[LLVMdev] [llvm-commits] [PATCH] A "very verbose" mode for FileCheck
eliben at google.com
Thu Jan 17 12:34:19 PST 2013
>>> The drawback to your approach is that we'd have to enable this feature
>>> unconditionally - rather than having the optimization advantage of
>>> only dumping files on failure (see my question earlier in the thread,
>>> Eli's concern that dumping output would be expensive, and Dmitri's
>>> response that we'd only be dumping on failure anyway). Given that it
>>> seems the vast majority of our failures aren't flakey, we could have
>>> lit setup to rerun failures in a "create all the temporary files"
>>> mode, though missing flakes would be unfortunate.
>> In my experience, there isn't actually that much performance difference (and
>> sometimes a loss) of using a file instead of a pipe. With all the other
>> stuff going on in the test suite I would be surprised if this added much
>> impact. Note that I personally often write my tests to use temporary files
>> instead of pipes anyway just because I like how the RUN lines look, so its
>> not exactly like we aren't already storing many of these files.
> Fair enough.
> Eli - any counterargument/other views on this issue? (since you'd
> mentioned some concern previously)
I agree with Daniel's suggestion to benchmark it. I guess that no one
really wants our regression tests to run much slower as a result of
this change, so it's OK to wait and actually measure the performance
impact of such a change.
More information about the llvm-dev