[llvm-branch-commits] [llvm] release/19.x: [AVR] Fix 16-bit LDDs with immediate overflows (#104923) (PR #106993)

[Patryk Wychowaniec via llvm-branch-commits]

Patryk27 wrote:

> Is this PR a fix for a regression or a critical issue?

It's a fix for critical issue within the AVR codegen which causes it to generate invalid binaries and/or crash llvm (depending on which concrete code path is taken).

It was spotted downstream in Rust, https://github.com/Rahix/avr-hal/issues/573.

> What is the risk of accepting this into the release branch?

The fix has been tested both within LLVM (as proven by the new tests added here) and by recompiling rustc and making sure the downstream project works correctly as well, so the risk is small.

What's more, the change is isolated to the AVR target - there's no way it could affect other targets.

> What is the risk of NOT accepting this into the release branch?

Not huge, I think - rustc has its own fork of llvm-project`where I could cherry-pick this commit anyway. Having backport is just the preferred way of the rustc team, but I think it's not a rule set in stone. 

https://github.com/llvm/llvm-project/pull/106993