[InstCombine] Preferred shift direction in bit test

[LebedevRI]

Bugzilla Link	42466
Version	trunk
OS	Linux
CC	@topperc,@davidbolvansky,@efriedma-quic,@RKSimon,@nikic,@rotateright

Extended Description

As discussed in https://reviews.llvm.org/D63829

These two patterns are identical:

define i1 @​v0(i32 %x, i32 %y, i32 %z) {
%t0 = lshr i32 %x, %y
%t2 = and i32 %t0, %z
%t3 = icmp ne i32 %t2, 0
ret i1 %t3
}
define i1 @​v1(i32 %x, i32 %y, i32 %z) {
%t0 = shl i32 %y, %y
%t2 = and i32 %t0, %x
%t3 = icmp ne i32 %t2, 0
ret i1 %t3
}

We should canonicalize to one of them.
Probably to @​v0.
With one obvious exception of (1 << y) & x.

https://godbolt.org/z/KPgPPQ
https://rise4fun.com/Alive/0avd

[LebedevRI]

Still trying to understand the llvm/test/Transforms/InstCombine/onehot_merge.ll
regression in preparatory patch llvm/llvm-bugzilla-archive#42470
(the failure can be seen in https://reviews.llvm.org/D62818#change-4wdEQMcvPYE5).

But it should be noted that we really should have a reversing fold in backend, regardless of this patch.
We currently already canonicalize "((X <> Y) & C) ==/!= 0" to "(X & (C l>>/<< Y)) ==/!= 0":
https://godbolt.org/z/OwZGrD
^ and that is the opposite of what backends want :)

https://reviews.llvm.org/D62871 contains the de-canonicalization.