Failure to propogate knowledge from comparison into branches

[Kmeakin]

This rust code produces sub-optimal assembly:

https://godbolt.org/z/Ex7jKGzj7

#[no_mangle]
pub fn starts_with_a(s: &str) -> bool {
    s.bytes().next() == Some(b'a')
}

starts_with_a:
        cbz     x1, .LBB0_2
        ldrb    w8, [x0]
        cmp     w8, #97
        cset    w8, eq
        cmp     x1, #0
        cset    w9, ne
        and     w0, w9, w8
        ret
.LBB0_2:
        cmp     x1, #0
        cset    w9, ne
        and     w0, w9, w8
        ret

The cmp x1, #0 in each branch is implied true/false by the cbz x1 test. The optimal assembly can be produced by removing the select in the final basic block:

https://godbolt.org/z/87MKvPjf6 https://alive2.llvm.org/ce/z/VofZ27

define i1 @src(ptr %s.0, i64 %s.1) {
bb1:
  %0 = icmp ne i64 %s.1, 0
  br i1 %0, label %bb2, label %bb3

bb2:
  %2 = load i8, ptr %s.0, align 1
  %3 = icmp eq i8 %2, 97
  br label %bb3

bb3:
  %4 = phi i1 [ %3, %bb2 ], [ undef, %bb1 ]
  %5 = and i1 %0, %4
  ret i1 %5
}

define i1 @tgt(ptr %s.0, i64 %s.1) {
bb1:
  %0 = icmp ne i64 %s.1, 0
  br i1 %0, label %bb2, label %bb3

bb2:                                       
  %2 = load i8, ptr %s.0, align 1
  %3 = icmp eq i8 %2, 97
  br label %bb3

bb3:
  %4 = phi i1 [ %3, %bb2 ], [ 0, %bb1 ]
  ret i1 %4
}

[zjaffal]

@fhahn is this something we can simplify in ConstraintElimination ?