[EXPERIMENT] GVN: Track known value ranges through assert and switchInt.

[dianqk]

r? ghost (for perf)

[dianqk]

@bors try @rust-timer queue

[rust-bors]

☀️ Try build successful (CI)
Build commit: f2bc50f (f2bc50f4689333c80bcd074203a0a387187d19d8, parent: d3f16d2f35d8d4b25bb15a85ae869eee70ffb4cd)

[rust-timer]

Finished benchmarking commit (f2bc50f): comparison URL.

Overall result: ❌✅ regressions and improvements - please read the text below

Benchmarking this pull request means it may be perf-sensitive – we'll automatically label it not fit for rolling up. You can override this, but we strongly advise not to, due to possible changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please do so in sufficient writing along with @rustbot label: +perf-regression-triaged. If not, please fix the regressions and do another perf run. If its results are neutral or positive, the label will be automatically removed.

@bors rollup=never
@rustbot label: -S-waiting-on-perf +perf-regression

Instruction count

Our most reliable metric. Used to determine the overall result above. However, even this metric can be noisy.

	mean	range	count
Regressions ❌ (primary)	0.3%	[0.1%, 0.4%]	3
Regressions ❌ (secondary)	0.5%	[0.5%, 0.5%]	1
Improvements ✅ (primary)	-0.4%	[-0.9%, -0.2%]	9
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-0.2%	[-0.9%, 0.4%]	12

Max RSS (memory usage)

Results (primary -1.1%, secondary 0.8%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	2.5%	[1.9%, 3.1%]	2
Regressions ❌ (secondary)	2.4%	[2.4%, 2.4%]	1
Improvements ✅ (primary)	-2.9%	[-4.2%, -1.4%]	4
Improvements ✅ (secondary)	-0.7%	[-0.7%, -0.7%]	1
All ❌✅ (primary)	-1.1%	[-4.2%, 3.1%]	6

Cycles

This benchmark run did not return any relevant results for this metric.

Binary size

Results (primary -0.0%, secondary -0.1%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	0.3%	[0.1%, 0.4%]	5
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.1%	[-0.8%, -0.0%]	16
Improvements ✅ (secondary)	-0.1%	[-0.8%, -0.0%]	19
All ❌✅ (primary)	-0.0%	[-0.8%, 0.4%]	21

Bootstrap: 481.777s -> 483.166s (0.29%)
Artifact size: 390.59 MiB -> 390.61 MiB (0.01%)

[bors]

☔ The latest upstream changes (presumably #150231) made this pull request unmergeable. Please resolve the merge conflicts.

[dianqk]

I don't think GVN is required here. A new pass with SSA locals may make sense. @cjgillot What do you think?

[cjgillot]

This is wrong. Codegen can turn overflow checks into nops.

[dianqk]

IIUC, they should be handled in instsimplify. This just does some useless work.

[cjgillot]

This PR is very interesting. From a cursory glance, I see 2 features in this PR:

1. computing value ranges in GVN;
2. handling assumes and other dataflow-dependent equations.

I have no issue about (1). My recommendation would be to copy what eval_to_const does.

I'm a bit skeptical about (2). GVN is designed to handle structural equality that is independent of dataflow. So I'm not sure how the two are going to fit together in full generality.

For this specific case, my first reflex would be to finish DataflowConstProp (= make it fast enough) and generalize it to ranges to make it worth the cost. This is pretty much what is advised in #76579.

But a separate SSA pass could be a much simpler and straightforward way to get some of the same improvements.

Do you have an idea where those improvements come from? Assertions disappearing? Switches in cleanup blocks disappearing?

[dianqk]

Do you have an idea where those improvements come from? Assertions disappearing? Switches in cleanup blocks disappearing?

Most matched patterns in std are bool. I guess they are from assertions.