Merge pull request #515 from vacantron/jit/cache

Drop unused structure "chain_entry_t"

Author: jserv

src/cache.c

@@ -196,33 +196,57 @@ void *cache_get(const cache_t *cache, uint32_t key, bool update)
 
 void *cache_put(cache_t *cache, uint32_t key, void *value)
 {
-    void *delete_value = NULL;
     assert(cache->list_size <= cache->capacity);
-    /* check the cache is full or not before adding a new entry */
-    if (cache->list_size == cache->capacity) {
+
+    lfu_entry_t *replaced_entry = NULL, *entry;
+    hlist_for_each_entry (entry, &cache->map->ht_list_head[cache_hash(key)],
+                          ht_list) {
+        if (entry->key != key)
+            continue;
+        /* update the existing cache */
+        if (entry->value != value) {
+            replaced_entry = entry;
+            break;
+        }
+        /* should not put an identical block to cache */
+        assert(NULL);
+        __UNREACHABLE;
+    }
+
+    /* get the entry to be replaced if cache is full */
+    if (!replaced_entry && cache->list_size == cache->capacity) {
         for (int i = 0; i < THRESHOLD; i++) {
             if (list_empty(cache->lists[i]))
                 continue;
-            lfu_entry_t *delete_target =
+            replaced_entry =
                 list_last_entry(cache->lists[i], lfu_entry_t, list);
-            list_del_init(&delete_target->list);
-            hlist_del_init(&delete_target->ht_list);
-            delete_value = delete_target->value;
-            cache->list_size--;
-            mpool_free(cache_mp, delete_target);
             break;
         }
+        assert(replaced_entry);
     }
+
+    void *replaced_value = NULL;
+    if (replaced_entry) {
+        replaced_value = replaced_entry->value;
+        list_del_init(&replaced_entry->list);
+        hlist_del_init(&replaced_entry->ht_list);
+        mpool_free(cache_mp, replaced_entry);
+        cache->list_size--;
+    }
+
     lfu_entry_t *new_entry = mpool_alloc(cache_mp);
+    INIT_LIST_HEAD(&new_entry->list);
+    INIT_HLIST_NODE(&new_entry->ht_list);
     new_entry->key = key;
     new_entry->value = value;
     new_entry->frequency = 0;
     list_add(&new_entry->list, cache->lists[new_entry->frequency++]);
-    cache->list_size++;
     hlist_add_head(&new_entry->ht_list,
                    &cache->map->ht_list_head[cache_hash(key)]);
+    cache->list_size++;
+
     assert(cache->list_size <= cache->capacity);
-    return delete_value;
+    return replaced_value;
 }
 
 void cache_free(cache_t *cache)

src/emulate.c

@@ -533,8 +533,37 @@ FORCE_INLINE bool insn_is_unconditional_branch(uint8_t opcode)
     case rv_insn_cebreak:
 #endif
         return true;
+    default:
+        return false;
+    }
+}
+
+FORCE_INLINE bool insn_is_direct_branch(uint8_t opcode)
+{
+    switch (opcode) {
+    case rv_insn_jal:
+#if RV32_HAS(EXT_C)
+    case rv_insn_cjal:
+    case rv_insn_cj:
+#endif
+        return true;
+    default:
+        return false;
+    }
+}
+
+FORCE_INLINE bool insn_is_indirect_branch(uint8_t opcode)
+{
+    switch (opcode) {
+    case rv_insn_jalr:
+#if RV32_HAS(EXT_C)
+    case rv_insn_cjalr:
+    case rv_insn_cjr:
+#endif
+        return true;
+    default:
+        return false;
     }
-    return false;
 }
 
 static void block_translate(riscv_t *rv, block_t *block)
@@ -571,11 +600,7 @@ static void block_translate(riscv_t *rv, block_t *block)
 #endif
         /* stop on branch */
         if (insn_is_branch(ir->opcode)) {
-            if (ir->opcode == rv_insn_jalr
-#if RV32_HAS(EXT_C)
-                || ir->opcode == rv_insn_cjalr || ir->opcode == rv_insn_cjr
-#endif
-            ) {
+            if (insn_is_indirect_branch(ir->opcode)) {
                 ir->branch_table = calloc(1, sizeof(branch_history_table_t));
                 assert(ir->branch_table);
                 memset(ir->branch_table->PC, -1,
@@ -768,95 +793,89 @@ static block_t *block_find_or_translate(riscv_t *rv)
 #if !RV32_HAS(JIT)
     block_map_t *map = &rv->block_map;
     /* lookup the next block in the block map */
-    block_t *next = block_find(map, rv->PC);
+    block_t *next_blk = block_find(map, rv->PC);
 #else
     /* lookup the next block in the block cache */
-    block_t *next = (block_t *) cache_get(rv->block_cache, rv->PC, true);
+    /*
+     * The function "cache_get()" gets the cached block by the given "key (PC)".
+     * In system simulation, the returned block might be dropped because it is
+     * not the one from the current process (by checking SATP CSR register).
+     */
+    block_t *next_blk = (block_t *) cache_get(rv->block_cache, rv->PC, true);
 #endif
 
-    if (!next) {
+    if (next_blk)
+        return next_blk;
+
 #if !RV32_HAS(JIT)
-        if (map->size * 1.25 > map->block_capacity) {
-            block_map_clear(rv);
-            prev = NULL;
-        }
+    /* clear block list if it is going to be filled */
+    if (map->size * 1.25 > map->block_capacity) {
+        block_map_clear(rv);
+        prev = NULL;
+    }
 #endif
-        /* allocate a new block */
-        next = block_alloc(rv);
-        block_translate(rv, next);
+    /* allocate a new block */
+    next_blk = block_alloc(rv);
+
+    block_translate(rv, next_blk);
 
-        optimize_constant(rv, next);
+    optimize_constant(rv, next_blk);
 #if RV32_HAS(GDBSTUB)
-        if (likely(!rv->debug_mode))
+    if (likely(!rv->debug_mode))
 #endif
-            /* macro operation fusion */
-            match_pattern(rv, next);
+        /* macro operation fusion */
+        match_pattern(rv, next_blk);
 
 #if !RV32_HAS(JIT)
-        /* insert the block into block map */
-        block_insert(&rv->block_map, next);
+    /* insert the block into block map */
+    block_insert(&rv->block_map, next_blk);
 #else
-        /* insert the block into block cache */
-        block_t *delete_target = cache_put(rv->block_cache, rv->PC, &(*next));
-        if (delete_target) {
-            if (prev == delete_target)
-                prev = NULL;
-            chain_entry_t *entry, *safe;
-            /* correctly remove deleted block from its chained block */
-            rv_insn_t *taken = delete_target->ir_tail->branch_taken,
-                      *untaken = delete_target->ir_tail->branch_untaken;
-            if (taken && taken->pc != delete_target->pc_start) {
-                block_t *target = cache_get(rv->block_cache, taken->pc, false);
-                bool flag = false;
-                list_for_each_entry_safe (entry, safe, &target->list, list) {
-                    if (entry->block == delete_target) {
-                        list_del_init(&entry->list);
-                        mpool_free(rv->chain_entry_mp, entry);
-                        flag = true;
-                    }
-                }
-                assert(flag);
-            }
-            if (untaken && untaken->pc != delete_target->pc_start) {
-                block_t *target =
-                    cache_get(rv->block_cache, untaken->pc, false);
-                assert(target);
-                bool flag = false;
-                list_for_each_entry_safe (entry, safe, &target->list, list) {
-                    if (entry->block == delete_target) {
-                        list_del_init(&entry->list);
-                        mpool_free(rv->chain_entry_mp, entry);
-                        flag = true;
-                    }
-                }
-                assert(flag);
-            }
-            /* correctly remove deleted block from the block chained to it */
-            list_for_each_entry_safe (entry, safe, &delete_target->list, list) {
-                if (entry->block == delete_target)
-                    continue;
-                rv_insn_t *target = entry->block->ir_tail;
-                if (target->branch_taken == delete_target->ir_head)
-                    target->branch_taken = NULL;
-                else if (target->branch_untaken == delete_target->ir_head)
-                    target->branch_untaken = NULL;
-                mpool_free(rv->chain_entry_mp, entry);
-            }
-            /* free deleted block */
-            uint32_t idx;
-            rv_insn_t *ir, *next;
-            for (idx = 0, ir = delete_target->ir_head;
-                 idx < delete_target->n_insn; idx++, ir = next) {
-                free(ir->fuse);
-                next = ir->next;
-                mpool_free(rv->block_ir_mp, ir);
-            }
-            mpool_free(rv->block_mp, delete_target);
+    list_add(&next_blk->list, &rv->block_list);
+
+    /* insert the block into block cache */
+    block_t *replaced_blk = cache_put(rv->block_cache, rv->PC, &(*next_blk));
+
+    if (!replaced_blk)
+        return next_blk;
+
+    list_del_init(&replaced_blk->list);
+
+    if (prev == replaced_blk)
+        prev = NULL;
+
+    /* remove the connection from parents */
+    rv_insn_t *replaced_blk_entry = replaced_blk->ir_head;
+
+    /* TODO: record parents of each block to avoid traversing all blocks */
+    block_t *entry;
+    list_for_each_entry (entry, &rv->block_list, list) {
+        rv_insn_t *taken = entry->ir_tail->branch_taken,
+                  *untaken = entry->ir_tail->branch_untaken;
+
+        if (taken == replaced_blk_entry) {
+            entry->ir_tail->branch_taken = NULL;
         }
-#endif
+        if (untaken == replaced_blk_entry) {
+            entry->ir_tail->branch_untaken = NULL;
+        }
+    }
+
+    /* free IRs in replaced block */
+    for (rv_insn_t *ir = replaced_blk->ir_head, *next_ir; ir != NULL;
+         ir = next_ir) {
+        next_ir = ir->next;
+
+        if (ir->fuse)
+            free(ir->fuse);
+
+        mpool_free(rv->block_ir_mp, ir);
     }
 
-    return next;
+    mpool_free(rv->block_mp, replaced_blk);
+#endif
+
+    assert(next_blk);
+    return next_blk;
 }
 
 #if RV32_HAS(JIT)
@@ -918,31 +937,12 @@ void rv_step(void *arg)
             if (!insn_is_unconditional_branch(last_ir->opcode)) {
                 if (is_branch_taken && !last_ir->branch_taken) {
                     last_ir->branch_taken = block->ir_head;
-#if RV32_HAS(JIT)
-                    chain_entry_t *new_entry = mpool_alloc(rv->chain_entry_mp);
-                    new_entry->block = prev;
-                    list_add(&new_entry->list, &block->list);
-#endif
                 } else if (!is_branch_taken && !last_ir->branch_untaken) {
                     last_ir->branch_untaken = block->ir_head;
-#if RV32_HAS(JIT)
-                    chain_entry_t *new_entry = mpool_alloc(rv->chain_entry_mp);
-                    new_entry->block = prev;
-                    list_add(&new_entry->list, &block->list);
-#endif
                 }
-            } else if (IF_insn(last_ir, jal)
-#if RV32_HAS(EXT_C)
-                       || IF_insn(last_ir, cj) || IF_insn(last_ir, cjal)
-#endif
-            ) {
+            } else if (insn_is_direct_branch(last_ir->opcode)) {
                 if (!last_ir->branch_taken) {
                     last_ir->branch_taken = block->ir_head;
-#if RV32_HAS(JIT)
-                    chain_entry_t *new_entry = mpool_alloc(rv->chain_entry_mp);
-                    new_entry->block = prev;
-                    list_add(&new_entry->list, &block->list);
-#endif
                 }
             }
         }

src/riscv.c

@@ -315,9 +315,7 @@ riscv_t *rv_create(riscv_user_t rv_attr)
     /* initialize the block map */
     block_map_init(&rv->block_map, BLOCK_MAP_CAPACITY_BITS);
 #else
-    rv->chain_entry_mp =
-        mpool_create(sizeof(chain_entry_t) << BLOCK_IR_MAP_CAPACITY_BITS,
-                     sizeof(chain_entry_t));
+    INIT_LIST_HEAD(&rv->block_list);
     rv->jit_state = jit_state_init(CODE_CACHE_SIZE);
     rv->block_cache = cache_create(BLOCK_MAP_CAPACITY_BITS);
     assert(rv->block_cache);
@@ -426,7 +424,6 @@ void rv_delete(riscv_t *rv)
     pthread_mutex_destroy(&rv->wait_queue_lock);
     jit_cache_exit(rv->jit_cache);
 #endif
-    mpool_destroy(rv->chain_entry_mp);
     jit_state_exit(rv->jit_state);
     cache_free(rv->block_cache);
 #endif

src/riscv_private.h

@@ -100,18 +100,12 @@ typedef struct block {
 #endif
 } block_t;
 
-#if RV32_HAS(JIT)
-typedef struct {
-    block_t *block;
-    struct list_head list;
-} chain_entry_t;
-#if RV32_HAS(T2C)
+#if RV32_HAS(JIT) && RV32_HAS(T2C)
 typedef struct {
     block_t *block;
     struct list_head list;
 } queue_entry_t;
 #endif
-#endif
 
 typedef struct {
     uint32_t block_capacity; /**< max number of entries in the block map */
@@ -178,7 +172,7 @@ struct riscv_internal {
     block_map_t block_map; /**< basic block map */
 #else
     struct cache *block_cache;
-    struct mpool *chain_entry_mp;
+    struct list_head block_list; /**< list of all translated blocks */
 #if RV32_HAS(T2C)
     struct list_head wait_queue;
     pthread_mutex_t wait_queue_lock;

src/utils.h

@@ -106,12 +106,22 @@ static inline void list_del_init(struct list_head *node)
     list_entry((head)->prev, type, member)
 
 #ifdef __HAVE_TYPEOF
+#define list_for_each_entry(entry, head, member)                       \
+    for (entry = list_entry((head)->next, __typeof__(*entry), member); \
+         &entry->member != (head);                                     \
+         entry = list_entry(entry->member.next, __typeof__(*entry), member))
+
 #define list_for_each_entry_safe(entry, safe, head, member)                \
     for (entry = list_entry((head)->next, __typeof__(*entry), member),     \
         safe = list_entry(entry->member.next, __typeof__(*entry), member); \
          &entry->member != (head); entry = safe,                           \
         safe = list_entry(safe->member.next, __typeof__(*entry), member))
 #else
+#define list_for_each_entry(entry, head, member, type)   \
+    for (entry = list_entry((head)->next, type, member); \
+         &entry->member != (head);                       \
+         entry = list_entry(entry->member.next, type, member))
+
 #define list_for_each_entry_safe(entry, safe, head, member, type) \
     for (entry = list_entry((head)->next, type, member),          \
         safe = list_entry(entry->member.next, type, member);      \