Hl prefetch

src/generators/genhl.ml

@@ -131,6 +131,7 @@ type access =
 	| AInstanceProto of texpr * field index
 	| AInstanceField of texpr * field index
 	| AArray of reg * (ttype * ttype) * reg
+	| ACArray of reg * ttype * reg
 	| AVirtualMethod of texpr * field index
 	| ADynamic of texpr * string index
 	| AEnum of tenum * field index
@@ -310,6 +311,12 @@ let unsigned_op e1 e2 =
 	in
 	is_unsigned e1 && is_unsigned e2
 
+let rec get_const e =
+	match e.eexpr with
+	| TConst c -> c
+	| TParenthesis e | TCast (e,_) -> get_const e
+	| _ -> abort "Should be a constant" e.epos
+
 let set_curpos ctx p =
 	ctx.m.mcurpos <- p
 
@@ -1377,6 +1384,13 @@ and get_access ctx e =
 				free ctx a;
 				let t = to_type ctx t in
 				AArray (a,(t,t),i)
+			| TInst ({ cl_path = ["hl"],"Abstract" },[TInst({ cl_kind = KExpr (EConst (String("hl_carray",_)),_) },_)]) ->
+				let a = eval_null_check ctx a in
+				hold ctx a;
+				let i = eval_to ctx i HI32 in
+				free ctx a;
+				let t = to_type ctx e.etype in
+				ACArray (a,t,i)
 			| TAbstract (a,pl) ->
 				loop (Abstract.get_underlying_type a pl)
 			| _ ->
@@ -1894,7 +1908,13 @@ and eval_expr ctx e =
 			r
 		| "$asize", [e] ->
 			let r = alloc_tmp ctx HI32 in
-			op ctx (OArraySize (r, eval_to ctx e HArray));
+			(match follow e.etype with
+			| TInst ({cl_path=["hl"],"Abstract"},[TInst({ cl_kind = KExpr (EConst (String("hl_carray",_)),_) },_)]) ->
+				let arr = eval_expr ctx e in
+				op ctx (ONullCheck arr);
+				op ctx (OArraySize (r, arr))
+			| _ ->
+				op ctx (OArraySize (r, eval_to ctx e HArray)));
 			r
 		| "$aalloc", [esize] ->
 			let et = (match follow e.etype with TAbstract ({ a_path = ["hl"],"NativeArray" },[t]) -> to_type ctx t | _ -> invalid()) in
@@ -2054,6 +2074,15 @@ and eval_expr ctx e =
 			free ctx rfile;
 			free ctx min;
 			r
+		| "$prefetch", [value; mode] ->
+			let mode = (match get_const mode with
+				| TInt m -> Int32.to_int m
+				| _ -> abort "Constant mode required" e.epos
+			) in
+			(match get_access ctx value with
+			| AInstanceField (f, index) -> op ctx (OPrefetch (eval_expr ctx f, index + 1, mode))
+			| _ -> op ctx (OPrefetch (eval_expr ctx value, 0, mode)));
+			alloc_tmp ctx HVoid
 		| _ ->
 			abort ("Unknown native call " ^ s) e.epos)
 	| TEnumIndex v ->
@@ -2208,7 +2237,7 @@ and eval_expr ctx e =
 				ignore(make_fun ctx ("","") fid f None None);
 			end;
 			op ctx (OStaticClosure (r,fid));
-		| ANone | ALocal _ | AArray _ | ACaptured _ ->
+		| ANone | ALocal _ | AArray _ | ACaptured _ | ACArray _ ->
 			abort "Invalid access" e.epos);
 		let to_t = to_type ctx e.etype in
 		(match to_t with
@@ -2452,7 +2481,7 @@ and eval_expr ctx e =
 				let r = value() in
 				op ctx (OSetEnumField (ctx.m.mcaptreg,index,r));
 				r
-			| AEnum _ | ANone | AInstanceFun _ | AInstanceProto _ | AStaticFun _ | AVirtualMethod _ ->
+			| AEnum _ | ANone | AInstanceFun _ | AInstanceProto _ | AStaticFun _ | AVirtualMethod _ | ACArray _ ->
 				die "" __LOC__)
 		| OpBoolOr ->
 			let r = alloc_tmp ctx HBool in
@@ -2727,6 +2756,10 @@ and eval_expr ctx e =
 		(match get_access ctx e with
 		| AArray (a,at,idx) ->
 			array_read ctx a at idx e.epos
+		| ACArray (a,t,idx) ->
+			let tmp = alloc_tmp ctx t in
+			op ctx (OGetArray (tmp,a,idx));
+			tmp
 		| _ ->
 			die "" __LOC__)
 	| TMeta (_,e) ->
@@ -3046,7 +3079,7 @@ and gen_assign_op ctx acc e1 f =
 		free ctx robj;
 		op ctx (ODynSet (robj,fid,r));
 		r
-	| ANone | ALocal _ | AStaticFun _ | AInstanceFun _ | AInstanceProto _ | AVirtualMethod _ | AEnum _ ->
+	| ANone | ALocal _ | AStaticFun _ | AInstanceFun _ | AInstanceProto _ | AVirtualMethod _ | AEnum _ | ACArray _ ->
 		die "" __LOC__
 
 and build_capture_vars ctx f =

src/generators/hl2c.ml

@@ -1087,6 +1087,15 @@ let generate_function ctx f =
 			sexpr "%s = %s + %s" (reg r) (reg r2) (reg off)
 		| ONop _ ->
 			()
+		| OPrefetch (r,fid,mode) ->
+			let expr = (if fid = 0 then reg r else (match rtype r with
+			| HObj o | HStruct o ->
+				let name, t = resolve_field o (fid - 1) in
+				Printf.sprintf "%s->%s" (reg r) name
+			| _ ->
+				Globals.die "" __LOC__
+			)) in
+			sexpr "__hl_prefetch_m%d(%s)" mode expr
 	) f.code;
 	flush_options (Array.length f.code);
 	unblock();

src/generators/hlcode.ml

@@ -201,6 +201,7 @@ type opcode =
 	| ORefData of reg * reg
 	| ORefOffset of reg * reg * reg
 	| ONop of string
+	| OPrefetch of reg * field index * int
 
 type fundecl = {
 	fpath : string * string;
@@ -572,6 +573,8 @@ let ostr fstr o =
 	| ORefData (r,d) -> Printf.sprintf "refdata %d, %d" r d
 	| ORefOffset (r,r2,off) -> Printf.sprintf "refoffset %d, %d, %d" r r2 off
 	| ONop s -> if s = "" then "nop" else "nop " ^ s
+	| OPrefetch (r,f,mode) -> Printf.sprintf "prefetch %d[%d] %d" r f mode
+
 let fundecl_name f = if snd f.fpath = "" then "fun$" ^ (string_of_int f.findex) else (fst f.fpath) ^ "." ^ (snd f.fpath)
 
 let dump pr code =

src/generators/hlinterp.ml

@@ -1154,7 +1154,7 @@ let interp ctx f args =
 			(match get r2, get off with
 			| VRef (RArray (a,pos),t), VInt i -> set r (VRef (RArray (a,pos + Int32.to_int i),t))
 			| _ -> Globals.die "" __LOC__)
-		| ONop _ ->
+		| ONop _ | OPrefetch _ ->
 			()
 		);
 		loop()
@@ -2436,7 +2436,7 @@ let check code macros =
 			| ORethrow r ->
 				reg r HDyn
 			| OGetArray (v,a,i) ->
-				reg a HArray;
+				(match rtype a with HAbstract ("hl_carray",_) -> () | _ -> reg a HArray);
 				reg i HI32;
 				ignore(rtype v);
 			| OGetUI8 (r,b,p) | OGetUI16(r,b,p) ->
@@ -2466,7 +2466,7 @@ let check code macros =
 				ignore(rtype a);
 				ignore(rtype b);
 			| OArraySize (r,a) ->
-				reg a HArray;
+				(match rtype a with HAbstract ("hl_carray",_) -> () | _ -> reg a HArray);
 				reg r HI32
 			| OType (r,_) ->
 				reg r HType
@@ -2547,6 +2547,8 @@ let check code macros =
 				reg off HI32;
 			| ONop _ ->
 				();
+			| OPrefetch (r,f,_) ->
+				if f = 0 then ignore(rtype r) else ignore(tfield r (f - 1) false)
 		) f.code
 		(* TODO : check that all path correctly initialize NULL values and reach a return *)
 	in

src/generators/hlopt.ml

@@ -164,6 +164,8 @@ let opcode_fx frw op =
 		write r;
 	| ONop _  ->
 		()
+	| OPrefetch (r,_,_) ->
+		read r
 
 let opcode_eq a b =
 	match a, b with
@@ -432,6 +434,9 @@ let opcode_map read write op =
 		ORefOffset (write r,r2,off);
 	| ONop _ ->
 		op
+	| OPrefetch (r, fid, mode) ->
+		let r2 = read r in
+		OPrefetch (r2, fid, mode)
 
 (* build code graph *)
 
@@ -875,6 +880,11 @@ let _optimize (f:fundecl) =
 			| OGetThis (r,fid) when (match f.regs.(r) with HStruct _ -> true | _ -> false) ->
 				do_write r;
 				if is_packed_field 0 fid then state.(r).rnullcheck <- true;
+			| OGetArray (r,arr,idx) ->
+				do_read arr;
+				do_read idx;
+				do_write r;
+				(match f.regs.(arr) with HAbstract _ -> state.(r).rnullcheck <- true | _ -> ());
 			| _ ->
 				opcode_fx (fun r read ->
 					if read then do_read r else do_write r

std/hl/CArray.hx

@@ -9,9 +9,14 @@ abstract CArray<T>(Abstract<"hl_carray">) {
 
 	public var length(get,never) : Int;
 
-	inline function get_length() return getLen(cast this);
 
+	#if (hl_ver >= version("1.14.0"))
+	inline function get_length() return untyped $asize(this);
+	@:arrayAccess inline function get( index : Int ) : T return untyped this[index];
+	#else
+	inline function get_length() return getLen(cast this);
 	@:arrayAccess inline function get( index : Int ) : T return getIndex(cast this, index);
+	#end
 
 	public static function alloc<T>( cl : Class<T>, size : Int ) : CArray<T> {
 		return cast alloc_carray( (cast cl:BaseType).__type__ , size );