testing: exit B.Loop early upon saturation

There's a cap of 1 billion benchmark iterations because more than that
is usually not going to give more useful data. Unfortunately, the
existing implementation neglected to check whether the 1e9 cap had
already been exceeded when it adjusted the number of iterations in the
B.Loop slow path (stopOrScaleBLoop), since it's only when that cap is hit
that it needed to terminate early.

As a result, for _very_ cheap benchmarks (e.g. testing assembly
implementations with just a few instructions), the B.Loop would stop
incrementing the number of iterations, but wouldn't terminate early,
making it re-enter the slow-path _every_ iteration until the benchmark
time was exhausted.

This wasn't normally visible with the default -benchtime 2s, but when
raised to 5s, it would cause benchmarks that took <5ns/op to be reported
as exactly 5ns/op. (which looks a bit suspicious)

Notably, one can use -count for larger groupings to compute statistics.
golang.org/x/perf/cmd/benchstat is valuable for coalescing larger
run-counts from -count into more useful statistics.

Add a test which allows for fewer iterations on slow/contended
platforms but guards against reintroducing a bug of this nature.

Fixes #75210

Change-Id: Ie7f0b2e6c737b064448434f3ed565bfef8c4f020
Reviewed-on: https://go-review.googlesource.com/c/go/+/700275
Reviewed-by: Junyang Shao <shaojunyang@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Damien Neil <dneil@google.com>
Reviewed-by: Sean Liao <sean@liao.dev>
Auto-Submit: Sean Liao <sean@liao.dev>

0 files changed, 0 insertions, 0 deletions