Change struct/obj alloc size (#682)

Author: ncannasse

libs/directx/dx/Dx12.hx

@@ -1044,6 +1044,7 @@ enum abstract InputClassification(Int) {
 @:struct class InputLayoutDesc {
 	public var inputElementDescs : hl.CArray<InputElementDesc>;
 	public var numElements : Int;
+	public var __padding : Int; // largest element
 	public function new() {
 	}
 }
@@ -1094,7 +1095,6 @@ enum abstract PipelineStateFlags(Int) {
 	@:packed public var rasterizerState(default,null) : RasterizerDesc;
 	@:packed public var depthStencilDesc(default,null) : DepthStencilDesc;
 	@:packed public var inputLayout(default,null) : InputLayoutDesc;
-	var __padding : Int; // ?
 	public var ibStripCutValue : IndexBufferStripCutValue;
 	public var primitiveTopologyType : PrimitiveTopologyType;
 	public var numRenderTargets : Int;

src/gc.c

@@ -1271,16 +1271,13 @@ vdynamic *hl_alloc_dynbool( bool b ) {
 
 vdynamic *hl_alloc_obj( hl_type *t ) {
 	vobj *o;
-	int size;
 	int i;
 	hl_runtime_obj *rt = t->obj->rt;
 	if( rt == NULL || rt->methods == NULL ) rt = hl_get_obj_proto(t);
-	size = rt->size;
-	if( size & (HL_WSIZE-1) ) size += HL_WSIZE - (size & (HL_WSIZE-1));
 	if( t->kind == HSTRUCT ) {
-		o = (vobj*)hl_gc_alloc_gen(t, size, (rt->hasPtr ? MEM_KIND_RAW : MEM_KIND_NOPTR) | MEM_ZERO);
+		o = (vobj*)hl_gc_alloc_gen(t, rt->size, (rt->hasPtr ? MEM_KIND_RAW : MEM_KIND_NOPTR) | MEM_ZERO);
 	} else {
-		o = (vobj*)hl_gc_alloc_gen(t, size, (rt->hasPtr ? MEM_KIND_DYNAMIC : MEM_KIND_NOPTR) | MEM_ZERO);
+		o = (vobj*)hl_gc_alloc_gen(t, rt->size, (rt->hasPtr ? MEM_KIND_DYNAMIC : MEM_KIND_NOPTR) | MEM_ZERO);
 		o->t = t;
 	}
 	for(i=0;i<rt->nbindings;i++) {

src/hl.h

@@ -533,6 +533,8 @@ struct hl_runtime_obj {
 	int size;
 	int nmethods;
 	int nbindings;
+	unsigned char pad_size;
+	unsigned char largest_field;
 	bool hasPtr;
 	void **methods;
 	int *fields_indexes;

src/jit.c

@@ -3822,7 +3822,6 @@ int hl_jit_function( jit_ctx *ctx, hl_module *m, hl_function *f ) {
 					else {
 						hl_runtime_obj *rt = hl_get_obj_rt(dst->t);
 						osize = rt->size;
-						if( osize & (HL_WSIZE-1) ) osize += HL_WSIZE - (osize & (HL_WSIZE-1));
 					}
 					preg *idx = alloc_cpu(ctx, rb, true);
 					op64(ctx, IMUL, idx, pconst(&p,osize));

src/std/array.c

@@ -57,13 +57,11 @@ HL_PRIM void *hl_alloc_carray( hl_type *at, int size ) {
 
 	hl_runtime_obj *rt = at->obj->rt;
 	if( rt == NULL || rt->methods == NULL ) rt = hl_get_obj_proto(at);
-	int osize = rt->size;
-	if( osize & (HL_WSIZE-1) ) osize += HL_WSIZE - (osize & (HL_WSIZE-1));
-	char *arr = hl_gc_alloc_gen(at, size * osize, (rt->hasPtr ? MEM_KIND_RAW : MEM_KIND_NOPTR) | MEM_ZERO);
+	char *arr = hl_gc_alloc_gen(at, size * rt->size, (rt->hasPtr ? MEM_KIND_RAW : MEM_KIND_NOPTR) | MEM_ZERO);
 	if( at->kind == HOBJ || rt->nbindings ) {
 		int i,k;
 		for(k=0;k<size;k++) {
-			char *o = arr + osize * k;
+			char *o = arr + rt->size * k;
 			if( at->kind == HOBJ )
 				((vobj*)o)->t = at;
 			for(i=0;i<rt->nbindings;i++) {

src/std/obj.c

@@ -236,18 +236,21 @@ HL_PRIM hl_runtime_obj *hl_get_obj_rt( hl_type *ot ) {
 		start = p->nfields;
 		memcpy(t->fields_indexes, p->fields_indexes, sizeof(int)*p->nfields);
 	}
-	size = p ? p->size : (ot->kind == HSTRUCT ? 0 : HL_WSIZE); // hl_type*
+	size = p ? p->size - p->pad_size : (ot->kind == HSTRUCT ? 0 : HL_WSIZE); // hl_type*
 	nlookup = 0;
+	int largest_field = p ? p->largest_field : size;
 	for(i=0;i<o->nfields;i++) {
 		hl_type *ft = o->fields[i].t;
-		hl_type *pad = ft;
-		while( pad->kind == HPACKED ) {
-			// align on first field
-			pad = pad->tparam;
-			while( pad->obj->super && hl_get_obj_rt(pad->obj->super)->nfields ) pad = pad->obj->super;
-			pad = pad->obj->fields[0].t;
-		}
-		size += hl_pad_struct(size,pad);
+		if( ft->kind == HPACKED ) {
+			// align on packed largest field
+			int large = hl_get_obj_rt(ft->tparam)->largest_field;
+			int pad = size % large;
+			if( pad != 0 )
+				size += large - pad;
+			if( large > largest_field )
+				largest_field = large;
+		} else
+			size += hl_pad_struct(size,ft);
 		t->fields_indexes[i+start] = size;
 		if( *o->fields[i].name )
 			hl_lookup_insert(t->lookup,nlookup++,o->fields[i].hashed_name,o->fields[i].t,size);
@@ -259,10 +262,22 @@ HL_PRIM hl_runtime_obj *hl_get_obj_rt( hl_type *ot ) {
 			if( rts->hasPtr ) t->hasPtr = true;
 			continue;
 		}
-		size += hl_type_size(ft);
+		int sz = hl_type_size(ft);
+		size += sz;
+		if( sz > largest_field ) largest_field = sz;
 		if( !t->hasPtr && hl_is_ptr(ft) ) t->hasPtr = true;
 	}
 	t->size = size;
+	t->pad_size = 0;
+	// align size on largest field size to match C, so arr[i].field is always aligned
+	if( largest_field > 0 ) {
+		int pad = size % largest_field;
+		if( pad != 0 ) {
+			t->pad_size = largest_field - pad;
+			t->size += t->pad_size;
+		}
+	}
+	t->largest_field = largest_field;
 	t->nmethods = p ? p->nmethods : o->nproto;
 	t->methods = NULL;
 	o->rt = t;