Merge branch 'dev' into merging-master-into-dev

Author: udesou

src/gc-interface.h

@@ -101,13 +101,26 @@ JL_DLLEXPORT void jl_gc_set_max_memory(uint64_t max_mem);
 JL_DLLEXPORT void jl_gc_collect(jl_gc_collection_t collection);
 // Returns whether the thread with `tid` is a collector thread
 JL_DLLEXPORT int gc_is_collector_thread(int tid) JL_NOTSAFEPOINT;
+// Pinning objects; Returns whether the object has been pinned by this call.
+JL_DLLEXPORT unsigned char jl_gc_pin_object(void* obj);
 // Returns which GC implementation is being used and possibly its version according to the list of supported GCs
 // NB: it should clearly identify the GC by including e.g. ‘stock’ or ‘mmtk’ as a substring.
 JL_DLLEXPORT const char* jl_gc_active_impl(void);
 // Sweep Julia's stack pools and mtarray buffers. Note that this function has been added to the interface as
 // each GC should implement it but it will most likely not be used by other code in the runtime.
 // It still needs to be annotated with JL_DLLEXPORT since it is called from Rust by MMTk.
 JL_DLLEXPORT void jl_gc_sweep_stack_pools_and_mtarraylist_buffers(jl_ptls_t ptls) JL_NOTSAFEPOINT;
+// Notifies the GC that the given thread is about to yield for a GC. ctx is the ucontext for the thread
+// if it is already fetched by the caller, otherwise it is NULL.
+JL_DLLEXPORT void jl_gc_notify_thread_yield(jl_ptls_t ptls, void* ctx);
+
+// TODO: The preserve hook functions may be temporary. We should see the performance impact of the change.
+
+// Runtime hook for gc preserve begin. The GC needs to make sure that the preserved objects and its children stay alive and won't move.
+JL_DLLEXPORT void jl_gc_preserve_begin_hook(int n, ...) JL_NOTSAFEPOINT;
+// Runtime hook for gc preserve end. The GC needs to make sure that the preserved objects and its children stay alive and won't move.
+JL_DLLEXPORT void jl_gc_preserve_end_hook(void) JL_NOTSAFEPOINT;
+
 
 // ========================================================================= //
 // Metrics
@@ -207,10 +220,22 @@ JL_DLLEXPORT void *jl_gc_perm_alloc(size_t sz, int zero, unsigned align,
 //              the allocated object. All objects stored in fields of this object
 //              must be either permanently allocated or have other roots.
 struct _jl_value_t *jl_gc_permobj(size_t sz, void *ty) JL_NOTSAFEPOINT;
+// permanently allocates a symbol (jl_sym_t). The object needs to be word aligned,
+// and tagged with jl_sym_tag.
+// FIXME: Ideally we should merge this with jl_gc_permobj, as symbol is an object.
+// Currently there are a few differences between the two functions, and refactoring is needed.
+// 1. sz for this function includes the object header, and sz for jl_gc_permobj excludes the header size.
+// 2. align for this function is word align, and align for jl_gc_permobj depends on the allocation size.
+// 3. ty for this function is jl_symbol_tag << 4, and ty for jl_gc_permobj is a datatype pointer.
+struct _jl_value_t *jl_gc_permsymbol(size_t sz) JL_NOTSAFEPOINT;
 // This function notifies the GC about memory addresses that are set when loading the boot image.
 // The GC may use that information to, for instance, determine that such objects should
 // be treated as marked and belonged to the old generation in nursery collections.
 void jl_gc_notify_image_load(const char* img_data, size_t len);
+// This function notifies the GC about memory addresses that are set when allocating the boot image.
+// The GC may use that information to, for instance, determine that all objects in that chunk of memory should
+// be treated as marked and belonged to the old generation in nursery collections.
+void jl_gc_notify_image_alloc(const char* img_data, size_t len);
 
 // ========================================================================= //
 // Runtime Write-Barriers

src/gc-mmtk.c

@@ -323,6 +323,24 @@ JL_DLLEXPORT void jl_gc_prepare_to_collect(void)
     errno = last_errno;
 }
 
+JL_DLLEXPORT unsigned char jl_gc_pin_object(void* obj) {
+    return mmtk_pin_object(obj);
+}
+
+JL_DLLEXPORT void jl_gc_notify_thread_yield(jl_ptls_t ptls, void* ctx) {
+    if (ctx == NULL) {
+        // Save the context for the thread as it was running at the time of the call
+        int r = getcontext(&ptls->gc_tls.ctx_at_the_time_gc_started);
+        if (r == -1) {
+            jl_safe_printf("Failed to save context for conservative scanning\n");
+            abort();
+        }
+        return;
+    }
+    memcpy(&ptls->gc_tls.ctx_at_the_time_gc_started, ctx, sizeof(ucontext_t));
+}
+
+
 // ========================================================================= //
 // GC Statistics
 // ========================================================================= //
@@ -847,24 +865,25 @@ STATIC_INLINE void* mmtk_immix_alloc_fast(MMTkMutatorContext* mutator, size_t si
     return bump_alloc_fast(mutator, (uintptr_t*)&allocator->cursor, (intptr_t)allocator->limit, size, align, offset, 0);
 }
 
-inline void mmtk_immix_post_alloc_slow(MMTkMutatorContext* mutator, void* obj, size_t size) {
-    mmtk_post_alloc(mutator, obj, size, 0);
-}
-
 STATIC_INLINE void mmtk_immix_post_alloc_fast(MMTkMutatorContext* mutator, void* obj, size_t size) {
-    // FIXME: for now, we do nothing
-    // but when supporting moving, this is where we set the valid object (VO) bit
+    if (MMTK_NEEDS_VO_BIT) {
+        mmtk_set_side_metadata(MMTK_SIDE_VO_BIT_BASE_ADDRESS, obj);
+    }
 }
 
 STATIC_INLINE void* mmtk_immortal_alloc_fast(MMTkMutatorContext* mutator, size_t size, size_t align, size_t offset) {
     BumpAllocator* allocator = &mutator->allocators.bump_pointer[MMTK_IMMORTAL_BUMP_ALLOCATOR];
     return bump_alloc_fast(mutator, (uintptr_t*)&allocator->cursor, (uintptr_t)allocator->limit, size, align, offset, 1);
 }
 
-STATIC_INLINE void mmtk_immortal_post_alloc_fast(MMTkMutatorContext* mutator, void* obj, size_t size) {
-    // FIXME: Similarly, for now, we do nothing
-    // but when supporting moving, this is where we set the valid object (VO) bit
-    // and log (old gen) bit
+STATIC_INLINE  void mmtk_immortal_post_alloc_fast(MMTkMutatorContext* mutator, void* obj, size_t size) {
+    if (MMTK_NEEDS_WRITE_BARRIER == MMTK_OBJECT_BARRIER) {
+        mmtk_set_side_metadata(MMTK_SIDE_LOG_BIT_BASE_ADDRESS, obj);
+    }
+
+    if (MMTK_NEEDS_VO_BIT) {
+        mmtk_set_side_metadata(MMTK_SIDE_VO_BIT_BASE_ADDRESS, obj);
+    }
 }
 
 JL_DLLEXPORT jl_value_t *jl_mmtk_gc_alloc_default(jl_ptls_t ptls, int osize, size_t align, void *ty)
@@ -1042,6 +1061,16 @@ jl_value_t *jl_gc_permobj(size_t sz, void *ty) JL_NOTSAFEPOINT
     return jl_valueof(o);
 }
 
+jl_value_t *jl_gc_permsymbol(size_t sz) JL_NOTSAFEPOINT
+{
+    jl_taggedvalue_t *tag = (jl_taggedvalue_t*)jl_gc_perm_alloc(sz, 0, sizeof(void*), 0);
+    jl_value_t *sym = jl_valueof(tag);
+    jl_ptls_t ptls = jl_current_task->ptls;
+    jl_set_typetagof(sym, jl_symbol_tag, 0);    // We need to set symbol tag. The GC tag doesnt matter.
+    mmtk_immortal_post_alloc_fast(&ptls->gc_tls.mmtk_mutator, sym, sz);
+    return sym;
+}
+
 JL_DLLEXPORT void *jl_gc_managed_malloc(size_t sz)
 {
     jl_ptls_t ptls = jl_current_task->ptls;
@@ -1079,6 +1108,47 @@ void jl_gc_notify_image_load(const char* img_data, size_t len)
     mmtk_set_vm_space((void*)img_data, len);
 }
 
+void jl_gc_notify_image_alloc(const char* img_data, size_t len)
+{
+    mmtk_immortal_region_post_alloc((void*)img_data, len);
+}
+
+// ========================================================================= //
+// Write Barriers
+// ========================================================================= //
+
+extern const void* MMTK_SIDE_LOG_BIT_BASE_ADDRESS;
+extern const void* MMTK_SIDE_VO_BIT_BASE_ADDRESS;
+
+// Directly call into MMTk for write barrier (debugging only)
+STATIC_INLINE void mmtk_gc_wb_full(const void *parent, const void *ptr) JL_NOTSAFEPOINT
+{
+    jl_task_t *ct = jl_current_task;
+    jl_ptls_t ptls = ct->ptls;
+    mmtk_object_reference_write_post(&ptls->gc_tls.mmtk_mutator, parent, ptr);
+}
+
+// Inlined fastpath
+STATIC_INLINE void mmtk_gc_wb_fast(const void *parent, const void *ptr) JL_NOTSAFEPOINT
+{
+    if (MMTK_NEEDS_WRITE_BARRIER == MMTK_OBJECT_BARRIER) {
+        intptr_t addr = (intptr_t) (void*) parent;
+        uint8_t* meta_addr = (uint8_t*) (MMTK_SIDE_LOG_BIT_BASE_ADDRESS) + (addr >> 6);
+        intptr_t shift = (addr >> 3) & 0b111;
+        uint8_t byte_val = *meta_addr;
+        if (((byte_val >> shift) & 1) == 1) {
+            jl_task_t *ct = jl_current_task;
+            jl_ptls_t ptls = ct->ptls;
+            mmtk_object_reference_write_slow(&ptls->gc_tls.mmtk_mutator, parent, ptr);
+        }
+    }
+}
+
+STATIC_INLINE void mmtk_gc_wb(const void *parent, const void *ptr) JL_NOTSAFEPOINT
+{
+    mmtk_gc_wb_fast(parent, ptr);
+}
+
 // ========================================================================= //
 // Code specific to stock that is not supported by MMTk
 // ========================================================================= //
@@ -1208,6 +1278,51 @@ JL_DLLEXPORT jl_value_t *jl_gc_internal_obj_base_ptr(void *p)
     return NULL;
 }
 
+// This macro currently uses malloc instead of alloca because this function will exit
+// after pushing the roots into the gc_preserve_stack, which means that the preserve_begin function's
+// stack frame will be destroyed (together with its alloca variables). When we support lowering this code
+// inside the same function that is doing the preserve_begin/preserve_end calls we should be able to simple use allocas.
+// Note also that we use a separate stack for gc preserve roots to avoid the possibility of calling free
+// on a stack that has been allocated with alloca instead of malloc, which could happen depending on the order in which
+// JL_GC_POP() and jl_gc_preserve_end_hook() occurs.
+
+#define JL_GC_PUSHARGS_PRESERVE_ROOT_OBJS(rts_var,n)                                                    \
+  rts_var = ((jl_value_t**)malloc(((n)+2)*sizeof(jl_value_t*)))+2;                                      \
+  ((void**)rts_var)[-2] = (void*)JL_GC_ENCODE_PUSHARGS(n);                                              \
+  ((void**)rts_var)[-1] = jl_p_gcpreserve_stack;                                                        \
+  memset((void*)rts_var, 0, (n)*sizeof(jl_value_t*));                                                   \
+  jl_p_gcpreserve_stack = (jl_gcframe_t*)&(((void**)rts_var)[-2]);                                      \
+
+#define JL_GC_POP_PRESERVE_ROOT_OBJS()                                                                  \
+    jl_gcframe_t *curr = jl_p_gcpreserve_stack;                                                         \
+    if(curr) {                                                                                          \
+        (jl_p_gcpreserve_stack = jl_p_gcpreserve_stack->prev);                                          \
+        free(curr);                                                                                     \
+    }
+
+// Add each argument as a tpin root object.
+// However, we cannot use JL_GC_PUSH and JL_GC_POP since the slots should live
+// beyond this function. Instead, we maintain a tpin stack by mallocing/freeing
+// the frames for each of the preserve regions we encounter
+JL_DLLEXPORT void jl_gc_preserve_begin_hook(int n, ...) JL_NOTSAFEPOINT
+{
+    jl_value_t** frame;
+    JL_GC_PUSHARGS_PRESERVE_ROOT_OBJS(frame, n);
+    if (n == 0) return;
+
+    va_list args;
+    va_start(args, n);
+    for (int i = 0; i < n; i++) {
+        frame[i] = va_arg(args, jl_value_t *);
+    }
+    va_end(args);
+}
+
+JL_DLLEXPORT void jl_gc_preserve_end_hook(void) JL_NOTSAFEPOINT
+{
+    JL_GC_POP_PRESERVE_ROOT_OBJS();
+}
+
 #ifdef __cplusplus
 }
 #endif

src/gc-stock.c

@@ -3458,6 +3458,11 @@ JL_DLLEXPORT void jl_gc_collect(jl_gc_collection_t collection)
         gc_cblist_pre_gc, (collection));
 
     if (!jl_atomic_load_acquire(&jl_gc_disable_counter)) {
+        // This thread will yield.
+        // jl_gc_notify_thread_yield does nothing for the stock GC at the point, but it may be non empty in the future,
+        // and this is a place where we should call jl_gc_notify_thread_yield.
+        // TODO: This call can be removed if requested.
+        jl_gc_notify_thread_yield(ptls, NULL);
         JL_LOCK_NOGC(&finalizers_lock); // all the other threads are stopped, so this does not make sense, right? otherwise, failing that, this seems like plausibly a deadlock
 #ifndef __clang_gcanalyzer__
         if (_jl_gc_collect(ptls, collection)) {
@@ -3940,6 +3945,15 @@ jl_value_t *jl_gc_permobj(size_t sz, void *ty) JL_NOTSAFEPOINT
     return jl_valueof(o);
 }
 
+jl_value_t *jl_gc_permsymbol(size_t sz) JL_NOTSAFEPOINT
+{
+    jl_taggedvalue_t *tag = (jl_taggedvalue_t*)jl_gc_perm_alloc(sz, 0, sizeof(void*), 0);
+    jl_value_t *sym = jl_valueof(tag);
+    // set to old marked so that we won't look at it in the GC or write barrier.
+    jl_set_typetagof(sym, jl_symbol_tag, GC_OLD_MARKED);
+    return sym;
+}
+
 JL_DLLEXPORT int jl_gc_enable_conservative_gc_support(void)
 {
     if (jl_is_initialized()) {
@@ -4076,6 +4090,47 @@ void jl_gc_notify_image_load(const char* img_data, size_t len)
     // Do nothing
 }
 
+void jl_gc_notify_image_alloc(const char* img_data, size_t len)
+{
+    // Do nothing
+}
+
+JL_DLLEXPORT unsigned char jl_gc_pin_object(void* obj) {
+    return 0;
+}
+
+// added for MMTk integration
+
+JL_DLLEXPORT void jl_gc_wb1_noinline(const void *parent) JL_NOTSAFEPOINT
+{
+}
+
+JL_DLLEXPORT void jl_gc_wb2_noinline(const void *parent, const void *ptr) JL_NOTSAFEPOINT
+{
+}
+
+JL_DLLEXPORT void jl_gc_wb1_slow(const void *parent) JL_NOTSAFEPOINT
+{
+}
+
+JL_DLLEXPORT void jl_gc_wb2_slow(const void *parent, const void* ptr) JL_NOTSAFEPOINT
+{
+}
+
+JL_DLLEXPORT void jl_gc_preserve_begin_hook(int n, ...) JL_NOTSAFEPOINT
+{
+    jl_unreachable();
+}
+
+JL_DLLEXPORT void jl_gc_preserve_end_hook(void) JL_NOTSAFEPOINT
+{
+    jl_unreachable();
+}
+
+JL_DLLEXPORT void jl_gc_notify_thread_yield(jl_ptls_t ptls, void* ctx) {
+    // Do nothing before a thread yields
+}
+
 JL_DLLEXPORT const char* jl_gc_active_impl(void) {
     return "Built with stock GC";
 }

src/gc-tls-mmtk.h

@@ -14,6 +14,7 @@ extern "C" {
 typedef struct {
     MMTkMutatorContext mmtk_mutator;
     _Atomic(size_t) malloc_sz_since_last_poll;
+    ucontext_t ctx_at_the_time_gc_started;
 } jl_gc_tls_states_t;
 
 #ifdef __cplusplus

src/genericmemory.c

@@ -111,6 +111,7 @@ JL_DLLEXPORT jl_genericmemory_t *jl_string_to_genericmemory(jl_value_t *str)
     m->length = jl_string_len(str);
     m->ptr = jl_string_data(str);
     jl_genericmemory_data_owner_field(m) = str;
+    PTR_PIN(str);
     return m;
 }
 
@@ -166,6 +167,7 @@ JL_DLLEXPORT jl_genericmemory_t *jl_ptr_to_genericmemory(jl_value_t *mtype, void
     m->length = nel;
     jl_genericmemory_data_owner_field(m) = own_buffer ? (jl_value_t*)m : NULL;
     if (own_buffer) {
+        PTR_PIN(m);
         int isaligned = 0;  // TODO: allow passing memalign'd buffers
         jl_gc_track_malloced_genericmemory(ct->ptls, m, isaligned);
         size_t allocated_bytes = memory_block_usable_size(data, isaligned);
@@ -235,6 +237,10 @@ JL_DLLEXPORT void jl_genericmemory_copyto(jl_genericmemory_t *dest, char* destda
         _Atomic(void*) * dest_p = (_Atomic(void*)*)destdata;
         _Atomic(void*) * src_p = (_Atomic(void*)*)srcdata;
         jl_value_t *owner = jl_genericmemory_owner(dest);
+        // FIXME: The following should be a write barrier impl provided by the GC.
+#ifdef MMTK_GC
+        jl_gc_wb(owner, NULL);
+#else
         if (__unlikely(jl_astaggedvalue(owner)->bits.gc == GC_OLD_MARKED)) {
             jl_value_t *src_owner = jl_genericmemory_owner(src);
             ssize_t done = 0;
@@ -265,6 +271,7 @@ JL_DLLEXPORT void jl_genericmemory_copyto(jl_genericmemory_t *dest, char* destda
                 n -= done;
             }
         }
+#endif
         return memmove_refs(dest_p, src_p, n);
     }
     size_t elsz = layout->size;

src/interpreter.c

@@ -891,6 +891,7 @@ jl_value_t *NOINLINE jl_interpret_toplevel_thunk(jl_module_t *m, jl_code_info_t
     unsigned nroots = jl_source_nslots(src) + jl_source_nssavalues(src);
     JL_GC_PUSHFRAME(s, s->locals, nroots);
     jl_array_t *stmts = src->code;
+    JL_GC_PUSH1(&stmts);
     assert(jl_typetagis(stmts, jl_array_any_type));
     s->src = src;
     s->module = m;
@@ -901,6 +902,7 @@ jl_value_t *NOINLINE jl_interpret_toplevel_thunk(jl_module_t *m, jl_code_info_t
     JL_GC_ENABLEFRAME(s);
     jl_value_t *r = eval_body(stmts, s, 0, 1);
     JL_GC_POP();
+    JL_GC_POP();
     return r;
 }