[Bug c/82765] [6/7/8 Regression] ICE at -Os on valid code on x86_64-linux-gnu: in tree_to_shwi, at tree.c:6611

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82765

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
middle-end "fix"

Index: gcc/varasm.c
===================================================================
--- gcc/varasm.c        (revision 254211)
+++ gcc/varasm.c        (working copy)
@@ -2879,7 +2879,7 @@ static void
 decode_addr_const (tree exp, struct addr_const *value)
 {
   tree target = TREE_OPERAND (exp, 0);
-  int offset = 0;
+  HOST_WIDE_INT offset = 0;
   rtx x;

   while (1)
@@ -2893,8 +2893,9 @@ decode_addr_const (tree exp, struct addr
       else if (TREE_CODE (target) == ARRAY_REF
               || TREE_CODE (target) == ARRAY_RANGE_REF)
        {
-         offset += (tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (target)))
-                    * tree_to_shwi (TREE_OPERAND (target, 1)));
+         /* Truncate big offset.  */
+         offset += (TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (target)))
+                    * TREE_INT_CST_LOW (TREE_OPERAND (target, 1)));
          target = TREE_OPERAND (target, 0);
        }
       else if (TREE_CODE (target) == MEM_REF

there's a lot more errors in it though.  For example it doesn't handle
an array domain low bound != 1 correctly ...  but maybe that's not
required (well, strictly speaking it probably is but it may be hard
to create a testcase that fails).