Lift a few limitations on ConstProp optimization pass #107856

cjgillot · 2023-02-09T19:34:47Z

ConstProp does not emit lints any more. Therefore, we can extend and simplify it.

This PR starts by fortifying the current implementation of ConstProp and of the interpreter.

In a second part, we enable propagation even if overflow checks are disabled. It's now codegen's responsibility to skip emitting overflow checks.

In a third part, we remove limitation on propagation.

Finally, we only disable the pass for mir-opt-level=1, since it's now purely an optimization pass.

rustbot · 2023-02-09T19:34:53Z

r? @lcnr

(rustbot has picked a reviewer for you, use r? to override)

rustbot · 2023-02-09T19:34:56Z

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

Some changes occurred to the CTFE / Miri engine

cc @rust-lang/miri

Some changes occurred to the CTFE / Miri engine

cc @rust-lang/miri

cjgillot · 2023-02-09T19:36:33Z

@bors try @rust-timer queue

bors · 2023-02-09T19:36:44Z

⌛ Trying commit b27fb59b1c018d8223f4ef35d5ae3eb687a8bbf3 with merge 85e382751271dbc84dc43390b8ce9624f3b96753...

RalfJung · 2023-02-09T19:42:50Z

compiler/rustc_const_eval/src/interpret/operand.rs

@@ -601,7 +601,9 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
        // that the `RevealAll` pass has happened and that the body's consts
        // are normalized, so any call to resolve before that needs to be
        // manually normalized.
-        let val = self.tcx.normalize_erasing_regions(self.param_env, *val);
+        let Ok(val) = self.tcx.try_normalize_erasing_regions(self.param_env, *val) else {
+            throw_unsup!(Normalization)


This is more invalid than unsupported I think, just like all the other const-prop-hacks we carry in the interpreter.

throw_inval!(TooGeneric) may be better suited then.

Ah right, we already have that variant. Agreed.

RalfJung · 2023-02-09T19:43:10Z

compiler/rustc_const_eval/src/interpret/cast.rs

@@ -356,7 +356,7 @@ impl<'mir, 'tcx: 'mir, M: Machine<'mir, 'tcx>> InterpCx<'mir, 'tcx, M> {
            }

            _ => {
-                span_bug!(self.cur_span(), "invalid unsizing {:?} -> {:?}", src.layout.ty, cast_ty)
+                throw_inval!(Unsize(src.layout.ty, cast_ty))


Even more const-prop induced technical debt in the interpreter. :(

How can this one even happen...?

We can have a generic context

fn foo<T: ?Sized, V: std::marker::Unsize<T>>(value: &V) -> &T { value }

Please add a comment explaining that (and that this is only needed for ConstProp).

Why does that not affect the similar bug! in unsize_into?

oli-obk · 2023-02-09T20:05:01Z

I don't think we should make const prop smarter. We should get the dataflow const prop to cover most of the const prop optimizations and then just remove const prop.

saethlin · 2023-02-09T20:44:29Z

Do we have a solution to DataFlowConstProp's runtime?

cjgillot · 2023-02-09T22:20:38Z

I don't think we should make const prop smarter. We should get the dataflow const prop to cover most of the const prop optimizations and then just remove const prop.

I don't agree with this vision. My suggestion is to have:

this ConstProp pass do the "easy" cases: locals assigned once, and intra-bb propagation;
some intermediary cleanup pass to shrink MIR and reduce state space (SimplifyConstSwitch and SimplifyLocals in particular);
let DataflowConstProp do the harder work of inter-bb propagation for what remains.

Whatever the decision on that discussion, there are 2 changes that I'd like to propose:

gate ConstProp on mir-opt-level >= 2, since it doesn't have a reason to run on mir-opt-level=1 any more;
make codegen alone handle the decisions on overflow checks, and always propagate checked operations (in both ConstProp and DataflowConstProp).

I can move them in separate PRs if needed.

bors · 2023-02-09T22:24:59Z

☀️ Try build successful - checks-actions
Build commit: 85e382751271dbc84dc43390b8ce9624f3b96753 (85e382751271dbc84dc43390b8ce9624f3b96753)

rust-timer · 2023-02-09T23:43:51Z

Finished benchmarking commit (85e382751271dbc84dc43390b8ce9624f3b96753): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

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

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	0.7%	[0.4%, 1.1%]	7
Regressions ❌ (secondary)	7.3%	[1.0%, 21.6%]	10
Improvements ✅ (primary)	-0.9%	[-1.9%, -0.3%]	12
Improvements ✅ (secondary)	-1.1%	[-2.0%, -0.2%]	16
All ❌✅ (primary)	-0.3%	[-1.9%, 1.1%]	19

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)	2.9%	[0.1%, 5.7%]	2
Regressions ❌ (secondary)	29.0%	[3.9%, 53.1%]	4
Improvements ✅ (primary)	-3.4%	[-4.4%, -2.4%]	10
Improvements ✅ (secondary)	-2.4%	[-4.1%, -0.7%]	50
All ❌✅ (primary)	-2.3%	[-4.4%, 5.7%]	12

Cycles

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)	13.0%	[2.8%, 23.3%]	7
Improvements ✅ (primary)	-1.5%	[-1.9%, -1.1%]	4
Improvements ✅ (secondary)	-1.9%	[-1.9%, -1.9%]	1
All ❌✅ (primary)	-1.5%	[-1.9%, -1.1%]	4

lcnr · 2023-02-10T09:00:20Z

r? @oli-obk @RalfJung

though re 7bbc67468dae49930263f44eb9fedd275acb912e, is there a test which actually ICEs without this change? While I can imagine cases where this ICEs with generic_const_exprs, that's a bug with that feature and in my ideal impl of CTFE this is actually infallible

RalfJung · 2023-02-10T09:50:09Z

I don't think we should make const prop smarter. We should get the dataflow const prop to cover most of the const prop optimizations and then just remove const prop.

I think that would be great, if it can be done in our perf budget. It means less code to maintain, and fewer issues with ConstProp using the interpreter in ways it was not intended to be used.

oli-obk · 2023-02-10T11:16:08Z

I'll benchmark removing the old pass and enabling the new one so we can make a decision here.

Even if we can't split it, we can still change const prop to stop using mir interpreter rvalue processing and instead make it only use the math APIs like the dataflow const prop does

cjgillot · 2023-02-10T16:22:59Z

On the perf report:

check builds: mostly non-relevant red and green, nothing obvious;
debug builds: mostly green, effect of not running ConstProp any more. An exception: keccak (20% regression), as generated MIR is 40% larger.
opt builds: mostly red. Largest (diesel) may be due to evaluation of ty unevaluated constants (to be confirmed).

bors · 2023-02-10T19:07:26Z

⌛ Trying commit 2131dae17d267c3c4cb53690cf00cdbc7a2e9de0 with merge 47d13e87e7c1c86fba1429c13a990657fa7a3561...

bors · 2023-02-10T21:24:30Z

☀️ Try build successful - checks-actions
Build commit: 47d13e87e7c1c86fba1429c13a990657fa7a3561 (47d13e87e7c1c86fba1429c13a990657fa7a3561)

rust-timer · 2023-02-10T22:45:51Z

Finished benchmarking commit (47d13e87e7c1c86fba1429c13a990657fa7a3561): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

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

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	0.8%	[0.3%, 1.6%]	19
Regressions ❌ (secondary)	1.2%	[0.9%, 1.6%]	6
Improvements ✅ (primary)	-0.6%	[-1.1%, -0.3%]	8
Improvements ✅ (secondary)	-1.0%	[-1.2%, -0.8%]	2
All ❌✅ (primary)	0.4%	[-1.1%, 1.6%]	27

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)	3.8%	[0.1%, 9.2%]	5
Regressions ❌ (secondary)	2.9%	[2.7%, 3.1%]	2
Improvements ✅ (primary)	-3.6%	[-3.9%, -3.4%]	2
Improvements ✅ (secondary)	-2.2%	[-2.2%, -2.2%]	1
All ❌✅ (primary)	1.7%	[-3.9%, 9.2%]	7

Cycles

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)	6.7%	[5.4%, 7.7%]	6
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-1.3%	[-1.3%, -1.3%]	1
All ❌✅ (primary)	-	-	0

cjgillot · 2023-02-11T10:09:59Z

@bors try @rust-timer queue

bors · 2023-02-11T10:10:09Z

⌛ Trying commit 9a58168 with merge 1cff2133ab3b7724841c768a38ec63160562386a...

bors · 2023-02-11T12:26:40Z

☀️ Try build successful - checks-actions
Build commit: 1cff2133ab3b7724841c768a38ec63160562386a (1cff2133ab3b7724841c768a38ec63160562386a)

rust-timer · 2023-02-11T14:46:58Z

Finished benchmarking commit (1cff2133ab3b7724841c768a38ec63160562386a): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

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

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	0.6%	[0.4%, 0.7%]	5
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.6%	[-0.6%, -0.6%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.4%	[-0.6%, 0.7%]	6

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)	3.4%	[2.4%, 4.1%]	6
Regressions ❌ (secondary)	2.7%	[0.7%, 4.2%]	49
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	3.4%	[2.4%, 4.1%]	6

Cycles

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

cjgillot · 2023-02-11T15:10:33Z

The last perf result confirm there is nothing to gain by extending this pass.
I'll close this PR, and propose the cleanups in another PR based on #107921.

cjgillot added 2 commits February 8, 2023 22:13

Move state fixup into a different method.

9e0bf2b

Merge if-let and match.

9ee5a2c

cjgillot added the A-mir-opt Area: MIR optimizations label Feb 9, 2023

rustbot assigned lcnr Feb 9, 2023

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. labels Feb 9, 2023

This comment has been minimized.

Sign in to view

rustbot added the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Feb 9, 2023

RalfJung reviewed Feb 9, 2023

View reviewed changes

This comment has been minimized.

Sign in to view

rustbot added perf-regression Performance regression. and removed S-waiting-on-perf Status: Waiting on a perf run to be completed. labels Feb 9, 2023

This comment has been minimized.

Sign in to view

rustbot assigned oli-obk and unassigned lcnr Feb 10, 2023

rustbot added the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Feb 10, 2023

cjgillot added 3 commits February 10, 2023 19:50

Remove overflow checks from ConstProp.

298a19d

Remove use_ecx.

294a797

Do not attempt to const-prop unsize.

1e739d7

This comment has been minimized.

Sign in to view

rustbot removed the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Feb 10, 2023

This comment has been minimized.

Sign in to view

cjgillot force-pushed the unleash-const-prop branch from 2131dae to d341cd3 Compare February 11, 2023 08:30

cjgillot mentioned this pull request Feb 11, 2023

Make codegen choose whether to emit overflow checks #107921

Merged

cjgillot added 3 commits February 11, 2023 09:09

Normalize constants before evaluating them.

9b71592

Also evaluate TY consts.

048f122

Bless ui tests.

9a58168

cjgillot force-pushed the unleash-const-prop branch from d341cd3 to 9a58168 Compare February 11, 2023 09:26

This comment has been minimized.

Sign in to view

rustbot added the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Feb 11, 2023

This comment has been minimized.

Sign in to view

rustbot removed the S-waiting-on-perf Status: Waiting on a perf run to be completed. label Feb 11, 2023

cjgillot closed this Feb 11, 2023

cjgillot deleted the unleash-const-prop branch February 11, 2023 15:10

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Lift a few limitations on ConstProp optimization pass #107856

Lift a few limitations on ConstProp optimization pass #107856

cjgillot commented Feb 9, 2023

rustbot commented Feb 9, 2023

rustbot commented Feb 9, 2023

cjgillot commented Feb 9, 2023

This comment has been minimized.

bors commented Feb 9, 2023

RalfJung Feb 9, 2023 •

edited

Loading

cjgillot Feb 9, 2023

RalfJung Feb 10, 2023

RalfJung Feb 9, 2023

cjgillot Feb 9, 2023

RalfJung Feb 10, 2023

oli-obk commented Feb 9, 2023

saethlin commented Feb 9, 2023

cjgillot commented Feb 9, 2023

bors commented Feb 9, 2023

This comment has been minimized.

rust-timer commented Feb 9, 2023

This comment has been minimized.

lcnr commented Feb 10, 2023 •

edited

Loading

RalfJung commented Feb 10, 2023

oli-obk commented Feb 10, 2023

cjgillot commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 10, 2023

This comment has been minimized.

rust-timer commented Feb 10, 2023

This comment has been minimized.

cjgillot commented Feb 11, 2023

This comment has been minimized.

bors commented Feb 11, 2023

bors commented Feb 11, 2023

This comment has been minimized.

rust-timer commented Feb 11, 2023

cjgillot commented Feb 11, 2023

Lift a few limitations on ConstProp optimization pass #107856

Lift a few limitations on ConstProp optimization pass #107856

Conversation

cjgillot commented Feb 9, 2023

rustbot commented Feb 9, 2023

rustbot commented Feb 9, 2023

cjgillot commented Feb 9, 2023

This comment has been minimized.

bors commented Feb 9, 2023

RalfJung Feb 9, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

oli-obk commented Feb 9, 2023

saethlin commented Feb 9, 2023

cjgillot commented Feb 9, 2023

bors commented Feb 9, 2023

This comment has been minimized.

rust-timer commented Feb 9, 2023

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

This comment has been minimized.

lcnr commented Feb 10, 2023 • edited Loading

RalfJung commented Feb 10, 2023

oli-obk commented Feb 10, 2023

cjgillot commented Feb 10, 2023

bors commented Feb 10, 2023

bors commented Feb 10, 2023

This comment has been minimized.

rust-timer commented Feb 10, 2023

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

This comment has been minimized.

cjgillot commented Feb 11, 2023

This comment has been minimized.

bors commented Feb 11, 2023

bors commented Feb 11, 2023

This comment has been minimized.

rust-timer commented Feb 11, 2023

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

cjgillot commented Feb 11, 2023

RalfJung Feb 9, 2023 •

edited

Loading

lcnr commented Feb 10, 2023 •

edited

Loading