arm64: Replace RSH/RSZ -> CAST nodes with clearing register by jonathandavies-arm · Pull Request #121007 · dotnet/runtime

jonathandavies-arm · 2025-10-23T09:12:17Z

In lowering change a down cast and right shift into a mov w0, wzr if the shift amount is constant and greater & equal to the size of the downcast type. e.g.

static int CastASR8_byte_int(byte x)
{
    //ARM64-FULL-LINE: mov {{w[0-9]+}}, wzr
    return x >> 8;
}

assembly changes from

uxtb    w0, w0
asr     w0, w0, #8

to

mov     w0, wzr

dotnet-policy-service · 2025-10-23T09:13:28Z

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

jonathandavies-arm · 2025-10-29T10:24:06Z

Please can I have a review? @dotnet/arm64-contrib @EgorBo

SwapnilGaikwad · 2025-11-17T12:00:58Z

cc: @a74nh @JulieLeeMSFT

EgorBo · 2025-12-03T13:27:28Z

/azp run Fuzzlyn

azure-pipelines · 2025-12-03T13:27:41Z

Azure Pipelines successfully started running 1 pipeline(s).

a74nh · 2025-12-09T14:05:05Z

/azp run Fuzzlyn

Looks like Fuzzlyn got stuck?

a74nh · 2025-12-17T16:30:32Z

Could someone run fuzzlyn again on this please. Previous one got cancelled by the CI, I think.

saucecontrol · 2025-12-22T22:03:14Z

src/coreclr/jit/lower.cpp

-        if (!cast->isContained() && !cast->IsRegOptional() && !cast->gtOverflow() &&
-            // Smaller CastOp is most likely an IND(X) node which is lowered to a zero-extend load
-            cast->CastOp()->TypeIs(TYP_LONG, TYP_INT))
+        // Try to recognize right shift with a CAST node that is equivilent to mov #0


This optimization should be useful for all platforms. Why restrict it to Arm64?

This would also likely be more beneficial if implemented in morph, where it could enable further downstream optimizations.

This optimization should be useful for all platforms. Why restrict it to Arm64?

This would also likely be more beneficial if implemented in morph, where it could enable further downstream optimizations.

Agreed. There is nothing fundamentally architecture specific here, just replacing an overflowing shift with zero.

My only concern would be if for some reason the casts weren't being introduced until after all the morph passes. But, I don't think that's going to happen.

We can always have it in both places, but I do think that morph is the more meaningful location here.

A more comprehensive version of this is #122533, which also handles other optimizations but is also doing it in lowering.

I checked out the work in #122533 and ran my tests in this PR and they don't pass. I think both of these PRs are doing different optimisations. The Fix section in the other PR doesn't describe the situation I'm trying to optimise.

I've moved the optimisation into morph.

dhartglassMSFT · 2026-01-26T07:59:33Z

src/coreclr/jit/morph.cpp

+                GenTreeIntCon* cns  = op2->AsIntCon();
+                if (!cast->gtOverflow() && cast->CastOp()->TypeIs(TYP_INT) && varTypeIsUnsigned(cast->CastToType()))
+                {
+                    ssize_t  shiftAmount = cns->IconValue();


Can you add a small comment around here saying what the transform does? Just similar to the one on line 11279.

Change lgtm otherwise

Added a comment.

jakobbotsch · 2026-01-26T12:39:30Z

No diffs in SPMI -- do we think this transformation is worth it?

dhartglassMSFT · 2026-02-02T18:07:53Z

No diffs in SPMI -- do we think this transformation is worth it?

Hi @jonathandavies-arm , echoing Jakob's question. Does this transform go after a specific scenario you were looking at?

jonathandavies-arm · 2026-02-04T09:32:55Z

No diffs in SPMI -- do we think this transformation is worth it?

Hi @jonathandavies-arm , echoing Jakob's question. Does this transform go after a specific scenario you were looking at?

There isn't a specific scenario we are looking at. This tasks was one of a list of general improvements that we thought we should do. The asmdiff throughput doesn't show any increase in instructions although the arm64 linux has failed to build. https://dev.azure.com/dnceng-public/public/_build/results?buildId=1273270&view=ms.vss-build-web.run-extensions-tab

Here is a link to compiler explorer https://godbolt.org/z/jf5455rcb that shows a similar optimisation done by GCC and Clang.

I do think this is a useful optimisation but we can close it if you feel it isn't worth it.

jakobbotsch · 2026-02-04T16:09:28Z

There isn't a specific scenario we are looking at. This tasks was one of a list of general improvements that we thought we should do. The asmdiff throughput doesn't show any increase in instructions although the arm64 linux has failed to build. https://dev.azure.com/dnceng-public/public/_build/results?buildId=1273270&view=ms.vss-build-web.run-extensions-tab

Here is a link to compiler explorer https://godbolt.org/z/jf5455rcb that shows a similar optimisation done by GCC and Clang.

I do think this is a useful optimisation but we can close it if you feel it isn't worth it.

It is not unusual for us to implement optimizations but to end up closing them when we find small or no impact, or even remove existing optimizations in those scenarios. It is not purely about JIT throughput but also about maintaining them in the future and consistency in the optimization phases. We like to see some motivating factors, like asm diffs improvements or simplifying upcoming changes. With that in mind I will close this, but feel free to reopen if you have any other motivation for the change.

arm64: Replace RSH/RSZ -> CAST nodes with clearing register

974f180

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Oct 23, 2025

github-actions bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Oct 23, 2025

build-analysis bot mentioned this pull request Oct 23, 2025

/root/helix/work/correlation/scripts/<hash>/execute.sh: Permission denied dotnet/dnceng#3412

Open

3 tasks

SwapnilGaikwad added the arch-arm64 label Oct 29, 2025

Only optimise on unsigned casts

1e0e05e

jonathandavies-arm added 2 commits November 17, 2025 11:47

Pass int into tests and perform casts in functions

6e345eb

Merge branch 'main' into upstream/ce/right-shift-cast

6aa1714

Merge branch 'main' into upstream/ce/right-shift-cast

aed793e

saucecontrol reviewed Dec 22, 2025

View reviewed changes

Move optimisation from lowering to morph

2f5e093

This was referenced Jan 7, 2026

Unable to pull image from mcr.microsoft.com #117164

Open

[mono] mono_thread_info_install_interrupt: previous_token should be INTERRUPT_STATE #122669

Open

iOS.Device test WorkItemExecutions #122874

Open

Merge branch 'main' into upstream/ce/right-shift-cast

8f98e15

dhartglassMSFT reviewed Jan 26, 2026

View reviewed changes

dhartglassMSFT approved these changes Jan 26, 2026

View reviewed changes

Add comment

578d286

build-analysis bot mentioned this pull request Jan 26, 2026

browser-wasm linux Release LibraryTests queues timing out #117974

Open

build-analysis bot mentioned this pull request Jan 26, 2026

[android][arm64] System.Net.Sockets.Tests.SendTo_SyncForceNonBlocking.Datagram_UDP_ShouldImplicitlyBindLocalEndpoint fails with NetworkUnreachable #120526

Open

Merge branch 'main' into upstream/ce/right-shift-cast

1927134

jakobbotsch closed this Feb 4, 2026

Conversation

jonathandavies-arm commented Oct 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dotnet-policy-service bot commented Oct 23, 2025

Uh oh!

jonathandavies-arm commented Oct 29, 2025

Uh oh!

SwapnilGaikwad commented Nov 17, 2025

Uh oh!

EgorBo commented Dec 3, 2025

Uh oh!

azure-pipelines bot commented Dec 3, 2025

Uh oh!

a74nh commented Dec 9, 2025

Uh oh!

a74nh commented Dec 17, 2025

Uh oh!

saucecontrol Dec 22, 2025

Choose a reason for hiding this comment

Uh oh!

a74nh Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

tannergooding Jan 5, 2026

Choose a reason for hiding this comment

Uh oh!

jonathandavies-arm Jan 7, 2026

Choose a reason for hiding this comment

Uh oh!

jonathandavies-arm Jan 7, 2026

Choose a reason for hiding this comment

Uh oh!

dhartglassMSFT Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

jonathandavies-arm Jan 26, 2026

Choose a reason for hiding this comment

Uh oh!

jakobbotsch commented Jan 26, 2026

Uh oh!

dhartglassMSFT commented Feb 2, 2026

Uh oh!

jonathandavies-arm commented Feb 4, 2026

Uh oh!

jakobbotsch commented Feb 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

jonathandavies-arm commented Oct 23, 2025 •

edited

Loading

jakobbotsch commented Feb 4, 2026 •

edited

Loading