Instruction pmaxsw not generated

[mariannems]

The code generated by 'x86-64 clang 13.0.1' in /O3 will generate pmaxsw instructions. With 'x86-64 clang 14.0.0' and 'x86-64 clang (trunk)', the pmaxw is not generated, and the code size is slightly bigger.

short vecreduce_smax_v2i16(int n, short* v)
{
  short p = 0;
  for (int i = 0; i < n; ++i)
    p = p < v[i] ? v[i] : p;
  return p;
}

[llvmbot]

@llvm/issue-subscribers-backend-x86

[davidbolvansky]

Regressed with new intrinsics?

@rotateright @nikic

[RKSimon]

Godbolt: https://godbolt.org/z/7vxqbMzaa

This looks to be generating pmaxsw instructions to me?

[mariannems]

Here is what I was seeing: https://godbolt.org/z/hxz1xhMPh

[davidbolvansky]

hm, lul. Difference is whether we are in C or C++ mode.

https://godbolt.org/z/8zeqzoff9

[davidbolvansky]

Interestingly, Clang in C mode creates phi i32, in C++ phi i16.

Caused by some language rules, @AaronBallman ?

[llvmbot]

@llvm/issue-subscribers-clang-codegen