On Apr 20, 2014, at 13:19, Mark Rowe <mr...@apple.com> wrote:
>
> On Apr 20, 2014, at 13:08, Filip Pizlo <fpi...@apple.com> wrote:
>
>>
>>
>> On Apr 20, 2014, at 12:56 PM, Mark Rowe <mr...@apple.com> wrote:
>>
>>>
>>> On Apr 19, 2014, at 13:09, Filip Pizlo <fpi...@apple.com> wrote:
>>>
>>>> Hey everyone,
>>>>
>>>> When guarding code with macros that are always defined, such as
>>>> ASSERT_DISABLED (it's always either 0 or 1), we have a choice between:
>>>>
>>>> if (!ASSERT_DISABLED) {
>>>> // do things
>>>> }
>>>>
>>>> and:
>>>>
>>>> #if !ASSERT_DISABLED
>>>> // do things
>>>> #endif
>>>>
>>>> I'd like to propose that anytime the normal if would be semantically
>>>> equivalent to the preprocessor #if, the normal if should be used.
>>>>
>>>> We don't lose any compiler optimization, since even at -O0, the compiler
>>>> will constant fold the normal if. We do gain a lot of clarity, since the
>>>> control flow of normal if statements is subject to proper indentation.
>>>>
>>>> The "semantically equivalent" requirement still allows for #if to be used
>>>> for thinngs like:
>>>>
>>>> - Guarding the placement of fields in a class.
>>>> - Guarding the definitions of other macros (in the case of
>>>> ASSERT_DISABLED, we define ASSERT in different ways guarded by #if's)
>>>> - Guarding the definition/declaration/inclusion of entire functions,
>>>> classes, and other top-level constructs.
>>>>
>>>> Thoughts?
>>>
>>> It’d be one thing if we could adopt this approach everywhere, but as you
>>> note there are numerous situations in which it won’t compile. What’s worse
>>> is that there are situations in which it’ll silently give unintended
>>> behavior.
>>
>> Can you give an example of this?
>
> Won’t compile:
>
> if (!MAYBE_DEFINED) {
> …
> }
>
> if (SOMETHING) {
> void doSomething() { }
> }
>
> if (SOMETHING) {
> #include “Something.h"
> }
It’s also not possible to determine from context alone whether #if or if should
be used. When #if is used to remove members or functions you’re then forced to
use #if, rather than if, around all uses of those members or functions since
the alternative won’t compile.
An example lifted from Vector.h:
#if !ASSERT_DISABLED
template<typename T> struct ValueCheck<Vector<T>> {
typedef Vector<T> TraitType;
static void checkConsistency(const Vector<T>& v)
{
v.checkConsistency();
}
};
#endif
template<typename T, size_t inlineCapacity, typename OverflowHandler>
inline void Vector<T, inlineCapacity, OverflowHandler>::checkConsistency()
{
if (!ASSERT_DISABLED) {
for (size_t i = 0; i < size(); ++i)
ValueCheck<T>::checkConsistency(at(i));
}
}
This won’t compile with assertions disabled because
ValueCheck::checkConsistency doesn’t exist.
- Mark
_______________________________________________
webkit-dev mailing list
webkit-dev@lists.webkit.org
https://lists.webkit.org/mailman/listinfo/webkit-dev
