Improve readability of spy redshift output

[paugier]

For my experiment on loop fusion with SPy, I need to get a nice, human-readable redshift output.

First, it should be possible to ask for a simplified short version of the redshift output. For the loop fusion demo, I used (https://github.com/paugier/spy-demos/blob/loop-fusion/loop-fusion/simplify_rs_output.py):

# `name` is the file name without `.spy`.
for name, content in files.items():
    files[name] = (
        content.replace(name + "::", "")
        .replace("::Self", "")
        .replace(str(Path.home()), "~")
    )

Self is the name of the concrete types created in type factories. We don't need ndarray[f64, 3]::Self, especially when it's repeated several times. This is about the representation of generic function returning a type. A simple solution would be to treat the word Self as special and not display "::Self" in this case.

Second, the redshift output should be formatted to avoid very long lines to obtain something like this rather that the one-line version:

`unsafe::gc_ptr[f64]::getitem_byval`(
    `unsafe::ptr_getfield_byval[unsafe::gc_ptr[f64]]`(
        `ndarray[f64]::__get___ll____`(
            `LazyAdd[f64, ndarray[f64], ndarray[f64]]::__get_left__`(
                `LazyAdd[f64, LazyAdd[f64, ndarray[f64], ndarray[f64]], ndarray[f64]]::__get_left__`(expr))), 'items', 8), i)
+ `unsafe::gc_ptr[f64]::getitem_byval`(
    `unsafe::ptr_getfield_byval[unsafe::gc_ptr[f64]]`(
        `ndarray[f64]::__get___ll____`(
            `LazyAdd[f64, ndarray[f64], ndarray[f64]]::__get_right__`(
                `LazyAdd[f64, LazyAdd[f64, ndarray[f64], ndarray[f64]], ndarray[f64]]::__get_left__`(expr))), 'items', 8), i)
+ `unsafe::gc_ptr[f64]::getitem_byval`(
    `unsafe::ptr_getfield_byval[unsafe::gc_ptr[f64]]`(
        `ndarray[f64]::__get___ll____`(
            `LazyAdd[f64, LazyAdd[f64, ndarray[f64], ndarray[f64]], ndarray[f64]]::__get_right__`(expr)), 'items', 8), i)

[antocuni]

I agree that very long FQNs are not very readable but I cannot think of a "general" fix to improve the situation.

One possible solution is to let the user define a "human-readable FQN map" (e.g. by passing spy redshift --fqn-map=... so you can decide what is the best way to shorten FQNs for the specific case at hand. Not a general solution but if the goal is to document a very specific example, it might be good enough.

+1 also to improve the formatting by generating multiline output.
It would be interesting to apply the same technique also to the C backend, which suffers of the same problem.

[paugier]

2 things:

• It's indeed possible to format with ruff redshift output by replacing few characters: https://github.com/paugier/spy-demos/blob/loop-fusion/loop-fusion/spy_ruff_format.py. It works well and give nice results (at least much nicer than the unformatted version).

• I think it would make sense to replace (by default) the absolute paths in Loc with relative paths.

With these two changes coupled with what is being prepared in #590, the redshift output can be much more readable.

[antocuni]

2 things:

• It's indeed possible to format with ruff redshift output by replacing few characters: https://github.com/paugier/spy-demos/blob/loop-fusion/loop-fusion/spy_ruff_format.py. It works well and give nice results (at least much nicer than the unformatted version).
• I think it would make sense to replace (by default) the absolute paths in Loc with relative paths.

With these two changes coupled with what is being prepared in #590, the redshift output can be much more readable.

+1 to both.

For ruff format, the best way to do that might be to have the SPy backend emitting stuff already in a ruff-compatible format, and THEN convert them into the format that we want for final output (e.g. using backticks for FQNs)