JIT: Make effect handling in lowering less conservative #92710

jakobbotsch · 2023-09-27T14:21:19Z

The interference checking in lowering bases some of its checks on GenTree::gtFlags. This is conservative since it includes effect flags of operands, which does not make much sense for LIR.

This PR replaces the relevant uses of gtFlags with a new GenTree::OperEffects() that computes the relevant effect flags for the node, excluding operands. We already know how to recompute effect flags other than GTF_GLOB_REF and GTF_ORDER_SIDEEFF. This PR adds functions for these as well (the GTF_GLOB_REF version
GenTree::OperRequiresGlobRefFlag is courtesy of @SingleAccretion).

For GTF_ORDER_SIDEEFF we add a GenTree::OperSupportsOrderingSideEffect which captures explicitly (and conservatively) the current cases where we are setting the flag, and only allows these cases to support the flag. Setting the flag for other cases may result in the flag being removed or ignored.

Fix #92699

Example:

public struct Point
{
    public long X, Y;

    public static unsafe bool Equals(in Point left, in Point right)
    {
        return (left.X == right.X) & (left.Y == right.Y);
    }
}

Codegen before:

G_M28859_IG02:  ;; offset=0x0008
            ldr     x2, [x0]
            ldr     x3, [x1]
            cmp     x2, x3
            cset    x2, eq
            ldr     x0, [x0, #0x08]
            ldr     x1, [x1, #0x08]
            cmp     x0, x1
            cset    x0, eq
            and     w0, w2, w0

Codegen after:

G_M28859_IG02:  ;; offset=0x0008
            ldr     x2, [x0]
            ldr     x3, [x1]
            ldr     x0, [x0, #0x08]
            ldr     x1, [x1, #0x08]
            cmp     x2, x3
            ccmp    x0, x1, 0, eq
            cset    x0, eq

@SingleAccretion

The interference checking in lowering bases some of its checks on GenTree::gtFlags. This is conservative since it includes effect flags of operands. For LIR this does not really make sense and ends up being conservative. This PR replaces the relevant uses of gtFlags with a new GenTree::OperEffects() that computes the relevant effect flags for the node, excluding operands. We already know how to recompute effect flags other than GTF_GLOB_REF and GTF_ORDER_SIDEEFF. This PR adds functions for these as well (the GTF_GLOB_REF version GenTree::OperRequiresGlobRefFlag is courtesy of @SingleAccretion). For GTF_ORDER_SIDEEFF we add a GenTree::OperSupportsOrderingSideEffect which captures explicitly (and conservatively) the current cases where we are setting the flag, and only allows these cases to support the flag. Setting the flag for other cases may result in the flag being removed or ignored. Fix dotnet#92699

ghost · 2023-09-27T14:21:32Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

The interference checking in lowering bases some of its checks on GenTree::gtFlags. This is conservative since it includes effect flags of operands. For LIR this does not really make sense and ends up being conservative.

This PR replaces the relevant uses of gtFlags with a new GenTree::OperEffects() that computes the relevant effect flags for the node, excluding operands. We already know how to recompute effect flags other than GTF_GLOB_REF and GTF_ORDER_SIDEEFF. This PR adds functions for these as well (the GTF_GLOB_REF version
GenTree::OperRequiresGlobRefFlag is courtesy of @SingleAccretion).

For GTF_ORDER_SIDEEFF we add a GenTree::OperSupportsOrderingSideEffect which captures explicitly (and conservatively) the current cases where we are setting the flag, and only allows these cases to support the flag. Setting the flag for other cases may result in the flag being removed or ignored.

Fix #92699

Author:	jakobbotsch
Assignees:	-
Labels:	`area-CodeGen-coreclr`
Milestone:	-

jakobbotsch · 2023-09-27T20:06:02Z

/azp run runtime-coreclr superpmi-diffs, runtime-coreclr superpmi-replay

azure-pipelines · 2023-09-27T20:06:24Z

Azure Pipelines successfully started running 2 pipeline(s).

jakobbotsch · 2023-09-28T11:06:32Z

cc @dotnet/jit-contrib PTAL @BruceForstall @kunalspathak

Diffs, mainly from new containment opportunities and ccmp opportunities like in the issue.
TP diffs are mostly positive, except for the native linux-x64 TP runs that seem to be somewhat worse affected, with a libraries.pmi outlier of up to +0.21%.

jakobbotsch · 2023-09-28T11:08:47Z

/azp run Fuzzlyn, runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

azure-pipelines · 2023-09-28T11:09:15Z

Azure Pipelines successfully started running 3 pipeline(s).

SingleAccretion · 2023-09-27T14:29:55Z

src/coreclr/jit/gentree.h

+    // True if this indirection is invariant.
+    bool IsInvariantLoad() const
+    {
+        return (gtFlags & GTF_IND_INVARIANT) != 0;


I also had an assert(!isInvariant || OperIs(GT_IND, GT_BLK)); here to exclude stores. Worth adding?

SingleAccretion · 2023-09-27T16:20:27Z

src/coreclr/jit/gentree.cpp

+//------------------------------------------------------------------------------
+// OperRequiresGlobRefFlag : Check whether the operation requires GTF_GLOB_REF
+//                           flag regardless of the children's flags.
+//


Suggested change

//

//

// Globally visible stores and loads, as well as some equivalently modeled

// operations, require the GLOB_REF flag to be set on the node.

//

Expanded on this a bit in the remarks.

src/coreclr/jit/gentree.h

src/coreclr/jit/gentree.cpp

SingleAccretion · 2023-09-27T16:28:19Z

src/coreclr/jit/gentree.cpp

+//   parent may be ok as long as it does not break ordering dependencies of the
+//   operand.
+//
+bool GenTree::OperSupportsOrderingSideEffect() const


Do you have ideas on how to make sure that this method does not silently drop ordering effects when we start using it for new nodes?

For example, say someone is implementing non-null assertion propagation for GT_ARR_LENGTH family of operators. They will set the flag, just as current code does for indirs, and it will show up in dumps and all, but the effect will be silently dropped.

Perhaps add a setter for GTF_ORDER_SIDEEFF with an assert (there is SetAllEffectsFlags & Co already, but they are used to propagate flags upwards, so, not very suitable)?

I can add assert in fgdiagnostic to ensure that GTF_ORDER_SIDEEFF implies that OperSupportsOrderingSideEffect() || (any operand has the flag).

Doing this requires making sure we propagate/unset these flags in all the same places as the other flags... That's probably a good thing to do down the road, but I think for now I will skip it and think about doing that separately.
Let me add something like SetHasOrderingSideEffect() with the assert in it for this PR.

Doing this did make me notice that I missed GT_NULLCHECK in OperSupportsOrderingSideEffect though, so definitely worth to have something.

SingleAccretion · 2023-09-28T11:38:09Z

src/coreclr/jit/gentree.cpp

+        case GT_CALL:
+            return AsCall()->HasSideEffects(comp, /* ignoreExceptions */ true);


This was apparently in my second commit - GT_ALLOCOBJ should be here as well with return AsAllocObj()->gtHelperHasSideEffects;

src/coreclr/jit/gentree.cpp

jakobbotsch · 2023-09-28T14:44:37Z

The failures are known according to build analysis. The Fuzzlyn failures look like #92537

ghost added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Sep 27, 2023

ghost assigned jakobbotsch Sep 27, 2023

Fix build

25324a2

This was referenced Sep 27, 2023

Tracking issue for CI build timeouts #76454

Closed

MSBuild crashing in the build #92290

Open

jakobbotsch mentioned this pull request Sep 28, 2023

JIT: Consider reordering loads on ARM64 to make them amenable to ldp optimization #92756

Closed

jakobbotsch marked this pull request as ready for review September 28, 2023 10:54

jakobbotsch requested review from BruceForstall, kunalspathak and SingleAccretion September 28, 2023 11:06

SingleAccretion reviewed Sep 28, 2023

View reviewed changes

build-analysis bot mentioned this pull request Sep 28, 2023

Test_EventSource_EtwManifestGeneration* tests failing in CI #48798

Closed

build-analysis bot mentioned this pull request Sep 28, 2023

Test failure: Loader\\binding\\tracing\\BinderTracingTest.ResolutionFlow\\BinderTracingTest.ResolutionFlow.cmd #92236

Closed

Address some feedback

d28c644

BruceForstall approved these changes Sep 28, 2023

View reviewed changes

jakobbotsch added 2 commits September 28, 2023 22:42

Add GT_NULLCHECK to OperSupportsOrderingSideEffect

a4d40cd

Add GenTree::SetHasOrderingSideEffect

52e0247

jakobbotsch merged commit 8ff1893 into dotnet:main Sep 29, 2023

jakobbotsch deleted the fix-92699 branch September 29, 2023 09:41

jakobbotsch mentioned this pull request Sep 29, 2023

JIT: Optimize SequenceEqual to use ccmp on ARM64 #92810

Merged

EgorBo mentioned this pull request Oct 3, 2023

[Perf] Linux/x64: 1 Improvement on 9/29/2023 10:19:41 AM dotnet/perf-autofiling-issues#22511

Closed

jakobbotsch mentioned this pull request Oct 4, 2023

jitstress failing widely with NRE #92990

Closed

jakobbotsch mentioned this pull request Oct 4, 2023

JIT: Add ordering side effects in fgMorphExpandInstanceField #92998

Merged

EgorBo mentioned this pull request Oct 5, 2023

[Perf] Linux/arm64: 5 Improvements on 9/30/2023 12:13:33 AM dotnet/perf-autofiling-issues#22699

Closed

EgorBo mentioned this pull request Oct 12, 2023

[Perf] Windows/arm64: 6 Improvements on 10/6/2023 11:50:29 PM dotnet/perf-autofiling-issues#23011

Closed

ghost locked as resolved and limited conversation to collaborators Oct 29, 2023

-//
+//
+// Globally visible stores and loads, as well as some equivalently modeled
+// operations, require the GLOB_REF flag to be set on the node.
+//

		case GT_CALL:
		return AsCall()->HasSideEffects(comp, /* ignoreExceptions */ true);

JIT: Make effect handling in lowering less conservative #92710

JIT: Make effect handling in lowering less conservative #92710

Uh oh!

Conversation

jakobbotsch commented Sep 27, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ghost commented Sep 27, 2023

Uh oh!

jakobbotsch commented Sep 27, 2023

Uh oh!

azure-pipelines bot commented Sep 27, 2023

Uh oh!

jakobbotsch commented Sep 28, 2023

Uh oh!

jakobbotsch commented Sep 28, 2023

Uh oh!

azure-pipelines bot commented Sep 28, 2023

Uh oh!

SingleAccretion Sep 27, 2023

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Sep 28, 2023

Choose a reason for hiding this comment

Uh oh!

SingleAccretion Sep 27, 2023

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Sep 28, 2023

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

SingleAccretion Sep 27, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Sep 28, 2023

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Sep 28, 2023

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Sep 28, 2023

Choose a reason for hiding this comment

Uh oh!

SingleAccretion Sep 28, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

jakobbotsch commented Sep 28, 2023

Uh oh!

Uh oh!

jakobbotsch commented Sep 27, 2023 •

edited

Loading

SingleAccretion Sep 27, 2023 •

edited

Loading

SingleAccretion Sep 28, 2023 •

edited

Loading