This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Fix bug with self-referencing constructor in Ada


This is a regression from the 3.x series present on all platforms.

The gimplifier features a sophisticated machinery to detect constructors on 
the rhs that overlap with the lhs, e.g.

  type R is record
    I1 : Integer;
    I2 : Integer;
  end record;

  R := (R.I1 - R.I2, R.I1 + R.I2);

If overlap is detected, the element will first be preevaluated in a 
temporary variable and then the result assigned to the component.  The 
function in charge is gimplify_init_ctor_preeval.

Now there are these lines in gimplify_init_ctor_preeval:

  /* We can't preevaluate if the type contains a placeholder.  */
  if (type_contains_placeholder_p (TREE_TYPE (*expr_p)))
    return;

And

/* Return true if any part of the computation of TYPE involves a
   PLACEHOLDER_EXPR.  This includes size, bounds, qualifiers
   (for QUAL_UNION_TYPE) and field positions.  */

static bool
type_contains_placeholder_1 (tree type)
{
  /* If the size contains a placeholder or the parent type (component type in
     the case of arrays) type involves a placeholder, this type does.  */
  if (CONTAINS_PLACEHOLDER_P (TYPE_SIZE (type))
      || CONTAINS_PLACEHOLDER_P (TYPE_SIZE_UNIT (type))
      || (TREE_TYPE (type) != 0
	  && type_contains_placeholder_p (TREE_TYPE (type))))
    return true;

Note that type_contains_placeholder_1 goes to the pointed-to type for 
pointers.  This means that gimplify_init_ctor_preeval will punt for records 
that contains pointers to unconstrained types, like in the attached testcase, 
thus causing a violation of the Write-After-Read dependency in the code.


The problematic lines are there to guard a further gimplification:

  /* Gimplify the constructor element to something appropriate for the rhs
     of a MODIFY_EXPR.  Given that we know the lhs is an aggregate, we know
     the gimplifier will consider this a store to memory.  Doing this
     gimplification now means that we won't have to deal with complicated
     language-specific trees, nor trees like SAVE_EXPR that can induce
     exponential search behavior.  */
  one = gimplify_expr (expr_p, pre_p, post_p, is_gimple_mem_rhs, fb_rvalue);

that is to say, if you remove them you end up with unmatched PLACEHOLDER_EXPRs 
during this gimplification.  It turns out that this problem is already solved 
in 2 other places in the gimplifier (gimplify_arg, gimplify_modify_expr) with

  /* If this is a variable sized type, we must remember the size.  */
  maybe_with_size_expr (expr_p);


Bootstrapped/regtested on x86_64-suse-linux, applied to mainline as obvious.


2006-09-28  Eric Botcazou  <ebotcazou@adacore.com>

	* gimplify.c (gimplify_init_ctor_preeval): Call maybe_with_size_expr
	on the element before gimplifying it, instead of punting if it is of
	variable size.


2006-09-28  Eric Botcazou  <ebotcazou@adacore.com>

	* gnat.dg/self_aggregate_with_pointer.adb: New test.


-- 
Eric Botcazou
Index: gimplify.c
===================================================================
--- gimplify.c	(revision 117190)
+++ gimplify.c	(working copy)
@@ -2670,9 +2670,8 @@ gimplify_init_ctor_preeval (tree *expr_p
       return;
     }
 
-  /* We can't preevaluate if the type contains a placeholder.  */
-  if (type_contains_placeholder_p (TREE_TYPE (*expr_p)))
-    return;
+  /* If this is a variable sized type, we must remember the size.  */
+  maybe_with_size_expr (expr_p);
 
   /* Gimplify the constructor element to something appropriate for the rhs
      of a MODIFY_EXPR.  Given that we know the lhs is an aggregate, we know
-- { dg-do run }

procedure self_aggregate_with_pointer is

  type Arr is array (Natural range <>) of Integer;

  type Rec (N : Natural) is record
    A : Arr (1..N);
  end record;

  type Acc_Rec is access all Rec;

  type SRec is record
    A : Acc_Rec;
    I1, I2, I3, I4, I5, I6, I7: Integer;
  end record;

  R : aliased Rec (1);
  S : Srec := (A => R'Access, others => 0);

begin
  S := (A => S.A, others => 0);
  if S.A /= R'Access then
    raise Program_Error;
  end if;
end;

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]