Regression in inlining performance from 1.82.0

[Guilucand]

Hi, I found a suboptimal code generation while inlining some struct functions starting from Rust 1.82.0 (including nightly).

Here's a link to a compiler explorer minimal example:
https://godbolt.org/z/WKqMr1fE7

In Rust < 1.82.0 the code is optimized correctly, while in versions above the code is different and an inefficient imul instruction appears, causing a major performance drop during execution.

It seems that the problem is already present in the generated llvm IR (generated with --emit=llvm-ir).

I tested various configurations in Rust 1.82.0:

• Debug symbols disabled => suboptimal
• TestStruct without #[repr(C)] and -C debuginfo=2 => suboptimal
• TestStruct with #[repr(C)] and -C debuginfo=2 => correct

I also uploaded the code to github:
https://github.com/Guilucand/rust-inlining-bug

[nikic]

Looks like LLVM 19 started "optimizing" this code in an interesting way...

https://llvm.godbolt.org/z/45cWbE3eY

define i8 @src(i8 noundef %_23) {
  %_17 = lshr i8 %_23, 4
  %_18 = and i8 %_23, 15
  %_16 = add nuw nsw i8 %_17, %_18
  ret i8 %_16
}
define i8 @tgt(i8 noundef %_23) {
  %_17 = lshr i8 %_23, 4
  %1 = mul i8 %_17, -15
  %_16 = add i8 %1, %_23
  ret i8 %_16
}

That's a regression both in the used instruction type and dependency length.

Upstream issue: llvm/llvm-project#147176