search

nim-works / phy

compiler and vm experiments
MIT License
3 stars 2 forks source link

add the `phy` CLI application #61

Closed zerbina closed 3 weeks ago

zerbina commented 3 weeks ago

Summary

Bundle the passes, source2il, the assembler, and the VM into a single CLI application. The idea with it is to help with development, by providing a convenient way to invoke the passes/VM/etc. and inspect the (intermediate) output.

Details

The phy program provides a small command line interface for specifying the source and target language are, what IRs to print, and what to do (only compile or also evaluate).

Much of the implementation is copied from the various test runner:

The program also provides the --runner option, which makes it behave in a way such that it's suitable as being used as a test runner, by lifting some input restrictions and altering the output formatting, so that it can maybe replace the currently used custom test runners at some point.

zerbina commented 3 weeks ago

Replacing the current runners with phy --tester would reduce code duplication (the pass10 - pass1 test runners are slightly adjusted copies of each other), though it also has the downside of introducing some per-test overhead, which could be a problem in practice. It's worth trying out, I think.

zerbina commented 3 weeks ago

@saem: I did consider embedding (i.e., importing the module) phy into the runners, but there's a problem with this approach: for every runner, all passes and their satellite modules would be recompiled, instead of only the parts needed by the runners (which is how it works at the moment).

What I have in mind is to add a feature to the tester where, in addition to the current dedicated programs, a directory can specify a generic command for running the tests. For example, the expr tests would use the following command bin/phy --source:source --target:vm --runner r $file. This approach has the benefit that the phy executable can be compiled once and then be reused for all pass tests.

Since creating the PR, I've also measured how long the command line processing takes for phy --source:L10 --target:vm --runner c file on my machine (where phy was compiled in debug mode with --opt:speed), and it's ~0.015 milliseconds. For comparison, compiling phy.nim without the command line processing and with an empty compilation cache takes ~14.5 seconds. I think it's quite safe to say that we won't reach the amount of tests per runner for embedding the phy module into runners to ever amortize.