interesting codegen difference between idiomatic check and pattern matching discriminant checks

https://godbolt.org/z/ebPdcdcfo

both functions have the same input Option type, one does the idiomatic if is_none() { init() } and the other does the check by pattern matching if let None = opt { init() }.

assembly

aaa generates an extra ret and compares a qword while bbb only compares a dword and has a single ret.

aaa:
        cmp     qword ptr [rdi], 0
        je      .LBB0_1
        ret
.LBB0_1:
        ; init
        ret

bbb:
        cmp     dword ptr [rdi], 1
        je      .LBB1_2
        ; init
.LBB1_2:
        ret

when compiled with -Cno-prepopulate-passes, the differences between the two functions disappear in terms of assembly.

LLVM-IR

we can see that LLVM resorted to swapping the labels in the branch for bbb. another note is that one does a compare to 0 (icmp) and branch on the result, the other truncates the load and branches on the truncated value.

define void @aaa(... %opt) {
start:
  %_4 = load i64, ptr %opt, align 8
  %_3.not = icmp eq i64 %_4, 0
  br i1 %_3.not, label %bb1, label %bb2

bb2:
  ret void

bb1:
  ; init
  br label %bb2
}

define void @bbb(... %opt) {
start:
  %_2 = load i64, ptr %opt, align 8
  %0 = trunc nuw i64 %_2 to i1
  br i1 %0, label %bb2, label %bb1

bb2:
  ret void

bb1:
  ; init
  br label %bb2
}

even with -Cno-prepopulate-passes, the trunc vs icmp difference remains.

Reproduce the example from the linked Compiler Explorer case with and without -Cno-prepopulate-passes. Compare the generated assembly and LLVM IR for the two Rust functions, focusing on the branch-label ordering and icmp versus trunc. Done means identifying the compiler stage responsible and either explaining the difference as expected or providing a verified correction.