Skip to content

DXC can sink and duplicate SetMeshOutputCounts if the inputs are computed within a scalar branch #8104

Open
#8438
Open
DXC can sink and duplicate SetMeshOutputCounts if the inputs are computed within a scalar branch#8104
#8438
Assignees
llvm-beanz
Labels
bugBug, regression, crashcorrectnessBugs that impact shader correctness
Milestone
Backlog

Description

@karltechno
Karl Davis (karltechno)
opened on Jan 27, 2026

Description
DXC can sink a call to SetMeshOutputCounts into a scalar branch if the branches are used to compute the number of vertices/indices. Anecdotally I have been able to fix it by wrapping the branch condition with WaveReadLaneFirst, despite the branch being scalar anyway.

Steps to Reproduce
See https://godbolt.org/z/oca77vqGh

Actual Behavior
DXC validation fails after SetMeshOutputCounts is duplicated in resulting DXIL:
<source>:32:5: error: SetMeshOUtputCounts cannot be called multiple times.

Environment

  • DXC version: DXC 1.8.2505 on compiler explorer
  • Host Operating System Observed on: windows and linux (via compiler explorer)

Activity

  1. Karl Davis (karltechno)
    added
    bugBug, regression, crash
    needs-triageAwaiting triage
    on Jan 27, 2026
  2. github-project-automation
    added this to HLSL Triageon Jan 27, 2026
  3. Damyan Pepper (damyanp)
    added
    correctnessBugs that impact shader correctness
    and removed
    needs-triageAwaiting triage
    on Feb 3, 2026
  4. Damyan Pepper (damyanp)
    added this to the Backlog milestone on Feb 3, 2026
  5. Damyan Pepper (damyanp)
    moved this to Triaged in HLSL Triageon Feb 3, 2026

  6. damyanp commented on Feb 3, 2026

    @damyanp
    Damyan Pepper (damyanp)
    on Feb 3, 2026
    Member

    A possible workaround for this is to change the SetMeshOutputs call to:

        SetMeshOutputCounts(WaveReadLaneFirst(nverts), nprims);

    See https://godbolt.org/z/W4bh6dErW

  7. llvm-beanz commented on Apr 24, 2026

    @llvm-beanz
    Chris B (llvm-beanz)
    on Apr 24, 2026
    Collaborator

    The correct fix here is that we should just annotate these operations with noduplicate.

  8. Chris B (llvm-beanz)
    assigned
    Copilot
    and
    Chris B (llvm-beanz)
    on Apr 24, 2026

  9. Copilot commented on Apr 24, 2026

    Copilot
    CopilotAI
    on Apr 24, 2026
    Contributor

    The agent encountered an error and was unable to start working on this issue: Please try again later, or contact support if the problem persists. (Request id: DC94:3F9BEF:447318F:5993951:69EBE042)

  10. Chris B (llvm-beanz)
    unassigned
    Copilot
    on Apr 24, 2026
  11. Chris (chrisbieneman)
    assigned
    Copilot
    and
    Chris (chrisbieneman)
    on Apr 24, 2026
  12. Copilot
    linked a pull request that will close this issueAdd `noduplicate` to SetMeshOutputCounts and DispatchMesh intrinsics #8407on Apr 24, 2026

  13. damyanp commented on Apr 24, 2026

    @damyanp
    Damyan Pepper (damyanp)
    on Apr 24, 2026
    Member

    Chris B (@llvm-beanz) - can you have copilot work on this in your fork please?

  14. Chris B (llvm-beanz)
    unassigned
    Chris (chrisbieneman)
    and
    Copilot
    on May 10, 2026
  15. Chris B (llvm-beanz)
    added a parent issue on Sep 8, 2026
  16. Chris B (llvm-beanz)
    removed a parent issue on Sep 8, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugBug, regression, crashcorrectnessBugs that impact shader correctness

Type

No type

Projects

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions