fix excess array object padding

src/array.c

@@ -114,20 +114,23 @@ static jl_array_t *_new_array_(jl_value_t *atype, uint32_t ndims, size_t *dims,
     }
 
     int ndimwords = jl_array_ndimwords(ndims);
-    int tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords*sizeof(size_t), JL_CACHE_BYTE_ALIGNMENT);
+    int tsz = sizeof(jl_array_t) + ndimwords*sizeof(size_t);
     if (tot <= ARRAY_INLINE_NBYTES) {
-        if (isunboxed && elsz >= 4)
-            tsz = JL_ARRAY_ALIGN(tsz, JL_SMALL_BYTE_ALIGNMENT); // align data area
+        // align data area
+        if (tot >= ARRAY_CACHE_ALIGN_THRESHOLD)
+            tsz = JL_ARRAY_ALIGN(tsz, JL_CACHE_BYTE_ALIGNMENT);
+        else if (isunboxed && elsz >= 4)
+            tsz = JL_ARRAY_ALIGN(tsz, JL_SMALL_BYTE_ALIGNMENT);
         size_t doffs = tsz;
         tsz += tot;
-        tsz = JL_ARRAY_ALIGN(tsz, JL_SMALL_BYTE_ALIGNMENT); // align whole object
+        // jl_array_t is large enough that objects will always be aligned 16
         a = (jl_array_t*)jl_gc_alloc(ct->ptls, tsz, atype);
+        assert(((size_t)a & 15) == 0);
         // No allocation or safepoint allowed after this
         a->flags.how = 0;
         data = (char*)a + doffs;
     }
     else {
-        tsz = JL_ARRAY_ALIGN(tsz, JL_CACHE_BYTE_ALIGNMENT); // align whole object
         data = jl_gc_managed_malloc(tot);
         // Allocate the Array **after** allocating the data
         // to make sure the array is still young
@@ -223,7 +226,7 @@ JL_DLLEXPORT jl_array_t *jl_reshape_array(jl_value_t *atype, jl_array_t *data,
     assert(jl_types_equal(jl_tparam0(jl_typeof(data)), jl_tparam0(atype)));
 
     int ndimwords = jl_array_ndimwords(ndims);
-    int tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords * sizeof(size_t) + sizeof(void*), JL_SMALL_BYTE_ALIGNMENT);
+    int tsz = sizeof(jl_array_t) + ndimwords * sizeof(size_t) + sizeof(void*);
     a = (jl_array_t*)jl_gc_alloc(ct->ptls, tsz, atype);
     // No allocation or safepoint allowed after this
     a->flags.pooled = tsz <= GC_MAX_SZCLASS;
@@ -304,7 +307,7 @@ JL_DLLEXPORT jl_array_t *jl_string_to_array(jl_value_t *str)
     jl_array_t *a;
 
     int ndimwords = jl_array_ndimwords(1);
-    int tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords*sizeof(size_t) + sizeof(void*), JL_SMALL_BYTE_ALIGNMENT);
+    int tsz = sizeof(jl_array_t) + ndimwords*sizeof(size_t) + sizeof(void*);
     a = (jl_array_t*)jl_gc_alloc(ct->ptls, tsz, jl_array_uint8_type);
     a->flags.pooled = tsz <= GC_MAX_SZCLASS;
     a->flags.ndims = 1;
@@ -351,7 +354,7 @@ JL_DLLEXPORT jl_array_t *jl_ptr_to_array_1d(jl_value_t *atype, void *data,
                       "unsafe_wrap: pointer %p is not properly aligned to %u bytes", data, align);
 
     int ndimwords = jl_array_ndimwords(1);
-    int tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords*sizeof(size_t), JL_CACHE_BYTE_ALIGNMENT);
+    int tsz = sizeof(jl_array_t) + ndimwords*sizeof(size_t);
     a = (jl_array_t*)jl_gc_alloc(ct->ptls, tsz, atype);
     // No allocation or safepoint allowed after this
     a->flags.pooled = tsz <= GC_MAX_SZCLASS;
@@ -418,7 +421,7 @@ JL_DLLEXPORT jl_array_t *jl_ptr_to_array(jl_value_t *atype, void *data,
                       "unsafe_wrap: pointer %p is not properly aligned to %u bytes", data, align);
 
     int ndimwords = jl_array_ndimwords(ndims);
-    int tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords*sizeof(size_t), JL_CACHE_BYTE_ALIGNMENT);
+    int tsz = sizeof(jl_array_t) + ndimwords*sizeof(size_t);
     a = (jl_array_t*)jl_gc_alloc(ct->ptls, tsz, atype);
     // No allocation or safepoint allowed after this
     a->flags.pooled = tsz <= GC_MAX_SZCLASS;

src/julia_internal.h

@@ -341,6 +341,7 @@ STATIC_INLINE uint8_t JL_CONST_FUNC jl_gc_szclass_align8(unsigned sz)
 // JL_HEAP_ALIGNMENT is the maximum alignment that the GC can provide
 #define JL_HEAP_ALIGNMENT JL_SMALL_BYTE_ALIGNMENT
 #define GC_MAX_SZCLASS (2032-sizeof(void*))
+static_assert(ARRAY_CACHE_ALIGN_THRESHOLD > GC_MAX_SZCLASS, "");
 
 STATIC_INLINE jl_value_t *jl_gc_alloc_(jl_ptls_t ptls, size_t sz, void *ty)
 {

src/options.h

@@ -12,10 +12,14 @@
 
 // object layout options ------------------------------------------------------
 
-// how much space we're willing to waste if an array outgrows its
-// original object
+// The data for an array this size or below will be allocated within the
+// Array object. If the array outgrows that space, it will be wasted.
 #define ARRAY_INLINE_NBYTES (2048*sizeof(void*))
 
+// Arrays at least this size will get larger alignment (JL_CACHE_BYTE_ALIGNMENT).
+// Must be bigger than GC_MAX_SZCLASS.
+#define ARRAY_CACHE_ALIGN_THRESHOLD 2048
+
 // codegen options ------------------------------------------------------------
 
 // (Experimental) Use MCJIT ELF, even where it's not the native format

src/staticdata.c

@@ -782,14 +782,23 @@ static void jl_write_values(jl_serializer_state *s)
 #define JL_ARRAY_ALIGN(jl_value, nbytes) LLT_ALIGN(jl_value, nbytes)
             jl_array_t *ar = (jl_array_t*)v;
             jl_value_t *et = jl_tparam0(jl_typeof(v));
+            size_t alen = jl_array_len(ar);
+            size_t datasize = alen * ar->elsize;
+            size_t tot = datasize;
+            int isbitsunion = jl_array_isbitsunion(ar);
+            if (isbitsunion)
+                tot += alen;
+            else if (ar->elsize == 1)
+                tot += 1;
             int ndimwords = jl_array_ndimwords(ar->flags.ndims);
-            size_t tsz = JL_ARRAY_ALIGN(sizeof(jl_array_t) + ndimwords * sizeof(size_t), JL_CACHE_BYTE_ALIGNMENT);
+            size_t headersize = sizeof(jl_array_t) + ndimwords*sizeof(size_t);
             // copy header
-            ios_write(s->s, (char*)v, tsz);
+            ios_write(s->s, (char*)v, headersize);
+            size_t alignment_amt = JL_SMALL_BYTE_ALIGNMENT;
+            if (tot >= ARRAY_CACHE_ALIGN_THRESHOLD)
+                alignment_amt = JL_CACHE_BYTE_ALIGNMENT;
             // make some header modifications in-place
             jl_array_t *newa = (jl_array_t*)&s->s->buf[reloc_offset];
-            size_t alen = jl_array_len(ar);
-            size_t tot = alen * ar->elsize;
             if (newa->flags.ndims == 1)
                 newa->maxsize = alen;
             newa->offset = 0;
@@ -799,8 +808,7 @@ static void jl_write_values(jl_serializer_state *s)
 
             // write data
             if (!ar->flags.ptrarray && !ar->flags.hasptr) {
-                uintptr_t data = LLT_ALIGN(ios_pos(s->const_data), 16);
-                // realign stream to max(data-align(array), sizeof(void*))
+                uintptr_t data = LLT_ALIGN(ios_pos(s->const_data), alignment_amt);
                 write_padding(s->const_data, data - ios_pos(s->const_data));
                 // write data and relocations
                 newa->data = NULL; // relocation offset
@@ -815,22 +823,27 @@ static void jl_write_values(jl_serializer_state *s)
                         write_pointer(s->const_data);
                 }
                 else {
-                    int isbitsunion = jl_array_isbitsunion(ar);
-                    if (ar->elsize == 1 && !isbitsunion)
-                        tot += 1;
-                    ios_write(s->const_data, (char*)jl_array_data(ar), tot);
-                    if (isbitsunion)
+                    if (isbitsunion) {
+                        ios_write(s->const_data, (char*)jl_array_data(ar), datasize);
                         ios_write(s->const_data, jl_array_typetagdata(ar), alen);
+                    }
+                    else {
+                        ios_write(s->const_data, (char*)jl_array_data(ar), tot);
+                    }
                 }
             }
             else {
-                newa->data = (void*)tsz; // relocation offset
+                size_t data = LLT_ALIGN(ios_pos(s->s), alignment_amt);
+                size_t padding_amt = data - ios_pos(s->s);
+                write_padding(s->s, padding_amt);
+                headersize += padding_amt;
+                newa->data = (void*)headersize; // relocation offset
                 arraylist_push(&s->relocs_list, (void*)(reloc_offset + offsetof(jl_array_t, data))); // relocation location
                 arraylist_push(&s->relocs_list, (void*)(((uintptr_t)DataRef << RELOC_TAG_OFFSET) + item)); // relocation target
                 if (ar->flags.hasptr) {
                     // copy all of the data first
                     const char *data = (const char*)jl_array_data(ar);
-                    ios_write(s->s, data, tot);
+                    ios_write(s->s, data, datasize);
                     // the rewrite all of the embedded pointers to null+relocation
                     uint16_t elsz = ar->elsize;
                     size_t j, np = ((jl_datatype_t*)et)->layout->npointers;
@@ -840,12 +853,12 @@ static void jl_write_values(jl_serializer_state *s)
                             size_t offset = i * elsz + jl_ptr_offset(((jl_datatype_t*)et), j) * sizeof(jl_value_t*);
                             jl_value_t *fld = *(jl_value_t**)&data[offset];
                             if (fld != NULL) {
-                                arraylist_push(&s->relocs_list, (void*)(uintptr_t)(reloc_offset + tsz + offset)); // relocation location
+                                arraylist_push(&s->relocs_list, (void*)(uintptr_t)(reloc_offset + headersize + offset)); // relocation location
                                 arraylist_push(&s->relocs_list, (void*)backref_id(s, fld)); // relocation target
-                                memset(&s->s->buf[reloc_offset + tsz + offset], 0, sizeof(fld)); // relocation offset (none)
+                                memset(&s->s->buf[reloc_offset + headersize + offset], 0, sizeof(fld)); // relocation offset (none)
                             }
                             else {
-                                assert(*(jl_value_t**)&s->s->buf[reloc_offset + tsz + offset] == NULL);
+                                assert(*(jl_value_t**)&s->s->buf[reloc_offset + headersize + offset] == NULL);
                             }
                         }
                     }

test/arrayops.jl

@@ -2660,7 +2660,7 @@ let TT = Union{UInt8, Int8}
     resize!(b, 1)
     @assert pointer(a) == pa
     @assert pointer(b) == pb
-    unsafe_store!(pa, 0x1, 2) # reset a[2] to 1
+    unsafe_store!(Ptr{UInt8}(pa), 0x1, 2) # reset a[2] to 1
     @test length(a) == length(b) == 1
     @test a[1] == b[1] == 0x0
     @test a == b

test/cmdlineargs.jl

@@ -324,7 +324,11 @@ let exename = `$(Base.julia_cmd()) --startup-file=no --color=no`
         rm(memfile)
         @test popfirst!(got) == "        0 g(x) = x + 123456"
         @test popfirst!(got) == "        - function f(x)"
-        @test popfirst!(got) == "       80     []"
+        if Sys.WORD_SIZE == 64
+            @test popfirst!(got) == "       48     []"
+        else
+            @test popfirst!(got) == "       32     []"
+        end
         if Sys.WORD_SIZE == 64
             # P64 pools with 64 bit tags
             @test popfirst!(got) == "       16     Base.invokelatest(g, 0)"
@@ -337,7 +341,11 @@ let exename = `$(Base.julia_cmd()) --startup-file=no --color=no`
             @test popfirst!(got) == "        8     Base.invokelatest(g, 0)"
             @test popfirst!(got) == "       32     Base.invokelatest(g, x)"
         end
-        @test popfirst!(got) == "       80     []"
+        if Sys.WORD_SIZE == 64
+            @test popfirst!(got) == "       48     []"
+        else
+            @test popfirst!(got) == "       32     []"
+        end
         @test popfirst!(got) == "        - end"
         @test popfirst!(got) == "        - f(1.23)"
         @test isempty(got) || got