Re: [PATCH][x86] Match movss and movsd "blend" instructions

[Allan Sandfeld Jensen]

Updated:

Match movss and movsd "blend" instructions

Adds the ability to match movss and movsd as blend patterns,
implemented in a new method to be able to match these before shuffles,
while keeping other blends after.

2018-08-11 Allan Sandfeld Jensen <allan.jen...@qt.io>

gcc/config/i386

    * i386.cc (expand_vec_perm_movs): New method matching movs
    patterns.
    * i386.cc (expand_vec_perm_1): Try the new method.

gcc/testsuite

    * gcc.target/i386/sse2-movs.c: New test.

diff --git a/gcc/config/i386/emmintrin.h b/gcc/config/i386/emmintrin.h
index b940a39d27b..6501638f619 100644
--- a/gcc/config/i386/emmintrin.h
+++ b/gcc/config/i386/emmintrin.h
@@ -113,7 +113,7 @@ _mm_setzero_pd (void)
 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_move_sd (__m128d __A, __m128d __B)
 {
-  return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
+  return __extension__ (__m128d) __builtin_shuffle((__v2df)__A, (__v2df)__B, (__v2di){2, 1});
 }
 
 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 7554fd1f659..485850096e9 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -46145,6 +46145,46 @@ expand_vselect_vconcat (rtx target, rtx op0, rtx op1,
   return ok;
 }
 
+/* A subroutine of ix86_expand_vec_perm_builtin_1.  Try to implement D
+   using movss or movsd.  */
+static bool
+expand_vec_perm_movs (struct expand_vec_perm_d *d)
+{
+  machine_mode vmode = d->vmode;
+  unsigned i, nelt = d->nelt;
+  rtx x;
+
+  if (d->one_operand_p)
+    return false;
+
+  if (TARGET_SSE2 && (vmode == V2DFmode || vmode == V4SFmode))
+    ;
+  else
+    return false;
+
+  /* Only the first element is changed. */
+  if (d->perm[0] != nelt && d->perm[0] != 0)
+    return false;
+  for (i = 1; i < nelt; ++i) {
+    {
+      if (d->perm[i] != i + nelt - d->perm[0])
+        return false;
+    }
+  }
+
+  if (d->testing_p)
+    return true;
+
+  if (d->perm[0] == nelt)
+    x = gen_rtx_VEC_MERGE (vmode, d->op1, d->op0, GEN_INT (1));
+  else
+    x = gen_rtx_VEC_MERGE (vmode, d->op0, d->op1, GEN_INT (1));
+
+  emit_insn (gen_rtx_SET (d->target, x));
+
+  return true;
+}
+
 /* A subroutine of ix86_expand_vec_perm_builtin_1.  Try to implement D
    in terms of blendp[sd] / pblendw / pblendvb / vpblendd.  */
 
@@ -46887,6 +46927,10 @@ expand_vec_perm_1 (struct expand_vec_perm_d *d)
 	}
     }
 
+  /* Try movss/movsd instructions.  */
+  if (expand_vec_perm_movs (d))
+    return true;
+
   /* Finally, try the fully general two operand permute.  */
   if (expand_vselect_vconcat (d->target, d->op0, d->op1, d->perm, nelt,
 			      d->testing_p))
diff --git a/gcc/config/i386/xmmintrin.h b/gcc/config/i386/xmmintrin.h
index f64f3f74a0b..f770570295c 100644
--- a/gcc/config/i386/xmmintrin.h
+++ b/gcc/config/i386/xmmintrin.h
@@ -1011,7 +1011,10 @@ _mm_storer_ps (float *__P, __m128 __A)
 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 _mm_move_ss (__m128 __A, __m128 __B)
 {
-  return (__m128) __builtin_ia32_movss ((__v4sf)__A, (__v4sf)__B);
+  return (__m128) __builtin_shuffle ((__v4sf)__A, (__v4sf)__B,
+                                     __extension__
+                                     (__attribute__((__vector_size__ (16))) int)
+                                     {4,1,2,3});
 }
 
 /* Extracts one of the four words of A.  The selector N must be immediate.  */