From 3d689c53de7fbcb8892c8d14ea5f1b8d0455af8e Mon Sep 17 00:00:00 2001
From: Andrea Mazzoleni <amadvance@gmail.com>
Date: Sun, 29 Sep 2013 15:24:15 +0200
Subject: [PATCH] Doesn't use extended SSE registers in the Bulldozer AMD
 family.

---
 cpu.h   | 43 ++++++++++++++++++++++++++++++++++++++++++-
 raid.c  |  9 +++++++--
 speed.c | 18 ++++++++++++------
 3 files changed, 61 insertions(+), 9 deletions(-)

diff --git a/cpu.h b/cpu.h
index 95ec94bf..b7ec1cc9 100644
--- a/cpu.h
+++ b/cpu.h
@@ -119,7 +119,7 @@ static inline int cpu_has_slowmult(void)
 		 * Memory is little-endian 64-bit
 		 * Speed test with 8 disk and 262144 buffer size...
 		 * memset0 1849 [MB/s]
-		 * HASH Murmur3 378 [MB/s]
+		 * HASH Murmur3 378 [MB/s] (extremely slower than Spooky2)
 		 * HASH Spooky2 3413 [MB/s]
 		 * RAID5 int32x2 707 [MB/s]
 		 * RAID5 mmxx2 1264 [MB/s]
@@ -138,6 +138,47 @@ static inline int cpu_has_slowmult(void)
 	return 0;
 }
 
+/**
+ * Check if the processor has a slow extended set of SSE registers.
+ * If yes, it's better to unroll without using the second part of registers.
+ */
+static inline int cpu_has_slowextendedreg(void)
+{
+	char vendor[CPU_VENDOR_MAX];
+	unsigned family;
+	unsigned model;
+
+	cpu_info(vendor, &family, &model);
+
+	if (strcmp(vendor, "AuthenticAMD") == 0) {
+		/* AMD Bulldozer (from user)
+		 * CPU AuthenticAMD, family 21, model 19, flags mmx sse2
+		 * Memory is little-endian 64-bit
+		 * Speed test with 8 disk and 262144 buffer size, for a total of 2560 KiB...
+		 * memset0 5721 [MB/s]
+		 * CRC table 1080 [MB/s]
+		 * CRC intel-crc32 2845 [MB/s]
+		 * HASH Murmur3 2970 [MB/s]
+		 * HASH Spooky2 7503 [MB/s]
+		 * RAID5 int32x2 4595 [MB/s]
+		 * RAID5 mmxx2 5856 [MB/s]
+		 * RAID5 mmxx4 6157 [MB/s]
+		 * RAID5 sse2x2 7151 [MB/s]
+		 * RAID5 sse2x4 8447 [MB/s]
+		 * RAID5 sse2x8 8155 [MB/s] (slower than sse2x4)
+		 * RAID6 int32x2 1892 [MB/s]
+		 * RAID6 mmxx2 3744 [MB/s]
+		 * RAID6 sse2x2 4922 [MB/s]
+		 * RAID6 sse2x4 4464 [MB/s] (slower than sse2x2)
+		 */
+		if (family == 21)
+			return 1;
+	}
+
+	return 0;
+}
+
+
 /* other cases */
 
 /* Intel(R) Core(TM) i7-3740QM CPU @ 2.70GHz
diff --git a/raid.c b/raid.c
index 25c6fc99..9c7fa10e 100644
--- a/raid.c
+++ b/raid.c
@@ -2505,8 +2505,13 @@ void raid_init(void)
 	}
 	if (cpu_has_sse2()) {
 #if defined(__x86_64__)
-		raid5_gen = raid5_sse2r8;
-		raid6_gen = raid6_sse2r4;
+		if (cpu_has_slowextendedreg()) {
+			raid5_gen = raid5_sse2r4;
+			raid6_gen = raid6_sse2r2;
+		} else {
+			raid5_gen = raid5_sse2r8;
+			raid6_gen = raid6_sse2r4;
+		}
 		raidTP_gen = raidTP_sse2r2;
 		raidQP_gen = raidQP_sse2r2;
 #else
diff --git a/speed.c b/speed.c
index 4ed50bfb..3c3525e9 100644
--- a/speed.c
+++ b/speed.c
@@ -132,13 +132,16 @@ void speed(void)
 
 	printf("Expected fastest RAID5 is ");
 #if defined(__i386__) || defined(__x86_64__)
-	if (cpu_has_sse2())
+	if (cpu_has_sse2()) {
 #if defined(__x86_64__)
-		printf("sse2x8");
+		if (cpu_has_slowextendedreg())
+			printf("sse2x4");
+		else
+			printf("sse2x8");
 #else
 		printf("sse2x4");
 #endif
-	else if (cpu_has_mmx())
+	} else if (cpu_has_mmx())
 		printf("mmxx4");
 	else
 		printf("int32x2");
@@ -152,13 +155,16 @@ void speed(void)
 
 	printf("Expected fastest RAID6 is ");
 #if defined(__i386__) || defined(__x86_64__)
-	if (cpu_has_sse2())
+	if (cpu_has_sse2()) {
 #if defined(__x86_64__)
-		printf("sse2x4");
+		if (cpu_has_slowextendedreg())
+			printf("sse2x2");
+		else
+			printf("sse2x4");
 #else
 		printf("sse2x2");
 #endif
-	else if (cpu_has_mmx())
+	} else if (cpu_has_mmx())
 		printf("mmxx2");
 	else
 		printf("int32x2");