arm.c (arm_evpc_neon_vext): New function.

2012-09-04  Christophe Lyon  <christophe.lyon@linaro.org>

	gcc/
	* config/arm/arm.c (arm_evpc_neon_vext): New
	function.
	(arm_expand_vec_perm_const_1): Add call to
	arm_evpc_neon_vext.

	gcc/testsuite/
	* gcc.target/arm/neon-vext.c: New test.
	* gcc.target/arm/neon-vext-execute.c: Ditto.

From-SVN: r190911

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 059967a8c3b7a83f1f2377b8237cbf538216317e..cda27fd74dfb06fbd66a38bdd753075df5f8ada6 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2012-09-04  Christophe Lyon  <christophe.lyon@linaro.org>
+
+       * config/arm/arm.c (arm_evpc_neon_vext): New
+       function.
+       (arm_expand_vec_perm_const_1): Add call to
+       arm_evpc_neon_vext.
+
 2012-09-04  Oleg Endo  <olegendo@gcc.gnu.org>
 
        PR target/51244

diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 9ce3c0f068add4a67d6213613bf9a1722a348842..36937d2b90d597dc878a8f4b253981c22c8fca58 100644 (file)
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -25937,6 +25937,72 @@ arm_evpc_neon_vtrn (struct expand_vec_perm_d *d)
   return true;
 }
 
+/* Recognize patterns for the VEXT insns.  */
+
+static bool
+arm_evpc_neon_vext (struct expand_vec_perm_d *d)
+{
+  unsigned int i, nelt = d->nelt;
+  rtx (*gen) (rtx, rtx, rtx, rtx);
+  rtx offset;
+
+  unsigned int location;
+
+  unsigned int next  = d->perm[0] + 1;
+
+  /* TODO: Handle GCC's numbering of elements for big-endian.  */
+  if (BYTES_BIG_ENDIAN)
+    return false;
+
+  /* Check if the extracted indexes are increasing by one.  */
+  for (i = 1; i < nelt; next++, i++)
+    {
+      /* If we hit the most significant element of the 2nd vector in
+        the previous iteration, no need to test further.  */
+      if (next == 2 * nelt)
+       return false;
+
+      /* If we are operating on only one vector: it could be a
+        rotation.  If there are only two elements of size < 64, let
+        arm_evpc_neon_vrev catch it.  */
+      if (d->one_vector_p && (next == nelt))
+       {
+         if ((nelt == 2) && (d->vmode != V2DImode))
+           return false;
+         else
+           next = 0;
+       }
+
+      if (d->perm[i] != next)
+       return false;
+    }
+
+  location = d->perm[0];
+
+  switch (d->vmode)
+    {
+    case V16QImode: gen = gen_neon_vextv16qi; break;
+    case V8QImode: gen = gen_neon_vextv8qi; break;
+    case V4HImode: gen = gen_neon_vextv4hi; break;
+    case V8HImode: gen = gen_neon_vextv8hi; break;
+    case V2SImode: gen = gen_neon_vextv2si; break;
+    case V4SImode: gen = gen_neon_vextv4si; break;
+    case V2SFmode: gen = gen_neon_vextv2sf; break;
+    case V4SFmode: gen = gen_neon_vextv4sf; break;
+    case V2DImode: gen = gen_neon_vextv2di; break;
+    default:
+      return false;
+    }
+
+  /* Success! */
+  if (d->testing_p)
+    return true;
+
+  offset = GEN_INT (location);
+  emit_insn (gen (d->target, d->op0, d->op1, offset));
+  return true;
+}
+
 /* The NEON VTBL instruction is a fully variable permuation that's even
    stronger than what we expose via VEC_PERM_EXPR.  What it doesn't do
    is mask the index operand as VEC_PERM_EXPR requires.  Therefore we
@@ -25976,6 +26042,12 @@ arm_evpc_neon_vtbl (struct expand_vec_perm_d *d)
 static bool
 arm_expand_vec_perm_const_1 (struct expand_vec_perm_d *d)
 {
+  /* Check if the input mask matches vext before reordering the
+     operands.  */
+  if (TARGET_NEON)
+    if (arm_evpc_neon_vext (d))
+      return true;
+
   /* The pattern matching functions above are written to look for a small
      number to begin the sequence (0, 1, N/2).  If we begin with an index
      from the second operand, we can swap the operands.  */

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 163eeab5897b75b4f9bb01e58914b8f5548b3cab..4a0532188cdd4e520af5beb5ffd7527af92a13c8 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2012-09-04  Christophe Lyon  <christophe.lyon@linaro.org>
+
+       * gcc.target/arm/neon-vext.c: New test.
+       * gcc.target/arm/neon-vext-execute.c: Ditto.
+
 2012-09-04  Janus Weil  <janus@gcc.gnu.org>
 
        PR fortran/54243

diff --git a/gcc/testsuite/gcc.target/arm/neon-vext-execute.c b/gcc/testsuite/gcc.target/arm/neon-vext-execute.c
new file mode 100644 (file)
index 0000000..3d6c28c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/neon-vext-execute.c
@@ -0,0 +1,340 @@
+/* { dg-do run } */
+/* { dg-require-effective-target arm_neon_ok } */
+/* { dg-require-effective-target arm_little_endian } */
+/* { dg-options "-O2" } */
+/* { dg-add-options arm_neon } */
+
+#include <arm_neon.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+uint8x8_t
+tst_vext_u8 (uint8x8_t __a, uint8x8_t __b)
+{
+  uint8x8_t __mask1 = {2, 3, 4, 5, 6, 7, 8, 9};
+
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint8x8_t
+tst_vext_u8_rotate (uint8x8_t __a)
+{
+  uint8x8_t __mask1 = {2, 3, 4, 5, 6, 7, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint16x4_t
+tst_vext_u16 (uint16x4_t __a, uint16x4_t __b)
+{
+  uint16x4_t __mask1 = {2, 3, 4, 5};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint16x4_t
+tst_vext_u16_rotate (uint16x4_t __a)
+{
+  uint16x4_t __mask1 = {2, 3, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint32x2_t
+tst_vext_u32 (uint32x2_t __a, uint32x2_t __b)
+{
+  uint32x2_t __mask1 = {1, 2};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+/* This one is mapped into vrev64.32.  */
+uint32x2_t
+tst_vext_u32_rotate (uint32x2_t __a)
+{
+  uint32x2_t __mask1 = {1, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint8x16_t
+tst_vextq_u8 (uint8x16_t __a, uint8x16_t __b)
+{
+  uint8x16_t __mask1 = {4, 5, 6, 7, 8, 9, 10, 11,
+                       12, 13, 14, 15, 16, 17, 18, 19};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint8x16_t
+tst_vextq_u8_rotate (uint8x16_t __a)
+{
+  uint8x16_t __mask1 = {4, 5, 6, 7, 8, 9, 10, 11,
+                       12, 13, 14, 15, 0, 1, 2, 3};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint16x8_t
+tst_vextq_u16 (uint16x8_t __a, uint16x8_t __b)
+{
+  uint16x8_t __mask1 = {2, 3, 4, 5, 6, 7, 8, 9};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint16x8_t
+tst_vextq_u16_rotate (uint16x8_t __a)
+{
+  uint16x8_t __mask1 = {2, 3, 4, 5, 6, 7, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint32x4_t
+tst_vextq_u32 (uint32x4_t __a, uint32x4_t __b)
+{
+  uint32x4_t __mask1 = {1, 2, 3, 4};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint32x4_t
+tst_vextq_u32_rotate (uint32x4_t __a)
+{
+  uint32x4_t __mask1 = {1, 2, 3, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint64x2_t
+tst_vextq_u64 (uint64x2_t __a, uint64x2_t __b)
+{
+  uint64x2_t __mask1 = {1, 2};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint64x2_t
+tst_vextq_u64_rotate (uint64x2_t __a)
+{
+  uint64x2_t __mask1 = {1, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+int main (void)
+{
+  uint8_t arr_u8x8[] = {0, 1, 2, 3, 4, 5, 6, 7};
+  uint8_t arr2_u8x8[] = {8, 9, 10, 11, 12, 13, 14, 15};
+  uint16_t arr_u16x4[] = {0, 1, 2, 3};
+  uint16_t arr2_u16x4[] = {4, 5, 6, 7};
+  uint32_t arr_u32x2[] = {0, 1};
+  uint32_t arr2_u32x2[] = {2, 3};
+  uint8_t arr_u8x16[] = {0, 1, 2, 3, 4, 5, 6, 7,
+                        8, 9, 10, 11, 12, 13, 14, 15};
+  uint8_t arr2_u8x16[] = {16, 17, 18, 19, 20, 21, 22, 23,
+                         24, 25, 26, 27, 28, 29, 30, 31};
+  uint16_t arr_u16x8[] = {0, 1, 2, 3, 4, 5, 6, 7};
+  uint16_t arr2_u16x8[] = {8, 9, 10, 11, 12, 13, 14, 15};
+  uint32_t arr_u32x4[] = {0, 1, 2, 3};
+  uint32_t arr2_u32x4[] = {4, 5, 6, 7};
+  uint64_t arr_u64x2[] = {0, 1};
+  uint64_t arr2_u64x2[] = {2, 3};
+
+  uint8_t expected_u8x8[] = {2, 3, 4, 5, 6, 7, 8, 9};
+  uint8_t expected_rot_u8x8[] = {2, 3, 4, 5, 6, 7, 0, 1};
+  uint16_t expected_u16x4[] = {2, 3, 4, 5};
+  uint16_t expected_rot_u16x4[] = {2, 3, 0, 1};
+  uint32_t expected_u32x2[] = {1, 2};
+  uint32_t expected_rot_u32x2[] = {1, 0};
+  uint8_t expected_u8x16[] = {4, 5, 6, 7, 8, 9, 10, 11,
+                             12, 13, 14, 15, 16, 17, 18, 19};
+  uint8_t expected_rot_u8x16[] = {4, 5, 6, 7, 8, 9, 10, 11,
+                                 12, 13, 14, 15, 0, 1, 2, 3,};
+  uint16_t expected_u16x8[] = {2, 3, 4, 5, 6, 7, 8, 9};
+  uint16_t expected_rot_u16x8[] = {2, 3, 4, 5, 6, 7, 0, 1};
+  uint32_t expected_u32x4[] = {1, 2, 3, 4};
+  uint32_t expected_rot_u32x4[] = {1, 2, 3, 0};
+  uint64_t expected_u64x2[] = {1, 2};
+  uint64_t expected_rot_u64x2[] = {1, 0};
+
+  uint8x8_t vec_u8x8 = vld1_u8 (arr_u8x8);
+  uint8x8_t vec2_u8x8 = vld1_u8 (arr2_u8x8);
+  uint16x4_t vec_u16x4 = vld1_u16 (arr_u16x4);
+  uint16x4_t vec2_u16x4 = vld1_u16 (arr2_u16x4);
+  uint32x2_t vec_u32x2 = vld1_u32 (arr_u32x2);
+  uint32x2_t vec2_u32x2 = vld1_u32 (arr2_u32x2);
+  uint8x16_t vec_u8x16 = vld1q_u8 (arr_u8x16);
+  uint8x16_t vec2_u8x16 = vld1q_u8 (arr2_u8x16);
+  uint16x8_t vec_u16x8 = vld1q_u16 (arr_u16x8);
+  uint16x8_t vec2_u16x8 = vld1q_u16 (arr2_u16x8);
+  uint32x4_t vec_u32x4 = vld1q_u32 (arr_u32x4);
+  uint32x4_t vec2_u32x4 = vld1q_u32 (arr2_u32x4);
+  uint64x2_t vec_u64x2 = vld1q_u64 (arr_u64x2);
+  uint64x2_t vec2_u64x2 = vld1q_u64 (arr2_u64x2);
+
+  uint8x8_t result_u8x8;
+  uint16x4_t result_u16x4;
+  uint32x2_t result_u32x2;
+  uint8x16_t result_u8x16;
+  uint16x8_t result_u16x8;
+  uint32x4_t result_u32x4;
+  uint64x2_t result_u64x2;
+
+  union {uint8x8_t v; uint8_t buf[8];} mem_u8x8;
+  union {uint16x4_t v; uint16_t buf[4];} mem_u16x4;
+  union {uint32x2_t v; uint32_t buf[2];} mem_u32x2;
+  union {uint8x16_t v; uint8_t buf[16];} mem_u8x16;
+  union {uint16x8_t v; uint16_t buf[8];} mem_u16x8;
+  union {uint32x4_t v; uint32_t buf[4];} mem_u32x4;
+  union {uint64x2_t v; uint64_t buf[2];} mem_u64x2;
+
+  int i;
+
+  result_u8x8 = tst_vext_u8 (vec_u8x8, vec2_u8x8);
+  vst1_u8 (mem_u8x8.buf, result_u8x8);
+
+  for (i=0; i<8; i++)
+      if (mem_u8x8.buf[i] != expected_u8x8[i])
+       {
+         printf ("tst_vext_u8[%d]=%d expected %d\n",
+                 i, mem_u8x8.buf[i], expected_u8x8[i]);
+         abort ();
+       }
+
+  result_u8x8 = tst_vext_u8_rotate (vec_u8x8);
+  vst1_u8 (mem_u8x8.buf, result_u8x8);
+
+  for (i=0; i<8; i++)
+      if (mem_u8x8.buf[i] != expected_rot_u8x8[i])
+       {
+         printf ("tst_vext_u8_rotate[%d]=%d expected %d\n",
+                 i, mem_u8x8.buf[i], expected_rot_u8x8[i]);
+         abort ();
+       }
+
+
+  result_u16x4 = tst_vext_u16 (vec_u16x4, vec2_u16x4);
+  vst1_u16 (mem_u16x4.buf, result_u16x4);
+
+  for (i=0; i<4; i++)
+      if (mem_u16x4.buf[i] != expected_u16x4[i])
+       {
+         printf ("tst_vext_u16[%d]=%d expected %d\n",
+                 i, mem_u16x4.buf[i], expected_u16x4[i]);
+         abort ();
+       }
+
+  result_u16x4 = tst_vext_u16_rotate (vec_u16x4);
+  vst1_u16 (mem_u16x4.buf, result_u16x4);
+
+  for (i=0; i<4; i++)
+      if (mem_u16x4.buf[i] != expected_rot_u16x4[i])
+       {
+         printf ("tst_vext_u16_rotate[%d]=%d expected %d\n",
+                 i, mem_u16x4.buf[i], expected_rot_u16x4[i]);
+         abort ();
+       }
+
+
+  result_u32x2 = tst_vext_u32 (vec_u32x2, vec2_u32x2);
+  vst1_u32 (mem_u32x2.buf, result_u32x2);
+
+  for (i=0; i<2; i++)
+      if (mem_u32x2.buf[i] != expected_u32x2[i])
+       {
+         printf ("tst_vext_u32[%d]=%d expected %d\n",
+                 i, mem_u32x2.buf[i], expected_u32x2[i]);
+         abort ();
+       }
+
+  result_u32x2 = tst_vext_u32_rotate (vec_u32x2);
+  vst1_u32 (mem_u32x2.buf, result_u32x2);
+
+  for (i=0; i<2; i++)
+      if (mem_u32x2.buf[i] != expected_rot_u32x2[i])
+       {
+         printf ("tst_vext_u32_rotate[%d]=%d expected %d\n",
+                 i, mem_u32x2.buf[i], expected_rot_u32x2[i]);
+         abort ();
+       }
+
+
+  result_u8x16 = tst_vextq_u8 (vec_u8x16, vec2_u8x16);
+  vst1q_u8 (mem_u8x16.buf, result_u8x16);
+
+  for (i=0; i<16; i++)
+      if (mem_u8x16.buf[i] != expected_u8x16[i])
+       {
+         printf ("tst_vextq_u8[%d]=%d expected %d\n",
+                 i, mem_u8x16.buf[i], expected_u8x16[i]);
+         abort ();
+       }
+
+  result_u8x16 = tst_vextq_u8_rotate (vec_u8x16);
+  vst1q_u8 (mem_u8x16.buf, result_u8x16);
+
+  for (i=0; i<16; i++)
+      if (mem_u8x16.buf[i] != expected_rot_u8x16[i])
+       {
+         printf ("tst_vextq_u8_rotate[%d]=%d expected %d\n",
+                 i, mem_u8x16.buf[i], expected_rot_u8x16[i]);
+         abort ();
+       }
+
+  result_u16x8 = tst_vextq_u16 (vec_u16x8, vec2_u16x8);
+  vst1q_u16 (mem_u16x8.buf, result_u16x8);
+
+  for (i=0; i<8; i++)
+      if (mem_u16x8.buf[i] != expected_u16x8[i])
+       {
+         printf ("tst_vextq_u16[%d]=%d expected %d\n",
+                 i, mem_u16x8.buf[i], expected_u16x8[i]);
+         abort ();
+       }
+
+  result_u16x8 = tst_vextq_u16_rotate (vec_u16x8);
+  vst1q_u16 (mem_u16x8.buf, result_u16x8);
+
+  for (i=0; i<8; i++)
+      if (mem_u16x8.buf[i] != expected_rot_u16x8[i])
+       {
+         printf ("tst_vextq_u16_rotate[%d]=%d expected %d\n",
+                 i, mem_u16x8.buf[i], expected_rot_u16x8[i]);
+         abort ();
+       }
+
+  result_u32x4 = tst_vextq_u32 (vec_u32x4, vec2_u32x4);
+  vst1q_u32 (mem_u32x4.buf, result_u32x4);
+
+  for (i=0; i<4; i++)
+      if (mem_u32x4.buf[i] != expected_u32x4[i])
+       {
+         printf ("tst_vextq_u32[%d]=%d expected %d\n",
+                 i, mem_u32x4.buf[i], expected_u32x4[i]);
+         abort ();
+       }
+
+  result_u32x4 = tst_vextq_u32_rotate (vec_u32x4);
+  vst1q_u32 (mem_u32x4.buf, result_u32x4);
+
+  for (i=0; i<4; i++)
+      if (mem_u32x4.buf[i] != expected_rot_u32x4[i])
+       {
+         printf ("tst_vextq_u32_rotate[%d]=%d expected %d\n",
+                 i, mem_u32x4.buf[i], expected_rot_u32x4[i]);
+         abort ();
+       }
+
+  result_u64x2 = tst_vextq_u64 (vec_u64x2, vec2_u64x2);
+  vst1q_u64 (mem_u64x2.buf, result_u64x2);
+
+  for (i=0; i<2; i++)
+      if (mem_u64x2.buf[i] != expected_u64x2[i])
+       {
+         printf ("tst_vextq_u64[%d]=%lld expected %lld\n",
+                 i, mem_u64x2.buf[i], expected_u64x2[i]);
+         abort ();
+       }
+
+  result_u64x2 = tst_vextq_u64_rotate (vec_u64x2);
+  vst1q_u64 (mem_u64x2.buf, result_u64x2);
+
+  for (i=0; i<2; i++)
+      if (mem_u64x2.buf[i] != expected_rot_u64x2[i])
+       {
+         printf ("tst_vextq_u64_rotate[%d]=%lld expected %lld\n",
+                 i, mem_u64x2.buf[i], expected_rot_u64x2[i]);
+         abort ();
+       }
+
+  return 0;
+}

diff --git a/gcc/testsuite/gcc.target/arm/neon-vext.c b/gcc/testsuite/gcc.target/arm/neon-vext.c
new file mode 100644 (file)
index 0000000..4a012a9
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/neon-vext.c
@@ -0,0 +1,115 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_neon_ok } */
+/* { dg-require-effective-target arm_little_endian } */
+/* { dg-options "-O2" } */
+/* { dg-add-options arm_neon } */
+
+#include <arm_neon.h>
+
+uint8x8_t
+tst_vext_u8 (uint8x8_t __a, uint8x8_t __b)
+{
+  uint8x8_t __mask1 = {2, 3, 4, 5, 6, 7, 8, 9};
+
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint8x8_t
+tst_vext_u8_rotate (uint8x8_t __a)
+{
+  uint8x8_t __mask1 = {2, 3, 4, 5, 6, 7, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint16x4_t
+tst_vext_u16 (uint16x4_t __a, uint16x4_t __b)
+{
+  uint16x4_t __mask1 = {2, 3, 4, 5};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint16x4_t
+tst_vext_u16_rotate (uint16x4_t __a)
+{
+  uint16x4_t __mask1 = {2, 3, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint32x2_t
+tst_vext_u32 (uint32x2_t __a, uint32x2_t __b)
+{
+  uint32x2_t __mask1 = {1, 2};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+/* This one is mapped into vrev64.32.  */
+uint32x2_t
+tst_vext_u32_rotate (uint32x2_t __a)
+{
+  uint32x2_t __mask1 = {1, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint8x16_t
+tst_vextq_u8 (uint8x16_t __a, uint8x16_t __b)
+{
+  uint8x16_t __mask1 = {4, 5, 6, 7, 8, 9, 10, 11,
+                       12, 13, 14, 15, 16, 17, 18, 19};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint8x16_t
+tst_vextq_u8_rotate (uint8x16_t __a)
+{
+  uint8x16_t __mask1 = {4, 5, 6, 7, 8, 9, 10, 11,
+                       12, 13, 14, 15, 0, 1, 2, 3};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint16x8_t
+tst_vextq_u16 (uint16x8_t __a, uint16x8_t __b)
+{
+  uint16x8_t __mask1 = {2, 3, 4, 5, 6, 7, 8, 9};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint16x8_t
+tst_vextq_u16_rotate (uint16x8_t __a)
+{
+  uint16x8_t __mask1 = {2, 3, 4, 5, 6, 7, 0, 1};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint32x4_t
+tst_vextq_u32 (uint32x4_t __a, uint32x4_t __b)
+{
+  uint32x4_t __mask1 = {1, 2, 3, 4};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint32x4_t
+tst_vextq_u32_rotate (uint32x4_t __a)
+{
+  uint32x4_t __mask1 = {1, 2, 3, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+uint64x2_t
+tst_vextq_u64 (uint64x2_t __a, uint64x2_t __b)
+{
+  uint64x2_t __mask1 = {1, 2};
+  return __builtin_shuffle ( __a, __b, __mask1) ;
+}
+
+uint64x2_t
+tst_vextq_u64_rotate (uint64x2_t __a)
+{
+  uint64x2_t __mask1 = {1, 0};
+  return __builtin_shuffle ( __a, __mask1) ;
+}
+
+/* { dg-final {scan-assembler-times "vext\.8\\t" 4} }  */
+/* { dg-final {scan-assembler-times "vext\.16\\t" 4} }  */
+/* { dg-final {scan-assembler-times "vext\.32\\t" 3} }  */
+/* { dg-final {scan-assembler-times "vrev64\.32\\t" 1} }  */
+/* { dg-final {scan-assembler-times "vext\.64\\t" 2} }  */