Missed optimization for bitset based equality

[bmaurer]

https://godbolt.org/z/nTseoW9z1

bool test(unsigned long long x) {
    auto mask = x & 0xf;
    return (mask == 10 || mask == 1 || mask == 2);
}

clang generates:

test(unsigned long long):                               # @test(unsigned long long)
        and     edi, 15
        cmp     rdi, 11
        setb    cl
        mov     eax, 1030
        bt      eax, edi
        setb    al
        and     al, cl
        ret

the comparison against 11 is unnecessary given the range of the mask operation. GCC emits shorter code:

test(unsigned long long):
        and     edi, 15
        mov     eax, 1030
        bt      rax, rdi
        setc    al
        ret

[RKSimon]

simplifycfg folds:

define i1 @test(i64 noundef %x) {
entry:
  %and = and i64 %x, 15
  switch i64 %and, label %lor.rhs [
    i64 10, label %lor.end
    i64 1, label %lor.end
    i64 2, label %lor.end
  ]

lor.rhs:                                          ; preds = %entry
  br label %lor.end

lor.end:                                          ; preds = %entry, %entry, %entry, %lor.rhs
  %0 = phi i1 [ true, %entry ], [ false, %lor.rhs ], [ true, %entry ], [ true, %entry ]
  ret i1 %0
}

to:

define i1 @test(i64 noundef %x) {
entry:
  %and = and i64 %x, 15
  %0 = icmp ult i64 %and, 11
  %switch.cast = trunc i64 %and to i11
  %switch.shiftamt = mul nuw nsw i11 %switch.cast, 1
  %switch.downshift = lshr i11 -1018, %switch.shiftamt
  %switch.masked = trunc i11 %switch.downshift to i1
  %1 = select i1 %0, i1 %switch.masked, i1 false
  ret i1 %1
}

which instcombine then folds to:

define i1 @test(i64 noundef %x){
entry:
  %and = and i64 %x, 15
  %0 = icmp ult i64 %and, 11
  %switch.cast = trunc i64 %and to i11
  %switch.downshift = lshr i11 -1018, %switch.cast
  %1 = and i11 %switch.downshift, 1
  %switch.masked = icmp ne i11 %1, 0
  %2 = select i1 %0, i1 %switch.masked, i1 false
  ret i1 %2
}