### Problem
ubuntu-latest was updated to 22.04 which removes clang++-10 we used for
coverage stats and creates a pre-compiled binary that requires a glibc
upgrade https://github.com/Roblox/luau/issues/773
### Solution
Pin to ubuntu-20.04 using multi-value matrix configurations.
In coverage configuration, we use clang++-10 once again.
We don't need to run any cachegrind benchmarks in benchmark-dev, since
benchmark uses our new callgrind setup instead.
Also removes prototyping filters that we no longer need from all builds.
Resolves#668
## The problem
Benchmarks jobs run concurrently for the different operating systems.
This means that when it comes time to push the benchmark results to [the
assigned benchmark results
repo](https://github.com/luau-lang/benchmark-data), there can be two
different jobs trying to push changes at the same time. In such a case,
one of the pushes will fail and we end up missing some benchmark results
data from the workflow run.
## The solution
Whenever a push fails, we need to retry the steps leading up to the push
(checking out the benchmark results repo, storing benchmark results,
pushing the results to [a specific
repo](https://github.com/luau-lang/benchmark-data)).
### Note
There are 3 push attempts before submitting to failure.
## TL;DR
This PR retries pushing benchmark results when they fail to get pushed
(often due to pushing from multiple jobs concurrently)
Co-authored-by: Jamie Kuppens <reshurum@gmail.com>
Co-authored-by: Ignacio Falk <flakolefluk@gmail.com>
This change adds another file for benchmarking luau-analyze and sets up
benchmarks for both non-strict/strict modes for analysis and all three
optimization levels for compilation performance.
To avoid issues with race conditions on repository update we do all this
in the same job in benchmark.yml.
To be able to benchmark both modes from a single file, luau-analyze
gains --mode argument which allows to override the default typechecking
mode. Not sure if we'll want this to be a hard override on top of the
module-specified mode in the future, but this works for now.
Since callgrind allows to control stats collection from the guest, this
allows us to reset the collection right before the benchmark starts.
This change exposes this to the benchmark runner and integrates
callgrind data parsing into bench.py, so that we can run bench.py with
--callgrind argument and, as long as the runner was built with callgrind
support, we get instruction counts from the run.
We convert instruction counts to seconds using 10G instructions/second
rate; there's no correct way to do this without simulating the full CPU
pipeline but it results in time units on a similar scale to real runs.
