In the frontend inlining used in GNATprove, inlining of a return statement
was using an unchecked type conversion, which could cause a necessary
run-time check on the conversion to be skipped. Now fixed.
There is no impact on compilation.
Tested on x86_64-pc-linux-gnu, committed on trunk
2018-05-21 Yannick Moy <m...@adacore.com>
gcc/ada/
* inline.adb (Expand_Inlined_Call.Process_Formals): Use a type
conversion instead of an unchecked type conversion when inlining a
return statement, unless type qualification is required (for character
and string literal) or no check can result from the conversion (for
access types).
* opt.ads: Update comment.--- gcc/ada/inline.adb
+++ gcc/ada/inline.adb
@@ -2482,26 +2482,42 @@ package body Inline is
end if;
-- Because of the presence of private types, the views of the
- -- expression and the context may be different, so place an
- -- unchecked conversion to the context type to avoid spurious
+ -- expression and the context may be different, so place
+ -- a type conversion to the context type to avoid spurious
-- errors, e.g. when the expression is a numeric literal and
-- the context is private. If the expression is an aggregate,
-- use a qualified expression, because an aggregate is not a
- -- legal argument of a conversion. Ditto for numeric literals
- -- and attributes that yield a universal type, because those
- -- must be resolved to a specific type.
-
- if Nkind_In (Expression (N), N_Aggregate, N_Null)
+ -- legal argument of a conversion. Ditto for numeric, character
+ -- and string literals, and attributes that yield a universal
+ -- type, because those must be resolved to a specific type.
+
+ if Nkind_In (Expression (N), N_Aggregate,
+ N_Null,
+ N_Character_Literal,
+ N_String_Literal)
or else Yields_Universal_Type (Expression (N))
then
Ret :=
Make_Qualified_Expression (Sloc (N),
Subtype_Mark => New_Occurrence_Of (Ret_Type, Sloc (N)),
Expression => Relocate_Node (Expression (N)));
- else
+
+ -- Use an unchecked type conversion between access types, for
+ -- which a type conversion would not always be valid, as no
+ -- check may result from the conversion.
+
+ elsif Is_Access_Type (Ret_Type) then
Ret :=
Unchecked_Convert_To
(Ret_Type, Relocate_Node (Expression (N)));
+
+ -- Otherwise use a type conversion, which may trigger a check
+
+ else
+ Ret :=
+ Make_Type_Conversion (Sloc (N),
+ Subtype_Mark => New_Occurrence_Of (Ret_Type, Sloc (N)),
+ Expression => Relocate_Node (Expression (N)));
end if;
if Nkind (Targ) = N_Defining_Identifier then
--- gcc/ada/opt.ads
+++ gcc/ada/opt.ads
@@ -727,7 +727,7 @@ package Opt is
-- Set True to activate inlining by front-end expansion (even on GCC
-- targets, where inlining is normally handled by the back end). Set by
-- the flag -gnatN (which is now considered obsolescent, since the GCC
- -- back end can do a better job of inlining than the front end these days.
+ -- back end can do a better job of inlining than the front end these days).
Full_Path_Name_For_Brief_Errors : Boolean := False;
-- PROJECT MANAGER
