Proposal for SentinelInputRange

[deadalnix]

On Thursday, 28 February 2013 at 16:36:44 UTC, Walter Bright 
wrote:
> On 2/28/2013 6:48 AM, Ary Borenszweig wrote:
>> On 2/28/13 1:57 AM, Walter Bright wrote:
>>> On 2/27/2013 8:01 PM, John Colvin wrote:
>>>> Why must sentinel be known at compile time? I don't see 
>>>> what's in the
>>>> way of it
>>>> being a runtime argument.
>>>
>>> Performance!
>>>
>>> It should be usable as a case in a switch statement.
>>
>> Isn't it possible for the optimizer to inline the function 
>> call and then combine
>> the next ifs?
>>
>> if (isSentinel(value)) {
>> } else {
>>   switch(value) {
>>   case ...
>>   }
>> }
>
> 1. I don't know of any C compiler that would do that.
>

As mentioned, LLVM is able to do this kind of things. I have to 
admit I was quite amazed when I discovered it.

> 2. There are other cases, as I pointed out to deadalnix.
>
> 3. You still can't do lookahead without extra checks
>

You need the actual check to do proper generic D code. But the 
compiler is able to optimize it away via inlining and mecanism 
described in 1.

> Once again, guys, have a look at lexer.c at the lines I pointed 
> out.

I see nothing that can't be done by an optimizing compiler. Maybe 
something is, but I can't find it. And the example you pointed me 
aren't.