[Ada] object.operation fix

[Arnaud Charlet <charlet at adacore dot com>]

Tested on i686-linux, committed on trunk

When an aliased object is used as the prefix in a call using Object.Operation
notation, and the corresponding formal is an access parameter, the code for
passing the actual's accessibility level mistakenly passes the level of its
type rather than the level of the named object. This can lead to dangling
references when the formal is converted or assigned within the subprogram
and the actual object is declared at a deeper level than the target type.
The wrong level is used because the aliased object name is rewritten as an
Access attribute, but the original node is used in determining the level.
We now check for this as a special case and ensure that the attribute name
is used to determine the level rather than the original object name.

gnat.dg/aliased_prefix_accessibility.adb must execute quietly.

Also, in Parameters of a scalar type are passed by value, and when inlining a
call where the actual for an in-parameter is a literal, a reference to the
formal can be replaced with the value of the actual. However, if the formal is
the prefix of an Address or Access attribute, the resulting block is not
semantically legal. In that case we must create a local variable to capture
the value of
the actual, and use the entity in the block. This is semantically correct, given
that the actual is an expression, and that using its address in the body cannot
mean more than the address of the location that holds the expression.

When generating validity checks for parameters (-gnatVi) set there was
already circuitry to handle the case of a packed array (unsetting the
analyzed bit to force a reanalyze in the validity check context), but
this was insufficient for the case where the packed array reference
appeared as a prefix of a selected or indexed component in this context.
This patch corrects that oversight, resetting analyzed bits on all
prefixes of selected or indexed components.

gnat.dg/valid1.adb should compile and execute silently.

2007-06-06  Gary Dismukes  <dismukes@adacore.com>
	    Ed Schonberg  <schonberg@adacore.com>

	* exp_ch6.ads, exp_ch6.adb (Expand_Call): When adding an extra
	accessibility actual, check for the case of an aliased object that has
	been rewritten as an Access attribute, and assign Prev to Prev_Orig so
	we fall into processing for the attribute rather than the name of the
	object.
	(Expand_Inline_Call): If an actual is a literal, and the corresponding
	formal has its address taken in the body, create a temporary to capture
	value.  If the return type is a limited interface, do not treat the
	return value as Controlled.
	(Is_In_Place_Function):  If the return type is a limited interface,
	treat as returning in place. The actual returned object may not always
	be limited, but the caller has to assume that it is returned in place.
	(Add_Final_List_Actual_To_Build_In_Place_Call): If the call is the
	context of an allocator, use the correct finalization chain (that is,
	the chain belonging to the access type, rather than the chain belonging
	to the current scope).
	(Add_Alloc_Form_Actual_To_Build_In_Place_Call): Test for a tagged
	result type rather than a controlling result as a precondition for
	adding an allocation form actual to a build-in-place call.
	(Add_Final_List_Actual_To_Build_In_Place_Call): Ditto.
	(Freeze_Subprogram): Code cleanup. Remove all the code that register the
	primitive in the dispatch tables. This work is now done by Make_DT when
	the type is frozen.
	(Register_Predefined_DT_Entry): Removed.
	(Add_Return): If end label is not present, use sloc of last statement
	for generated return statement in procedure, for better gdb behavior
	on expanded code.
	(Add_Access_Actual_To_Build_In_Place_Call): Set parent fields of the
	object address nodes to ensure proper processing by routines like
	Insert_After*.
	(Expand_Call): Fix generation of validity check for parameter
	(Add_Alloc_Form_Actual_To_Build_In_Place_Call): Return without passing
	the allocation form parameter if the result subtype is constrained,
	except when the function has a controlling result.
	(Add_Final_List_Actual_To_Build_In_Place_Call): Test Controlled_Type
	rather than Is_Controlled and Has_Controlled_Component, since we want to
	include class-wide result types in this treatment. Also test for a
	controlling result, since that also requires passing a finalization
	list.
	(Make_Build_In_Place_Call_In_Allocator): Call Add_Alloc_Form_Actual_*
	even when the result subtype is constrained, to handle calls involving
	controlling results.
	(Make_Build_In_Place_Call_In_Anonymous_Context): Add_Alloc_Form_Actual_*
	is now called even when the result subtype is constrained, to handle
	calls involving controlling results.
	(Make_Build_In_Place_Call_In_Assignment): Remove test for Is_Constrained
	on call to Add_Alloc_Form_Actual_To_Build_In_Place_Call (that procedure
	now performs the test).
	(Make_Build_In_Place_Call_In_Object_Declaration):
	Add_Alloc_Form_Actual_* is now called even when the result subtype is
	constrained, to handle calls involving controlling results.
	(Add_Return): Accomodate rewritten pattern from local raise to goto
	transformation, so that we still recognize an transfer statement
	and do the right thing here in that case.
	(Expand_N_Subprogram_Body): Add dummy Push/Pop_xxx_Label nodes at start
	and end of subprogram code.
	(Register_Interface_DT_Entry, Register_Predefined_DT_Entry): Add missing
	support for primitives that are functions (without formals) with a
	controlling result.
	(Inherited_From_Formal): If the actual subtype has not generic parent
	type, it is not an actual for a formal derived type, and there is no
	operation to inherit from the formal.

Attachment: difs
Description: Text document