[llvm] [InstCombine]: Replace overflow calculation with intrinsic (PR #168195)

[via llvm-commits]

mikaseianatsu wrote:

I noticed the same pattern also appears in vector form, but it is not folded yet.
Scalar:

```llvm
define i1 @src(i16 %x, i16 %y) {
  %1 = sext i16 %x to i32
  %2 = sext i16 %y to i32
  %3 = add i32 %1, -32768
  %4 = sub i32 %3, %2
  %5 = icmp ult i32 %4, -65536
  ret i1 %5
}
```

can be optimized to:

```llvm
define i1 @tgt(i16 %x, i16 %y) {
  %1 = call { i16, i1 } @llvm.ssub.with.overflow.i16(i16 %x, i16 %y)
  %2 = extractvalue { i16, i1 } %1, 1
  ret i1 %2
}
```

But the vector version:

```llvm
define <8 x i1> @src(<8 x i16> %0, <8 x i16> %1) {
entry:
  %2 = sext <8 x i16> %0 to <8 x i32>
  %3 = sext <8 x i16> %1 to <8 x i32>
  %4 = add <8 x i32> %2, splat (i32 -32768)
  %5 = sub <8 x i32> %4, %3
  %6 = icmp ult <8 x i32> %5, splat (i32 -65536)
  ret <8 x i1> %6
}
```

currently remains unchanged
So this pattern seems to be recognized for scalar, but not yet for vectors.

https://github.com/llvm/llvm-project/pull/168195