From 35f668bb14dbc86c367e38d57188747b7a2a942d Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Sat, 28 Jun 2014 10:01:56 +0200
Subject: [PATCH] bugfix for cgemm_kernel_8x2_sandy.S

---
 kernel/x86_64/cgemm_kernel_8x2_sandy.S | 43 ++++++++++++++------------
 make.inc                               |  2 +-
 2 files changed, 24 insertions(+), 21 deletions(-)

diff --git a/kernel/x86_64/cgemm_kernel_8x2_sandy.S b/kernel/x86_64/cgemm_kernel_8x2_sandy.S
index ccd04e742..564b73380 100644
--- a/kernel/x86_64/cgemm_kernel_8x2_sandy.S
+++ b/kernel/x86_64/cgemm_kernel_8x2_sandy.S
@@ -812,16 +812,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #ifdef WINDOWS_ABI
 	movq	%rdi,    48(%rsp)
 	movq	%rsi,    56(%rsp)
-	movups	%xmm6,   64(%rsp)
-	movups	%xmm7,   80(%rsp)
-	movups	%xmm8,   96(%rsp)
-	movups	%xmm9,  112(%rsp)
-	movups	%xmm10, 128(%rsp)
-	movups	%xmm11, 144(%rsp)
-	movups	%xmm12, 160(%rsp)
-	movups	%xmm13, 176(%rsp)
-	movups	%xmm14, 192(%rsp)
-	movups	%xmm15, 208(%rsp)
+	vmovups	%xmm6,   64(%rsp)
+	vmovups	%xmm7,   80(%rsp)
+	vmovups	%xmm8,   96(%rsp)
+	vmovups	%xmm9,  112(%rsp)
+	vmovups	%xmm10, 128(%rsp)
+	vmovups	%xmm11, 144(%rsp)
+	vmovups	%xmm12, 160(%rsp)
+	vmovups	%xmm13, 176(%rsp)
+	vmovups	%xmm14, 192(%rsp)
+	vmovups	%xmm15, 208(%rsp)
 
 	movq	ARG1,      OLD_M
 	movq	ARG2,      OLD_N
@@ -834,6 +834,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	movsd	OLD_OFFSET, %xmm12
 #endif
 	vmovaps	%xmm3, %xmm0
+	vmovsd   OLD_ALPHA_I, %xmm1
 
 #else
 	movq	STACKSIZE +  8(%rsp), LDC
@@ -2249,6 +2250,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 
 .L999:
+	vzeroupper
+
 	movq   		SP, %rsp
 	movq	   (%rsp), %rbx
 	movq	  8(%rsp), %rbp
@@ -2260,16 +2263,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #ifdef WINDOWS_ABI
 	movq	 48(%rsp), %rdi
 	movq	 56(%rsp), %rsi
-	movups	 64(%rsp), %xmm6
-	movups	 80(%rsp), %xmm7
-	movups	 96(%rsp), %xmm8
-	movups	112(%rsp), %xmm9
-	movups	128(%rsp), %xmm10
-	movups	144(%rsp), %xmm11
-	movups	160(%rsp), %xmm12
-	movups	176(%rsp), %xmm13
-	movups	192(%rsp), %xmm14
-	movups	208(%rsp), %xmm15
+	vmovups	 64(%rsp), %xmm6
+	vmovups	 80(%rsp), %xmm7
+	vmovups	 96(%rsp), %xmm8
+	vmovups	112(%rsp), %xmm9
+	vmovups	128(%rsp), %xmm10
+	vmovups	144(%rsp), %xmm11
+	vmovups	160(%rsp), %xmm12
+	vmovups	176(%rsp), %xmm13
+	vmovups	192(%rsp), %xmm14
+	vmovups	208(%rsp), %xmm15
 #endif
 
 	addq	$ STACKSIZE, %rsp
diff --git a/make.inc b/make.inc
index affae3a2d..da430b7d6 100644
--- a/make.inc
+++ b/make.inc
@@ -1,7 +1,7 @@
 SHELL = /bin/sh
 PLAT = _LINUX
 DRVOPTS  = $(OPTS)
-LOADER   = $(FORTRAN) -pthread
+LOADER   = $(FORTRAN)
 ARCHFLAGS= -ru
 #RANLIB   = ranlib