r226870 - Make the ?: precedence warning handle pointers to the left of ?

[Hans Wennborg]

On Tue, Jan 27, 2015 at 3:15 PM, David Blaikie <dblaikie at gmail.com> wrote:
>
>
> On Tue, Jan 27, 2015 at 9:29 AM, Hans Wennborg <hans at chromium.org> wrote:
>>
>> On Thu, Jan 22, 2015 at 4:55 PM, Hans Wennborg <hans at chromium.org> wrote:
>> > On Thu, Jan 22, 2015 at 2:43 PM, David Blaikie <dblaikie at gmail.com>
>> > wrote:
>> >>
>> >>
>> >> On Thu, Jan 22, 2015 at 2:11 PM, Hans Wennborg <hans at hanshq.net> wrote:
>> >>>
>> >>> Author: hans
>> >>> Date: Thu Jan 22 16:11:56 2015
>> >>> New Revision: 226870
>> >>>
>> >>> URL: http://llvm.org/viewvc/llvm-project?rev=226870&view=rev
>> >>> Log:
>> >>> Make the ?: precedence warning handle pointers to the left of ?
>> >>>
>> >>> Previously, Clang would fail to warn on:
>> >>>
>> >>>   int n = x + foo ? 1 : 2;
>> >>>
>> >>> when foo is a pointer.
>> >>>
>> >>> Modified:
>> >>>     cfe/trunk/lib/Sema/SemaExpr.cpp
>> >>>     cfe/trunk/test/Sema/parentheses.c
>> >>>
>> >>> Modified: cfe/trunk/lib/Sema/SemaExpr.cpp
>> >>> URL:
>> >>>
>> >>> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Sema/SemaExpr.cpp?rev=226870&r1=226869&r2=226870&view=diff
>> >>>
>> >>>
>> >>> ==============================================================================
>> >>> --- cfe/trunk/lib/Sema/SemaExpr.cpp (original)
>> >>> +++ cfe/trunk/lib/Sema/SemaExpr.cpp Thu Jan 22 16:11:56 2015
>> >>> @@ -6129,6 +6129,8 @@ static bool ExprLooksBoolean(Expr *E) {
>> >>>      return IsLogicOp(OP->getOpcode());
>> >>>    if (UnaryOperator *OP = dyn_cast<UnaryOperator>(E))
>> >>>      return OP->getOpcode() == UO_LNot;
>> >>> +  if (E->getType()->isPointerType())
>> >>
>> >>
>> >> Could we generalize this a bit further, somehow? (I haven't looked at
>> >> the
>> >> code in question, but it sounds like this should use some more general
>> >> tool
>> >> of "try to apply contextual conversion to bool" so that it matches the
>> >> actual semantic situation we're interested in here)
>> >
>> > It's tricky, because we don't really want to match the actual
>> > semantics, we want to figure out if the intention was to use 'foo' as
>> > a conditional expression. That's what 'ExprLooksBoolean' does, and
>> > it's erring on the side of caution.
>> >
>> > For example, we don't want to warn if 'foo' is int, even if that could
>> > be used as a conditional expression. But we do want to warn if 'foo'
>> > is 'a==b', even in C where the type of that expression is int.
>> >
>> > Having said that, I now realised I might have made the warning a
>> > little too broad.. we want to warn here:
>> >
>> >   int x = x + ptr ? 1 : 0;
>> >
>> > because 'x + ptr' seems like a pretty unlikely condition expression.
>> >
>> > But we don't want to warn here:
>> >
>> >   int y = ptr1 - ptr2 ? 1 : 0;
>> >
>> > I'll think about that.
>>
>> The last example actually showed up in code yesterday:
>>
>> http://git.savannah.gnu.org/cgit/emacs.git/tree/src/emacs.c?id=emacs-24.4#n2307
>>
>> The reason I said I didn't want to warn here is because the expression
>> can only be interpreted one way. For example,
>>
>> int y = ptr1 - (ptr2 ? 1 : 0);
>>
>> would not compile because the type has changed.
>>
>> On the other hand, parentheses would certainly make it more readable,
>> and that is what the warning suggests:
>>
>> int y = (ptr1 - ptr2) ? 1 : 0;
>>
>> So I'm now thinking maybe warning here is the right thing, just as
>> -Wparentheses warns about 'a && b || c'.
>
>
> Perhaps - though &&|| is trickier because it could easily be either way. As
> you point out with ?: it can't actually be the other way sometimes. So
> perhaps we could/should have the smarts to detect that case, but I'm not
> sure where the effort tipping point is.

I think it would be tricky. For

  int y = ptr1 - ptr2 ? 1 : 0;

we'd essentially have to tentatively build the 'ptr1 - (ptr2 ? 1 : 0)'
expression and check if that's legal. I don't think Clang provides an
easy way to do that :-/