--- SDL-1.2.8/src/hermes/HeadMMX.h.orig	2001-04-26 18:45:45.000000000 +0200
+++ SDL-1.2.8/src/hermes/HeadMMX.h	2005-05-05 22:32:19.000000000 +0200
@@ -10,24 +10,15 @@
 #ifndef __HERMES_HEAD_MMX__
 #define __HERMES_HEAD_MMX__
 
-
 /* If you cannot stand ifdefs, then please do not look into this file, it's
    going to end your life :) */
 
-#ifdef X86_ASSEMBLER
-
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 void STACKCALL ConvertMMX(HermesConverterInterface *);
 
-void STACKCALL ClearMMX_32(HermesClearInterface *);
-void STACKCALL ClearMMX_24(HermesClearInterface *);
-void STACKCALL ClearMMX_16(HermesClearInterface *);
-void STACKCALL ClearMMX_8(HermesClearInterface *);
-
 void ConvertMMXpII32_24RGB888();
 void ConvertMMXpII32_16RGB565();
 void ConvertMMXpII32_16BGR565();
@@ -41,40 +32,6 @@
 }
 #endif
 
-
-
-/* Fix the underscore business with ELF compilers */
-
-#if defined(__ELF__) && defined(__GNUC__)
-  #ifdef __cplusplus 
-  extern "C" {   
-  #endif
-
-  void ConvertMMX(HermesConverterInterface *) __attribute__ ((alias ("_ConvertMMX")));
-#if 0
-  void ClearMMX_32(HermesClearInterface *) __attribute__ ((alias ("_ClearMMX_32")));
-  void ClearMMX_24(HermesClearInterface *)  __attribute__ ((alias ("_ClearMMX_24")));
-  void ClearMMX_16(HermesClearInterface *)  __attribute__ ((alias ("_ClearMMX_16")));
-  void ClearMMX_8(HermesClearInterface *)  __attribute__ ((alias ("_ClearMMX_8")));
-
-  void ConvertMMXp32_16RGB555() __attribute__ ((alias ("_ConvertMMXp32_16RGB555")));
-#endif
-
-  void ConvertMMXpII32_24RGB888() __attribute__ ((alias ("_ConvertMMXpII32_24RGB888")));
-  void ConvertMMXpII32_16RGB565() __attribute__ ((alias ("_ConvertMMXpII32_16RGB565")));
-  void ConvertMMXpII32_16BGR565() __attribute__ ((alias ("_ConvertMMXpII32_16BGR565")));
-  void ConvertMMXpII32_16RGB555() __attribute__ ((alias ("_ConvertMMXpII32_16RGB555")));
-  void ConvertMMXpII32_16BGR555() __attribute__ ((alias ("_ConvertMMXpII32_16BGR555")));
-
-  #ifdef __cplusplus
-  }
-  #endif
-
-#endif /* ELF and GNUC */
-
-
-
-
 /* Make it work with Watcom */
 #ifdef __WATCOMC__
 #pragma warning 601 9
@@ -95,7 +52,4 @@
 
 #endif /* WATCOM */
 
-#endif /* X86_ASSEMBLER */
-
-
 #endif
--- SDL-1.2.8/src/hermes/HeadX86.h.orig	2002-06-08 19:45:39.000000000 +0200
+++ SDL-1.2.8/src/hermes/HeadX86.h	2005-05-05 22:32:11.000000000 +0200
@@ -10,9 +10,6 @@
 #ifndef __HERMES_HEAD_X86__
 #define __HERMES_HEAD_X86__
 
-
-#ifdef X86_ASSEMBLER
-
 /* If you can't stand IFDEFS, then close your eyes now, please :) */
 
 /* Ok, we start with normal function definitions */
@@ -20,12 +17,7 @@
 extern "C" {
 #endif
 
-
 void STACKCALL ConvertX86(HermesConverterInterface *);
-void STACKCALL ClearX86_32(HermesClearInterface *);
-void STACKCALL ClearX86_24(HermesClearInterface *);
-void STACKCALL ClearX86_16(HermesClearInterface *);
-void STACKCALL ClearX86_8(HermesClearInterface *);
 
 int STACKCALL Hermes_X86_CPU();
 
@@ -40,104 +32,15 @@
 void ConvertX86p32_16BGR555();
 void ConvertX86p32_8RGB332();
 
-void ConvertX86p16_32RGB888();
-void ConvertX86p16_32BGR888();
-void ConvertX86p16_32RGBA888();
-void ConvertX86p16_32BGRA888();
-void ConvertX86p16_24RGB888();
-void ConvertX86p16_24BGR888();
 void ConvertX86p16_16BGR565();
 void ConvertX86p16_16RGB555();
 void ConvertX86p16_16BGR555();
 void ConvertX86p16_8RGB332();
 
-void CopyX86p_4byte();
-void CopyX86p_3byte();
-void CopyX86p_2byte();
-void CopyX86p_1byte();
-
-void ConvertX86pI8_32();
-void ConvertX86pI8_24();
-void ConvertX86pI8_16();
-
-extern int ConvertX86p16_32RGB888_LUT_X86[512];
-extern int ConvertX86p16_32BGR888_LUT_X86[512];
-extern int ConvertX86p16_32RGBA888_LUT_X86[512];
-extern int ConvertX86p16_32BGRA888_LUT_X86[512];
-  
 #ifdef __cplusplus
 }
 #endif
 
-
-
-
-/* Now fix up the ELF underscore problem */
-
-#if defined(__ELF__) && defined(__GNUC__)
-  #ifdef __cplusplus
-  extern "C" {
-  #endif
-
-  int Hermes_X86_CPU() __attribute__ ((alias ("_Hermes_X86_CPU")));
-
-  void ConvertX86(HermesConverterInterface *) __attribute__ ((alias ("_ConvertX86")));
-
-#if 0
-  void ClearX86_32(HermesClearInterface *) __attribute__ ((alias ("_ClearX86_32")));
-  void ClearX86_24(HermesClearInterface *)  __attribute__ ((alias ("_ClearX86_24")));
-  void ClearX86_16(HermesClearInterface *)  __attribute__ ((alias ("_ClearX86_16")));
-  void ClearX86_8(HermesClearInterface *)  __attribute__ ((alias ("_ClearX86_8")));
-#endif
-
-  void ConvertX86p32_32BGR888() __attribute__ ((alias ("_ConvertX86p32_32BGR888")));
-  void ConvertX86p32_32RGBA888() __attribute__ ((alias ("_ConvertX86p32_32RGBA888")));
-  void ConvertX86p32_32BGRA888() __attribute__ ((alias ("_ConvertX86p32_32BGRA888")));
-  void ConvertX86p32_24RGB888() __attribute__ ((alias ("_ConvertX86p32_24RGB888")));
-  void ConvertX86p32_24BGR888() __attribute__ ((alias ("_ConvertX86p32_24BGR888")));
-  void ConvertX86p32_16RGB565() __attribute__ ((alias ("_ConvertX86p32_16RGB565")));
-  void ConvertX86p32_16BGR565() __attribute__ ((alias ("_ConvertX86p32_16BGR565")));
-  void ConvertX86p32_16RGB555() __attribute__ ((alias ("_ConvertX86p32_16RGB555")));
-  void ConvertX86p32_16BGR555() __attribute__ ((alias ("_ConvertX86p32_16BGR555")));
-  void ConvertX86p32_8RGB332() __attribute__ ((alias ("_ConvertX86p32_8RGB332")));
-
-#if 0
-  void ConvertX86p16_32RGB888() __attribute__ ((alias ("_ConvertX86p16_32RGB888")));
-  void ConvertX86p16_32BGR888() __attribute__ ((alias ("_ConvertX86p16_32BGR888")));
-  void ConvertX86p16_32RGBA888() __attribute__ ((alias ("_ConvertX86p16_32RGBA888")));
-  void ConvertX86p16_32BGRA888() __attribute__ ((alias ("_ConvertX86p16_32BGRA888")));
-  void ConvertX86p16_24RGB888() __attribute__ ((alias ("_ConvertX86p16_24RGB888")));
-  void ConvertX86p16_24BGR888() __attribute__ ((alias ("_ConvertX86p16_24BGR888")));
-#endif
-  void ConvertX86p16_16BGR565() __attribute__ ((alias ("_ConvertX86p16_16BGR565")));
-  void ConvertX86p16_16RGB555() __attribute__ ((alias ("_ConvertX86p16_16RGB555")));
-  void ConvertX86p16_16BGR555() __attribute__ ((alias ("_ConvertX86p16_16BGR555")));
-  void ConvertX86p16_8RGB332() __attribute__ ((alias ("_ConvertX86p16_8RGB332")));
-
-#if 0
-  void CopyX86p_4byte() __attribute__ ((alias ("_CopyX86p_4byte")));
-  void CopyX86p_3byte() __attribute__ ((alias ("_CopyX86p_3byte")));
-  void CopyX86p_2byte() __attribute__ ((alias ("_CopyX86p_2byte")));
-  void CopyX86p_1byte() __attribute__ ((alias ("_CopyX86p_1byte")));
-
-  void ConvertX86pI8_32() __attribute__ ((alias ("_ConvertX86pI8_32")));
-  void ConvertX86pI8_24() __attribute__ ((alias ("_ConvertX86pI8_24")));
-  void ConvertX86pI8_16() __attribute__ ((alias ("_ConvertX86pI8_16")));
-
-  extern int ConvertX86p16_32RGB888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32RGB888_LUT_X86")));
-  extern int ConvertX86p16_32BGR888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32BGR888_LUT_X86")));
-  extern int ConvertX86p16_32RGBA888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32RGBA888_LUT_X86")));
-  extern int ConvertX86p16_32BGRA888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32BGRA888_LUT_X86")));
-#endif
-
-  #ifdef __cplusplus
-  }
-  #endif
-
-#endif /* ELF & GNU */
-
-
-
 /* Make it run with WATCOM C */
 #ifdef __WATCOMC__
 #pragma warning 601 9
@@ -188,8 +91,4 @@
 
 #endif /* __WATCOMC__ */
 
-
-#endif /* X86_ASSEMBLER */
-
-
 #endif 
--- SDL-1.2.8/src/hermes/mmx_main.asm.orig	2005-05-05 22:38:42.000000000 +0200
+++ SDL-1.2.8/src/hermes/mmx_main.asm	2005-05-05 22:38:46.000000000 +0200
@@ -10,6 +10,7 @@
 BITS 32
 
 GLOBAL _ConvertMMX
+GLOBAL ConvertMMX
 GLOBAL _mmxreturn
 
 SECTION .text
@@ -30,6 +31,7 @@
 ;;  36: int32 *lookup
 	
 _ConvertMMX:
+ConvertMMX:
 	push ebp
 	mov ebp,esp
 
--- SDL-1.2.8/src/hermes/mmxp2_32.asm.orig	2002-06-08 19:45:39.000000000 +0200
+++ SDL-1.2.8/src/hermes/mmxp2_32.asm	2005-05-05 22:41:00.000000000 +0200
@@ -26,6 +26,11 @@
 GLOBAL _ConvertMMXpII32_16BGR565
 GLOBAL _ConvertMMXpII32_16RGB555
 GLOBAL _ConvertMMXpII32_16BGR555
+GLOBAL ConvertMMXpII32_24RGB888
+GLOBAL ConvertMMXpII32_16RGB565
+GLOBAL ConvertMMXpII32_16BGR565
+GLOBAL ConvertMMXpII32_16RGB555
+GLOBAL ConvertMMXpII32_16BGR555
 
 EXTERN _mmxreturn
  
@@ -51,6 +56,7 @@
 SECTION .text
 
 _ConvertMMXpII32_24RGB888:
+ConvertMMXpII32_24RGB888:
 
         ; set up mm6 as the mask, mm7 as zero
         movq mm6, qword [mmx32_rgb888_mask]
@@ -113,6 +119,7 @@
 
 
 _ConvertMMXpII32_16RGB565:
+ConvertMMXpII32_16RGB565:
 
         ; set up masks
         movq mm5, [mmx32_rgb565_b]
@@ -180,6 +187,7 @@
 
 	
 _ConvertMMXpII32_16BGR565:
+ConvertMMXpII32_16BGR565:
 
         movq mm5, [mmx32_rgb565_r]
         movq mm6, [mmx32_rgb565_g]
@@ -248,6 +256,7 @@
         jmp _mmxreturn
 
 _ConvertMMXpII32_16BGR555:
+ConvertMMXpII32_16BGR555:
 
         ; the 16BGR555 converter is identical to the RGB555 one,
         ; except it uses a different multiplier for the pmaddwd
@@ -261,7 +270,9 @@
 ; would almost certainly be faster, even if only a little.
 ; I did rename 'mmx32_rgb555_add' to 'mmx32_rgb555_mul', which is
 ; (I think) a more accurate name..
+
 _ConvertMMXpII32_16RGB555:
+ConvertMMXpII32_16RGB555:
 
         movq mm7,qword [mmx32_rgb555_mul]
 _convert_bgr555_cheat:
--- SDL-1.2.8/src/hermes/x86_main.asm.orig	2001-04-26 18:45:45.000000000 +0200
+++ SDL-1.2.8/src/hermes/x86_main.asm	2005-05-05 22:42:06.000000000 +0200
@@ -12,10 +12,11 @@
 BITS 32
 
 GLOBAL _ConvertX86
+GLOBAL ConvertX86
 GLOBAL _x86return
 	
 GLOBAL _Hermes_X86_CPU
-
+GLOBAL Hermes_X86_CPU
 
 SECTION .data
 	
@@ -40,6 +41,7 @@
 ;;  36: int32 *lookup
 	
 _ConvertX86:
+ConvertX86:
 	push ebp
 	mov ebp,esp
 
@@ -81,6 +83,7 @@
 ;; Hermes_X86_CPU returns the CPUID flags in eax
 	
 _Hermes_X86_CPU:
+Hermes_X86_CPU:
 	pushfd
 	pop eax
 	
--- SDL-1.2.8/src/hermes/x86p_16.asm.orig	2001-04-26 18:45:45.000000000 +0200
+++ SDL-1.2.8/src/hermes/x86p_16.asm	2005-05-05 22:42:58.000000000 +0200
@@ -23,6 +23,16 @@
 GLOBAL _ConvertX86p16_16RGB555
 GLOBAL _ConvertX86p16_16BGR555
 GLOBAL _ConvertX86p16_8RGB332
+GLOBAL ConvertX86p16_32RGB888
+GLOBAL ConvertX86p16_32BGR888
+GLOBAL ConvertX86p16_32RGBA888
+GLOBAL ConvertX86p16_32BGRA888
+GLOBAL ConvertX86p16_24RGB888
+GLOBAL ConvertX86p16_24BGR888
+GLOBAL ConvertX86p16_16BGR565
+GLOBAL ConvertX86p16_16RGB555
+GLOBAL ConvertX86p16_16BGR555
+GLOBAL ConvertX86p16_8RGB332
 
 EXTERN _ConvertX86
 EXTERN _x86return
@@ -33,6 +43,7 @@
 
 
 _ConvertX86p16_16BGR565:
+ConvertX86p16_16BGR565:
 
     ; check short
     cmp ecx,BYTE 16
@@ -143,6 +154,7 @@
 
 
 _ConvertX86p16_16RGB555:
+ConvertX86p16_16RGB555:
 
     ; check short
     cmp ecx,BYTE 32
@@ -251,6 +263,7 @@
 
 
 _ConvertX86p16_16BGR555:
+ConvertX86p16_16BGR555:
 
     ; check short
     cmp ecx,BYTE 16
@@ -369,6 +382,7 @@
 
 
 _ConvertX86p16_8RGB332:
+ConvertX86p16_8RGB332:
 
     ; check short
     cmp ecx,BYTE 16
--- SDL-1.2.8/src/hermes/x86p_32.asm.orig	2001-04-26 18:45:45.000000000 +0200
+++ SDL-1.2.8/src/hermes/x86p_32.asm	2005-05-05 22:44:07.000000000 +0200
@@ -22,6 +22,16 @@
 GLOBAL _ConvertX86p32_16RGB555
 GLOBAL _ConvertX86p32_16BGR555
 GLOBAL _ConvertX86p32_8RGB332
+GLOBAL ConvertX86p32_32BGR888
+GLOBAL ConvertX86p32_32RGBA888
+GLOBAL ConvertX86p32_32BGRA888
+GLOBAL ConvertX86p32_24RGB888	
+GLOBAL ConvertX86p32_24BGR888
+GLOBAL ConvertX86p32_16RGB565
+GLOBAL ConvertX86p32_16BGR565
+GLOBAL ConvertX86p32_16RGB555
+GLOBAL ConvertX86p32_16BGR555
+GLOBAL ConvertX86p32_8RGB332
 
 EXTERN _x86return
 		
@@ -38,6 +48,7 @@
 
 
 _ConvertX86p32_32BGR888:
+ConvertX86p32_32BGR888:
 
     ; check short
     cmp ecx,BYTE 32
@@ -119,6 +130,7 @@
 	
 		
 _ConvertX86p32_32RGBA888:
+ConvertX86p32_32RGBA888:
 	
     ; check short
     cmp ecx,BYTE 32
@@ -190,6 +202,7 @@
 
 
 _ConvertX86p32_32BGRA888:
+ConvertX86p32_32BGRA888:
 
     ; check short
     cmp ecx,BYTE 32
@@ -265,6 +278,7 @@
 ;; 32 bit RGB 888 to 24 BIT RGB 888
 
 _ConvertX86p32_24RGB888:
+ConvertX86p32_24RGB888:
 
 	; check short
 	cmp ecx,BYTE 32
@@ -362,6 +376,7 @@
 ;; 32 bit RGB 888 to 24 bit BGR 888
 
 _ConvertX86p32_24BGR888:
+ConvertX86p32_24BGR888:
 
 	; check short
 	cmp ecx,BYTE 32
@@ -463,6 +478,8 @@
 ;; 32 bit RGB 888 to 16 BIT RGB 565 
 
 _ConvertX86p32_16RGB565:
+ConvertX86p32_16RGB565:
+
 	; check short
 	cmp ecx,BYTE 16
 	ja .L3
@@ -578,6 +595,7 @@
 ;; 32 bit RGB 888 to 16 BIT BGR 565 
 
 _ConvertX86p32_16BGR565:
+ConvertX86p32_16BGR565:
 	
 	; check short
 	cmp ecx,BYTE 16
@@ -692,6 +710,7 @@
 ;; 32 BIT RGB TO 16 BIT RGB 555
 
 _ConvertX86p32_16RGB555:
+ConvertX86p32_16RGB555:
 
 	; check short
 	cmp ecx,BYTE 16
@@ -803,6 +822,7 @@
 ;; 32 BIT RGB TO 16 BIT BGR 555
 	
 _ConvertX86p32_16BGR555:
+ConvertX86p32_16BGR555:
 	
 	; check short
 	cmp ecx,BYTE 16
@@ -919,8 +939,9 @@
 ;; FROM 32 BIT RGB to 8 BIT RGB (rrrgggbbb)
 ;; This routine writes FOUR pixels at once (dword) and then, if they exist
 ;; the trailing three pixels
-_ConvertX86p32_8RGB332:
 
+_ConvertX86p32_8RGB332:
+ConvertX86p32_8RGB332:
 	
 .L_ALIGNED
 	push ecx