[Bug middle-end/19164] New: ICE in digest_init or build_vector

["jakub at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

--- gcc/testsuite/gcc.dg/20041227-1.c.jj        2004-12-27 13:57:19.126504511
+0100
+++ gcc/testsuite/gcc.dg/20041227-1.c   2004-12-27 13:57:15.864087556 +0100
@@ -0,0 +1,5 @@
+/* { dg-do compile { target i?86-*-* } } */
+/* { dg-options "-mmmx" } */
+
+typedef short int V __attribute__ ((vector_size (8)));
+static V v = (V) 0x00FF00FF00FF00FFLL;

gives ICE in both GCC 3.4.3-RH and on HEAD.  Depending whether checking is
enabled or not, it either results in checking failure in digest_init or
in a segfault in build_vector.

The problematic hunk is (c-typeck.c (digest_init)):
  /* Build a VECTOR_CST from a *constant* vector constructor.  If the
     vector constructor is not constant (e.g. {1,2,3,foo()}) then punt
     below and handle as a constructor.  */
    if (code == VECTOR_TYPE
        && TREE_CODE (TREE_TYPE (inside_init)) == VECTOR_TYPE
        && vector_types_convertible_p (TREE_TYPE (inside_init), type)
        && TREE_CONSTANT (inside_init))
      {
        if (TREE_CODE (inside_init) == VECTOR_CST
            && comptypes (TYPE_MAIN_VARIANT (TREE_TYPE (inside_init)),
                          TYPE_MAIN_VARIANT (type)))
          return inside_init;
        else
          return build_vector (type, CONSTRUCTOR_ELTS (inside_init));
      }

Changing it to else if (TREE_CODE (inside_init) == CONSTRUCTOR)
  return build_vector (type, CONSTRUCTOR_ELTS (inside_init));
else
  return inside_init;
fixes this, but I have trouble to understand what the code really wants to do.
If inside_init is a VECTOR_CST, then if comptypes returns 0, this results
in a checking failure too, as a VECTOR_CST is not a CONSTRUCTOR and
CONSTRUCTOR_ELTS is solely for CONSTRUCTORs.  So, either the inner test should
be only if (TREE_CODE (inside_init) == CONSTRUCTOR) return build_vector ...; else
return inside_init; with no checking for VECTOR_CST etc., or there needs to
be different actions for VECTOR_CST that are compatible, for those where
comptypes on TREE_TYPE (inside_init) and type gives non-zero, but for their
main variants gives zero (is that even possible?), for CONSTRUCTORS and for
other trees that can end up in that place (the testcase here has there a
NOP_EXPR with INTEGER_CST inside).

-- 
           Summary: ICE in digest_init or build_vector
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: jakub at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org,rth at redhat dot com
GCC target triplet: i386-redhat-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19164