[llvm] [X86] combineAdd - fold (add (sub (shl x, c), y), z) -> (sub (add (shl x, c), z), y) (PR #142734)

[Daniel Thornburgh via llvm-commits]

mysterymath wrote:

> > Hey there! The Fuchsia team is seeing dramatic slowdowns on our Mac AArch64 toolchain builders, from 3.5 hours to a timeout hit at 5 hours. This is one of two plausible commits in the blamelist. It only seems to appear in our LTO 2-stage prod builds. We do build builtins for X86 and other platforms. Do you think something in this change could cause this kind of slowdown? We haven't been able to directly determine the culprit beyond a rough range.
> 
> Is it a slowdown or is the compiler hung?

It's definitely a compiler hang; it occurs during the first CMake ABI check against the just-built clang. It only occurs in LTO mode, which I didn't realize. I'm running through a fresh round of trials with reverts for the 4 changes in the blamelist (this is still one of them); I'll post back with either an exoneration or confirmation that this is causing the hang.

https://github.com/llvm/llvm-project/pull/142734