[PATCH] Fix PR/24599, setting overflow flags on shared constants

[Paolo Bonzini]

The current c-typeck.c code for build_c_cast is

         if (CONSTANT_CLASS_P (ovalue))
            {
              TREE_OVERFLOW (value) = TREE_OVERFLOW (ovalue);
              TREE_CONSTANT_OVERFLOW (value) = TREE_CONSTANT_OVERFLOW 
(ovalue);
            }
         else
           TREE_OVERFLOW (value) = 0;

Given Andrew and Zdenek's analysis in the PR, a quick fix for this bug 
could just to add a `value = copy_node (value)' in the `then' branch.  
The patch is only slightly more complex than that.

      if (CONSTANT_CLASS_P (ovalue)
          && (TREE_OVERFLOW (ovalue)
              || TREE_CONSTANT_OVERFLOW (ovalue)))
        {
          /* Avoid clobbering a shared constant.  */
          value = copy_node (value);
           TREE_OVERFLOW (value) = TREE_OVERFLOW (ovalue);
           TREE_CONSTANT_OVERFLOW (value) = TREE_CONSTANT_OVERFLOW (ovalue);
        }
      else if (TREE_OVERFLOW (value) || TREE_CONSTANT_OVERFLOW (value))
        /* Reset VALUE's overflow flags, ensuring constant sharing.  */
        value = build_int_cst_wide (TREE_TYPE (value),
                                    TREE_INT_CST_LOW (value),
                                    TREE_INT_CST_HIGH (value));

1) The `else' branch uses build_int_cst_wide to ensure that the constant 
is shared.

2) The copying is only done if the overflow flags have to be set.  If 
CONSTANT_CLASS_P is set but the ovalue had both overflow flags reset, 
the two branches are functionally equivalent but the `else' is more 
memory efficient: so, this is what the patch does.

Bootstrapped/tested powerpc-apple-darwin8.2.0, also checked that the 
original reporter's testcase is fixed.  Ok for mainline?  Ok for 4.0 
after successful testing there?

Paolo
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: pr24599.patch
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20051105/c187b437/attachment.ksh>