From 8e97b7a248403ca9e9b38a70997fdb8ff93dfe94 Mon Sep 17 00:00:00 2001
From: Andrea Mazzoleni <amadvance@gmail.com>
Date: Sun, 14 Sep 2014 19:47:43 +0200
Subject: [PATCH] Adds proper detection for AVX2.

---
 cmdline/speed.c |  3 +--
 raid/cpu.h      | 28 ++++++++++++++++++++++++++--
 2 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/cmdline/speed.c b/cmdline/speed.c
index e4867669..a2e83d05 100644
--- a/cmdline/speed.c
+++ b/cmdline/speed.c
@@ -122,12 +122,11 @@ void speed(int period)
 
 		raid_cpu_info(vendor, &family, &model);
 
-		printf("CPU %s, family %u, model %u, flags%s%s%s%s%s%s%s%s%s\n", vendor, family, model,
+		printf("CPU %s, family %u, model %u, flags%s%s%s%s%s%s%s%s\n", vendor, family, model,
 			raid_cpu_has_mmx() ? " mmx" : "",
 			raid_cpu_has_sse2() ? " sse2" : "",
 			raid_cpu_has_ssse3() ? " ssse3" : "",
 			raid_cpu_has_sse42() ? " sse42" : "",
-			raid_cpu_has_avx() ? " avx" : "",
 			raid_cpu_has_avx2() ? "avx2" : "",
 			raid_cpu_has_slowmult() ? " slowmult" : "",
 			raid_cpu_has_slowpshufb() ? " slowpshufb" : "",
diff --git a/raid/cpu.h b/raid/cpu.h
index 5511e916..164321f2 100644
--- a/raid/cpu.h
+++ b/raid/cpu.h
@@ -106,19 +106,43 @@ static inline int raid_cpu_has_sse42(void)
 	return (reg[2] >> 20) & 1;
 }
 
-static inline int raid_cpu_has_avx(void)
+static inline int raid_cpu_has_xsavexrstor(void)
 {
 	uint32_t reg[4];
 
 	raid_cpuid(1, 0, reg);
 
-	return (reg[2] >> 28) & 1;
+	return (reg[2] >> 27) & 1;
+}
+
+static inline int raid_os_support_256bit_regs(void)
+{
+	uint32_t reg[4];
+
+	/* get the value of the Extended Control Register ecx=0 */
+	asm volatile(
+		"xgetbv\n"
+		: "=a" (reg[0]), "=d" (reg[3])
+		: "c" (0)
+	);
+
+	/* check if the OS save all the 256 bits registers */
+	return (reg[0] & 0x6) == 0x6;
 }
 
 static inline int raid_cpu_has_avx2(void)
 {
 	uint32_t reg[4];
 
+	/* check if the CPU supports XSAVE and XRSTOR, and also XGETBV */
+	if (!raid_cpu_has_xsavexrstor())
+		return 0;
+
+	/* check if the OS suports AVX2 registers */
+	if (!raid_os_support_256bit_regs())
+		return 0;
+
+	/* now check for AVX2 support */
 	raid_cpuid(7, 0, reg);
 
 	return (reg[1] >> 5) & 1;