Compiler no longer producing single horizontal add instruction after ab2c499d

[dyung]

Bugzilla Link	50392
Version	trunk
OS	All
CC	@alexey-bataev,@anton-afanasyev,@RKSimon

Extended Description

One of our internal tests compiles the following code and verifies that it generates a horizontal add instruction in the resulting assembly when optimizations are enabled and targeting btver2.

__attribute__((noinline))
__m256d add_pd_002(__m256d a, __m256d b) {
  __m256d r = (__m256d){ a[0] + a[1], a[2] + a[3], b[0] + b[1], b[2] + b[3] };
  return __builtin_shufflevector(r, a, 0, -1, 2, 3);
}

Prior to commit ab2c499, when compiled with "-g0 -O3 -march=btver2", the compiler would produce the following code:

        vinsertf128     $1, %xmm1, %ymm0, %ymm0
        vhaddpd %ymm1, %ymm0, %ymm0

But following the mentioned change, the compiler is now producing the following code instead:

        vextractf128    $1, %ymm1, %xmm2
        vhaddpd %xmm0, %xmm0, %xmm0
        vhaddpd %xmm2, %xmm1, %xmm1
        vperm2f128      $2, %ymm0, %ymm1, %ymm0 # ymm0 = ymm0[0,1],ymm1[0,1]

[dyung]

Link to review of the change that caused the regression: https://reviews.llvm.org/D98714

[anton-afanasyev]

To add more details: this should be fixed after Alexey's patch (https://reviews.llvm.org/D57059) for the non-power-of-two vector sizes landing.