regression: `@divExact` no longer emits `*div exact` LLVM IR

[xdBronch]

Zig Version

0.12.0-dev.3518+d2be725e4

Steps to Reproduce and Observed Behavior

export fn div(a: i32) i32 {
    return @divExact(a, 2);
}

https://godbolt.org/z/zEhYTY1h9
because of @divExacts semantics with safety disabled this should be turned into a shift which happened until 0.12 where it emits %1 = sdiv i32 %0, 2 LLVM IR rather than %1 = ashr exact i32 %0, 1

Expected Behavior

zig should emit a right shift to llvm in unsafe modes

[xdBronch]

caused by #19031

[xdBronch]

actually this isnt exclusive to powers of 2, https://godbolt.org/z/9czo197Mh other denominators used to emit sdiv exact now its just sdiv

[xdBronch]

im investigating this a little more and im a bit confused. the ir from -femit-llvm-ir gives sdiv while the ir from --verbose-llvm-ir gives sdiv exact which i believe is somewhat expected since the latter is post optimizations? i dont fully understand when/how this is optimized, is it zig or an earlier stage in llvm? whats confusing me is that manually running them thru llc, the output from --verbose-llvm-ir turns this into a shift as you'd expect, so why isn't that llvm ir getting to llvm during the normal compilation process? is that the underlying cause of the bug or am i misunderstanding something?