Replace vpermpd with vpermilpd

[platform/upstream/openblas.git] / common_x86_64.h

diff --git a/common_x86_64.h b/common_x86_64.h
index e0a6c4c..c05998d 100644 (file)
--- a/common_x86_64.h
+++ b/common_x86_64.h
@@ -41,6 +41,10 @@
 
 #ifndef ASSEMBLER
 
+#ifdef C_MSVC
+#include <intrin.h>
+#endif
+
 #ifdef C_SUN
 #define        __asm__ __asm
 #define        __volatile__
@@ -56,35 +60,55 @@
 #endif
 */
 
-#define MB
-#define WMB
+#ifdef __GNUC__
+#define MB do { __asm__ __volatile__("": : :"memory"); } while (0)
+#define WMB do { __asm__ __volatile__("": : :"memory"); } while (0)
+#else
+#define MB do {} while (0)
+#define WMB do {} while (0)
+#endif
 
 static void __inline blas_lock(volatile BLASULONG *address){
 
+#ifndef C_MSVC
   int ret;
+#else
+  BLASULONG ret;
+#endif
 
   do {
     while (*address) {YIELDING;};
 
+#ifndef C_MSVC
     __asm__ __volatile__(
                         "xchgl %0, %1\n"
                         : "=r"(ret), "=m"(*address)
                         : "0"(1), "m"(*address)
                         : "memory");
-
+#else
+    ret=InterlockedExchange64((volatile LONG64 *)(address), 1);
+#endif
   } while (ret);
+
 }
+#define BLAS_LOCK_DEFINED
 
 static __inline BLASULONG rpcc(void){
+#ifdef C_MSVC
+  return __rdtsc();
+#else
   BLASULONG a, d;
 
   __asm__ __volatile__ ("rdtsc" : "=a" (a), "=d" (d));
 
   return ((BLASULONG)a + ((BLASULONG)d << 32));
+#endif
 }
+#define RPCC_DEFINED
 
 #define RPCC64BIT
 
+#ifndef C_MSVC
 static __inline BLASULONG getstackaddr(void){
   BLASULONG addr;
 
@@ -93,22 +117,33 @@ static __inline BLASULONG getstackaddr(void){
 
   return addr;
 }
+#endif
 
 static __inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
 
-        __asm__ __volatile__("cpuid"
+#ifdef C_MSVC
+  int cpuinfo[4];
+  __cpuid(cpuinfo, op);
+  *eax=cpuinfo[0];
+  *ebx=cpuinfo[1];
+  *ecx=cpuinfo[2];
+  *edx=cpuinfo[3];
+#else
+        __asm__ __volatile__("mov $0, %%ecx;"
+                            "cpuid"
                             : "=a" (*eax),
                             "=b" (*ebx),
                             "=c" (*ecx),
                             "=d" (*edx)
                             : "0" (op));
+#endif
 }
 
 /*
 #define WHEREAMI
 */
 
-static inline int WhereAmI(void){
+static __inline int WhereAmI(void){
   int eax, ebx, ecx, edx;
   int apicid;
 
@@ -150,10 +185,14 @@ static inline int WhereAmI(void){
 #define GET_IMAGE_CANCEL
 
 #ifdef SMP
-#ifdef USE64BITINT
+#if defined(USE64BITINT)
 static __inline blasint blas_quickdivide(blasint x, blasint y){
   return x / y;
 }
+#elif defined (C_MSVC)
+static __inline BLASLONG blas_quickdivide(BLASLONG x, BLASLONG y){
+  return x / y;
+}
 #else
 extern unsigned int blas_quick_divide_table[];
 
@@ -163,9 +202,16 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 
   if (y <= 1) return x;
 
+#if (MAX_CPU_NUMBER > 64)  
+  if (y > 64) { 
+         result = x / y;
+         return result;
+  }
+#endif
+       
   y = blas_quick_divide_table[y];
 
-  __asm__ __volatile__  ("mull %0" :"=d" (result) :"a"(x), "0" (y));
+  __asm__ __volatile__  ("mull %0" :"=d" (result), "+a"(x) : "0" (y));
 
   return result;
 }
@@ -212,6 +258,10 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 #define RETURN_BY_STACK
 #endif
 
+#ifdef F_INTERFACE_FLANG
+#define RETURN_BY_STACK
+#endif
+
 #ifdef F_INTERFACE_PGI
 #define RETURN_BY_STACK
 #endif
@@ -226,8 +276,8 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 
 #ifdef ASSEMBLER
 
-#if defined(PILEDRIVER) || defined(BULLDOZER) || defined(STEAMROLLER)
-//Enable some optimazation for barcelona.
+#if defined(PILEDRIVER) || defined(BULLDOZER) || defined(STEAMROLLER) || defined(EXCAVATOR)
+//Enable some optimization for barcelona.
 #define BARCELONA_OPTIMIZATION
 #endif
 
@@ -363,10 +413,10 @@ REALNAME:
 
 #define PROFCODE
 
-#define EPILOGUE .end   REALNAME
+#define EPILOGUE .end
 #endif
 
-#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(__ELF__) || defined(C_PGI)
+#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(__ELF__) || defined(C_PGI)
 #define PROLOGUE \
        .text; \
        .align 512; \