[CVP] Missed opportunity to do slt to ult conversion

[vladimirradosavljevic]

For the following example, CVP is not converting slt to ult:

define i64 @test_select(i64 %arg, i64 %arg1) {
entry:
  %cmp1 = icmp ult i64 %arg, 64424509440
  %and = and i64 %arg1, 4294967295
  %cmp2 = icmp slt i64 %arg, %and
  %select = select i1 %cmp1, i1 %cmp2, i1 false
  br i1 %select, label %good, label %bad

bad:
  ret i64 1

good:
  ret i64 0
}

Proof that this transformation is valid.
On the other hand, it is able to convert it if there are 2 BBs and no select instruction.

[XChy]

CVP leverages only getConstantRange(Value *V, Instruction *CxtI, bool UndefAllowed) to flip signedness predicate, based on def level.
Here is a possible solution. We could get the range in use level. But I'm not sure it's profitable.
If it has no much impact on compile-time, I may post a pr to solve it.

[vladimirradosavljevic]

CVP leverages only getConstantRange(Value *V, Instruction *CxtI, bool UndefAllowed) to flip signedness predicate, based on def level. Here is a possible solution. We could get the range in use level. But I'm not sure it's profitable. If it has no much impact on compile-time, I may post a pr to solve it.

Thanks for working on this!
I saw in you patch that you are using noundef for arguments in tests, and without that, transformation is not working. Do you know why is that so?

[XChy]

I saw in you patch that you are using noundef for arguments in tests, and without that, transformation is not working. Do you know why is that so?

getConstantRangeAtUse requires noundef. There is the comment:

// If the value is undef, a different value may be chosen in
// the select condition and at use.
if (!isGuaranteedNotToBeUndefOrPoison(SI->getCondition(), AC))
        break;

Some at-use transforms may be unsafe with undef values. I think it may be OK in this issue, but LVI Analysis does not provide such option.