[Cranelift] rule restore from #13081

[myunbin]

This PR restores the rules that were reverted in #13081.

Those rules were reverted because #13068 reported a significant compile-time slowdown. However, after commit f2807a1 (#13063), that severe slowdown no longer reproduces.

On the current commit, d5f1b973, the median real time is 0.17s on plain main and 0.14s on the restored version. For comparison, the historical slow configuration took about 29.89s median real time on the same input. Full benchmark results are included below.

One likely factor is the vector-heavy section of foo.clif, especially around lines 199-242. The original slowdown was observed before vector constant support landed in #13063, so vector self-subtraction patterns such as x - x -> 0, as well as related rewrite rules like (x - z) - (x - y) -> y - z, may have interacted poorly in the earlier configuration. This does not fully explain the entire e-graph blowup, and we're working on to figure it out.

At least from the current measurements, restoring the rules on top of the current codebase does not bring back the compile-time regression. Based on that result, this PR proposes restoring the deleted rules.

Benchmark results

Environments

• plain_current: commit d5f1b973
• restored_current: d5f1b973 plus the rules / expected outputs removed by #13081
• historical_1578325: historical commit 1578325, where the rules were originally present

real Time Comparison

scenario	commit / branch	runs (s)	median	mean	min	max	stdev	stalled	slow	vs plain
plain_current	d5f1b973	0.18, 0.14, 0.17, 0.15, 0.17	0.17	0.162	0.14	0.18	0.016	0	0	1.00x
restored_current	d5f1b973+restored-13081	0.14, 0.16, 0.14, 0.13, 0.13	0.14	0.140	0.13	0.16	0.012	0	0	0.82x
historical_1578325	1578325	29.59, 29.89, 29.71, 30.65, 30.08	29.89	29.984	29.59	30.65	0.416	5	5	175.82x

user Time Comparison

scenario	commit / branch	runs (s)	median	mean	min	max	stdev
plain_current	d5f1b973	0.17, 0.14, 0.17, 0.13, 0.16	0.16	0.154	0.13	0.17	0.018
restored_current	d5f1b973+restored-13081	0.13, 0.15, 0.14, 0.13, 0.14	0.14	0.138	0.13	0.15	0.008
historical_1578325	1578325	28.87, 29.08, 29.00, 29.90, 29.33	29.08	29.236	28.87	29.90	0.407

sys Time Comparison

scenario	commit / branch	runs (s)	median	mean	min	max	stdev
plain_current	d5f1b973	0.01, 0.00, 0.00, 0.02, 0.01	0.01	0.008	0.00	0.02	0.008
restored_current	d5f1b973+restored-13081	0.01, 0.01, 0.00, 0.00, 0.00	0.00	0.004	0.00	0.01	0.005
historical_1578325	1578325	0.72, 0.81, 0.71, 0.76, 0.76	0.76	0.752	0.71	0.81	0.040

[cfallin]

Thanks! As mentioned in the revert, I think I'd like to see us understand fully what blew up before putting the rules back in -- in general we prefer affirmative understanding of bugs (to make sure we've fixed them) rather than no-longer-reproduces. (We can fall back on the latter eventually if needed, but in this case, since the blowup is deterministic at a certain commit, it should be possible to work out what rule interaction is happening). Once you've got that, I'm happy to put these back in!

[github-actions]

Subscribe to Label Action

cc @cfallin, @fitzgen

Details

This issue or pull request has been labeled: "cranelift", "isle"

Thus the following users have been cc'd because of the following labels:

• cfallin: isle
• fitzgen: isle

To subscribe or unsubscribe from this label, edit the .github/subscribe-to-label.json configuration file.

Learn more.