[gcc r12-5420] rs6000: Add optimizations for _mm_sad_epu8

[gcc r12-5420] rs6000: Add optimizations for _mm_sad_epu8

[Paul Clarke]

https://gcc.gnu.org/g:fc6c6f64ecff376902e7e1ef295f2d8518407ab5

commit r12-5420-gfc6c6f64ecff376902e7e1ef295f2d8518407ab5
Author: Paul A. Clarke <pc@us.ibm.com>
Date:   Fri Oct 22 12:09:43 2021 -0500

    rs6000: Add optimizations for _mm_sad_epu8
    
    Power9 ISA added `vabsdub` instruction which is realized in the
    `vec_absd` instrinsic.
    
    Use `vec_absd` for `_mm_sad_epu8` compatibility intrinsic, when
    `_ARCH_PWR9`.
    
    Also, the realization of `vec_sum2s` on little-endian includes
    two rotates in order to position the input and output to match
    the semantics of `vec_sum2s`:
    - Rotate the second input vector left 12 bytes. In the current usage,
      that vector is `{0}`, so this shift is unnecessary, but is currently
      not eliminated under optimization.
    - Rotate the vector produced by the `vsum2sws` instruction left 4 bytes.
      The two words within each doubleword of this (rotated) result must then
      be explicitly swapped to match the semantics of `_mm_sad_epu8`,
      effectively reversing this rotate.  So, this rotate (and a susequent
      swap) are unnecessary, but not currently removed under optimization.
    
    Using `__builtin_altivec_vsum2sws` retains both rotates, so is not an
    option for removing the rotates.
    
    For little-endian, use the `vsum2sws` instruction directly, and
    eliminate the explicit rotate (swap).
    
    2021-11-19  Paul A. Clarke  <pc@us.ibm.com>
    
    gcc
            * config/rs6000/emmintrin.h (_mm_sad_epu8): Use vec_absd when
            _ARCH_PWR9, optimize vec_sum2s when LE.

Diff:
---
 gcc/config/rs6000/emmintrin.h | 24 +++++++++++++++++-------
 1 file changed, 17 insertions(+), 7 deletions(-)

diff --git a/gcc/config/rs6000/emmintrin.h b/gcc/config/rs6000/emmintrin.h
index 32ad72b4cc3..4125b12fa09 100644
--- a/gcc/config/rs6000/emmintrin.h
+++ b/gcc/config/rs6000/emmintrin.h
@@ -2189,27 +2189,37 @@ extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __arti
 _mm_sad_epu8 (__m128i __A, __m128i __B)
 {
   __v16qu a, b;
-  __v16qu vmin, vmax, vabsdiff;
+  __v16qu vabsdiff;
   __v4si vsum;
   const __v4su zero = { 0, 0, 0, 0 };
   __v4si result;
 
   a = (__v16qu) __A;
   b = (__v16qu) __B;
-  vmin = vec_min (a, b);
-  vmax = vec_max (a, b);
+#ifndef _ARCH_PWR9
+  __v16qu vmin = vec_min (a, b);
+  __v16qu vmax = vec_max (a, b);
   vabsdiff = vec_sub (vmax, vmin);
+#else
+  vabsdiff = vec_absd (a, b);
+#endif
   /* Sum four groups of bytes into integers.  */
   vsum = (__vector signed int) vec_sum4s (vabsdiff, zero);
+#ifdef __LITTLE_ENDIAN__
+  /* Sum across four integers with two integer results.  */
+  asm ("vsum2sws %0,%1,%2" : "=v" (result) : "v" (vsum), "v" (zero));
+  /* Note: vec_sum2s could be used here, but on little-endian, vector
+     shifts are added that are not needed for this use-case.
+     A vector shift to correctly position the 32-bit integer results
+     (currently at [0] and [2]) to [1] and [3] would then need to be
+     swapped back again since the desired results are two 64-bit
+     integers ([1]|[0] and [3]|[2]).  Thus, no shift is performed.  */
+#else
   /* Sum across four integers with two integer results.  */
   result = vec_sum2s (vsum, (__vector signed int) zero);
   /* Rotate the sums into the correct position.  */
-#ifdef __LITTLE_ENDIAN__
-  result = vec_sld (result, result, 4);
-#else
   result = vec_sld (result, result, 6);
 #endif
-  /* Rotate the sums into the correct position.  */
   return (__m128i) result;
 }