[PATCH 1/2] lib/hash: initialize __m128i data type in a portable way

[PATCH 1/2] lib/hash: initialize __m128i data type in a portable way

[Andre Muezerie]

The mechanism used to initialize an __m128i data type in rte_thash.h is
non-portable and MSVC does not like it. It clearly is not doing what
is desired:

..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
    truncation from 'unsigned __int64' to 'char'
..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
    truncation from 'unsigned __int64' to 'char'

A more portable approach is to use compiler intrinsics to perform the
initialization. This patch uses a single compiler intrinsic to
initialize the data type using a sequence of 16 bytes stored in
memory.

There should be no perf degradation due to this change.

Signed-off-by: Andre Muezerie <andremue@linux.microsoft.com>
---
 lib/hash/rte_thash.h | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/lib/hash/rte_thash.h b/lib/hash/rte_thash.h
index c0af5968df..3512639792 100644
--- a/lib/hash/rte_thash.h
+++ b/lib/hash/rte_thash.h
@@ -34,8 +34,9 @@ extern "C" {
 /* Byte swap mask used for converting IPv6 address
  * 4-byte chunks to CPU byte order
  */
-static const __m128i rte_thash_ipv6_bswap_mask = {
-		0x0405060700010203ULL, 0x0C0D0E0F08090A0BULL};
+static const uint8_t rte_thash_ipv6_bswap_mask[] = {
+		0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04,
+		0x0B, 0x0A, 0x09, 0x08, 0x0F, 0x0E, 0x0D, 0x0C};
 #endif
 
 /**
@@ -152,12 +153,14 @@ rte_thash_load_v6_addrs(const struct rte_ipv6_hdr *orig,
 			union rte_thash_tuple *targ)
 {
 #ifdef RTE_ARCH_X86
+	const __m128i ipv6_bswap_mask =
+			_mm_loadu_si128((const __m128i*)&rte_thash_ipv6_bswap_mask);
 	__m128i ipv6 = _mm_loadu_si128((const __m128i *)&orig->src_addr);
 	*(__m128i *)&targ->v6.src_addr =
-			_mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
+			_mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
 	ipv6 = _mm_loadu_si128((const __m128i *)&orig->dst_addr);
 	*(__m128i *)&targ->v6.dst_addr =
-			_mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
+			_mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
 #elif defined(__ARM_NEON)
 	uint8x16_t ipv6 = vld1q_u8(orig->src_addr.a);
 	vst1q_u8(targ->v6.src_addr.a, vrev32q_u8(ipv6));
-- 
2.34.1

[PATCH 2/2] app/test: add test_init_m128i using compiler intrinsic

[Andre Muezerie]

This test initializes an __m128i data type using the old
non-portable way used until now and the more portable way
using compiler intrinsics. The test ensures the resulting
values after initialization match.

Signed-off-by: Andre Muezerie <andremue@linux.microsoft.com>
---
 app/test/test_thash.c | 37 +++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/app/test/test_thash.c b/app/test/test_thash.c
index b9c6e9118e..c121b1f43f 100644
--- a/app/test/test_thash.c
+++ b/app/test/test_thash.c
@@ -1030,6 +1030,38 @@ test_keygen(void)
 	return TEST_SUCCESS;
 }
 
+#ifdef RTE_ARCH_X86
+#ifndef RTE_TOOLCHAIN_MSVC
+static int
+test_init_m128i(void)
+{
+	/* When initializing __m128i with two constant values like below
+	 * MSVC issues warning C4305:
+	 *     'initializing': truncation from 'unsigned __int64' to 'char'
+	 */
+	static const __m128i a = {
+			0x0405060700010203ULL, 0x0C0D0E0F08090A0BULL};
+
+	/* Using compiler intrinsics to initialize __m128i is therefore
+	 * preferred, like below
+	 */
+	static const uint8_t b_bytes[] = {
+			0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04,
+			0x0B, 0x0A, 0x09, 0x08, 0x0F, 0x0E, 0x0D, 0x0C};
+	const __m128i b =
+			_mm_loadu_si128((const __m128i *)&b_bytes);
+
+	if (memcmp(&a, &b, sizeof(a)) != 0) {
+		printf("Same value was expected when initializing data "
+				"type using compiler intrinsic\n");
+		return -1;
+	}
+
+	return TEST_SUCCESS;
+}
+#endif
+#endif
+
 static struct unit_test_suite thash_tests = {
 	.suite_name = "thash autotest",
 	.setup = NULL,
@@ -1052,6 +1084,11 @@ static struct unit_test_suite thash_tests = {
 	TEST_CASE(test_adjust_tuple),
 	TEST_CASE(test_adjust_tuple_mult_reta),
 	TEST_CASE(test_keygen),
+#ifdef RTE_ARCH_X86
+#ifndef RTE_TOOLCHAIN_MSVC
+	TEST_CASE(test_init_m128i),
+#endif
+#endif
 	TEST_CASES_END()
 	}
 };
-- 
2.34.1

[PATCH 1/2] lib/hash: initialize __m128i data type in a portable way

[Andre Muezerie]

The mechanism used to initialize an __m128i data type in rte_thash.h is
non-portable and MSVC does not like it. It clearly is not doing what
is desired:

..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
    truncation from 'unsigned __int64' to 'char'
..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
    truncation from 'unsigned __int64' to 'char'

A more portable approach is to use compiler intrinsics to perform the
initialization. This patch uses a single compiler intrinsic to
initialize the data type using a sequence of 16 bytes stored in
memory.

There should be no perf degradation due to this change.

Signed-off-by: Andre Muezerie <andremue@linux.microsoft.com>
---
 lib/hash/rte_thash.h | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/lib/hash/rte_thash.h b/lib/hash/rte_thash.h
index c0af5968df..3512639792 100644
--- a/lib/hash/rte_thash.h
+++ b/lib/hash/rte_thash.h
@@ -34,8 +34,9 @@ extern "C" {
 /* Byte swap mask used for converting IPv6 address
  * 4-byte chunks to CPU byte order
  */
-static const __m128i rte_thash_ipv6_bswap_mask = {
-		0x0405060700010203ULL, 0x0C0D0E0F08090A0BULL};
+static const uint8_t rte_thash_ipv6_bswap_mask[] = {
+		0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04,
+		0x0B, 0x0A, 0x09, 0x08, 0x0F, 0x0E, 0x0D, 0x0C};
 #endif
 
 /**
@@ -152,12 +153,14 @@ rte_thash_load_v6_addrs(const struct rte_ipv6_hdr *orig,
 			union rte_thash_tuple *targ)
 {
 #ifdef RTE_ARCH_X86
+	const __m128i ipv6_bswap_mask =
+			_mm_loadu_si128((const __m128i*)&rte_thash_ipv6_bswap_mask);
 	__m128i ipv6 = _mm_loadu_si128((const __m128i *)&orig->src_addr);
 	*(__m128i *)&targ->v6.src_addr =
-			_mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
+			_mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
 	ipv6 = _mm_loadu_si128((const __m128i *)&orig->dst_addr);
 	*(__m128i *)&targ->v6.dst_addr =
-			_mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
+			_mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
 #elif defined(__ARM_NEON)
 	uint8x16_t ipv6 = vld1q_u8(orig->src_addr.a);
 	vst1q_u8(targ->v6.src_addr.a, vrev32q_u8(ipv6));
-- 
2.34.1