feat(veir-opt): teach optimizer about collections of passes

[regehr]

I would like the optimizer to understand about common groups of passes that are used together to achieve some effect, so we don't have to have different versions of this sitting around in different parts of the repo.

I think two things are up for debate here:

• what are the actual passes we want to invoke for "-O" and also for the RISC-V backend? I've taken some guesses and am open to feedback
• do we want this option to compose with other pass options, so that we can build larger pipelines? for now I've gone with "no" to keep things simple

[github-actions]

VeIR Benchmarks

Details

Benchmark suite	Current: 964b643	Previous: 1971543	Ratio
add-fold-worklist/create	1839000 ns (± 68678)	2264000 ns (± 89652)	0.81
add-fold-worklist/rewrite	3672000 ns (± 73387)	4027000 ns (± 79991)	0.91
add-fold-worklist-local/create	1862000 ns (± 128621)	2225000 ns (± 106345)	0.84
add-fold-worklist-local/rewrite	3494000 ns (± 116613)	3716000 ns (± 61682)	0.94
add-zero-worklist/create	1852000 ns (± 14018)	2226000 ns (± 51046)	0.83
add-zero-worklist/rewrite	2304000 ns (± 31196)	2600000 ns (± 51267)	0.89
add-zero-reuse-worklist/create	1563000 ns (± 22554)	1889000 ns (± 90843)	0.83
add-zero-reuse-worklist/rewrite	1950000 ns (± 36074)	2127000 ns (± 49138)	0.92
mul-two-worklist/create	1849000 ns (± 15758)	2209000 ns (± 62608)	0.84
mul-two-worklist/rewrite	5092000 ns (± 50251)	5620000 ns (± 248024)	0.91
add-fold-forwards/create	1865000 ns (± 14839)	2279000 ns (± 74547)	0.82
add-fold-forwards/rewrite	2702000 ns (± 32445)	3063000 ns (± 51863)	0.88
add-zero-forwards/create	1842000 ns (± 87634)	2163000 ns (± 99039)	0.85
add-zero-forwards/rewrite	1741000 ns (± 28726)	1940000 ns (± 16682)	0.90
add-zero-reuse-forwards/create	1544000 ns (± 9290)	1902000 ns (± 49987)	0.81
add-zero-reuse-forwards/rewrite	1387000 ns (± 12268)	1574000 ns (± 33890)	0.88
mul-two-forwards/create	1987000 ns (± 80341)	2239000 ns (± 86176)	0.89
mul-two-forwards/rewrite	3476000 ns (± 103840)	3723000 ns (± 102624)	0.93
add-zero-reuse-first/create	1545500 ns (± 91981)	1837000 ns (± 97092)	0.84
add-zero-reuse-first/rewrite	10000 ns (± 2289)	8000 ns (± 1821)	1.25
add-zero-lots-of-reuse-first/create	1608000 ns (± 74869)	1924000 ns (± 83197)	0.84
add-zero-lots-of-reuse-first/rewrite	781500 ns (± 24897)	801000 ns (± 13405)	0.98

This comment was automatically generated by workflow using github-action-benchmark.

[tobiasgrosser]

I love it. Let's see what @luisacicolini @math-fehr and @nchappe say.

[math-fehr]

I'm not a big fan of this particular design, because as you said you cannot compose it with other groups.
Maybe it would be better to just create passes that are collection of passes, so that we can compose everything? Also, it should be fairly easy to implement

[regehr]

I'm not a big fan of this particular design, because as you said you cannot compose it with other groups. Maybe it would be better to just create passes that are collection of passes, so that we can compose everything? Also, it should be fairly easy to implement

take another look please @math-fehr ?

[math-fehr]

Sorry, what I meant is having both -p and -pgrp in the same option (like -p).
So you would have -p=canonicalize,O,canonicalize.
I think it would make this more readable and easy to understand?

[regehr]

@math-fehr ok look one more time please!