I could _not_ reproduce the problems I previously saw when building the Rust compiler, so seems to be all good now :).

In D141712#4348919, @Carrot wrote:

@TimNN, I have reverted it, LLVM buildbot also reported a sanitized bootstrap failure.

If you have simple repro, it will be a great help to me.

I strongly suspect that this patch is breaking the Rust compiler when building against LLVM head:

The nofpclass-invalid.ll test added by this patch fails for me locally, and on the buildbot as well: https://lab.llvm.org/buildbot/#/builders/109/builds/58503

The current code is failing for a situation like this:

The ThinLTO related breakage I mentioned above should be fixed as of https://github.com/llvm/llvm-project/commit/451799bb8261bde52bbfef226d019caf1d82aa42.

Rebase onto HEAD.

Fix definition/declaration wording in comment.

Don't use anonymous values in tests

Adapt to changes from D137360.

Thanks a lot @nikic! I've verified that this unbreaks Rust's codegen test.

Just as a heads-up, this regresses one of Rust's codegen tests: fn check_foo2 in [issue-45222.rs](https://github.com/rust-lang/rust/blob/master/src/test/codegen/issue-45222.rs) is no longer being constant-folded, AFAICT because IndVarSimplifyPass no longer works on the %a = tail call i64 @llvm.usub.sat.i64(i64 %b, i64 1) generated by this patch (instead of %a = add i64 %b, -1 without this patch).

In D139209#4013911, @gulfem wrote:

I'm thinking that I can land D137360, and you can extend updateAttributes() function to include a src and dest arguments, and call your ensureFunctionCompatibility() function from there.

Only call ensureFunctionCompatibility if linkGlobalValueBody isn't called.

Sync to HEAD to check if that resolves the (IMO likely unrelated) test timeout.

Thanks for the review and also your input on https://github.com/llvm/llvm-project/issues/59347!

Address review comments

I'm sorry for the noise. Further investigation has shown that this happens when Rust is doing (thin) LTO, and I don't think this patch can be considered in any way "at fault" here, so this is the last you'll hear from me on the topic here.

I didn't manage to repro with opt, so still no compilable IR. I did some more debugging, though:

I've included excerpts from the IR below. It will take me a bit to provide something compilable. Though you are right, the noundef did indeed get removed from the call.

I'm still trying to properly minimize this, but this definitely interacts badly with other optimizations (which triggers the Rust CI failure I mentioned above):

This also broke Rust when compiled at LLVM head: https://buildkite.com/llvm-project/rust-llvm-integrate-prototype/builds/13166#0182fb4a-0f2d-4f2e-830f-f62b463b8d48. (I don't know whether this was some existing UB that got only exposed by this patch or not, but wanted to make you aware. Reproduces with 2+ codegen units and a 1+ optimization level).

Sorry for taking so long to reply. The original example I shared (https://gist.github.com/TimNN/501819422631149c3ab2b8cc0b15c98d) reproduces the issue for the AVR target. I'm not sure about the terminology here, but I would consider AVR "in-tree", even though it is still experimental. I don't know if / how the issue can be reproduced on other targets.

Friendly Ping! Is there anything I can do to help this patch along? It looks like no reviewers have been assigned yet, do you know of anyone who would be appropriate, @shiva0217?

I don't think I can help here any more. The changes to hasPredecessorHelper look good to me an definitely fix the original problem. However I don't know enough about the surrounding code to comment on the changes in AssignTopologicalOrder.

The approach in general looks good to me, however I don't think this interacts correctly with the TopologicalPrune optimization.

Another fix has been commited.

I cannot commit this myself, could someone please do it for me?

@hfinkel: Thanks for the feedback!

Updated to a more targeted fix and added a test as suggested.

I cannot commit this myself, could someone please do it for me?

I have updated the patch as suggested.