[x86][reg][performance] `addsubpd` not generated in complex multiplication since LLVM 13

[p0nce]

Consider the following program that multiplies double complex numbers.

#include <pmmintrin.h>

__m128d _mm_complexmult_pd(__m128d a, __m128d b)
{
    __m128d A;
    A[0] = a[0]*b[0];
    A[1] = a[0]*b[1];

    __m128d B;
    B[0] = a[1]*b[1];
    B[1] = a[1]*b[0];

    return _mm_addsub_pd(A, B);
}

__m128d _mm_complexmult_pd_naive(__m128d a, __m128d b)
{
    __m128d A;
    A[0] = a[0]*b[0] - a[1] * b[1];
    A[1] = a[0]*b[1] + a[1] * b[0];
    return A;
}

The first one use explicit addsubpd, the second one not.
In LLVM 12, _mm_complexmult_pd_naive is faster.
In LLVM 13, _mm_complexmult_pd is faster because addsubpd is not generated anymore unless in simpler cases.

Godbolt: https://cpp.godbolt.org/z/sMdh7eG4s

[p0nce]

LDC issue => ldc-developers/ldc#4227

[p0nce]

Found that because some loops would execute in 164% the previous time when upgrading from LLVM 12 to 13; leading to a -10% performance loss over the whole program.

[llvmbot]

@llvm/issue-subscribers-backend-x86

[davidbolvansky]

cc @RKSimon

[rotateright]

We could try something like this in IR (VectorCombine):

define <2 x half> @src(<2 x half> %x) {
  %e1 = extractelement <2 x half> %x, i64 1
  %n1 = fneg half %e1
  %i0 = insertelement <2 x half> poison, half %e1, i64 0
  %i1 = insertelement <2 x half> %i0, half %n1, i64 1
  ret <2 x half> %i1
}

define <2 x half> @tgt(<2 x half> %x) {
  %n = fneg <2 x half> %x
  %r = shufflevector <2 x half> %x, <2 x half> %n, <2 x i32> <i32 1, i32 3>
  ret <2 x half> %r
}

https://alive2.llvm.org/ce/z/UMLVyH
...or the equivalent in the backend (I haven't looked at the example enough to know if either change solves the original problem).

[rotateright]

Proposal for a VectorCombine fold (won't get the requested x86 codegen, but should get us closer):
https://reviews.llvm.org/D135278

[ParkHanbum]

@rotateright is this issue will be resolved with the PR I submitted? #126408