From d0d304b3422938044ff15cd3659c9b64b18735a4 Mon Sep 17 00:00:00 2001
From: Timo Kreuzer <timo.kreuzer@reactos.org>
Date: Wed, 12 Oct 2011 21:11:02 +0000
Subject: [PATCH] [RTL] On x86 builds use the asm versions of the byte swap
 functions. This is neccessary for RtlUlonglongByteswap, because the function
 breaks the calling convention on Windows, not cleaning up the stack. Make the
 asm compatible to ML and improve it a bit. Investigated by Thomas Faber.

svn path=/trunk/; revision=54097
---
 reactos/lib/rtl/CMakeLists.txt |  5 ++++-
 reactos/lib/rtl/byteswap.c     | 11 +++++++---
 reactos/lib/rtl/i386/rtlswap.S | 39 +++++++++++++++++++---------------
 reactos/lib/rtl/rtl.rbuild     |  4 +++-
 4 files changed, 37 insertions(+), 22 deletions(-)

diff --git a/reactos/lib/rtl/CMakeLists.txt b/reactos/lib/rtl/CMakeLists.txt
index 83e38da00c3..7dd2fe8c14e 100644
--- a/reactos/lib/rtl/CMakeLists.txt
+++ b/reactos/lib/rtl/CMakeLists.txt
@@ -14,7 +14,6 @@ list(APPEND SOURCE
     avltable.c
     bitmap.c
     bootdata.c
-    byteswap.c
     compress.c
     condvar.c
     crc32.c
@@ -71,10 +70,12 @@ if(ARCH MATCHES i386)
         i386/except.c
         i386/interlck.S
         i386/rtlmem.s
+        i386/rtlswap.s
         i386/res_asm.s
         i386/thread.c)
 elseif(ARCH MATCHES amd64)
     list(APPEND SOURCE
+        byteswap.c
         amd64/debug_asm.S
         amd64/except_asm.S
         amd64/slist.S
@@ -83,10 +84,12 @@ elseif(ARCH MATCHES amd64)
         mem.c)
 elseif(ARCH MATCHES arm)
     list(APPEND SOURCE
+        byteswap.c
         arm/debug_asm.S
         mem.c)
 elseif(ARCH MATCHES powerpc)
     list(APPEND SOURCE
+        byteswap.c
         powerpc/debug.c
         powerpc/except.c
         powerpc/interlocked.c
diff --git a/reactos/lib/rtl/byteswap.c b/reactos/lib/rtl/byteswap.c
index b11715ec86f..b58f6cf05f6 100644
--- a/reactos/lib/rtl/byteswap.c
+++ b/reactos/lib/rtl/byteswap.c
@@ -12,6 +12,11 @@
 #define NDEBUG
 #include <debug.h>
 
+#if defined(_M_IX86)
+/* RtlUlonglongByteSwap is broken and cannot be done in C on x86 */
+#error "Use rtlswap.S!"
+#endif
+
 #undef RtlUlonglongByteSwap
 #undef RtlUlongByteSwap
 #undef RtlUshortByteSwap
@@ -31,7 +36,7 @@ FASTCALL
 RtlUshortByteSwap(
     IN USHORT Source)
 {
-#if defined(_M_IX86) || defined(_M_AMD64)
+#if defined(_M_AMD64)
     return _byteswap_ushort(Source);
 #else
     return (Source >> 8) | (Source << 8);
@@ -55,7 +60,7 @@ FASTCALL
 RtlUlongByteSwap(
    IN ULONG Source)
 {
-#if defined(_M_IX86) || defined(_M_AMD64)
+#if defined(_M_AMD64)
     return _byteswap_ulong(Source);
 #else
     return ((ULONG)RtlUshortByteSwap((USHORT)Source) << 16) | RtlUshortByteSwap((USHORT)(Source >> 16));
@@ -80,7 +85,7 @@ ULONGLONG FASTCALL
 RtlUlonglongByteSwap(
     IN ULONGLONG Source)
 {
-#if defined(_M_IX86) || defined(_M_AMD64)
+#if defined(_M_AMD64)
     return _byteswap_uint64(Source);
 #else
     return ((ULONGLONG) RtlUlongByteSwap (Source) << 32) | RtlUlongByteSwap (Source>>32);
diff --git a/reactos/lib/rtl/i386/rtlswap.S b/reactos/lib/rtl/i386/rtlswap.S
index 6ded910eb85..cd37de3a249 100644
--- a/reactos/lib/rtl/i386/rtlswap.S
+++ b/reactos/lib/rtl/i386/rtlswap.S
@@ -4,50 +4,55 @@
  * PURPOSE:           Byte swap functions
  * FILE:              lib/rtl/i386/rtlswap.S
  * PROGRAMER:         Alex Ionescu (alex.ionescu@reactos.org)
+ *                    Timo Kreuzer (timo.kreuzer@reactos.org)
  */
 
-.intel_syntax noprefix
+#include <asm.inc>
 
-.globl @RtlUshortByteSwap@4
-.globl @RtlUlongByteSwap@4
-.globl @RtlUlonglongByteSwap@8
+PUBLIC @RtlUshortByteSwap@4
+PUBLIC @RtlUlongByteSwap@4
+PUBLIC @RtlUlonglongByteSwap@8
 
 /* FUNCTIONS ***************************************************************/
+.code
 
-.func @RtlUshortByteSwap@4, @RtlUshortByteSwap@4
 @RtlUshortByteSwap@4:
+FUNC RtlUshortByteSwap
+    FPO 0, 0, 0, 0, 0, FRAME_FPO
 
     /* Swap high and low bits */
     mov ah, cl
     mov al, ch
     ret
-.endfunc
+ENDFUNC RtlUshortByteSwap
 
-.func @RtlUlongByteSwap@4, @RtlUlongByteSwap@4
 @RtlUlongByteSwap@4:
+FUNC RtlUlongByteSwap
+    FPO 0, 0, 0, 0, 0, FRAME_FPO
 
     /* Swap high and low bits */
     mov eax, ecx
     bswap eax
     ret
-.endfunc
+ENDFUNC RtlUlongByteSwap
 
-.func @RtlUlonglongByteSwap@8, @RtlUlonglongByteSwap@8
 @RtlUlonglongByteSwap@8:
+FUNC RtlUlonglongByteSwap
+    FPO 0, 2, 0, 0, 0, FRAME_FPO
 
     /* Get 64-bit integer */
-    mov edx, [esp+8]
-    mov eax, [esp+4]
+    mov eax, [esp+8]
+    mov edx, [esp+4]
 
     /* Swap it */
     bswap edx
     bswap eax
 
-    /* Return it */
-    mov ecx, eax
-    mov eax, edx
-    mov edx, ecx
+    /* Return it (NOTE: this might look wrong, since fastcall functions
+       should clean up the stack, even if the first parameter is an ULONGLONG,
+       and therefore put on tthe stack instead of in ecx and edx,
+       but thats exactly how the function behaves on Windows! */
     ret
-.endfunc
-
+ENDFUNC RtlUlonglongByteSwap
 
+END
diff --git a/reactos/lib/rtl/rtl.rbuild b/reactos/lib/rtl/rtl.rbuild
index 348a7a31350..75758ab5353 100644
--- a/reactos/lib/rtl/rtl.rbuild
+++ b/reactos/lib/rtl/rtl.rbuild
@@ -20,6 +20,7 @@
 		</directory>
 	</if>
 	<if property="ARCH" value="powerpc">
+		<file>byteswap.c</file>
 	<directory name="powerpc">
 		<file>debug.c</file>
 		<file>except.c</file>
@@ -30,12 +31,14 @@
 	</directory>
 	</if>
 	<if property="ARCH" value="arm">
+		<file>byteswap.c</file>
 		<directory name="arm">
 			<file>debug_asm.S</file>
 		</directory>
 		<file>mem.c</file>
 	</if>
 	<if property="ARCH" value="amd64">
+		<file>byteswap.c</file>
 		<directory name="amd64">
 			<file>debug_asm.S</file>
 			<file>except_asm.S</file>
@@ -51,7 +54,6 @@
 	<file>assert.c</file>
 	<file>atom.c</file>
 	<file>avltable.c</file>
-	<file>byteswap.c</file>
 	<file>bitmap.c</file>
 	<file>bootdata.c</file>
 	<file>compress.c</file>