[llvm-commits] llvm-test patch for review

[Chris Lattner]

>>> I would prefer better implementations of the "millisec" functions  
>>> you
>>> added in some tests. Counting the number of calls isn't accurate  
>>> given
>>> that you could at least call time(0) and multiply by 1000 (after
>>> subtracting
>>> the baseline value).
>>
>> I don't think that we need a precise version of millisec in the
>> benchmarks, we use external timing instead of internal timing, so the
>> timer functions in these programs are dynamically dead or their
>> output is ignored.
>
> The functions involved don't have anything to do with our timing
> requirements, they are part of the benchmark. On Unix they will do
> getrusage. I'm simply trying to make the MINGW32 version do  
> something at
> least comparable. The MINGW32 replacements that Anton provided just
> returned a monotonic counter. I'm concerned that the benchmark's
> functioning could be skewed by this.
>
> The timer functions are not dead, the benchmarks use them.

I believe the benchmarks do call them, but if they never print out or  
change behavior based on the functions, then we don't care what they  
do.  If the benchmark *does* act on the value returned, then the  
benchmark should be fixed.

-Chris