JIT: Avoid resolving def-use conflicts by changing use registers by jakobbotsch · Pull Request #125333 · dotnet/runtime

jakobbotsch · 2026-03-09T14:02:46Z

The only benefit is that it saves insertion of some explicit copies, but the trade off is an implicit contract between LSRA and codegen that is not very faithful. When LSRA defines in a different register than the constrained register, it believes that codegen can insert a copy and simultaneously use it from the new location, which is not true. This difference in the model has to be handled by inserting additional kill constraints during LSRA build, for example for shifts on xarch.

This PR additionally fixes a few issues uncovered while making this change:

GT_CATCH_ARG codegen was straight up wrong, not properly producing its register value
Multireg definitions cannot have their register assignment changed since that could override another defined register.

Fix #125432
Fix #125685

The only benefit is that it saves insertion of some explicit copies, but the trade off is an implicit contract between LSRA and codegen that is not very faithful. When LSRA defines in a different register than the constrained register, it believes that codegen can insert a copy and simultaneously use it from the new location, which is not true.

Copilot

Pull request overview

This PR refactors LinearScan::resolveConflictingDefAndUse in the JIT's register allocator (LSRA) to eliminate the practice of changing use-side register assignments when resolving def-use conflicts. The change ensures that only the definition RefPosition's register assignment is ever modified, while explicit copies are inserted for cases where the use constraint cannot be satisfied. This removes a fragile implicit contract between LSRA and codegen where LSRA would assign a def to a different register than constrained and expect codegen to simultaneously insert a copy and use from the new location.

Changes:

Simplifies resolveConflictingDefAndUse in lsrabuild.cpp by removing old Cases 1 and 4 (which changed the USE register assignment), consolidating the remaining cases into four cleaner ones that only modify the DEF's register assignment.
Updates the LsraDumpEvent enum in lsra.h to replace numbered case events (CASE1–CASE6) with descriptive names.
Updates the debug dump handler in lsra.cpp to use the new event names and removes the now-dead LSRA_EVENT_DEFUSE_FIXED_DELAY_USE no-op case.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
`src/coreclr/jit/lsrabuild.cpp`	Core logic change: removes use-register-changing cases, simplifies conflict resolution to only modify def registers, updates comments
`src/coreclr/jit/lsra.h`	Replaces numbered DEFUSE case enum values with descriptive names
`src/coreclr/jit/lsra.cpp`	Updates debug dump switch/case to use new enum names, removes dead `LSRA_EVENT_DEFUSE_FIXED_DELAY_USE` case

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

jakobbotsch · 2026-04-01T09:43:12Z

Another related issue here is #10196

…changes

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

jakobbotsch · 2026-04-01T14:25:07Z

/azp run runtime-coreclr jitstressregs

azure-pipelines · 2026-04-01T14:25:42Z

Azure Pipelines successfully started running 1 pipeline(s).

jakobbotsch · 2026-04-08T12:18:03Z

/azp run runtime-coreclr jitstressregs

azure-pipelines · 2026-04-08T12:18:22Z

Azure Pipelines successfully started running 1 pipeline(s).

…changes

Copilot

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 2 comments.

jakobbotsch · 2026-04-09T15:04:40Z

/azp run runtime-coreclr jitstressregs

azure-pipelines · 2026-04-09T15:04:53Z

Azure Pipelines successfully started running 1 pipeline(s).

jakobbotsch · 2026-04-10T07:11:19Z

jitstressregs failures looks like the same that are present in main.

Rest looks like infra failures, will rerun CI.

jakobbotsch · 2026-04-14T08:29:56Z

/azp run runtime-coreclr jitstressregs, runtime-coreclr jitstress2-jitstressregs

azure-pipelines · 2026-04-14T08:30:15Z

Azure Pipelines successfully started running 2 pipeline(s).

Copilot

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated no new comments.

jakobbotsch · 2026-04-14T09:39:45Z

/azp run runtime-coreclr jitstressregs

azure-pipelines · 2026-04-14T09:40:03Z

Azure Pipelines successfully started running 1 pipeline(s).

jakobbotsch · 2026-04-15T08:10:12Z

cc @dotnet/jit-contrib PTAL @AndyAyersMS

Diffs. Some minor regressions mostly. I analyzed the arm64 ones and they happen "by chance": EH write-thru variables that cross calls are not strongly preferred into callee saves (unlike non EH-write thru variables).
Before this change, however, we would often have x0 taken by some interval that went through the def-use conflict resolution, and that would result in some unrelated EH write-thru variable being allocated into a callee save because x0 was taken and the next best register was a callee save.
With this change we end up with x0 being free and LSRA then allocates some EH write-thru variable to that register. That results in now needing additional reloads after some calls.
I tried a few heuristics to solve this but I was not able to come up with something that was a win in the majority of cases.

jakobbotsch · 2026-04-15T08:11:33Z

-            // for cmpxchg we need to keep the original value in RAX
-            assert(node->GetRegNum() == REG_RAX);


Nothing was stopping LSRA from changing the definition register on this node before, it just never happened because LSRA was preferring to change the user instead.

jakobbotsch · 2026-04-15T08:13:30Z

+            assert(node->GetRegNumByIdx(0) == REG_EAX);
+            assert(node->GetRegNumByIdx(1) == REG_EDX);


There is a "new" invariant that LSRA does not change definition registers on multireg nodes, since that would require a parallel store that is very hard to accomplish for codegen without a temporary. E.g. in this case we could need to do (eax, edx) = (edx, eax).

Copilot AI review requested due to automatic review settings March 9, 2026 14:02

dotnet-policy-service bot assigned jakobbotsch Mar 9, 2026

github-actions bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Mar 9, 2026

Copilot started reviewing on behalf of jakobbotsch March 9, 2026 14:04 View session

Copilot AI reviewed Mar 9, 2026

View reviewed changes

Comment thread src/coreclr/jit/lsrabuild.cpp Outdated

jakobbotsch added 3 commits March 9, 2026 15:17

Feedback

d3795f9

Fix GT_CATCH_ARG codegen

891feb0

Codegen fixes

6c82625

Copilot AI review requested due to automatic review settings March 10, 2026 13:04

Copilot started reviewing on behalf of jakobbotsch March 10, 2026 13:05 View session

Copilot AI reviewed Mar 10, 2026

View reviewed changes

Comment thread src/coreclr/jit/codegenlinear.cpp Outdated

jakobbotsch mentioned this pull request Mar 26, 2026

JIT: Fix register move ordering in genConsumeBlockOp when srcAddr is in dstReg #126133

Closed

jakobbotsch added 2 commits April 1, 2026 11:44

Merge branch 'main' of github.com:dotnet/runtime into remove-use-reg-…

bd1b361

…changes

Avoid changing use regs to all-in-use registers

ffdccca

Copilot AI review requested due to automatic review settings April 1, 2026 11:01

Copilot started reviewing on behalf of jakobbotsch April 1, 2026 11:03 View session

Copilot AI reviewed Apr 1, 2026

View reviewed changes

Comment thread src/coreclr/jit/lsrabuild.cpp

Merge branch 'main' of github.com:dotnet/runtime into remove-use-reg-…

8887784

…changes

jakobbotsch mentioned this pull request Apr 8, 2026

JIT miscompilation: Bmi1.X64.TrailingZeroCount returns wrong result with optimizations enabled (.NET 11 regression) #126667

Open

Do not change multireg defs

c7a1c5f

Fixes

27f6338

Copilot AI review requested due to automatic review settings April 9, 2026 14:16

Copilot started reviewing on behalf of jakobbotsch April 9, 2026 14:17 View session

Copilot AI reviewed Apr 9, 2026

View reviewed changes

Comment thread src/coreclr/jit/lsrabuild.cpp

Comment thread src/coreclr/jit/lsrabuild.cpp

Merge branch 'main' into remove-use-reg-changes

f88e5b5

Copilot AI mentioned this pull request Apr 11, 2026

Fix WasmTestOnChrome: exclude browser/wasi from gen-debug-dump-docs.py helix post-command #126789

Open

1 task

Merge branch 'main' into remove-use-reg-changes

042802e

Copilot AI review requested due to automatic review settings April 14, 2026 08:29

Copilot started reviewing on behalf of jakobbotsch April 14, 2026 08:30 View session

Copilot AI reviewed Apr 14, 2026

View reviewed changes

jakobbotsch marked this pull request as ready for review April 15, 2026 08:04

jakobbotsch requested a review from AndyAyersMS April 15, 2026 08:10

jakobbotsch commented Apr 15, 2026

View reviewed changes

		// for cmpxchg we need to keep the original value in RAX
		assert(node->GetRegNum() == REG_RAX);

		assert(node->GetRegNumByIdx(0) == REG_EAX);
		assert(node->GetRegNumByIdx(1) == REG_EDX);

Conversation

jakobbotsch commented Mar 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

jakobbotsch commented Apr 1, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

jakobbotsch commented Apr 1, 2026

Uh oh!

azure-pipelines bot commented Apr 1, 2026

Uh oh!

jakobbotsch commented Apr 8, 2026

Uh oh!

azure-pipelines bot commented Apr 8, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

jakobbotsch commented Apr 9, 2026

Uh oh!

azure-pipelines bot commented Apr 9, 2026

Uh oh!

jakobbotsch commented Apr 10, 2026

Uh oh!

jakobbotsch commented Apr 14, 2026

Uh oh!

azure-pipelines bot commented Apr 14, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

jakobbotsch commented Apr 14, 2026

Uh oh!

azure-pipelines bot commented Apr 14, 2026

Uh oh!

jakobbotsch commented Apr 15, 2026

Uh oh!

jakobbotsch Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

jakobbotsch Apr 15, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jakobbotsch commented Mar 9, 2026 •

edited

Loading