Thanks, committed as 5c18444289.

Ping.

Thanks, committed as 203b6f31bb71.

Ping.

Realised I hadn't updated the DomTree correctly first time round.

Convert RISC-V tests to use their preferred style.

Ping.

Thanks. 2002c82278f6 on main.

Reword comment to avoid confusion.

Thanks, resynthesized a test from a much smaller, single, switch and committed as dc700f1e4daf.

Thanks, it turns out 1c10d5b175992 actually implemented support (including windows) so I was on a stale branch. Sorry about that.

Ping.

Ping.

Thanks for the warning. Now that's settled down a bit I've updated the patch to the new regime.

Oh, thanks. I thought I'd seen something like that but only turned up the IR-level ones in my search. Switched. Also sanitized some of the strings in the test.

Thanks, pushed to main (3ed58d4df65e).

Thanks. Committed as 150595ab4be3.

Thanks, committed as 5b24d4210647

Thanks. Looks like I pushed it a while back (6c85668d28), which is weird.

Tests weren't running on Darwin properly for some reason so things were a bit broken there.

Belated ping for this one. It looks like it's come up again recently and fixes https://github.com/llvm/llvm-project/issues/59715.

Thanks, committed as 8ba9a5218782

Remove COMPILER_RT_HAS_FLOAT16 check. It's actually a soft/float ABI check rather than whether to support the type entirely (at least for conversions) so it meant the new functions didn't get built on Darwin (which is still soft-float for _Float16 compiler-rt calls for historic reasons).

Thanks Eli. I filed https://github.com/llvm/llvm-project/issues/59516 for the moves and pushed it with the changes you asked for (10d34f5538e0).

Updating diff to ToT.

I know I came up with the patch originally, but I think more emphasis needs to be put on the fact that it violates the required semantics of alwaysinline. I think this is a dirty hack that worked for us but isn't up to OSS quality.

Thanks. Committed as 6b98824a587a.

Ping.

Thanks, committed as b32280baf9ef

Yep, looks dangling to me.

Put the issue number in summary or test comments?

Thanks. Committed as 2bcf51c7f82c.

I don't think this is how division works. It mixes bits up far more than add/sub/mul for which a bit in the input can only affect equal or lower ones in the output.

Ping.

I think this looks reasonable now.

Thanks for working on this, I think this change is reasonable.

Ping.

In fact I noticed this problem when switching the ABI and tried to fix it in the compiler-rt side, see D128872.

Yep, basically the same thing. Abandoning this since it's already fixed.

Oops, yes. Thanks for that.

Sorry about that, committed as e62b3a9375d8.

I tried a few minor modifications to put things back, but in the end decided the interactions here were too complicated and even if I managed the result would be no neater than before this patch. So I've put up https://reviews.llvm.org/D134640 that essentially reverts to the original Clang.cpp, but expressed in terms of getDefaultUnwindTableLevel rather than the old callback.

Would you be okay with reverting this? I don't think these breaking changes are worth it just for a refactor.

Thanks, committed as 677da09d0259.

Thanks Florian. LGTM!

Accidentally clang-formatted the ampere1 test. Undo that.

Not really intended, no. Kind of correct in that with -fsanitize=thread it doesn't matter what unwind tables you requested, you're getting them. But also I think we consider the arguement to the extent we need to.

Just noticed it'd already been approved with those suggestions, so pushed the revised version: 4388b56d and 58f9abaed4aa.