update submodules for llvm debug info and fancy assertions

[glguy]

No description provided.

[robdockins]

The changes here in SAW look good to me.

However, based on the CI results, it looks like we need to be a bit more careful about how we parse the metadata intrinsics. I'm guessing clang has changed the signatures of those over time, so we'll need to figure out what the range of possibilities are.

[glguy]

The changes here in SAW look good to me.

However, based on the CI results, it looks like we need to be a bit more careful about how we parse the metadata intrinsics. I'm guessing clang has changed the signatures of those over time, so we'll need to figure out what the range of possibilities are.

What range of LLVM versions do we need to support well?

I can imagine a couple of options:

• ignore outdated metadata
• update the test cases to use a more recent LLVM
• add a bunch of metadata patterns we can recognize.

[robdockins]

According to the SAW readme, we're aiming to support LLVM versions from 3.5 to 11.0.

[glguy]

According to the SAW readme, we're aiming to support LLVM versions from 3.5 to 11.0.

I tried recompiling the salsa20 case with llvm 11 and saw couldn't handle it (independently of the metadata changes). I had to roll back to 9. I might have done something wrong but I don't think we actually have that complete of LLVM coverage.

LLVM 12 is out now, that'll need to be on the radar.

[robdockins]

We should definitely fix whatever is broken there with LLVM 11.

For the metadata stuff, I'd chose some reasonable version cutoff where the metadata looks more or less the way you need it to (they've changed their format quite a bit) and make sure to handle all formats from there forward, and gracefully ignore older ones.