[SM6.9] Derivative Calls Incorrectly Sunk Into Conditional Branches

[VladM1076]

Encountered this issue when running the the LongVector derivative tests in ExecHLSLTests, so marking it as SM6.9 despite it reproducing with SM66

DXC's optimizer sinks ddx/ddy calls into conditional branches when the result is only used inside that branch. This breaks quad semantics when only one lane executes the conditional branch - derivatives require all 4 lanes to participate.

Impact: Derivatives return 0 instead of correct values.
Affected: ddx, ddy, ddx_fine, ddy_fine, ddx_coarse, ddy_coarse in compute shaders (SM 6.6+).
Repros for both vector and scalar cases.
Also impacts other wave ops.

Steps to Reproduce
SM66 Scalar repro: https://godbolt.org/z/PMK9EoTnK
SM69 Vector repro: https://godbolt.org/z/xv4e5hjrx

Actual Behavior
ddx sunk into conditional branch only executed by lane 3, not all 4 lanes.

Environment

• DXC version dxcompiler.dll: 1.10 - 1.8.0.15125 (main, aa9b462); dxil.dll: 1.10(1.8.0.15125)
• Host Operating System Microsoft Windows [Version 10.0.19045.5011]

[VladM1076]

From my quick look into the code, it looks like at the HLSL level the results are not marked convergent which allows sinking them. Marking the results convergent resolves the test failures.
Not entirely sure if this is a sound solution, but it works :)

[alsepkow]

Vlad shared this with me and I asked him to open an issue.