Mir-Opt for copying enums with large discrepancies #85158

JulianKnodt · 2021-05-10T20:40:26Z

I have been meaning to make this for quite a while, based off of this hackmd.

I'm not sure where to put this opt now that I've made it, so I'd appreciate suggestions on that!
It's also one long chain of statements, not sure if there's a more friendly format to make it.

r? @tmiasko
I would r oli but he's on leave so he suggested I r tmiasko or wesleywiser.

compiler/rustc_mir/src/transform/large_enums.rs

tmiasko

Hmm, so the idea is to transform: _a = _b; into:

_index = discriminant(_b);
_bytes  = [size of each variant in bytes][_index]
CopyNonOverlapping(src: &_b as *const u8, dst: &_a as *mut u8, count: _bytes);

The transformation is localized to a single assignment, requires layout information, some aspects of it feel non-trivial to express directly in MIR (e.g., preserving the alignment). Maybe codegen would be a better fit?

compiler/rustc_mir/src/transform/large_enums.rs

tmiasko · 2021-05-12T15:46:39Z

This general pattern for copying the data appears to be opaque to memcpy optimizations, i.e., every single transformed copy remains. Any ideas how this could be improved?

scottmcm · 2021-05-12T16:31:32Z

Maybe codegen would be a better fit?

That would mean it would have the monomorphized versions of the enum too, right? Could be handy for anything using a generic Result and such...

JulianKnodt · 2021-05-12T18:04:35Z

Mmmm I've not written anything for codegen at all, but if that makes more sense I can close this and move it to that, altho seeing a guide before it would be helpful.

tmiasko · 2021-05-12T19:15:48Z

As you prefer. The current implementation is not that far from the point where we could run some tests on it. The one missing component is a mapping from a discriminant to a variant index (or limiting the transformation to the cases where there is direct correspondence between the two).

compiler/rustc_mir/src/transform/large_enums.rs

JulianKnodt · 2021-05-12T22:39:46Z

ah I thought reading from the array of sizes was handled around line 132, but I guess mapping discriminants to variant idxes is not guaranteed to be 1-1, is there anyway to check that?

bjorn3 · 2021-05-13T07:54:12Z

but I guess mapping discriminants to variant idxes is not guaranteed to be 1-1, is there anyway to check that?

I believe

enum Foo {
    Bar = 2,
    Baz = 0,
}

has variant index 0 for Bar, but discriminant 2 and variant index 1 for Baz, but discriminant 0.

compiler/rustc_mir/src/transform/large_enums.rs

tmiasko · 2021-05-13T09:49:30Z

mapping discriminants to variant idxes is not guaranteed to be 1-1, is there anyway to check that?

The InterpCx::read_discriminant exercises related API and contains explanatory comments, might be good starting point.

camelid · 2021-05-28T00:41:00Z

Does this fix #54360?

JulianKnodt · 2021-05-28T02:12:50Z

I believe so? I don't remember the original issue.

tmiasko · 2021-05-28T16:27:50Z

compiler/rustc_mir/src/transform/large_enums.rs

+                    Variants::Multiple { variants, .. } => {
+                        let min = variants.iter().map(|v| v.size).min().unwrap();
+                        let max = variants.iter().map(|v| v.size).max().unwrap();
+                        if max.bytes() - min.bytes() < D {


The max - min criterion is based on the most optimistic scenario.

Consider two enums with the same maximum absolute deviation in variant sizes. The first enum has one large outlier and remaining variants are small, the second enum conversely has one small outlier and remaining variants are large. Both enums would be be equally good candidates under current criterion, but the first one would seem like a much better candidate.

What about assuming that variants are uniformly distributed and calculating expected reduction in the number of bytes copied?

oops I never directly addressed this, I'm not sure if uniform distribution makes a ton of sense, as I'd expect if this is hit that one variant is probably significantly larger than the others. Probably something to experiment with.

tmiasko · 2021-05-28T16:33:53Z

In terms of placement in MIR pipeline, I would put it towards the end, somewhere after SimplifyLocals, maybe just last? I wouldn't expect it to create any new optimization opportunities.

compiler/rustc_mir/src/transform/large_enums.rs

Changing a bunch of struct constructors to `from`, no extra destructuring, getting the type of the discriminant.

Since we're changing a bunch of stuff, necessary to remove some codegen tests which look for specific things. Also attempting to restart a test which timed out, maybe due to fastly failing?

Instead of storing an extra array for discriminant values, create an allocation there and store those in an allocation immediately.

There is a distinction between running this on wasm and i686, even though they should be identical. This technically is not _incorrect_, it's just an unexpected difference, which is worth investigating, but not for correctness.

compiler/rustc_mir_transform/src/large_enums.rs

JulianKnodt · 2023-02-08T05:48:42Z

Should be ok to retry submitting again at this point

cjgillot · 2023-02-10T21:01:14Z

@bors r+

bors · 2023-02-10T21:01:16Z

📌 Commit 15d4728 has been approved by cjgillot

It is now in the queue for this repository.

bors · 2023-02-10T21:49:34Z

⌛ Testing commit 15d4728 with merge 5a8dfd9...

bors · 2023-02-11T00:50:54Z

☀️ Test successful - checks-actions
Approved by: cjgillot
Pushing 5a8dfd9 to master...

rust-timer · 2023-02-11T02:45:19Z

Finished benchmarking commit (5a8dfd9): comparison URL.

Overall result: no relevant changes - no action needed

@rustbot label: -perf-regression

Instruction count

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

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-2.2%	[-2.2%, -2.2%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-2.2%	[-2.2%, -2.2%]	1

Cycles

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

This comment has been minimized.

Sign in to view

LingMan reviewed May 11, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

scottmcm reviewed May 11, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

JulianKnodt force-pushed the array_const_val branch from 28da2d2 to 8d8c374 Compare May 11, 2021 15:08

tmiasko reviewed May 12, 2021

View reviewed changes

JulianKnodt force-pushed the array_const_val branch from 8d8c374 to b65bb73 Compare May 12, 2021 15:38

tmiasko reviewed May 12, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

JulianKnodt force-pushed the array_const_val branch from b65bb73 to 6eaa543 Compare May 12, 2021 22:28

tmiasko reviewed May 13, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

camelid added the S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. label May 28, 2021

camelid assigned tmiasko May 28, 2021

camelid added the A-mir-opt Area: MIR optimizations label May 28, 2021

JulianKnodt force-pushed the array_const_val branch from 6eaa543 to d6b74be Compare May 28, 2021 04:18

This comment has been minimized.

Sign in to view

JulianKnodt force-pushed the array_const_val branch from d6b74be to 90f8c1c Compare May 28, 2021 04:37

This comment has been minimized.

Sign in to view

JulianKnodt force-pushed the array_const_val branch from 90f8c1c to ad67765 Compare May 28, 2021 04:51

tmiasko reviewed May 28, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

tmiasko reviewed May 28, 2021

View reviewed changes

compiler/rustc_mir/src/transform/large_enums.rs Outdated Show resolved Hide resolved

JulianKnodt added 6 commits February 7, 2023 09:37

Clean up MIR transform

33b4d20

Update with comments

f7cbf2e

Changing a bunch of struct constructors to `from`, no extra destructuring, getting the type of the discriminant.

Set mir-opt-level = 0 on some codegen tests

3e97cef

Since we're changing a bunch of stuff, necessary to remove some codegen tests which look for specific things. Also attempting to restart a test which timed out, maybe due to fastly failing?

Rm allocation in candidate

5d9f514

Instead of storing an extra array for discriminant values, create an allocation there and store those in an allocation immediately.

Add tag for ignoring wasm

610e1a1

Leave FIXME for wasm layout difference.

15f4eec

There is a distinction between running this on wasm and i686, even though they should be identical. This technically is not _incorrect_, it's just an unexpected difference, which is worth investigating, but not for correctness.

JulianKnodt force-pushed the array_const_val branch 5 times, most recently from c67e312 to e5352d9 Compare February 7, 2023 10:43

cjgillot reviewed Feb 7, 2023

View reviewed changes

compiler/rustc_mir_transform/src/large_enums.rs Outdated Show resolved Hide resolved

Add de-init to destination place

15d4728

JulianKnodt force-pushed the array_const_val branch from e5352d9 to 15d4728 Compare February 8, 2023 02:04

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Feb 10, 2023

bors added the merged-by-bors This PR was explicitly merged by bors. label Feb 11, 2023

bors merged commit 5a8dfd9 into rust-lang:master Feb 11, 2023

rustbot added this to the 1.69.0 milestone Feb 11, 2023

JulianKnodt deleted the array_const_val branch February 11, 2023 01:26

rustbot removed the perf-regression Performance regression. label Feb 11, 2023

jackh726 mentioned this pull request Feb 15, 2023

Implement deferred_projection_equality for erica solver #107507

Merged

matthiaskrgr mentioned this pull request Nov 14, 2023

ICE: resolve: unreachable!() #117920

Closed

saethlin mentioned this pull request Nov 25, 2023

Segfault from mir-opt-level >= 3 (EnumSizeOpt) #118283

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Mir-Opt for copying enums with large discrepancies #85158

Mir-Opt for copying enums with large discrepancies #85158

JulianKnodt commented May 10, 2021

This comment has been minimized.

tmiasko left a comment

tmiasko commented May 12, 2021

scottmcm commented May 12, 2021

JulianKnodt commented May 12, 2021

tmiasko commented May 12, 2021

JulianKnodt commented May 12, 2021

bjorn3 commented May 13, 2021

tmiasko commented May 13, 2021

camelid commented May 28, 2021

JulianKnodt commented May 28, 2021

This comment has been minimized.

This comment has been minimized.

tmiasko May 28, 2021

JulianKnodt Jun 9, 2021

tmiasko commented May 28, 2021

JulianKnodt commented Feb 8, 2023

cjgillot commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 11, 2023

rust-timer commented Feb 11, 2023

Mir-Opt for copying enums with large discrepancies #85158

Mir-Opt for copying enums with large discrepancies #85158

Conversation

JulianKnodt commented May 10, 2021

This comment has been minimized.

tmiasko left a comment

Choose a reason for hiding this comment

tmiasko commented May 12, 2021

scottmcm commented May 12, 2021

JulianKnodt commented May 12, 2021

tmiasko commented May 12, 2021

JulianKnodt commented May 12, 2021

bjorn3 commented May 13, 2021

tmiasko commented May 13, 2021

camelid commented May 28, 2021

JulianKnodt commented May 28, 2021

This comment has been minimized.

This comment has been minimized.

tmiasko May 28, 2021

Choose a reason for hiding this comment

JulianKnodt Jun 9, 2021

Choose a reason for hiding this comment

tmiasko commented May 28, 2021

JulianKnodt commented Feb 8, 2023

cjgillot commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 11, 2023

rust-timer commented Feb 11, 2023

Overall result: no relevant changes - no action needed

Instruction count

Max RSS (memory usage)

Cycles