Proposal for SentinelInputRange

[Dmitry Olshansky]

28-Feb-2013 21:38, deadalnix пишет:
> On Thursday, 28 February 2013 at 17:13:36 UTC, Dmitry Olshansky wrote:
>> No as a compiler will take it (or may depending on its brain) that 0
>> is what you want to test *first*. It may speed things up if branch is
>> almost always taken but its not the case with sentinel. Thus its jsut
>> dead code that needs to be decoded, evalutated and skipped (as
>> predicated) *before* doing switch jump.
>>
>> In fact some people avoid the overhead of switch by placing one or two
>> of highly-frequent branches with tests before the switch (thus
>> avoiding indirect branch it entails in these frequent cases).
>
> That won't work as expected with LLVM and full optimizations, as it will
> combine everything in a switch, unless you use branch weight, in which
> case it can do the reverse : extract common cases from the switch.
>
> See : http://llvm.org/docs/BranchWeightMetadata.html
>
> GCC does something similar.
>
> This is another example of how people think they are doing high level
> assembly when in fact, the compiler is changing everything behind their
> back, and not doing at all what they think.

Depends on the brains of compiler like I said. BTW it can't know the 
most frequent branches so jump-table may be actually slower (in some 
cases though they are niche).

-- 
Dmitry Olshansky