[BDCE] Drop nsw/nuw/exact from users of trivialized values

[Natalie Chouinard (sudonatalie)]

BDCE replaces values that have no live bits with zero, but left nsw/nuw/exact on their users. Those flags no longer held, so later passes could fold the users to wrong constants.

Backports llvm/llvm-project@fe346f9f5be9 (D36592) and its follow-ups llvm/llvm-project@a1067d9baecd and llvm/llvm-project@9e54b7093abd, adapted to BDCE's current state in DXC.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

Repeated user-chain traversal introduces quadratic work on chains of zero-demanded instructions.

1 open finding

• Reuse visited set to avoid repeated dead-code scans · New

What changed in this PR

Backports LLVM BDCE fixes to prevent shader miscompilation when dead-bit values are replaced with zero.

Changes:

• Clears invalid nsw, nuw, and exact flags along affected user chains.
• Adds an HLSL regression test for incorrect groupshared indexing.

File	Description
tools/​clang/​test/​DXC/​Passes/​BDCE/​bdce-trivialized-clears-nuw.hlsl	Tests the shader miscompilation regression.
lib/​Transforms/​Scalar/​BDCE.cpp	Clears arithmetic flags before replacing dead-bit values.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Damyan Pepper (damyanp)]

/azp run

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).

[Copilot]

🔵 Needs a closer look

The user-visible miscompilation fix needs release-note coverage under the repository policy.

0 open findings

1 resolved since last review

• Reuse visited set to avoid repeated dead-code scans

Previously missed (1)

In code that hasn't changed since last review

Add release note for shader miscompilation fix

lib/​Transforms/​Scalar/​BDCE.cpp:436

Please add a one-sentence entry under ### Upcoming Release → #### Bug Fixes in docs/ReleaseNotes.md. This fixes user-visible shader miscompilation, including an incorrect out-of-range array index, and meets the release-note criteria in CONTRIBUTING.md:115–133. For example: “Fixed incorrect shader results caused by stale overflow and exactness flags after bit-tracking dead code elimination.”

🧠 Review effort: Balanced

[Copilot]

🟢 Approval recommended

No blocking issues remain, and the previously reported repeated-traversal problem is resolved.

0 open findings

🧠 Review effort: Balanced

[Chris B (llvm-beanz)]

Using an end-to-end test to verify correctness of an IR pass change is not ideal. I know this is how DXC had been tested in the past, but for a good while now we've been pushing people to do IR->IR tests for IR pass chagnes.

There is a script ExtractIRForPassTest.py that you can use to extract the IR from an HLSL source prior to a given pass. That should make it straightforward to convert this test case to an IR test.

[Natalie Chouinard (sudonatalie)]

Switched it to an IR test, thanks!

[Copilot]

🟢 Approval recommended

No unresolved issues remain in the focused implementation, regression coverage, or release notes.

0 open findings

🧠 Review effort: Balanced