fix #38224, set and use DataType.zeroinit in arrays (#38444)

Author: JeffBezanson

src/array.c

@@ -75,7 +75,7 @@ size_t jl_arr_xtralloc_limit = 0;
 #define MAXINTVAL (((size_t)-1)>>1)
 
 static jl_array_t *_new_array_(jl_value_t *atype, uint32_t ndims, size_t *dims,
-                               int isunboxed, int hasptr, int isunion, int elsz)
+                               int8_t isunboxed, int8_t hasptr, int8_t isunion, int8_t zeroinit, int elsz)
 {
     jl_ptls_t ptls = jl_get_ptls_states();
     size_t i, tot, nel=1;
@@ -123,8 +123,6 @@ static jl_array_t *_new_array_(jl_value_t *atype, uint32_t ndims, size_t *dims,
         // No allocation or safepoint allowed after this
         a->flags.how = 0;
         data = (char*)a + doffs;
-        if (tot > 0 && (!isunboxed || hasptr || isunion)) // TODO: check for zeroinit
-            memset(data, 0, tot);
     }
     else {
         tsz = JL_ARRAY_ALIGN(tsz, JL_CACHE_BYTE_ALIGNMENT); // align whole object
@@ -135,12 +133,11 @@ static jl_array_t *_new_array_(jl_value_t *atype, uint32_t ndims, size_t *dims,
         // No allocation or safepoint allowed after this
         a->flags.how = 2;
         jl_gc_track_malloced_array(ptls, a);
-        if (tot > 0 && (!isunboxed || hasptr || isunion)) // TODO: check for zeroinit
-            // need to zero out isbits union array selector bytes to ensure a valid type index
-            memset(data, 0, tot);
     }
     a->flags.pooled = tsz <= GC_MAX_SZCLASS;
 
+    if (zeroinit)
+        memset(data, 0, tot);
     a->data = data;
     if (JL_ARRAY_IMPL_NUL && elsz == 1)
         ((char*)data)[tot - 1] = '\0';
@@ -186,14 +183,15 @@ static inline jl_array_t *_new_array(jl_value_t *atype, uint32_t ndims, size_t *
     else {
         elsz = LLT_ALIGN(elsz, al);
     }
+    int zi = !isunboxed || hasptr || isunion || (jl_is_datatype(eltype) && ((jl_datatype_t*)eltype)->zeroinit);
 
-    return _new_array_(atype, ndims, dims, isunboxed, hasptr, isunion, elsz);
+    return _new_array_(atype, ndims, dims, isunboxed, hasptr, isunion, zi, elsz);
 }
 
 jl_array_t *jl_new_array_for_deserialization(jl_value_t *atype, uint32_t ndims, size_t *dims,
                                              int isunboxed, int hasptr, int isunion, int elsz)
 {
-    return _new_array_(atype, ndims, dims, isunboxed, hasptr, isunion, elsz);
+    return _new_array_(atype, ndims, dims, isunboxed, hasptr, isunion, 0, elsz);
 }
 
 #ifndef JL_NDEBUG
@@ -784,7 +782,7 @@ STATIC_INLINE void jl_array_grow_at_beg(jl_array_t *a, size_t idx, size_t inc,
     char *data = (char*)a->data;
     char *newdata;
     char *typetagdata;
-    char *newtypetagdata;
+    char *newtypetagdata = NULL;
     int isbitsunion = jl_array_isbitsunion(a);
     if (isbitsunion) typetagdata = jl_array_typetagdata(a);
     if (a->offset >= inc) {
@@ -864,10 +862,10 @@ STATIC_INLINE void jl_array_grow_at_beg(jl_array_t *a, size_t idx, size_t inc,
 #endif
     a->nrows = newnrows;
     a->data = newdata;
-    if (a->flags.ptrarray || a->flags.hasptr) { // TODO: check for zeroinit
+    if (jl_is_array_zeroinit(a)) {
         memset(newdata + idx * elsz, 0, nbinc);
     }
-    else if (isbitsunion) {
+    if (newtypetagdata) {
         memset(newtypetagdata + idx, 0, inc);
     }
 }
@@ -945,7 +943,7 @@ STATIC_INLINE void jl_array_grow_at_end(jl_array_t *a, size_t idx,
     a->length = newnrows;
 #endif
     a->nrows = newnrows;
-    if (a->flags.ptrarray || a->flags.hasptr) { // TODO: check for zeroinit
+    if (jl_is_array_zeroinit(a)) {
         memset(data + idx * elsz, 0, inc * elsz);
     }
 }
@@ -1201,7 +1199,7 @@ JL_DLLEXPORT jl_array_t *jl_array_copy(jl_array_t *ary)
     int isunion = jl_is_uniontype(jl_tparam0(jl_typeof(ary)));
     jl_array_t *new_ary = _new_array_(jl_typeof(ary), jl_array_ndims(ary),
                                       &ary->nrows, !ary->flags.ptrarray,
-                                      ary->flags.hasptr, isunion, elsz);
+                                      ary->flags.hasptr, isunion, 0, elsz);
     memcpy(new_ary->data, ary->data, len * elsz);
     // ensure isbits union arrays copy their selector bytes correctly
     if (jl_array_isbitsunion(ary))

src/datatype.c

@@ -544,6 +544,7 @@ void jl_compute_field_offsets(jl_datatype_t *st)
             if (npointers)
                 free(pointers);
         }
+        st->zeroinit = zeroinit;
     }
     // now finish deciding if this instantiation qualifies for special properties
     assert(!isbitstype || st->layout->npointers == 0); // the definition of isbits

src/dump.c

@@ -1346,7 +1346,7 @@ static jl_value_t *jl_deserialize_value_array(jl_serializer_state *s, uint8_t ta
         isptr = (elsize >> 15) & 1;
         hasptr = (elsize >> 14) & 1;
         isunion = (elsize >> 13) & 1;
-        elsize = elsize & 0x3fff;
+        elsize = elsize & 0x1fff;
     }
     uintptr_t pos = backref_list.len;
     arraylist_push(&backref_list, NULL);

src/ircode.c

@@ -418,7 +418,7 @@ static jl_value_t *jl_decode_value_array(jl_ircode_state *s, uint8_t tag) JL_GC_
         isptr = (elsize >> 15) & 1;
         hasptr = (elsize >> 14) & 1;
         isunion = (elsize >> 13) & 1;
-        elsize = elsize & 0x3fff;
+        elsize = elsize & 0x1fff;
     }
     size_t *dims = (size_t*)alloca(ndims * sizeof(size_t));
     for (i = 0; i < ndims; i++) {

src/julia.h

@@ -1250,6 +1250,14 @@ STATIC_INLINE int jl_is_type_type(jl_value_t *v) JL_NOTSAFEPOINT
             ((jl_datatype_t*)(v))->name == ((jl_datatype_t*)jl_type_type->body)->name);
 }
 
+STATIC_INLINE int jl_is_array_zeroinit(jl_array_t *a) JL_NOTSAFEPOINT
+{
+    if (a->flags.ptrarray || a->flags.hasptr)
+        return 1;
+    jl_value_t *elty = jl_tparam0(jl_typeof(a));
+    return jl_is_datatype(elty) && ((jl_datatype_t*)elty)->zeroinit;
+}
+
 // object identity
 JL_DLLEXPORT int jl_egal(jl_value_t *a JL_MAYBE_UNROOTED, jl_value_t *b JL_MAYBE_UNROOTED) JL_NOTSAFEPOINT;
 JL_DLLEXPORT uintptr_t jl_object_id(jl_value_t *v) JL_NOTSAFEPOINT;

test/core.jl

@@ -7508,3 +7508,14 @@ let array = Int[]
 end
 @test compare_union37557(Ref{Union{Int,Vector{Int}}}(1),
                          Ref{Union{Int,Vector{Int}}}(1))
+
+# issue #38224
+struct S38224
+    i::Union{Int,Missing}
+end
+@test S38224.zeroinit
+for _ in 1:5
+    let a = Vector{S38224}(undef, 1000000)
+        @test all(x->ismissing(x.i), a)
+    end
+end