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 22439


The observable problem is that size_binop gets fed something that
isn't a TYPE_IS_SIZETYPE, and aborts.  Given that the value fed
into size_binop came straight from TYPE_SIZE_UNIT, I don't think
that section of code is wrong.  I think we have a right to assume
that TYPE_SIZE_UNIT has type sizetype and TYPE_SIZE has type
bitsizetype.

The type mismatch was introduced during gimplify_type_sizes.  We
had a NOP_EXPR<sizetype, SAVE_EXPR<n>>, where N is size_t.  For
C, the only difference between size_t and sizetype is the actual
TYPE_IS_SIZETYPE bit, and so we discard the cast considering the
two types compatible.  Which leads to TYPE_SIZE_UNIT being set to N.

Given that gimple code itself does not care about the difference
between sizetype and size_t, it didn't seem prudent to enforce this
distinction in tree_ssa_useless_type_conversion or the like.  
Instead, we can notice the change in the one place that it matters
and enforce type correctness by hand.

Tested on i686 and x86_64 linux.


r~


        * gimplify.c (gimplify_one_sizepos): Preserve the original type.

Index: gimplify.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/gimplify.c,v
retrieving revision 2.143
diff -u -p -d -r2.143 gimplify.c
--- gimplify.c	28 Jul 2005 16:29:51 -0000	2.143
+++ gimplify.c	8 Aug 2005 21:36:40 -0000
@@ -4610,18 +4610,44 @@ gimplify_type_sizes (tree type, tree *li
 void
 gimplify_one_sizepos (tree *expr_p, tree *stmt_p)
 {
+  tree type, expr = *expr_p;
+
   /* We don't do anything if the value isn't there, is constant, or contains
      A PLACEHOLDER_EXPR.  We also don't want to do anything if it's already
      a VAR_DECL.  If it's a VAR_DECL from another function, the gimplifier
      will want to replace it with a new variable, but that will cause problems
      if this type is from outside the function.  It's OK to have that here.  */
-  if (*expr_p == NULL_TREE || TREE_CONSTANT (*expr_p)
-      || TREE_CODE (*expr_p) == VAR_DECL
-      || CONTAINS_PLACEHOLDER_P (*expr_p))
+  if (expr == NULL_TREE || TREE_CONSTANT (expr)
+      || TREE_CODE (expr) == VAR_DECL
+      || CONTAINS_PLACEHOLDER_P (expr))
     return;
 
-  *expr_p = unshare_expr (*expr_p);
+  type = TREE_TYPE (expr);
+  *expr_p = unshare_expr (expr);
+
   gimplify_expr (expr_p, stmt_p, NULL, is_gimple_val, fb_rvalue);
+  expr = *expr_p;
+
+  /* Verify that we've an exact type match with the original expression.
+     In particular, we do not wish to drop a "sizetype" in favour of a
+     type of similar dimensions.  We don't want to pollute the generic
+     type-stripping code with this knowledge because it doesn't matter
+     for the bulk of GENERIC/GIMPLE.  It only matters that TYPE_SIZE_UNIT
+     and friends retain their "sizetype-ness".  */
+  if (TREE_TYPE (expr) != type && TYPE_IS_SIZETYPE (type))
+    {
+      tree tmp;
+
+      *expr_p = create_tmp_var (type, NULL);
+      tmp = build1 (NOP_EXPR, type, expr);
+      tmp = build2 (MODIFY_EXPR, type, *expr_p, expr);
+      if (EXPR_HAS_LOCATION (expr))
+	SET_EXPR_LOCUS (tmp, EXPR_LOCUS (expr));
+      else
+	SET_EXPR_LOCATION (tmp, input_location);
+
+      gimplify_and_add (tmp, stmt_p);
+    }
 }
 
 #ifdef ENABLE_CHECKING


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