Merge pull request #5668 from kumpera/wasm-work-p4

[mono.git] / mono / sgen / sgen-marksweep.c

diff --git a/mono/sgen/sgen-marksweep.c b/mono/sgen/sgen-marksweep.c
index 04feb680f2d77b9f6b239c0046cf4abf165cb589..3f14d5ee7acab33c1d35ccb0b84b3640bc9eb7e8 100644 (file)
--- a/mono/sgen/sgen-marksweep.c
+++ b/mono/sgen/sgen-marksweep.c
@@ -189,10 +189,15 @@ enum {
        SWEEP_STATE_COMPACTING
 };
 
+typedef enum {
+       SGEN_SWEEP_SERIAL = FALSE,
+       SGEN_SWEEP_CONCURRENT = TRUE,
+} SgenSweepMode;
+
 static volatile int sweep_state = SWEEP_STATE_SWEPT;
 
 static gboolean concurrent_mark;
-static gboolean concurrent_sweep = TRUE;
+static gboolean concurrent_sweep = DEFAULT_SWEEP_MODE;
 
 int sweep_pool_context = -1;
 
@@ -212,6 +217,7 @@ static SgenArrayList allocated_blocks = SGEN_ARRAY_LIST_INIT (NULL, sgen_array_l
 /* non-allocated block free-list */
 static void *empty_blocks = NULL;
 static size_t num_empty_blocks = 0;
+static gboolean compact_blocks = FALSE;
 
 /*
  * We can iterate the block list also while sweep is in progress but we
@@ -234,6 +240,16 @@ static size_t num_empty_blocks = 0;
                (bl) = BLOCK_UNTAG ((bl));
 #define END_FOREACH_BLOCK_NO_LOCK      } SGEN_ARRAY_LIST_END_FOREACH_SLOT; }
 
+#define FOREACH_BLOCK_RANGE_HAS_REFERENCES_NO_LOCK(bl,begin,end,index,hr) {    \
+       volatile gpointer *slot;                                        \
+       SGEN_ARRAY_LIST_FOREACH_SLOT_RANGE (&allocated_blocks, begin, end, slot, index) { \
+               (bl) = (MSBlockInfo *) (*slot);                         \
+               if (!(bl))                                              \
+                       continue;                                       \
+               (hr) = BLOCK_IS_TAGGED_HAS_REFERENCES ((bl));           \
+               (bl) = BLOCK_UNTAG ((bl));
+#define END_FOREACH_BLOCK_RANGE_NO_LOCK        } SGEN_ARRAY_LIST_END_FOREACH_SLOT_RANGE; }
+
 static volatile size_t num_major_sections = 0;
 /*
  * One free block list for each block object size.  We add and remove blocks from these
@@ -1209,7 +1225,7 @@ major_get_and_reset_num_major_objects_marked (void)
 #endif
 
 /* gcc 4.2.1 from xcode4 crashes on sgen_card_table_get_card_address () when this is enabled */
-#if defined(PLATFORM_MACOSX)
+#if defined(HOST_DARWIN)
 #if MONO_GNUC_VERSION <= 40300
 #undef PREFETCH_CARDS
 #endif
@@ -1601,6 +1617,8 @@ sweep_start (void)
 
        sgen_workers_foreach (GENERATION_NURSERY, sgen_worker_clear_free_block_lists);
        sgen_workers_foreach (GENERATION_OLD, sgen_worker_clear_free_block_lists);
+
+       compact_blocks = TRUE;
 }
 
 static void sweep_finish (void);
@@ -1974,6 +1992,18 @@ major_start_nursery_collection (void)
 #endif
 
        old_num_major_sections = num_major_sections;
+
+       /* Compact the block list if it hasn't been compacted in a while and nobody is using it */
+       if (compact_blocks && !sweep_in_progress () && !sweep_blocks_job && !sgen_concurrent_collection_in_progress ()) {
+               /*
+                * We support null elements in the array but do regular compaction to avoid
+                * excessive traversal of the array and to facilitate splitting into well
+                * balanced sections for parallel modes. We compact as soon as possible after
+                * sweep.
+                */
+               sgen_array_list_remove_nulls (&allocated_blocks);
+               compact_blocks = FALSE;
+       }
 }
 
 static void
@@ -2137,7 +2167,7 @@ major_free_swept_blocks (size_t section_reserve)
 {
        SGEN_ASSERT (0, sweep_state == SWEEP_STATE_SWEPT, "Sweeping must have finished before freeing blocks");
 
-#if defined(HOST_WIN32) || defined(HOST_ORBIS)
+#if defined(HOST_WIN32) || defined(HOST_ORBIS) || defined (HOST_WASM)
                /*
                 * sgen_free_os_memory () asserts in mono_vfree () because windows doesn't like freeing the middle of
                 * a VirtualAlloc ()-ed block.
@@ -2603,10 +2633,24 @@ scan_card_table_for_block (MSBlockInfo *block, CardTableScanType scan_type, Scan
 }
 
 static void
-major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job_index, int job_split_count)
+major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job_index, int job_split_count, int block_count)
 {
        MSBlockInfo *block;
        gboolean has_references, was_sweeping, skip_scan;
+       int first_block, last_block, index;
+
+       /*
+        * The last_block's index is at least (num_major_sections - 1) since we
+        * can have nulls in the allocated_blocks list. The last worker will
+        * scan the left-overs of the list. We expect few null entries in the
+        * allocated_blocks list, therefore using num_major_sections for computing
+        * block_count shouldn't affect work distribution.
+        */
+       first_block = block_count * job_index;
+       if (job_index == job_split_count - 1)
+               last_block = allocated_blocks.next_slot;
+       else
+               last_block = block_count * (job_index + 1);
 
        if (!concurrent_mark)
                g_assert (scan_type == CARDTABLE_SCAN_GLOBAL);
@@ -2616,11 +2660,9 @@ major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job
        was_sweeping = sweep_in_progress ();
 
        binary_protocol_major_card_table_scan_start (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
-       FOREACH_BLOCK_HAS_REFERENCES_NO_LOCK (block, has_references) {
-               if (__index % job_split_count != job_index)
-                       continue;
+       FOREACH_BLOCK_RANGE_HAS_REFERENCES_NO_LOCK (block, first_block, last_block, index, has_references) {
 #ifdef PREFETCH_CARDS
-               int prefetch_index = __index + 6 * job_split_count;
+               int prefetch_index = index + 6;
                if (prefetch_index < allocated_blocks.next_slot) {
                        MSBlockInfo *prefetch_block = BLOCK_UNTAG (*sgen_array_list_get_slot (&allocated_blocks, prefetch_index));
                        PREFETCH_READ (prefetch_block);
@@ -2631,7 +2673,6 @@ major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job
                        }
                 }
 #endif
-
                if (!has_references)
                        continue;
                skip_scan = FALSE;
@@ -2655,16 +2696,16 @@ major_scan_card_table (CardTableScanType scan_type, ScanCopyContext ctx, int job
                                 * sweep start since we are in a nursery collection. Also avoid CAS-ing
                                 */
                                if (sweep_in_progress ()) {
-                                       skip_scan = !ensure_block_is_checked_for_sweeping (__index, TRUE, NULL);
+                                       skip_scan = !ensure_block_is_checked_for_sweeping (index, TRUE, NULL);
                                } else if (was_sweeping) {
                                        /* Recheck in case sweep finished after dereferencing the slot */
-                                       skip_scan = *sgen_array_list_get_slot (&allocated_blocks, __index) == 0;
+                                       skip_scan = *sgen_array_list_get_slot (&allocated_blocks, index) == 0;
                                }
                        }
                }
                if (!skip_scan)
                        scan_card_table_for_block (block, scan_type, ctx);
-       } END_FOREACH_BLOCK_NO_LOCK;
+       } END_FOREACH_BLOCK_RANGE_NO_LOCK;
        binary_protocol_major_card_table_scan_end (sgen_timestamp (), scan_type & CARDTABLE_SCAN_MOD_UNION);
 }