Re: C++ PATCH for c++/91740 - ICE with constexpr call and ?: in ARRAY_REF

[Jason Merrill]

On 9/11/19 4:15 PM, Marek Polacek wrote:
This ICEs since r267272 - more location wrapper nodes, but not because we can't
cope with new location wrappers, but because that commit introduced a call to
maybe_constant_value in cp_build_array_ref.  In this testcase we call it with
   f (VIEW_CONVERT_EXPR<const char[4]>("BAR")) ? 1 : 0

argument and that crashes in fold_convert because we end up trying to convert
"BAR" of type const char[4] to const char * when evaluating the call.  At this
point, decay_conversion hasn't turned the argument into (const char *) "BAR"
yet.

The ICE doesn't occur without :?, because then the call will be wrapped in
NON_DEPENDENT_EXPR and constexpr throws its hands (I'm anthropomorphizing) up
when it encounters such an expression.

I noticed that build_non_dependent_expr doesn't wrap op0 of ?: in N_D_E.  This
is so since r70606 -- Nathan, is there a reason not to do it?  Doing it fixes
this problem.

Bootstrapped/regtested on x86_64-linux, ok for trunk?

2019-09-11  Marek Polacek  <pola...@redhat.com>

        PR c++/91740 - ICE with constexpr call and ?: in ARRAY_REF.
        * pt.c (build_non_dependent_expr): Call build_non_dependent_expr for
        the first operand.

        * g++.dg/cpp1y/var-templ63.C: New test.

diff --git gcc/cp/pt.c gcc/cp/pt.c
index c5915a5ecd0..775389d8245 100644
--- gcc/cp/pt.c
+++ gcc/cp/pt.c
@@ -26709,7 +26709,7 @@ build_non_dependent_expr (tree expr)
    if (TREE_CODE (expr) == COND_EXPR)
      return build3 (COND_EXPR,
                   TREE_TYPE (expr),
-                  TREE_OPERAND (expr, 0),
+                  build_non_dependent_expr (TREE_OPERAND (expr, 0)),
                   (TREE_OPERAND (expr, 1)
                    ? build_non_dependent_expr (TREE_OPERAND (expr, 1))
                    : build_non_dependent_expr (TREE_OPERAND (expr, 0))),

OK.  I wonder why this code copies op0 for the ?: extension rather than 
leave it null?

Jason