On 08/26/2016 03:48 PM, David Malcolm wrote:
I'm nervous about the build_random_cfg function: randomness in
selftests seems like a double-edged sword. On the one hand, we can use
it to fuzz-test an optimization to rapidly gain a lot of coverage. On
the other hand, does every host generate the same sequence of values?
Presumably we want everyone to be effectively running the same
selftests.
I intended to put a srandom call in there at the start but forgot. One
could argue that there's value in having different tests on different
hosts, as long as they are stable between runs, but...
Maybe there's a need for some kind of selftest::rng class here?
That might be a good idea too.
On a unrelated note, should the various vfunc implementations be marked
with "FINAL OVERRIDE" (from coretypes.h), so that the compiler has a
better chance of devirtualizing them in a release build?
Wasn't aware of that - will have a look.
Maybe we could also use a long-running "-fselftest=extensive" for
stress-testing algorithms such as this (increasing the number of random
cfgs it tries)?
Bernd
