[C++] Fix constexpr ICE with poor man's offsetof (PR c++/86738)

Hi!

Some time ago, I've moved the poor man's offsetof recognizing hack to
cp_fold.  On the following testcase that means we don't fold it early during
parsing.  Now, if we try to evaluate those inside of constexpr contexts
with !ctx->quiet, it is diagnosed as invalid (but *non_constant_p is not
set).  Worse, with ctx->quiet, we pretend it is a constant expression but
don't actually fold it, and then we have e.g. in array index evaluation
  VERIFY_CONSTANT (index);
...
  VERIFY_CONSTANT (nelts);
  if ((lval
       ? !tree_int_cst_le (index, nelts)
       : !tree_int_cst_lt (index, nelts))
      || tree_int_cst_sgn (index) < 0)
    {
      diag_array_subscript (ctx, ary, index);
      *non_constant_p = true;
      return t;
    }
where VERIFY_CONSTANT is happy about it, even when index is not an
INTEGER_CST, but large complex TREE_CONSTANT expression.  The above though
assumes INTEGER_CST.  Perhaps we should check for INTEGER_CST somewhere (and
in other similar code too), but it isn't clear to me what exactly we should
do if those trees aren't INTEGER_CSTs, especially with !ctx->quiet.

This patch changes a different thing, the usual case (including other spots
for NULL pointer dereferences or arith) in constexpr.c is
  if (some condition)
    {
      if (!ctx->quiet)
	error* (...);
      *non_constant_p = true;
      return t;
    }
but the following two spots were different and that caused the array
handling to see those complex unsimplified constant expressions.
With this, it is not treated as constant for maybe_constant_value etc.
purposes, though following cp_fold can still fold it into a constant.

Bootstrapped/regtested on x86_64-linux and i686-linux (including
check-c++-all testing on both), ok for trunk and 8.3 after a while?

2018-08-03  Jakub Jelinek  <jakub@redhat.com>

	PR c++/86738
	* constexpr.c (cxx_eval_binary_expression): For arithmetics involving
	NULL pointer set *non_constant_p to true.
	(cxx_eval_component_reference): For dereferencing of a NULL pointer,
	set *non_constant_p to true and return t.

	* g++.dg/opt/pr86738.C: New test.


	Jakub

OK

On Sat, Aug 4, 2018 at 2:18 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> Some time ago, I've moved the poor man's offsetof recognizing hack to
> cp_fold.  On the following testcase that means we don't fold it early during
> parsing.  Now, if we try to evaluate those inside of constexpr contexts
> with !ctx->quiet, it is diagnosed as invalid (but *non_constant_p is not
> set).  Worse, with ctx->quiet, we pretend it is a constant expression but
> don't actually fold it, and then we have e.g. in array index evaluation
>   VERIFY_CONSTANT (index);
> ...
>   VERIFY_CONSTANT (nelts);
>   if ((lval
>        ? !tree_int_cst_le (index, nelts)
>        : !tree_int_cst_lt (index, nelts))
>       || tree_int_cst_sgn (index) < 0)
>     {
>       diag_array_subscript (ctx, ary, index);
>       *non_constant_p = true;
>       return t;
>     }
> where VERIFY_CONSTANT is happy about it, even when index is not an
> INTEGER_CST, but large complex TREE_CONSTANT expression.  The above though
> assumes INTEGER_CST.  Perhaps we should check for INTEGER_CST somewhere (and
> in other similar code too), but it isn't clear to me what exactly we should
> do if those trees aren't INTEGER_CSTs, especially with !ctx->quiet.
>
> This patch changes a different thing, the usual case (including other spots
> for NULL pointer dereferences or arith) in constexpr.c is
>   if (some condition)
>     {
>       if (!ctx->quiet)
>         error* (...);
>       *non_constant_p = true;
>       return t;
>     }
> but the following two spots were different and that caused the array
> handling to see those complex unsimplified constant expressions.
> With this, it is not treated as constant for maybe_constant_value etc.
> purposes, though following cp_fold can still fold it into a constant.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux (including
> check-c++-all testing on both), ok for trunk and 8.3 after a while?
>
> 2018-08-03  Jakub Jelinek  <jakub@redhat.com>
>
>         PR c++/86738
>         * constexpr.c (cxx_eval_binary_expression): For arithmetics involving
>         NULL pointer set *non_constant_p to true.
>         (cxx_eval_component_reference): For dereferencing of a NULL pointer,
>         set *non_constant_p to true and return t.
>
>         * g++.dg/opt/pr86738.C: New test.
>
> --- gcc/cp/constexpr.c.jj       2018-07-31 23:57:24.193432388 +0200
> +++ gcc/cp/constexpr.c  2018-08-03 14:54:13.302817282 +0200
> @@ -2082,6 +2082,7 @@ cxx_eval_binary_expression (const conste
>      {
>        if (!ctx->quiet)
>         error ("arithmetic involving a null pointer in %qE", lhs);
> +      *non_constant_p = true;
>        return t;
>      }
>    else if (code == POINTER_PLUS_EXPR)
> @@ -2522,9 +2523,13 @@ cxx_eval_component_reference (const cons
>                                              lval,
>                                              non_constant_p, overflow_p);
>    if (INDIRECT_REF_P (whole)
> -      && integer_zerop (TREE_OPERAND (whole, 0))
> -      && !ctx->quiet)
> -    error ("dereferencing a null pointer in %qE", orig_whole);
> +      && integer_zerop (TREE_OPERAND (whole, 0)))
> +    {
> +      if (!ctx->quiet)
> +       error ("dereferencing a null pointer in %qE", orig_whole);
> +      *non_constant_p = true;
> +      return t;
> +    }
>
>    if (TREE_CODE (whole) == PTRMEM_CST)
>      whole = cplus_expand_constant (whole);
> --- gcc/testsuite/g++.dg/opt/pr86738.C.jj       2018-08-03 15:03:51.477358712 +0200
> +++ gcc/testsuite/g++.dg/opt/pr86738.C  2018-08-03 15:02:51.940201694 +0200
> @@ -0,0 +1,12 @@
> +// PR c++/86738
> +// { dg-do compile }
> +
> +struct S { int s; };
> +unsigned char a[20];
> +unsigned char *p = &a[(__UINTPTR_TYPE__) &((S *) 0)->s];
> +
> +void
> +foo ()
> +{
> +  __builtin_memcpy (&a[15], &a[(__UINTPTR_TYPE__) &((S *) 0)->s], 2);
> +}
>
>         Jakub