GC support for stack allocated objects

Only works with copy_stacks for now. The code is not used anywhere in
codegen right now, it can be tested by using (:stknew typ args...)
instead of :new. It is essentially the GC side of #8134 but without the
global alloca stack & unmarking.

Stack objects are kept marked (+young) permanently and are only scanned
when coming from the owning task.

Author: carnaval

base/inference.jl

@@ -961,7 +961,7 @@ function abstract_eval(e::ANY, vtypes, sv::StaticVarInfo)
         t = abstract_eval_call(e, vtypes, sv)
     elseif is(e.head,:null)
         t = Void
-    elseif is(e.head,:new)
+    elseif is(e.head,:new) || is(e.head,:stknew)
         t = abstract_eval(e.args[1], vtypes, sv)
         if isType(t)
             t = t.parameters[1]

src/alloc.c

@@ -99,7 +99,7 @@ jl_sym_t *compositetype_sym; jl_sym_t *type_goto_sym;
 jl_sym_t *global_sym; jl_sym_t *tuple_sym;
 jl_sym_t *dot_sym;    jl_sym_t *newvar_sym;
 jl_sym_t *boundscheck_sym; jl_sym_t *copyast_sym;
-jl_sym_t *fastmath_sym;
+jl_sym_t *fastmath_sym; jl_sym_t *stknew_sym;
 jl_sym_t *simdloop_sym; jl_sym_t *meta_sym;
 jl_sym_t *arrow_sym;  jl_sym_t *inert_sym;
 jl_sym_t *vararg_sym;

src/cgutils.cpp

@@ -1790,7 +1790,7 @@ static void emit_write_barrier(jl_codectx_t* ctx, Value *parent, Value *ptr)
 {
     Value* parenttag = builder.CreateBitCast(emit_typeptr_addr(parent), T_psize);
     Value* parent_type = builder.CreateLoad(parenttag);
-    Value* parent_mark_bits = builder.CreateAnd(parent_type, 1);
+    Value* parent_mark_bits = builder.CreateAnd(parent_type, 3);
 
     // the branch hint does not seem to make it to the generated code
     //builder.CreateCall(expect_func, {parent_marked, ConstantInt::get(T_int1, 0)});
@@ -1857,7 +1857,7 @@ static void emit_setfield(jl_datatype_t *sty, const jl_cgval_t &strct, size_t id
     }
 }
 
-static jl_cgval_t emit_new_struct(jl_value_t *ty, size_t nargs, jl_value_t **args, jl_codectx_t *ctx)
+static jl_cgval_t emit_new_struct(jl_value_t *ty, size_t nargs, jl_value_t **args, jl_codectx_t *ctx, bool on_stack = false)
 {
     assert(jl_is_datatype(ty));
     assert(jl_is_leaf_type(ty));
@@ -1913,10 +1913,18 @@ static jl_cgval_t emit_new_struct(jl_value_t *ty, size_t nargs, jl_value_t **arg
             if (might_need_root(args[1]) || !fval_info.isboxed)
                 make_gcroot(f1, ctx);
         }
-        Value *strct = emit_allocobj(sty->size);
+        Value *strct;
+        Value *tag = literal_pointer_val((jl_value_t*)ty);
+        if (on_stack) {
+            strct = builder.CreateConstGEP1_32(builder.CreateBitCast(builder.CreateAlloca(T_int8, ConstantInt::get(T_size, sty->size + sizeof(void*))), jl_pvalue_llvmt), 1);
+            // set the gc bits as marked & young
+            tag = builder.CreateIntToPtr(builder.CreateOr(builder.CreatePtrToInt(tag, T_size), 3), jl_pvalue_llvmt);
+        }
+        else {
+            strct = emit_allocobj(sty->size);
+        }
         jl_cgval_t strctinfo = mark_julia_type(strct, ty);
-        builder.CreateStore(literal_pointer_val((jl_value_t*)ty),
-                            emit_typeptr_addr(strct));
+        builder.CreateStore(tag, emit_typeptr_addr(strct));
         if (f1) {
             jl_cgval_t f1info = mark_julia_type(f1, jl_any_type);
             if (!jl_subtype(expr_type(args[1],ctx), jl_field_type(sty,0), 0))

src/codegen.cpp

@@ -3443,13 +3443,13 @@ static jl_cgval_t emit_expr(jl_value_t *expr, jl_codectx_t *ctx, bool isboxed, b
             jl_add_linfo_root(ctx->linfo, extype);
         return mark_julia_const(extype);
     }
-    else if (head == new_sym) {
+    else if (head == new_sym || head == stknew_sym) {
         jl_value_t *ty = expr_type(args[0], ctx);
         size_t nargs = jl_array_len(ex->args);
         if (jl_is_type_type(ty) &&
             jl_is_datatype(jl_tparam0(ty)) &&
             jl_is_leaf_type(jl_tparam0(ty))) {
-            return emit_new_struct(jl_tparam0(ty),nargs,args,ctx);
+            return emit_new_struct(jl_tparam0(ty),nargs,args,ctx, head == stknew_sym);
         }
         Value *typ = boxed(emit_expr(args[0], ctx), ctx);
         Value *val = emit_jlcall(jlnew_func, typ, &args[1], nargs-1, ctx);

src/gc.c

@@ -516,67 +516,66 @@ static inline void objprofile_count(void* ty, int old, int sz)
 
 #define inc_sat(v,s) v = (v) >= s ? s : (v)+1
 
-static inline int gc_setmark_big(void *o, int mark_mode)
+static inline int gc_setmark_big(void *o, int bits)
 {
+    int mark_mode = GC_MARKED_NOESC;
 #ifdef GC_VERIFY
     if (verifying) {
         _gc_setmark(o, mark_mode);
         return 0;
     }
 #endif
     bigval_t* hdr = bigval_header(o);
-    int bits = gc_bits(o);
-    if (bits == GC_QUEUED || bits == GC_MARKED)
-        mark_mode = GC_MARKED;
-    if ((mark_mode == GC_MARKED) & (bits != GC_MARKED)) {
-        // Move hdr from big_objects list to big_objects_marked list
-        *hdr->prev = hdr->next;
-        if (hdr->next)
-            hdr->next->prev = hdr->prev;
-        hdr->next = big_objects_marked;
-        hdr->prev = &big_objects_marked;
-        if (big_objects_marked)
-            big_objects_marked->prev = &hdr->next;
-        big_objects_marked = hdr;
-    }
     if (!(bits & GC_MARKED)) {
+        if (bits == GC_QUEUED)
+            mark_mode = GC_MARKED;
+        if ((mark_mode == GC_MARKED) & (bits != GC_MARKED)) {
+            // Move hdr from big_objects list to big_objects_marked list
+            *hdr->prev = hdr->next;
+            if (hdr->next)
+                hdr->next->prev = hdr->prev;
+            hdr->next = big_objects_marked;
+            hdr->prev = &big_objects_marked;
+            if (big_objects_marked)
+                big_objects_marked->prev = &hdr->next;
+            big_objects_marked = hdr;
+        }
         if (mark_mode == GC_MARKED)
             perm_scanned_bytes += hdr->sz&~3;
         else
             scanned_bytes += hdr->sz&~3;
 #ifdef OBJPROFILE
         objprofile_count(jl_typeof(o), mark_mode == GC_MARKED, hdr->sz&~3);
 #endif
+        _gc_setmark(o, mark_mode);
     }
-    _gc_setmark(o, mark_mode);
     verify_val(jl_valueof(o));
     return mark_mode;
 }
 
-static inline int gc_setmark_pool(void *o, int mark_mode)
+static inline int gc_setmark_pool(void *o, int bits)
 {
+    int mark_mode = GC_MARKED_NOESC;
 #ifdef GC_VERIFY
     if (verifying) {
         _gc_setmark(o, mark_mode);
         return mark_mode;
     }
 #endif
-    gcpage_t* page = page_metadata(o);
-    int bits = gc_bits(o);
-    if (bits == GC_QUEUED || bits == GC_MARKED) {
-        mark_mode = GC_MARKED;
-    }
     if (!(bits & GC_MARKED)) {
+        if (bits == GC_QUEUED)
+            mark_mode = GC_MARKED;
+        gcpage_t* page = page_metadata(o);
         if (mark_mode == GC_MARKED)
             perm_scanned_bytes += page->osize;
         else
             scanned_bytes += page->osize;
 #ifdef OBJPROFILE
         objprofile_count(jl_typeof(o), mark_mode == GC_MARKED, page->osize);
 #endif
+        _gc_setmark(o, mark_mode);
+        page->gc_bits |= mark_mode;
     }
-    _gc_setmark(o, mark_mode);
-    page->gc_bits |= mark_mode;
     verify_val(jl_valueof(o));
     return mark_mode;
 }
@@ -601,14 +600,19 @@ static inline int gc_setmark(jl_value_t *v, int sz, int mark_mode)
 inline void gc_setmark_buf(void *o, int mark_mode)
 {
     buff_t *buf = gc_val_buf(o);
+    int bits = gc_bits(buf);
+    if (mark_mode == GC_MARKED && bits != GC_MARKED) {
+        _gc_setmark(buf, GC_QUEUED);
+        bits = GC_QUEUED;
+    }
 #ifdef MEMDEBUG
-    gc_setmark_big(buf, mark_mode);
+    gc_setmark_big(buf, bits);
     return;
 #endif
     if (buf->pooled)
-        gc_setmark_pool(buf, mark_mode);
+        gc_setmark_pool(buf, bits);
     else
-        gc_setmark_big(buf, mark_mode);
+        gc_setmark_big(buf, bits);
 }
 
 static NOINLINE void *malloc_page(void)
@@ -1450,16 +1454,20 @@ void jl_gc_setmark(jl_value_t *v) // TODO rename this as it is misleading now
     if (!gc_marked(o)) {
         //        objprofile_count(jl_typeof(v), 1, 16);
 #ifdef MEMDEBUG
-        gc_setmark_big(o, GC_MARKED_NOESC);
+        gc_setmark_big(o, gc_bits(o));
 #else
-        gc_setmark_pool(o, GC_MARKED_NOESC);
+        gc_setmark_pool(o, gc_bits(o));
 #endif
     }
     //    perm_scanned_bytes = s;
 }
-
+static char* gc_stack_top;
+static char* gc_stack_bot;
 static void gc_mark_stack(jl_value_t* ta, jl_gcframe_t *s, ptrint_t offset, int d)
 {
+    jl_task_t *task = (jl_task_t*)ta;
+    char *stkbuf = task == jl_current_task ? gc_stack_top : task->stkbuf;
+    char *stkend = task == jl_current_task ? gc_stack_bot : stkbuf + task->ssize;
     while (s != NULL) {
         s = (jl_gcframe_t*)((char*)s + offset);
         jl_value_t ***rts = (jl_value_t***)(((void**)s)+2);
@@ -1473,9 +1481,17 @@ static void gc_mark_stack(jl_value_t* ta, jl_gcframe_t *s, ptrint_t offset, int
         }
         else {
             for(size_t i=0; i < nr; i++) {
-                if (rts[i] != NULL) {
+                void *v = rts[i];
+                if (v != NULL) {
                     verify_parent2("task", ta, &rts[i], "stack(%d)", (int)i);
-                    gc_push_root(rts[i], d);
+                    if (stkbuf <= (char*)v && (char*)v <= stkend) {
+                        // if v is on the stack it is kept permanently marked
+                        // but we still need to scan it once
+                        push_root(v, d, gc_bits(jl_astaggedvalue(v)));
+                    }
+                    else {
+                        gc_push_root(v, d);
+                    }
                 }
             }
         }
@@ -1576,12 +1592,12 @@ static int push_root(jl_value_t *v, int d, int bits)
     int refyoung = 0;
 
     if (vt == (jl_value_t*)jl_weakref_type) {
-        bits = gc_setmark(v, sizeof(jl_weakref_t), GC_MARKED_NOESC);
+        bits = gc_setmark(v, sizeof(jl_weakref_t), bits);
         goto ret;
     }
     if ((jl_is_datatype(vt) && ((jl_datatype_t*)vt)->pointerfree)) {
         int sz = jl_datatype_size(vt);
-        bits = gc_setmark(v, sz, GC_MARKED_NOESC);
+        bits = gc_setmark(v, sz, bits);
         goto ret;
     }
 #define MARK(v, s) do {                         \
@@ -1597,7 +1613,7 @@ static int push_root(jl_value_t *v, int d, int bits)
     // some values have special representations
     if (vt == (jl_value_t*)jl_simplevector_type) {
         size_t l = jl_svec_len(v);
-        MARK(v, bits = gc_setmark(v, l*sizeof(void*) + sizeof(jl_svec_t), GC_MARKED_NOESC));
+        MARK(v, bits = gc_setmark(v, l*sizeof(void*) + sizeof(jl_svec_t), bits));
         jl_value_t **data = ((jl_svec_t*)v)->data;
         for(size_t i=0; i < l; i++) {
             jl_value_t *elt = data[i];
@@ -1618,7 +1634,7 @@ static int push_root(jl_value_t *v, int d, int bits)
 #define _gc_setmark_pool gc_setmark_pool
 #endif
             MARK(a,
-                 bits = _gc_setmark_pool(o, GC_MARKED_NOESC);
+                 bits = _gc_setmark_pool(o, bits);
                  if (a->how == 2 && todo) {
                      objprofile_count(MATY, gc_bits(o) == GC_MARKED, array_nbytes(a));
                      if (gc_bits(o) == GC_MARKED)
@@ -1628,7 +1644,7 @@ static int push_root(jl_value_t *v, int d, int bits)
                  });
         else
             MARK(a,
-                 bits = gc_setmark_big(o, GC_MARKED_NOESC);
+                 bits = gc_setmark_big(o, bits);
                  if (a->how == 2 && todo) {
                      objprofile_count(MATY, gc_bits(o) == GC_MARKED, array_nbytes(a));
                      if (gc_bits(o) == GC_MARKED)
@@ -1670,11 +1686,11 @@ static int push_root(jl_value_t *v, int d, int bits)
         }
     }
     else if (vt == (jl_value_t*)jl_module_type) {
-        MARK(v, bits = gc_setmark(v, sizeof(jl_module_t), GC_MARKED_NOESC));
+        MARK(v, bits = gc_setmark(v, sizeof(jl_module_t), bits));
         refyoung |= gc_mark_module((jl_module_t*)v, d);
     }
     else if (vt == (jl_value_t*)jl_task_type) {
-        MARK(v, bits = gc_setmark(v, sizeof(jl_task_t), GC_MARKED_NOESC));
+        MARK(v, bits = gc_setmark(v, sizeof(jl_task_t), bits));
         gc_mark_task((jl_task_t*)v, d);
         // tasks should always be remarked since we do not trigger the write barrier
         // for stores to stack slots
@@ -1697,7 +1713,7 @@ static int push_root(jl_value_t *v, int d, int bits)
             dtsz = NWORDS(sizeof(jl_datatype_t) + jl_datatype_nfields(v)*sizeof(jl_fielddesc_t))*sizeof(void*);
         else
             dtsz = jl_datatype_size(dt);
-        MARK(v, bits = gc_setmark(v, dtsz, GC_MARKED_NOESC));
+        MARK(v, bits = gc_setmark(v, dtsz, bits));
         int nf = (int)jl_datatype_nfields(dt);
         // TODO check if there is a perf improvement for objects with a lot of fields
         // int fdsz = sizeof(void*)*nf;
@@ -1956,6 +1972,7 @@ void jl_gc_collect(int full)
     JL_SIGATOMIC_BEGIN();
     jl_in_gc = 1;
     uint64_t t0 = jl_hrtime();
+    gc_stack_top = (char*)&t0;
     int recollect = 0;
 #if defined(GC_TIME)
     int wb_activations = mark_sp - saved_mark_sp;
@@ -2366,6 +2383,8 @@ void jl_gc_init(void)
     if (maxmem > max_collect_interval)
         max_collect_interval = maxmem;
 #endif
+    char _dummy;
+    gc_stack_bot = &_dummy;
 }
 
 // GC summary stats

src/interpreter.c

@@ -244,7 +244,7 @@ static jl_value_t *eval(jl_value_t *e, jl_value_t **locals, size_t nl, size_t ng
         JL_GC_POP();
         return rhs;
     }
-    else if (ex->head == new_sym) {
+    else if (ex->head == new_sym || ex->head == stknew_sym) {
         jl_value_t *thetype = eval(args[0], locals, nl, ngensym);
         jl_value_t *v=NULL;
         JL_GC_PUSH2(&thetype, &v);

src/jltypes.c

@@ -3534,6 +3534,7 @@ void jl_init_types(void)
     dot_sym = jl_symbol(".");
     boundscheck_sym = jl_symbol("boundscheck");
     fastmath_sym = jl_symbol("fastmath");
+    stknew_sym = jl_symbol("stknew");
     newvar_sym = jl_symbol("newvar");
     copyast_sym = jl_symbol("copyast");
     simdloop_sym = jl_symbol("simdloop");

src/julia.h

@@ -486,7 +486,7 @@ extern jl_sym_t *abstracttype_sym; extern jl_sym_t *bitstype_sym;
 extern jl_sym_t *compositetype_sym; extern jl_sym_t *type_goto_sym;
 extern jl_sym_t *global_sym;  extern jl_sym_t *tuple_sym;
 extern jl_sym_t *boundscheck_sym; extern jl_sym_t *copyast_sym;
-extern jl_sym_t *fastmath_sym;
+extern jl_sym_t *fastmath_sym; extern jl_sym_t *stknew_sym;
 extern jl_sym_t *simdloop_sym; extern jl_sym_t *meta_sym;
 extern jl_sym_t *arrow_sym; extern jl_sym_t *inert_sym;
 
@@ -580,13 +580,13 @@ void gc_setmark_buf(void *buf, int);
 
 static inline void jl_gc_wb_binding(jl_binding_t *bnd, void *val) // val isa jl_value_t*
 {
-    if (__unlikely((*((uintptr_t*)bnd-1) & 1) == 1 && (*(uintptr_t*)jl_astaggedvalue(val) & 1) == 0))
+    if (__unlikely((*((uintptr_t*)bnd-1) & 3) == 1 && (*(uintptr_t*)jl_astaggedvalue(val) & 1) == 0))
         gc_queue_binding(bnd);
 }
 
 static inline void jl_gc_wb(void *parent, void *ptr) // parent and ptr isa jl_value_t*
 {
-    if (__unlikely((*((uintptr_t*)jl_astaggedvalue(parent)) & 1) == 1 &&
+    if (__unlikely((*((uintptr_t*)jl_astaggedvalue(parent)) & 3) == 1 &&
                    (*((uintptr_t*)jl_astaggedvalue(ptr)) & 1) == 0))
         jl_gc_queue_root((jl_value_t*)parent);
 }
@@ -602,7 +602,7 @@ static inline void jl_gc_wb_buf(void *parent, void *bufptr) // parent isa jl_val
 static inline void jl_gc_wb_back(void *ptr) // ptr isa jl_value_t*
 {
     // if ptr is marked
-    if(__unlikely((*((uintptr_t*)jl_astaggedvalue(ptr)) & 1) == 1)) {
+    if(__unlikely((*((uintptr_t*)jl_astaggedvalue(ptr)) & 3) == 1)) {
         jl_gc_queue_root((jl_value_t*)ptr);
     }
 }