@ void lightrec_register_block(struct blockcache *cache, struct block *block)
 		block->next = old;
 
 	cache->lut[(pc >> 2) & (LUT_SIZE - 1)] = block;
-	cache->tiny_lut[(pc >> 2) & (TINY_LUT_SIZE - 1)] = block;
 
 	remove_from_code_lut(cache, block);
 }
@@ -91,9 +85,7 @@ void lightrec_unregister_block(struct blockcache *cache, struct block *block)
 	u32 pc = kunseg(block->pc);
 	struct block *old = cache->lut[(pc >> 2) & (LUT_SIZE - 1)];
 
-	block->state->code_lut[lut_offset(pc)] = NULL;
-
-	cache->tiny_lut[(pc >> 2) & (TINY_LUT_SIZE - 1)] = NULL;
+	remove_from_code_lut(cache, block);
 
 	if (old == block) {
 		cache->lut[(pc >> 2) & (LUT_SIZE - 1)] = old->next;
@@ -138,7 +130,51 @@ struct blockcache * lightrec_blockcache_init(struct lightrec_state *state)
 	return cache;
 }
 
+u32 lightrec_calculate_block_hash(const struct block *block)
+{
+	const struct lightrec_mem_map *map = block->map;
+	u32 pc, hash = 0xffffffff;
+	const u32 *code;
+	unsigned int i;
+
+	pc = kunseg(block->pc) - map->pc;
+
+	while (map->mirror_of)
+		map = map->mirror_of;
+
+	code = map->address + pc;
+
+	/* Jenkins one-at-a-time hash algorithm */
+	for (i = 0; i < block->nb_ops; i++) {
+		hash += *code++;
+		hash += (hash << 10);
+		hash ^= (hash >> 6);
+	}
+
+	hash += (hash << 3);
+	hash ^= (hash >> 11);
+	hash += (hash << 15);
+
+	return hash;
+}
+
 bool lightrec_block_is_outdated(struct block *block)
 {
-	return !block->state->code_lut[lut_offset(block->pc)];
+	void **lut_entry = &block->state->code_lut[lut_offset(block->pc)];
+	bool outdated;
+
+	if (*lut_entry)
+		return false;
+
+	outdated = block->hash != lightrec_calculate_block_hash(block);
+	if (likely(!outdated)) {
+		/* The block was marked as outdated, but the content is still
+		 * the same */
+		if (block->function)
+			*lut_entry = block->function;
+		else
+			*lut_entry = block->state->get_next_block;
+	}
+
+	return outdated;
 }

deps/lightrec/blockcache.h

@@ -26,7 +26,7 @@ void lightrec_unregister_block(struct blockcache *cache, struct block *block);
 struct blockcache * lightrec_blockcache_init(struct lightrec_state *state);
 void lightrec_free_block_cache(struct blockcache *cache);
 
-u32 calculate_block_hash(const struct block *block);
+u32 lightrec_calculate_block_hash(const struct block *block);
 _Bool lightrec_block_is_outdated(struct block *block);
 
 void lightrec_mark_for_recompilation(struct blockcache *cache,

deps/lightrec/emitter.c

@@ -86,7 +86,8 @@ static void lightrec_emit_end_of_block(const struct block *block,
 		pr_debug("EOB: %u cycles\n", cycles);
 	}
 
-	state->branches[state->nb_branches++] = jit_jmpi();
+	if (op->next && ((op->flags & LIGHTREC_NO_DS) || op->next->next))
+		state->branches[state->nb_branches++] = jit_jmpi();
 }
 
 void lightrec_emit_eob(const struct block *block,
@@ -1365,15 +1366,12 @@ static void rec_meta_unload(const struct block *block,
 	struct lightrec_state *state = block->state;
 	struct regcache *reg_cache = state->reg_cache;
 	jit_state_t *_jit = block->_jit;
-	u8 reg;
 
 	jit_name(__func__);
 	jit_note(__FILE__, __LINE__);
 
-	reg = lightrec_alloc_reg_in(reg_cache, _jit, op->i.rs);
-
 	pr_debug("Unloading reg %s\n", lightrec_reg_name(op->i.rs));
-	lightrec_unload_reg(reg_cache, _jit, reg);
+	lightrec_clean_reg_if_loaded(reg_cache, _jit, op->i.rs, true);
 }
 
 static void rec_meta_BEQZ(const struct block *block,

deps/lightrec/lightrec-private.h

@@ -73,6 +73,7 @@ struct block {
 	struct opcode *opcode_list;
 	void (*function)(void);
 	u32 pc;
+	u32 hash;
 #if ENABLE_THREADED_COMPILER
 	atomic_flag op_list_freed;
 #endif
@@ -131,6 +132,8 @@ u32 lightrec_rw(struct lightrec_state *state, union code op,
 
 void lightrec_free_block(struct block *block);
 
+void remove_from_code_lut(struct blockcache *cache, struct block *block);
+
 static inline u32 kunseg(u32 addr)
 {
 	if (unlikely(addr >= 0xa0000000))

deps/lightrec/lightrec.c

@@ -396,6 +396,7 @@ static void * get_next_block_func(struct lightrec_state *state, u32 pc)
 			if (ENABLE_THREADED_COMPILER)
 				lightrec_recompiler_remove(state->rec, block);
 
+			remove_from_code_lut(state->block_cache, block);
 			lightrec_unregister(MEM_FOR_CODE, block->code_size);
 			if (block->_jit)
 				_jit_destroy_state(block->_jit);
@@ -784,6 +785,8 @@ static struct block * lightrec_precompile_block(struct lightrec_state *state,
 	if (list->flags & LIGHTREC_EMULATE_BRANCH)
 		block->flags |= BLOCK_NEVER_COMPILE;
 
+	block->hash = lightrec_calculate_block_hash(block);
+
 	return block;
 }