Fix suboptimal codegen for short or patterns in is-expressions by stevenelliottjr · Pull Request #82429 · dotnet/roslyn

stevenelliottjr · 2026-02-17T12:40:39Z

Summary

For short or patterns (≤4 tests), the compiler now uses an inverted linear test sequence instead of the general DAG lowering. This eliminates unnecessary boolean temporaries and BoundSpillSequence wrapping that caused IL bloat.

For example, c is '/' or '\\' now produces:

!(c != '/' && c != '\\')

instead of the general DAG path which created a bool temp with true/false assignments and gotos.

Large or patterns (5+ tests) continue to use the general DAG lowering with switch dispatch (binary search trees), which is more efficient at scale.

IL size improvements

Pattern	Before	After
`o is int or long` (Debug)	26 bytes	21 bytes
`o is int or long` (Release)	24 bytes	20 bytes
`(o is int or long) ? "True" : "False"` (Debug)	40 bytes	29 bytes
`x > 0 && c is '/' or '\\'` (Release)	—	19 bytes (no bool temp)

Why this approach vs previous attempts

This problem was previously addressed in PR #68694 (reverted in #69582) and PR #72273 (reverted in #72827 due to compiler crash #72753). Both previous attempts modified the general DAG lowering to avoid the result temp, which was fragile because it interacted with exception handler scopes (await using, try/finally), causing NullReferenceException in ILBuilder.BlockedBranchDestinationSlow.

This PR takes a fundamentally different approach:

Does not modify the general DAG lowering at all
Routes short or patterns to the existing, well-tested linear rewriter with swapped labels
Adds a maxTests parameter to canProduceLinearSequence to limit the optimization to short sequences
The linear rewriter doesn't use SpillSequence and doesn't interact with exception handler scopes, avoiding the crashes that killed previous attempts

Changes

LocalRewriter_IsPatternOperator.cs: Added inverted linear sequence path for short or patterns; added maxTests parameter to canProduceLinearSequence
PatternTests.cs: Updated IL expectations for IsPatternDisjunct_01 and _02; added new test IsPatternDisjunct_OrPatternInAndExpression for the exact scenario from Suboptimal codegen for is pattern compared to .NET 6/7 #80052

Test plan

All 6 IsPatternDisjunct tests pass (5 existing + 1 new)
IsWarningSwitchEmit passes (large or pattern still uses switch dispatch)
Full CodeGen test suite: 6245 passed, 0 failed
Full Emit test suite: 6986 passed, 0 failed
Verify no regressions in Semantic test suite
Verify the crash repro from [dotnet-sdk-9.0.100-preview.3.24175.24] Microsoft.CSharp.Core.targets thrown errors when using "dotnet build" #72753 (is X or Y inside await using) works correctly

…net#80052) For short `or` patterns (up to 4 tests), use an inverted linear test sequence instead of the general DAG lowering. The general path introduces unnecessary bool temporaries and a SpillSequence, producing larger and slower IL. The inverted linear path produces clean boolean expressions (e.g., `!(x != '/' && x != '\\')`) that the IL emitter handles efficiently. For large `or` patterns, the general DAG lowering with switch dispatch (binary search trees) is still preferred.

stevenelliottjr · 2026-02-17T12:43:17Z

@dotnet-policy-service agree

AlekseyTs · 2026-02-17T18:29:36Z

@stevenelliottjr It looks like there are legitimate test failures. #Closed

stevenelliottjr · 2026-02-17T20:26:30Z

Apologies for the noise, @AlekseyTs. I'll investigate the test failures and get them resolved.

The inverted linear sequence optimization for short `or` patterns followed "failure" branches, skipping BoundWhenDecisionDagNodes that contain variable bindings. This left captured variables uninitialized, causing NullReferenceException for patterns like `(B or C) and var x`. Add dagContainsBindings guard to fall back to general DAG lowering when the pattern contains variable captures. Update IL baselines for Span pattern tests to reflect the new (correct, smaller) codegen.

stevenelliottjr · 2026-02-17T20:56:19Z

@AlekseyTs I've pushed a fix. The issue was that the new inverted linear sequence optimization was skipping BoundWhenDecisionDagNodes with variable bindings — patterns like (B or C) and var x would leave captured variables uninitialized. I added a dagContainsBindings guard so those patterns fall back to the general DAG lowering. Also updated the Span pattern IL baselines. All 1350 PatternMatchingTests pass with zero regressions.

AlekseyTs · 2026-02-18T14:54:35Z

Large or patterns (5+ tests) continue to use the general DAG lowering with switch dispatch (binary search trees), which is more efficient at scale.

Is this just an assumption, or this statement is supported by some real numbers? #Closed

AlekseyTs · 2026-02-18T14:59:46Z

src/Compilers/CSharp/Portable/Lowering/LocalRewriter/LocalRewriter_IsPatternOperator.cs

            // linear tests with a single "golden" path to the true label and all other paths leading
            // to the false label?  This occurs with an is-pattern expression that uses no "or" or "not"
            // pattern forms.
+            // When maxTests is specified, the sequence is limited to at most that many test nodes.


When maxTests is specified, the sequence is limited to at most that many test nodes.

I think "When maxTests is specified" part is somewhat confusing. Consider adjusting to: "maxTests limits the amount BoundTestDecisionDagNode nodes in the sequence."

AlekseyTs · 2026-02-18T15:09:41Z

Large or patterns (5+ tests) continue to use the general DAG lowering with switch dispatch (binary search trees), which is more efficient at scale.

Is this just an assumption, or this statement is supported by some real numbers?

I guess this is just a rephrase of an existing comment in implementation #Closed

AlekseyTs · 2026-02-18T15:15:16Z

            isPatternRewriter.Free();

It looks like this code is a copy of the previous block body. Consider extracting and reusing a common helper.

Refers to: src/Compilers/CSharp/Portable/Lowering/LocalRewriter/LocalRewriter_IsPatternOperator.cs:57 in 8e13d09. [](commit_id = 8e13d09, deletion_comment = False)

AlekseyTs · 2026-02-18T15:24:35Z

src/Compilers/CSharp/Portable/Lowering/LocalRewriter/LocalRewriter_IsPatternOperator.cs

+        {
+            foreach (var node in decisionDag.TopologicallySortedNodes)
+            {
+                if (node is BoundWhenDecisionDagNode { Bindings.Length: > 0 })


Length: > 0

IsEmpty: false?

AlekseyTs · 2026-02-18T15:44:02Z

src/Compilers/CSharp/Test/Emit/CodeGen/PatternTests.cs

+            compilation.VerifyDiagnostics();
+            var expectedOutput = @"TrueTrueFalseFalse";
+            var compVerifier = CompileAndVerify(compilation, expectedOutput: expectedOutput);
+            // The `is '/' or '\\'` pattern should produce comparable IL to `c == '/' || c == '\\'`


// The is '/' or '\\' pattern should produce comparable IL to c == '/' || c == '\\'

It feels like we should assert that M2 results in the same IL

AlekseyTs · 2026-02-18T15:51:19Z

Done with review pass (commit 2)

stevenelliottjr requested a review from a team as a code owner February 17, 2026 12:40

github-actions bot added the Area-Compilers label Feb 17, 2026

dotnet-policy-service bot added the Community The pull request was submitted by a contributor who is not a Microsoft employee. label Feb 17, 2026

This was referenced Feb 17, 2026

The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#3008

Open

[android-arm64] The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#6408

Open

stevenelliottjr force-pushed the fix/pattern-match-codegen-regression branch from f37fda5 to 8e13d09 Compare February 17, 2026 20:55

AlekseyTs reviewed Feb 18, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix suboptimal codegen for short or patterns in is-expressions#82429

Fix suboptimal codegen for short or patterns in is-expressions#82429
stevenelliottjr wants to merge 2 commits intodotnet:mainfrom
stevenelliottjr:fix/pattern-match-codegen-regression

stevenelliottjr commented Feb 17, 2026

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

AlekseyTs commented Feb 17, 2026 •

edited

Loading

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

AlekseyTs commented Feb 18, 2026 •

edited

Loading

Uh oh!

AlekseyTs Feb 18, 2026

Uh oh!

AlekseyTs commented Feb 18, 2026 •

edited

Loading

Uh oh!

AlekseyTs commented Feb 18, 2026

Uh oh!

AlekseyTs Feb 18, 2026

Uh oh!

AlekseyTs Feb 18, 2026

Uh oh!

AlekseyTs commented Feb 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Comments

Conversation

stevenelliottjr commented Feb 17, 2026

Summary

IL size improvements

Why this approach vs previous attempts

Changes

Test plan

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

AlekseyTs commented Feb 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

stevenelliottjr commented Feb 17, 2026

Uh oh!

AlekseyTs commented Feb 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AlekseyTs Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

AlekseyTs commented Feb 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

AlekseyTs commented Feb 18, 2026

Uh oh!

AlekseyTs Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

AlekseyTs Feb 18, 2026

Choose a reason for hiding this comment

Uh oh!

AlekseyTs commented Feb 18, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Comments

AlekseyTs commented Feb 17, 2026 •

edited

Loading

AlekseyTs commented Feb 18, 2026 •

edited

Loading

AlekseyTs commented Feb 18, 2026 •

edited

Loading