- Move misplaced RtlPrefetchMemoryNonTemporal from rtl into kernel where it belongs (it only exists as an export for use by drivers to safely try to use prefetchnta even if it may not be available). Use existing global KePrefetchNTAGranularity that is set to the correct value via cpuid instead of hardcoded "Ke386CacheAlignment".

svn path=/trunk/; revision=68903

reactos/lib/rtl/i386/rtlmem.s

@@ -16,7 +16,6 @@ PUBLIC _RtlFillMemory@12
 PUBLIC _RtlFillMemoryUlong@12
 PUBLIC _RtlMoveMemory@12
 PUBLIC _RtlZeroMemory@8
-PUBLIC @RtlPrefetchMemoryNonTemporal@8
 
 /* FUNCTIONS *****************************************************************/
 .code
@@ -328,35 +327,4 @@ _RtlMoveMemory@12:
 	dec edi
 	jmp .CopyDownBytes
 
-
-
-@RtlPrefetchMemoryNonTemporal@8:
-
-    /*
-     * Overwritten by ntoskrnl/ke/i386/kernel.c if SSE is supported
-     * (see Ki386SetProcessorFeatures())
-     */
-    ret
-
-    /* Get granularity */
-    mov eax, [_Ke386CacheAlignment]
-
-FetchLine:
-
-    /* Prefetch this line */
-    prefetchnta byte ptr [ecx]
-
-    /* Update address and count */
-    add ecx, eax
-    sub edx, eax
-
-    /* Keep looping for the next line, or return if done */
-    ja FetchLine
-    ret
-
-
-/* FIXME: HACK */
-_Ke386CacheAlignment:
-    .long   64
-
 END

reactos/ntoskrnl/include/internal/amd64/ke.h

@@ -81,7 +81,6 @@ typedef struct _KI_INTERRUPT_DISPATCH_ENTRY
 } KI_INTERRUPT_DISPATCH_ENTRY, *PKI_INTERRUPT_DISPATCH_ENTRY;
 #include <poppack.h>
 
-extern ULONG Ke386CacheAlignment;
 extern ULONG KeI386NpxPresent;
 extern ULONG KeI386XMMIPresent;
 extern ULONG KeI386FxsrPresent;

reactos/ntoskrnl/include/internal/i386/ke.h

@@ -529,7 +529,6 @@ extern ULONG KeI386FxsrPresent;
 extern ULONG KiMXCsrMask;
 extern ULONG KeI386CpuType;
 extern ULONG KeI386CpuStep;
-extern ULONG Ke386CacheAlignment;
 extern ULONG KiFastSystemCallDisable;
 extern UCHAR KiDebugRegisterTrapOffsets[9];
 extern UCHAR KiDebugRegisterContextOffsets[9];

reactos/ntoskrnl/ke/i386/kiinit.c

@@ -162,7 +162,7 @@ KiInitMachineDependent(VOID)
             /* FIXME: Implement and enable XMM Page Zeroing for Mm */
 
             /* Patch the RtlPrefetchMemoryNonTemporal routine to enable it */
-            *(PCHAR)RtlPrefetchMemoryNonTemporal = 0x90;
+            *(PCHAR)RtlPrefetchMemoryNonTemporal = 0x90; // NOP
         }
     }
 

reactos/ntoskrnl/rtl/i386/stack.S

@@ -11,6 +11,8 @@
 #include <asm.inc>
 #include <ks386.inc>
 
+EXTERN _KePrefetchNTAGranularity:DWORD
+
 /* FUNCTIONS *****************************************************************/
 .code
 
@@ -35,4 +37,30 @@ _RtlpGetStackLimits@8:
     /* return */
     ret 8
 
+
+PUBLIC @RtlPrefetchMemoryNonTemporal@8
+@RtlPrefetchMemoryNonTemporal@8:
+
+    /*
+     * Kernel will overwrite this to 'nop' during init
+     * if prefetchnta is available. Slight optimization
+     * as compared to checking KeI386XMMIPresent for every call.
+     */
+    ret
+
+    /* Get granularity */
+    mov eax, [_KePrefetchNTAGranularity]
+
+    /* Prefetch this line */
+FetchLine:
+    prefetchnta byte ptr [ecx]
+
+    /* Update address and count */
+    add ecx, eax
+    sub edx, eax
+
+    /* Keep looping for the next line, or return if done */
+    ja FetchLine
+    ret
+
 END