[Middle-end] Moving truncate before right-shift

[LebedevRI]

Bugzilla Link	46234
Version	trunk
OS	Linux
CC	@RKSimon,@nikic,@rotateright

Extended Description

int src(long int num) {
num &= 0xFFFFFFF;
return num >> 2;
}
int tgt(long int num) {
return (((unsigned int)num) << 4) >> 6;
}

---

define i32 @​src(i64 %0) {
%1:
%2 = lshr i64 %0, 2
%3 = trunc i64 %2 to i32
%4 = and i32 %3, 67108863
ret i32 %4
}
=>
define i32 @​tgt(i64 %0) {
%1:
%2 = trunc i64 %0 to i32
%3 = lshr i32 %2, 2
%4 = and i32 %3, 67108863
ret i32 %4
}
Transformation seems to be correct!

https://godbolt.org/z/V7xRAy
^ we do that in backend, but not middle-end.

[rotateright]

The backend is using demanded bits to get this, and that seems to rely on knowing that narrowing and extending back to the destination type is free.

Probably just need to add a specific fold for this. I did something similar here:
https://reviews.llvm.org/rG2552f65183eb

[LebedevRI]

Thanks!

[LebedevRI]

Is there anything else left to do here, or can this be closed?

[rotateright]

We're still not getting this in instcombine. The patch I pointed to in comment 1 was for a different pattern.