4 files changed, 44 insertions, 3 deletions
diff --git a/src/utils/cpuid.cpp b/src/utils/cpuid.cpp
index 4f22ef867..d5149e2e5 100644
--- a/src/utils/cpuid.cpp
+++ b/src/utils/cpuid.cpp
@@ -31,12 +31,12 @@ u64bit x86_processor_flags()
 
 #if defined(BOTAN_TARGET_ARCH_IS_X86) || defined(BOTAN_TARGET_ARCH_IS_AMD64)
 
-#if defined(__GNUG__)
+#if defined(BOTAN_BUILD_COMPILER_IS_GCC)
 
    u32bit a = 1, b = 0, c = 0, d = 0;
 
 #if defined(__i386__) && defined(__PIC__)
-   // ebx is used in PIC on 32-bit x86, so save and restore it
+   // ebx is reserved for PIC on 32-bit x86, so save and restore it
    asm("xchgl %%ebx, %1\n\t"
        "cpuid\n\t"
        "xchgl %%ebx, %1\n\t"
@@ -49,7 +49,7 @@ u64bit x86_processor_flags()
 
    proc_flags = ((u64bit)c << 32) | d;
 
-#elif defined(_MSC_VER)
+#elif defined(BOTAN_BUILD_COMPILER_IS_MSVC)
 
    int cpuinfo[4] = { 0 };
    __cpuid(cpuinfo, 1);
diff --git a/src/utils/info.txt b/src/utils/info.txt
index f530fb03a..c9648af51 100644
--- a/src/utils/info.txt
+++ b/src/utils/info.txt
@@ -23,6 +23,7 @@ mlock.h
 mlock.cpp
 parsing.cpp
 parsing.h
+prefetch.h
 rotate.h
 rounding.h
 stl_util.h
diff --git a/src/utils/loadstor.h b/src/utils/loadstor.h
index 77ed1554e..8c64deaee 100644
--- a/src/utils/loadstor.h
+++ b/src/utils/loadstor.h
@@ -12,6 +12,7 @@
 #include <botan/types.h>
 #include <botan/bswap.h>
 #include <botan/rotate.h>
+#include <botan/prefetch.h>
 
 #if BOTAN_TARGET_UNALIGNED_LOADSTOR_OK
 
diff --git a/src/utils/prefetch.h b/src/utils/prefetch.h
new file mode 100644
index 000000000..100829dce
--- /dev/null
+++ b/src/utils/prefetch.h
@@ -0,0 +1,39 @@
+/*
+* Prefetching Operations
+* (C) 2009 Jack Lloyd
+*
+* Distributed under the terms of the Botan license
+*/
+
+#ifndef BOTAN_PREFETCH_H__
+#define BOTAN_PREFETCH_H__
+
+#include <botan/cpuid.h>
+
+namespace Botan {
+
+inline void prefetch_readonly(const void* addr_void, u32bit length)
+   {
+#if defined(__GNUG__)
+   const byte* addr = static_cast<const byte*>(addr_void);
+   const u32bit cl_size = CPUID::cache_line_size();
+
+   for(u32bit i = 0; i <= length; i += cl_size)
+      __builtin_prefetch(addr + i, 0);
+#endif
+   }
+
+inline void prefetch_readwrite(const void* addr_void, u32bit length)
+   {
+#if defined(__GNUG__)
+   const byte* addr = static_cast<const byte*>(addr_void);
+   const u32bit cl_size = CPUID::cache_line_size();
+
+   for(u32bit i = 0; i <= length; i += cl_size)
+      __builtin_prefetch(addr + i, 1);
+#endif
+   }
+
+}
+
+#endif