[corlib] Icall for Buffer.Memcpy

Jitted code is too slow. For larger sizes, doing an icall to memcpy is 4-5 times faster than using the managed implementation.

5 files changed, 18 insertions, 2 deletions

diff --git a/configure.ac b/configure.ac
index 19e70ff460a..fa7a2d939ad 100644
--- a/configure.ac
+++ b/configure.ac
@@ -46,7 +46,7 @@ MONO_VERSION_BUILD=`echo $VERSION | cut -d . -f 3`
 # There is no ordering of corlib versions, no old or new,
 # the runtime expects an exact match.
 #
-MONO_CORLIB_VERSION=893F4575-474B-47A4-831E-E36181251FCB
+MONO_CORLIB_VERSION=109afe51-a298-4c14-a076-cfc83a5e755a
 
 #
 # Put a quoted #define in config.h.
diff --git a/mcs/class/corlib/ReferenceSources/Buffer.cs b/mcs/class/corlib/ReferenceSources/Buffer.cs
index b04390195ed..5de72a7bc32 100644
--- a/mcs/class/corlib/ReferenceSources/Buffer.cs
+++ b/mcs/class/corlib/ReferenceSources/Buffer.cs
@@ -11,6 +11,9 @@ namespace System
 {
 	partial class Buffer
 	{
+		[MethodImplAttribute(MethodImplOptions.InternalCall)]
+		internal static extern unsafe void InternalMemcpy (byte *dest, byte *src, int count);
+
 		public static int ByteLength (Array array)
 		{
 			// note: the other methods in this class also use ByteLength to test for
@@ -191,6 +194,11 @@ namespace System
 		}
 
 		internal static unsafe void Memcpy (byte *dest, byte *src, int len) {
+			// For bigger lengths, we use the heavily optimized native code
+			if (len > 32) {
+				InternalMemcpy (dest, src, len);
+				return;
+			}
 			// FIXME: if pointers are not aligned, try to align them
 			// so a faster routine can be used. Handle the case where
 			// the pointers can't be reduced to have the same alignment
@@ -246,4 +254,4 @@ namespace System
 			}
 		}
 	}
-}
\ No newline at end of file
+}
diff --git a/mono/metadata/icall-decl.h b/mono/metadata/icall-decl.h
index 4beade9a29b..65b924f4042 100644
--- a/mono/metadata/icall-decl.h
+++ b/mono/metadata/icall-decl.h
@@ -167,6 +167,7 @@ ICALL_EXPORT void mono_ArgIterator_Setup (MonoArgIterator*, char*, char*);
 ICALL_EXPORT void ves_icall_Mono_Runtime_RegisterReportingForNativeLib (const char*, const char*);
 ICALL_EXPORT void ves_icall_System_Array_GetGenericValueImpl (MonoArray*, guint32, gpointer);
 ICALL_EXPORT void ves_icall_System_Array_SetGenericValueImpl (MonoArray*, guint32, gpointer);
+ICALL_EXPORT void ves_icall_System_Buffer_MemcpyInternal (gpointer dest, gconstpointer src, gint32 count);
 ICALL_EXPORT void ves_icall_System_Buffer_SetByteInternal (MonoArray*, gint32, gint8);
 ICALL_EXPORT void ves_icall_System_Environment_Exit (int);
 ICALL_EXPORT void ves_icall_System_GCHandle_FreeHandle (guint32 handle);
diff --git a/mono/metadata/icall-def.h b/mono/metadata/icall-def.h
index 4ba68dee1ae..452c4777ac7 100644
--- a/mono/metadata/icall-def.h
+++ b/mono/metadata/icall-def.h
@@ -219,6 +219,7 @@ HANDLES(ARRAY_13, "SetValueImpl",  ves_icall_System_Array_SetValueImpl, void, 3,
 
 ICALL_TYPE(BUFFER, "System.Buffer", BUFFER_1)
 ICALL(BUFFER_1, "InternalBlockCopy", ves_icall_System_Buffer_BlockCopyInternal)
+NOHANDLES(ICALL(BUFFER_5, "InternalMemcpy", ves_icall_System_Buffer_MemcpyInternal))
 ICALL(BUFFER_2, "_ByteLength", ves_icall_System_Buffer_ByteLengthInternal)
 ICALL(BUFFER_3, "_GetByte", ves_icall_System_Buffer_GetByteInternal)
 ICALL(BUFFER_4, "_SetByte", ves_icall_System_Buffer_SetByteInternal)
diff --git a/mono/metadata/icall.c b/mono/metadata/icall.c
index 9d82771f0e3..715229ba843 100644
--- a/mono/metadata/icall.c
+++ b/mono/metadata/icall.c
@@ -6634,6 +6634,12 @@ ves_icall_System_Buffer_SetByteInternal (MonoArray *array, gint32 idx, gint8 val
 	mono_array_set_internal (array, gint8, idx, value);
 }
 
+void
+ves_icall_System_Buffer_MemcpyInternal (gpointer dest, gconstpointer src, gint32 count)
+{
+	memcpy (dest, src, count);
+}
+
 MonoBoolean
 ves_icall_System_Buffer_BlockCopyInternal (MonoArray *src, gint32 src_offset, MonoArray *dest, gint32 dest_offset, gint32 count) 
 {