Hi!

ICC apparently has these two intrinsics (why it doesn't have a maskz_ one
is unclear to me) which are like _mm512_{,mask_}mullo_epi64, except they are
available in AVX512F rather than just AVX512DQ and if AVX512DQ is not
enabled they expand to 3 vpmuludq instructions + 3 shifts + 2 adds; for
AVX512DQ they are the same as mullo without x.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2018-04-26  Jakub Jelinek  <ja...@redhat.com>

        PR target/85530
        * config/i386/avx512fintrin.h (_mm512_mullox_epi64,
        _mm512_mask_mullox_epi64): New intrinsics.

        * gcc.target/i386/avx512f-vpmullq-1.c: New test.
        * gcc.target/i386/avx512f-vpmullq-2.c: New test.
        * gcc.target/i386/avx512dq-vpmullq-3.c: New test.
        * gcc.target/i386/avx512dq-vpmullq-4.c: New test.

--- gcc/config/i386/avx512fintrin.h.jj  2018-02-12 19:17:40.087215130 +0100
+++ gcc/config/i386/avx512fintrin.h     2018-04-26 11:51:09.176953712 +0200
@@ -567,6 +567,20 @@ _mm512_mask_mullo_epi32 (__m512i __W, __
 
 extern __inline __m512i
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm512_mullox_epi64 (__m512i __A, __m512i __B)
+{
+  return (__m512i) ((__v8du) __A * (__v8du) __B);
+}
+
+extern __inline __m512i
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
+_mm512_mask_mullox_epi64 (__m512i __W, __mmask8 __M, __m512i __A, __m512i __B)
+{
+  return _mm512_mask_mov_epi64 (__W, __M, _mm512_mullox_epi64 (__A, __B));
+}
+
+extern __inline __m512i
+__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm512_sllv_epi32 (__m512i __X, __m512i __Y)
 {
   return (__m512i) __builtin_ia32_psllv16si_mask ((__v16si) __X,
--- gcc/testsuite/gcc.target/i386/avx512f-vpmullq-1.c.jj        2018-04-26 
12:01:51.049333280 +0200
+++ gcc/testsuite/gcc.target/i386/avx512f-vpmullq-1.c   2018-04-26 
12:29:04.957253010 +0200
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx512f -mno-avx512dq -O2" } */
+/* { dg-final { scan-assembler-times "vpmullq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  0 } } */
+/* { dg-final { scan-assembler-times "vpmullq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\{%k\[1-7\]\}(?:\n|\[
 \\t\]+#)" 0 } } */
+/* { dg-final { scan-assembler-times "vpmuludq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  6 } } */
+/* { dg-final { scan-assembler-times "vpsrlq\[ 
\\t\]+\[^\{\n\]*\\\$32\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  4 } } */
+/* { dg-final { scan-assembler-times "vpsllq\[ 
\\t\]+\[^\{\n\]*\\\$32\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  2 } } */
+/* { dg-final { scan-assembler-times "vpaddq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  3 } } */
+/* { dg-final { scan-assembler-times "vpaddq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\{%k\[1-7\]\}(?:\n|\[
 \\t\]+#)" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m512i _x1, _y1, _z1;
+
+void extern
+avx512f_test (void)
+{
+  _x1 = _mm512_mullox_epi64 (_y1, _z1);
+  _x1 = _mm512_mask_mullox_epi64 (_x1, 3, _y1, _z1);
+}
--- gcc/testsuite/gcc.target/i386/avx512f-vpmullq-2.c.jj        2018-04-26 
12:01:54.545335345 +0200
+++ gcc/testsuite/gcc.target/i386/avx512f-vpmullq-2.c   2018-04-26 
12:01:09.491308704 +0200
@@ -0,0 +1,45 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -mavx512dq" } */
+/* { dg-require-effective-target avx512dq } */
+
+#define AVX512DQ
+#include "avx512f-helper.h"
+
+#define SIZE (AVX512F_LEN / 64)
+#include "avx512f-mask-type.h"
+
+void
+CALC (long long *src1, long long *src2, long long *dst)
+{
+  int i;
+
+  for (i = 0; i < SIZE; i++)
+    dst[i] = src1[i] * src2[i];
+}
+
+void
+TEST (void)
+{
+  UNION_TYPE (AVX512F_LEN, i_q) src1, src2, dst1, dst2;
+  long long dst_ref[SIZE];
+  int i;
+  MASK_TYPE mask = MASK_VALUE;
+
+  for (i = 0; i < SIZE; i++)
+    {
+      src1.a[i] = i + 50;
+      src2.a[i] = i + 100;
+      dst2.a[i] = DEFAULT_VALUE;
+    }
+
+  dst1.x = INTRINSIC (_mullox_epi64) (src1.x, src2.x);
+  dst2.x = INTRINSIC (_mask_mullox_epi64) (dst2.x, mask, src1.x, src2.x);
+  CALC (src1.a, src2.a, dst_ref);
+
+  if (UNION_CHECK (AVX512F_LEN, i_q) (dst1, dst_ref))
+    abort ();
+
+  MASK_MERGE (i_q) (dst_ref, mask, SIZE);
+  if (UNION_CHECK (AVX512F_LEN, i_q) (dst2, dst_ref))
+    abort ();
+}
--- gcc/testsuite/gcc.target/i386/avx512dq-vpmullq-3.c.jj       2018-04-26 
11:58:43.604222431 +0200
+++ gcc/testsuite/gcc.target/i386/avx512dq-vpmullq-3.c  2018-04-26 
11:59:49.585261449 +0200
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-mavx512dq -O2" } */
+/* { dg-final { scan-assembler-times "vpmullq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+(?:\n|\[ 
\\t\]+#)"  1 } } */
+/* { dg-final { scan-assembler-times "vpmullq\[ 
\\t\]+\[^\{\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\[^\n\]*%zmm\[0-9\]+\{%k\[1-7\]\}(?:\n|\[
 \\t\]+#)" 1 } } */
+
+#include <immintrin.h>
+
+volatile __m512i _x1, _y1, _z1;
+
+void extern
+avx512dq_test (void)
+{
+  _x1 = _mm512_mullox_epi64 (_y1, _z1);
+  _x1 = _mm512_mask_mullox_epi64 (_x1, 3, _y1, _z1);
+}
--- gcc/testsuite/gcc.target/i386/avx512dq-vpmullq-4.c.jj       2018-04-26 
11:59:59.526267331 +0200
+++ gcc/testsuite/gcc.target/i386/avx512dq-vpmullq-4.c  2018-04-26 
12:01:09.491308704 +0200
@@ -0,0 +1,45 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -mavx512dq" } */
+/* { dg-require-effective-target avx512dq } */
+
+#define AVX512DQ
+#include "avx512f-helper.h"
+
+#define SIZE (AVX512F_LEN / 64)
+#include "avx512f-mask-type.h"
+
+void
+CALC (long long *src1, long long *src2, long long *dst)
+{
+  int i;
+
+  for (i = 0; i < SIZE; i++)
+    dst[i] = src1[i] * src2[i];
+}
+
+void
+TEST (void)
+{
+  UNION_TYPE (AVX512F_LEN, i_q) src1, src2, dst1, dst2;
+  long long dst_ref[SIZE];
+  int i;
+  MASK_TYPE mask = MASK_VALUE;
+
+  for (i = 0; i < SIZE; i++)
+    {
+      src1.a[i] = i + 50;
+      src2.a[i] = i + 100;
+      dst2.a[i] = DEFAULT_VALUE;
+    }
+
+  dst1.x = INTRINSIC (_mullox_epi64) (src1.x, src2.x);
+  dst2.x = INTRINSIC (_mask_mullox_epi64) (dst2.x, mask, src1.x, src2.x);
+  CALC (src1.a, src2.a, dst_ref);
+
+  if (UNION_CHECK (AVX512F_LEN, i_q) (dst1, dst_ref))
+    abort ();
+
+  MASK_MERGE (i_q) (dst_ref, mask, SIZE);
+  if (UNION_CHECK (AVX512F_LEN, i_q) (dst2, dst_ref))
+    abort ();
+}

        Jakub

Reply via email to