RFA: PATCH to initializer_constant_valid_p for c++/38880

[Ian Lance Taylor]

Jason Merrill <jason@redhat.com> writes:

> Fixing this bug in either fold or the front end had the same problem
> of making the pseudo-offsetof expression look like a valid C++
> constant-expression, when it should not be.  We just want it to be
> accepted by initializer_constant_valid_p.  As it turns out, we just
> need to make the existing code accept PLUS_EXPR as well as
> POINTER_PLUS_EXPR, now that we no longer fold the former into the
> latter in this case.
>
> Ian, was there any particular reason for limiting narrowing support to
> POINTER_PLUS_EXPR and MINUS_EXPR, but not PLUS_EXPR?  The testcase was
> failing because we fold the MINUS_EXPR to a PLUS_EXPR with the second
> operand negated.

Only recognizing POINTER_PLUS_EXPR was a conservative change.  The old
code only handled MINUS_EXPR.  With the introduction of
POINTER_PLUS_EXPR, ptr - constant was handled as POINTER_PLUS_EXPR
rather than MINUS_EXPR.  So I extended the old code to handle
POINTER_PLUS_EXPR as well as MINUS_EXPR.  I did not handle PLUS_EXPR
because the old code didn't handle it.

If we can get a PLUS_EXPR for something which subtracts a constant from
a pointer (adds a negative constant to a pointer) then handling
PLUS_EXPR makes sense to me.

However, your patch does mean that
    char *p1;
    short a[] = { p1 + p1 };
may do something unexpected.  It would be conservative to only handle
PLUS_EXPR when the second operand is a constant.  And to then wonder why
POINTER_PLUS_EXPR wasn't used instead of PLUS_EXPR.

Ian