| /contrib/sdk/sources/SDL-1.2.2_newlib/src/Tupfile.lua |
|---|
| 27,7 → 27,7 |
| for i,v in ipairs(FOLDERS) do |
| compile_gcc(v .. "*.c", v .. "%B.o") |
| tup.append_table(OBJS, |
| tup.foreach_rule(v .. "*.asm", "nasm -i hermes -f coff -o %o %f", v .. "%B.o") |
| tup.foreach_rule(v .. "*.asm", "nasm -f coff -o %o %f", v .. "%B.o") |
| ) |
| end |
| tup.rule(OBJS, "kos32-ar rcs %o %f", {"../../../lib/libSDLn.a", "../../../lib/<libSDLn>"}) |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/common.inc |
|---|
| File deleted |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/HeadMMX.h |
|---|
| 6,6 → 6,7 |
| Please refer to the file COPYING.LIB contained in the distribution for |
| licensing conditions |
| */ |
| #ifndef __HERMES_HEAD_MMX__ |
| #define __HERMES_HEAD_MMX__ |
| 44,25 → 45,27 |
| /* Fix the underscore business with ELF compilers */ |
| #if (defined(__ELF__) && defined(__GNUC__)) || defined(__SUNPRO_C) |
| #if defined(__ELF__) && defined(__GNUC__) |
| #ifdef __cplusplus |
| extern "C" { |
| #endif |
| extern void _ConvertMMX(HermesConverterInterface *); |
| extern void _ConvertMMXpII32_24RGB888(); |
| extern void _ConvertMMXpII32_16RGB565(); |
| extern void _ConvertMMXpII32_16BGR565(); |
| extern void _ConvertMMXpII32_16RGB555(); |
| extern void _ConvertMMXpII32_16BGR555(); |
| 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"))); |
| #define ConvertMMX _ConvertMMX |
| #define ConvertMMXpII32_24RGB888 _ConvertMMXpII32_24RGB888 |
| #define ConvertMMXpII32_16RGB565 _ConvertMMXpII32_16RGB565 |
| #define ConvertMMXpII32_16BGR565 _ConvertMMXpII32_16BGR565 |
| #define ConvertMMXpII32_16RGB555 _ConvertMMXpII32_16RGB555 |
| #define ConvertMMXpII32_16BGR555 _ConvertMMXpII32_16BGR555 |
| 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 |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/HeadX86.h |
|---|
| 60,10 → 60,10 |
| 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]; |
| extern int32 ConvertX86p16_32RGB888_LUT_X86[512]; |
| extern int32 ConvertX86p16_32BGR888_LUT_X86[512]; |
| extern int32 ConvertX86p16_32RGBA888_LUT_X86[512]; |
| extern int32 ConvertX86p16_32BGRA888_LUT_X86[512]; |
| #ifdef __cplusplus |
| } |
| 74,53 → 74,62 |
| /* Now fix up the ELF underscore problem */ |
| #if (defined(__ELF__) && defined(__GNUC__)) || defined(__SUNPRO_C) |
| #if defined(__ELF__) && defined(__GNUC__) |
| #ifdef __cplusplus |
| extern "C" { |
| #endif |
| extern int _Hermes_X86_CPU(); |
| int Hermes_X86_CPU() __attribute__ ((alias ("_Hermes_X86_CPU"))); |
| extern void _ConvertX86(HermesConverterInterface *); |
| void ConvertX86(HermesConverterInterface *) __attribute__ ((alias ("_ConvertX86"))); |
| extern void _ConvertX86p32_32BGR888(); |
| extern void _ConvertX86p32_32RGBA888(); |
| extern void _ConvertX86p32_32BGRA888(); |
| extern void _ConvertX86p32_24RGB888(); |
| extern void _ConvertX86p32_24BGR888(); |
| extern void _ConvertX86p32_16RGB565(); |
| extern void _ConvertX86p32_16BGR565(); |
| extern void _ConvertX86p32_16RGB555(); |
| extern void _ConvertX86p32_16BGR555(); |
| extern void _ConvertX86p32_8RGB332(); |
| #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 |
| extern void _ConvertX86p16_16BGR565(); |
| extern void _ConvertX86p16_16RGB555(); |
| extern void _ConvertX86p16_16BGR555(); |
| extern void _ConvertX86p16_8RGB332(); |
| 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"))); |
| #define Hermes_X86_CPU _Hermes_X86_CPU |
| #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"))); |
| #define ConvertX86 _ConvertX86 |
| void ConvertX86pI8_32() __attribute__ ((alias ("_ConvertX86pI8_32"))); |
| void ConvertX86pI8_24() __attribute__ ((alias ("_ConvertX86pI8_24"))); |
| void ConvertX86pI8_16() __attribute__ ((alias ("_ConvertX86pI8_16"))); |
| #define ConvertX86p32_32BGR888 _ConvertX86p32_32BGR888 |
| #define ConvertX86p32_32RGBA888 _ConvertX86p32_32RGBA888 |
| #define ConvertX86p32_32BGRA888 _ConvertX86p32_32BGRA888 |
| #define ConvertX86p32_24RGB888 _ConvertX86p32_24RGB888 |
| #define ConvertX86p32_24BGR888 _ConvertX86p32_24BGR888 |
| #define ConvertX86p32_16RGB565 _ConvertX86p32_16RGB565 |
| #define ConvertX86p32_16BGR565 _ConvertX86p32_16BGR565 |
| #define ConvertX86p32_16RGB555 _ConvertX86p32_16RGB555 |
| #define ConvertX86p32_16BGR555 _ConvertX86p32_16BGR555 |
| #define ConvertX86p32_8RGB332 _ConvertX86p32_8RGB332 |
| extern int32 ConvertX86p16_32RGB888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32RGB888_LUT_X86"))); |
| extern int32 ConvertX86p16_32BGR888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32BGR888_LUT_X86"))); |
| extern int32 ConvertX86p16_32RGBA888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32RGBA888_LUT_X86"))); |
| extern int32 ConvertX86p16_32BGRA888_LUT_X86[512] __attribute__ ((alias ("_ConvertX86p16_32BGRA888_LUT_X86"))); |
| #endif |
| #define ConvertX86p16_16BGR565 _ConvertX86p16_16BGR565 |
| #define ConvertX86p16_16RGB555 _ConvertX86p16_16RGB555 |
| #define ConvertX86p16_16BGR555 _ConvertX86p16_16BGR555 |
| #define ConvertX86p16_8RGB332 _ConvertX86p16_8RGB332 |
| #ifdef __cplusplus |
| } |
| #endif |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/mmx_main.asm |
|---|
| 9,9 → 9,9 |
| BITS 32 |
| %include "common.inc" |
| GLOBAL _ConvertMMX |
| GLOBAL _mmxreturn |
| SDL_FUNC _ConvertMMX |
| SECTION .text |
| 50,8 → 50,9 |
| y_loop: |
| mov ecx,[ebp+4] |
| call [ebp+32] |
| jmp [ebp+32] |
| _mmxreturn: |
| add esi,[ebp+12] |
| add edi,[ebp+28] |
| 69,6 → 70,5 |
| ret |
| %ifidn __OUTPUT_FORMAT__,elf32 |
| section .note.GNU-stack noalloc noexec nowrite progbits |
| %endif |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/mmxp2_32.asm |
|---|
| 20,51 → 20,40 |
| BITS 32 |
| %include "common.inc" |
| SDL_FUNC _ConvertMMXpII32_24RGB888 |
| SDL_FUNC _ConvertMMXpII32_16RGB565 |
| SDL_FUNC _ConvertMMXpII32_16BGR565 |
| SDL_FUNC _ConvertMMXpII32_16RGB555 |
| SDL_FUNC _ConvertMMXpII32_16BGR555 |
| GLOBAL _ConvertMMXpII32_24RGB888 |
| GLOBAL _ConvertMMXpII32_16RGB565 |
| GLOBAL _ConvertMMXpII32_16BGR565 |
| GLOBAL _ConvertMMXpII32_16RGB555 |
| GLOBAL _ConvertMMXpII32_16BGR555 |
| ;; Macros for conversion routines |
| EXTERN _mmxreturn |
| %macro _push_immq_mask 1 |
| push dword %1 |
| push dword %1 |
| %endmacro |
| SECTION .data |
| %macro load_immq 2 |
| _push_immq_mask %2 |
| movq %1, [esp] |
| %endmacro |
| ALIGN 8 |
| %macro pand_immq 2 |
| _push_immq_mask %2 |
| pand %1, [esp] |
| %endmacro |
| ;; Constants for conversion routines |
| %define CLEANUP_IMMQ_LOADS(num) \ |
| add esp, byte 8 * num |
| mmx32_rgb888_mask dd 00ffffffh,00ffffffh |
| %define mmx32_rgb888_mask 00ffffffh |
| %define mmx32_rgb565_b 000000f8h |
| %define mmx32_rgb565_g 0000fc00h |
| %define mmx32_rgb565_r 00f80000h |
| mmx32_rgb565_b dd 000000f8h, 000000f8h |
| mmx32_rgb565_g dd 0000fc00h, 0000fc00h |
| mmx32_rgb565_r dd 00f80000h, 00f80000h |
| %define mmx32_rgb555_rb 00f800f8h |
| %define mmx32_rgb555_g 0000f800h |
| %define mmx32_rgb555_mul 20000008h |
| %define mmx32_bgr555_mul 00082000h |
| mmx32_rgb555_rb dd 00f800f8h,00f800f8h |
| mmx32_rgb555_g dd 0000f800h,0000f800h |
| mmx32_rgb555_mul dd 20000008h,20000008h |
| mmx32_bgr555_mul dd 00082000h,00082000h |
| SECTION .text |
| _ConvertMMXpII32_24RGB888: |
| ; set up mm6 as the mask, mm7 as zero |
| load_immq mm6, mmx32_rgb888_mask |
| CLEANUP_IMMQ_LOADS(1) |
| movq mm6, qword [mmx32_rgb888_mask] |
| pxor mm7, mm7 |
| mov edx, ecx ; save ecx |
| 119,7 → 108,7 |
| dec ecx |
| jnz .L3 |
| .L4: |
| retn |
| jmp _mmxreturn |
| 126,10 → 115,9 |
| _ConvertMMXpII32_16RGB565: |
| ; set up masks |
| load_immq mm5, mmx32_rgb565_b |
| load_immq mm6, mmx32_rgb565_g |
| load_immq mm7, mmx32_rgb565_r |
| CLEANUP_IMMQ_LOADS(3) |
| movq mm5, [mmx32_rgb565_b] |
| movq mm6, [mmx32_rgb565_g] |
| movq mm7, [mmx32_rgb565_r] |
| mov edx, ecx |
| shr ecx, 2 |
| 188,15 → 176,14 |
| jnz .L3 |
| .L4: |
| retn |
| jmp _mmxreturn |
| _ConvertMMXpII32_16BGR565: |
| load_immq mm5, mmx32_rgb565_r |
| load_immq mm6, mmx32_rgb565_g |
| load_immq mm7, mmx32_rgb565_b |
| CLEANUP_IMMQ_LOADS(3) |
| movq mm5, [mmx32_rgb565_r] |
| movq mm6, [mmx32_rgb565_g] |
| movq mm7, [mmx32_rgb565_b] |
| mov edx, ecx |
| shr ecx, 2 |
| 258,7 → 245,7 |
| jnz .L3 |
| .L4: |
| retn |
| jmp _mmxreturn |
| _ConvertMMXpII32_16BGR555: |
| 266,7 → 253,7 |
| ; except it uses a different multiplier for the pmaddwd |
| ; instruction. cool huh. |
| load_immq mm7, mmx32_bgr555_mul |
| movq mm7, qword [mmx32_bgr555_mul] |
| jmp _convert_bgr555_cheat |
| ; This is the same as the Intel version.. they obviously went to |
| 276,16 → 263,15 |
| ; (I think) a more accurate name.. |
| _ConvertMMXpII32_16RGB555: |
| load_immq mm7, mmx32_rgb555_mul |
| movq mm7,qword [mmx32_rgb555_mul] |
| _convert_bgr555_cheat: |
| load_immq mm6, mmx32_rgb555_g |
| CLEANUP_IMMQ_LOADS(2) |
| movq mm6,qword [mmx32_rgb555_g] |
| mov edx,ecx ; Save ecx |
| and ecx,DWORD 0fffffff8h ; clear lower three bits |
| and ecx,BYTE 0fffffff8h ; clear lower three bits |
| jnz .L_OK |
| jmp near .L2 |
| jmp .L2 |
| .L_OK: |
| 294,14 → 280,12 |
| movq mm0,[esi] |
| movq mm3,mm2 |
| pand_immq mm3, mmx32_rgb555_rb |
| pand mm3,qword [mmx32_rgb555_rb] |
| movq mm1,mm0 |
| pand_immq mm1, mmx32_rgb555_rb |
| pand mm1,qword [mmx32_rgb555_rb] |
| pmaddwd mm3,mm7 |
| CLEANUP_IMMQ_LOADS(2) |
| pmaddwd mm1,mm7 |
| pand mm2,mm6 |
| 318,13 → 302,13 |
| movq mm0,mm4 |
| psrld mm1,6 |
| pand_immq mm0, mmx32_rgb555_rb |
| pand mm0,qword [mmx32_rgb555_rb] |
| packssdw mm1,mm3 |
| movq mm3,mm5 |
| pmaddwd mm0,mm7 |
| pand_immq mm3, mmx32_rgb555_rb |
| pand mm3,qword [mmx32_rgb555_rb] |
| pand mm4,mm6 |
| movq [edi],mm1 |
| 345,14 → 329,12 |
| movq mm3,mm2 |
| movq mm1,mm0 |
| pand_immq mm3, mmx32_rgb555_rb |
| pand mm3,qword [mmx32_rgb555_rb] |
| packssdw mm5,mm4 |
| pand_immq mm1, mmx32_rgb555_rb |
| pand mm1,qword [mmx32_rgb555_rb] |
| pand mm2,mm6 |
| CLEANUP_IMMQ_LOADS(4) |
| movq [edi+8],mm5 |
| pmaddwd mm3,mm7 |
| 398,8 → 380,7 |
| jnz .L3 |
| .L4: |
| retn |
| jmp _mmxreturn |
| %ifidn __OUTPUT_FORMAT__,elf32 |
| section .note.GNU-stack noalloc noexec nowrite progbits |
| %endif |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/x86_main.asm |
|---|
| 11,14 → 11,17 |
| BITS 32 |
| %include "common.inc" |
| GLOBAL _ConvertX86 |
| GLOBAL _x86return |
| SDL_FUNC _ConvertX86 |
| SDL_FUNC _Hermes_X86_CPU |
| GLOBAL _Hermes_X86_CPU |
| SECTION .data |
| cpu_flags dd 0 |
| SECTION .text |
| ;; _ConvertX86: |
| 56,8 → 59,9 |
| y_loop: |
| mov ecx,[ebp+4] |
| call [ebp+32] |
| jmp [ebp+32] |
| _x86return: |
| add esi,[ebp+12] |
| add edi,[ebp+28] |
| 73,6 → 77,9 |
| ret |
| ;; Hermes_X86_CPU returns the CPUID flags in eax |
| _Hermes_X86_CPU: |
| pushfd |
| pop eax |
| 115,9 → 122,5 |
| mov eax,[cpu_flags] |
| .L1: |
| xor eax,eax |
| ret |
| %ifidn __OUTPUT_FORMAT__,elf32 |
| section .note.GNU-stack noalloc noexec nowrite progbits |
| %endif |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/x86p_16.asm |
|---|
| 10,19 → 10,28 |
| ; Used with permission. |
| ; |
| BITS 32 |
| %include "common.inc" |
| 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 |
| SDL_FUNC _ConvertX86p16_16BGR565 |
| SDL_FUNC _ConvertX86p16_16RGB555 |
| SDL_FUNC _ConvertX86p16_16BGR555 |
| SDL_FUNC _ConvertX86p16_8RGB332 |
| EXTERN _ConvertX86 |
| EXTERN _x86return |
| SECTION .text |
| _ConvertX86p16_16BGR565: |
| ; check short |
| 30,7 → 39,7 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov al,[esi] |
| mov ah,[esi+1] |
| mov ebx,eax |
| 47,10 → 56,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov eax,edi |
| and eax,BYTE 11b |
| jz .L4 |
| 70,7 → 79,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save count |
| .L4 ; save count |
| push ecx |
| ; unroll twice |
| 84,8 → 93,8 |
| neg ecx |
| jmp SHORT .L6 |
| .L5: mov [edi+ecx*4-4],eax |
| .L6: mov eax,[esi+ecx*4] |
| .L5 mov [edi+ecx*4-4],eax |
| .L6 mov eax,[esi+ecx*4] |
| mov ebx,[esi+ecx*4] |
| and eax,07E007E0h |
| 125,8 → 134,8 |
| add esi,BYTE 2 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| .L7 |
| jmp _x86return |
| 140,7 → 149,7 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov al,[esi] |
| mov ah,[esi+1] |
| mov ebx,eax |
| 154,10 → 163,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov eax,edi |
| and eax,BYTE 11b |
| jz .L4 |
| 174,7 → 183,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save ebp |
| .L4 ; save ebp |
| push ebp |
| ; save count |
| 191,7 → 200,7 |
| xor ebp,ebp |
| sub ebp,ecx |
| .L5: mov eax,[esi+ebp*8] ; agi? |
| .L5 mov eax,[esi+ebp*8] ; agi? |
| mov ecx,[esi+ebp*8+4] |
| mov ebx,eax |
| 217,7 → 226,7 |
| ; tail |
| pop ecx |
| .L6: and ecx,BYTE 11b |
| .L6 and ecx,BYTE 11b |
| jz .L7 |
| mov al,[esi] |
| mov ah,[esi+1] |
| 233,8 → 242,8 |
| dec ecx |
| jmp SHORT .L6 |
| .L7: pop ebp |
| retn |
| .L7 pop ebp |
| jmp _x86return |
| 248,7 → 257,7 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov al,[esi] |
| mov ah,[esi+1] |
| mov ebx,eax |
| 267,10 → 276,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov eax,edi |
| and eax,BYTE 11b |
| jz .L4 |
| 292,7 → 301,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save count |
| .L4 ; save count |
| push ecx |
| ; unroll twice |
| 306,8 → 315,8 |
| neg ecx |
| jmp SHORT .L6 |
| .L5: mov [edi+ecx*4-4],eax |
| .L6: mov eax,[esi+ecx*4] |
| .L5 mov [edi+ecx*4-4],eax |
| .L6 mov eax,[esi+ecx*4] |
| shr eax,1 |
| mov ebx,[esi+ecx*4] |
| 351,8 → 360,8 |
| add esi,BYTE 2 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| .L7 |
| jmp _x86return |
| 366,7 → 375,7 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov al,[esi+0] |
| mov ah,[esi+1] |
| mov ebx,eax |
| 384,10 → 393,10 |
| inc edi |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: mov eax,edi |
| .L3 mov eax,edi |
| and eax,BYTE 11b |
| jz .L4 |
| mov al,[esi+0] |
| 408,7 → 417,7 |
| dec ecx |
| jmp SHORT .L3 |
| .L4: ; save ebp |
| .L4 ; save ebp |
| push ebp |
| ; save count |
| 422,7 → 431,7 |
| mov bl,[esi+1] |
| mov dh,[esi+2] |
| .L5: shl edx,16 |
| .L5 shl edx,16 |
| mov bh,[esi+3] |
| shl ebx,16 |
| 463,7 → 472,7 |
| and ecx,BYTE 11b |
| jz .L7 |
| .L6: ; tail |
| .L6 ; tail |
| mov al,[esi+0] |
| mov ah,[esi+1] |
| mov ebx,eax |
| 482,9 → 491,6 |
| dec ecx |
| jnz .L6 |
| .L7: pop ebp |
| retn |
| .L7 pop ebp |
| jmp _x86return |
| %ifidn __OUTPUT_FORMAT__,elf32 |
| section .note.GNU-stack noalloc noexec nowrite progbits |
| %endif |
| /contrib/sdk/sources/SDL-1.2.2_newlib/src/hermes/x86p_32.asm |
|---|
| 9,23 → 9,25 |
| ; Most routines are (c) Glenn Fiedler (ptc@gaffer.org), used with permission |
| ; |
| BITS 32 |
| %include "common.inc" |
| 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 |
| SDL_FUNC _ConvertX86p32_32BGR888 |
| SDL_FUNC _ConvertX86p32_32RGBA888 |
| SDL_FUNC _ConvertX86p32_32BGRA888 |
| SDL_FUNC _ConvertX86p32_24RGB888 |
| SDL_FUNC _ConvertX86p32_24BGR888 |
| SDL_FUNC _ConvertX86p32_16RGB565 |
| SDL_FUNC _ConvertX86p32_16BGR565 |
| SDL_FUNC _ConvertX86p32_16RGB555 |
| SDL_FUNC _ConvertX86p32_16BGR555 |
| SDL_FUNC _ConvertX86p32_8RGB332 |
| EXTERN _x86return |
| SECTION .text |
| ;; _Convert_* |
| ;; Paramters: |
| ;; ESI = source |
| 41,7 → 43,7 |
| cmp ecx,BYTE 32 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov edx,[esi] |
| bswap edx |
| ror edx,8 |
| 50,10 → 52,10 |
| add edi,BYTE 4 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; save ebp |
| .L3 ; save ebp |
| push ebp |
| ; unroll four times |
| 63,7 → 65,7 |
| ; save count |
| push ecx |
| .L4: mov eax,[esi] |
| .L4 mov eax,[esi] |
| mov ebx,[esi+4] |
| bswap eax |
| 100,7 → 102,7 |
| and ecx,BYTE 11b |
| jz .L6 |
| .L5: ; tail loop |
| .L5 ; tail loop |
| mov edx,[esi] |
| bswap edx |
| ror edx,8 |
| 110,8 → 112,8 |
| dec ecx |
| jnz .L5 |
| .L6: pop ebp |
| retn |
| .L6 pop ebp |
| jmp _x86return |
| 122,7 → 124,7 |
| cmp ecx,BYTE 32 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov edx,[esi] |
| rol edx,8 |
| mov [edi],edx |
| 130,10 → 132,10 |
| add edi,BYTE 4 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; save ebp |
| .L3 ; save ebp |
| push ebp |
| ; unroll four times |
| 143,7 → 145,7 |
| ; save count |
| push ecx |
| .L4: mov eax,[esi] |
| .L4 mov eax,[esi] |
| mov ebx,[esi+4] |
| rol eax,8 |
| 172,7 → 174,7 |
| and ecx,BYTE 11b |
| jz .L6 |
| .L5: ; tail loop |
| .L5 ; tail loop |
| mov edx,[esi] |
| rol edx,8 |
| mov [edi],edx |
| 181,8 → 183,8 |
| dec ecx |
| jnz .L5 |
| .L6: pop ebp |
| retn |
| .L6 pop ebp |
| jmp _x86return |
| 193,7 → 195,7 |
| cmp ecx,BYTE 32 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov edx,[esi] |
| bswap edx |
| mov [edi],edx |
| 201,10 → 203,10 |
| add edi,BYTE 4 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; save ebp |
| .L3 ; save ebp |
| push ebp |
| ; unroll four times |
| 214,7 → 216,7 |
| ; save count |
| push ecx |
| .L4: mov eax,[esi] |
| .L4 mov eax,[esi] |
| mov ebx,[esi+4] |
| mov ecx,[esi+8] |
| 245,7 → 247,7 |
| and ecx,BYTE 11b |
| jz .L6 |
| .L5: ; tail loop |
| .L5 ; tail loop |
| mov edx,[esi] |
| bswap edx |
| mov [edi],edx |
| 254,8 → 256,8 |
| dec ecx |
| jnz .L5 |
| .L6: pop ebp |
| retn |
| .L6 pop ebp |
| jmp _x86return |
| 268,7 → 270,7 |
| cmp ecx,BYTE 32 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov al,[esi] |
| mov bl,[esi+1] |
| mov dl,[esi+2] |
| 279,10 → 281,10 |
| add edi,BYTE 3 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov edx,edi |
| and edx,BYTE 11b |
| jz .L4 |
| 297,7 → 299,7 |
| dec ecx |
| jmp SHORT .L3 |
| .L4: ; unroll 4 times |
| .L4 ; unroll 4 times |
| push ebp |
| mov ebp,ecx |
| shr ebp,2 |
| 305,7 → 307,7 |
| ; save count |
| push ecx |
| .L5: mov eax,[esi] ; first dword eax = [A][R][G][B] |
| .L5 mov eax,[esi] ; first dword eax = [A][R][G][B] |
| mov ebx,[esi+4] ; second dword ebx = [a][r][g][b] |
| shl eax,8 ; eax = [R][G][B][.] |
| 339,7 → 341,7 |
| and ecx,BYTE 11b |
| jz .L7 |
| .L6: ; tail loop |
| .L6 ; tail loop |
| mov al,[esi] |
| mov bl,[esi+1] |
| mov dl,[esi+2] |
| 351,8 → 353,8 |
| dec ecx |
| jnz .L6 |
| .L7: pop ebp |
| retn |
| .L7 pop ebp |
| jmp _x86return |
| 365,7 → 367,8 |
| cmp ecx,BYTE 32 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov dl,[esi] |
| mov bl,[esi+1] |
| mov al,[esi+2] |
| 376,10 → 379,10 |
| add edi,BYTE 3 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov edx,edi |
| and edx,BYTE 11b |
| jz .L4 |
| 394,7 → 397,7 |
| dec ecx |
| jmp SHORT .L3 |
| .L4: ; unroll 4 times |
| .L4 ; unroll 4 times |
| push ebp |
| mov ebp,ecx |
| shr ebp,2 |
| 402,7 → 405,7 |
| ; save count |
| push ecx |
| .L5: |
| .L5 |
| mov eax,[esi] ; first dword eax = [A][R][G][B] |
| mov ebx,[esi+4] ; second dword ebx = [a][r][g][b] |
| 438,7 → 441,7 |
| and ecx,BYTE 11b |
| jz .L7 |
| .L6: ; tail loop |
| .L6 ; tail loop |
| mov dl,[esi] |
| mov bl,[esi+1] |
| mov al,[esi+2] |
| 450,9 → 453,9 |
| dec ecx |
| jnz .L6 |
| .L7: |
| .L7 |
| pop ebp |
| retn |
| jmp _x86return |
| 464,7 → 467,7 |
| cmp ecx,BYTE 16 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov bl,[esi+0] ; blue |
| mov al,[esi+1] ; green |
| mov ah,[esi+2] ; red |
| 481,10 → 484,10 |
| jnz .L1 |
| .L2: ; End of short loop |
| retn |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov ebx,edi |
| and ebx,BYTE 11b |
| jz .L4 |
| 567,7 → 570,7 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| jmp _x86return |
| 580,7 → 583,7 |
| cmp ecx,BYTE 16 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov ah,[esi+0] ; blue |
| mov al,[esi+1] ; green |
| mov bl,[esi+2] ; red |
| 595,10 → 598,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov ebx,edi |
| and ebx,BYTE 11b |
| jz .L4 |
| 616,7 → 619,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save count |
| .L4 ; save count |
| push ecx |
| ; unroll twice |
| 630,9 → 633,9 |
| neg ecx |
| jmp SHORT .L6 |
| .L5: |
| .L5 |
| mov [edi+ecx*4-4],eax |
| .L6: |
| .L6 |
| mov edx,[esi+ecx*8+4] |
| mov bh,[esi+ecx*8+4] |
| 680,8 → 683,8 |
| add esi,BYTE 4 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| .L7 |
| jmp _x86return |
| 694,7 → 697,7 |
| cmp ecx,BYTE 16 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov bl,[esi+0] ; blue |
| mov al,[esi+1] ; green |
| mov ah,[esi+2] ; red |
| 709,10 → 712,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov ebx,edi |
| and ebx,BYTE 11b |
| jz .L4 |
| 730,7 → 733,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save count |
| .L4 ; save count |
| push ecx |
| ; unroll twice |
| 744,9 → 747,9 |
| neg ecx |
| jmp SHORT .L6 |
| .L5: |
| .L5 |
| mov [edi+ecx*4-4],eax |
| .L6: |
| .L6 |
| mov eax,[esi+ecx*8] |
| shr ah,3 |
| 791,8 → 794,8 |
| add esi,BYTE 4 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| .L7 |
| jmp _x86return |
| 806,7 → 809,7 |
| ja .L3 |
| .L1: ; short loop |
| .L1 ; short loop |
| mov ah,[esi+0] ; blue |
| mov al,[esi+1] ; green |
| mov bl,[esi+2] ; red |
| 821,10 → 824,10 |
| add edi,BYTE 2 |
| dec ecx |
| jnz .L1 |
| .L2: |
| retn |
| .L2 |
| jmp _x86return |
| .L3: ; head |
| .L3 ; head |
| mov ebx,edi |
| and ebx,BYTE 11b |
| jz .L4 |
| 842,7 → 845,7 |
| add edi,BYTE 2 |
| dec ecx |
| .L4: ; save count |
| .L4 ; save count |
| push ecx |
| ; unroll twice |
| 856,9 → 859,9 |
| neg ecx |
| jmp SHORT .L6 |
| .L5: |
| .L5 |
| mov [edi+ecx*4-4],eax |
| .L6: |
| .L6 |
| mov edx,[esi+ecx*8+4] |
| mov bh,[esi+ecx*8+4] |
| 906,8 → 909,8 |
| add esi,BYTE 4 |
| add edi,BYTE 2 |
| .L7: |
| retn |
| .L7 |
| jmp _x86return |
| 919,7 → 922,7 |
| _ConvertX86p32_8RGB332: |
| .L_ALIGNED: |
| .L_ALIGNED |
| push ecx |
| shr ecx,2 ; We will draw 4 pixels at once |
| 1037,8 → 1040,4 |
| jnz .L3 |
| .L4: |
| retn |
| %ifidn __OUTPUT_FORMAT__,elf32 |
| section .note.GNU-stack noalloc noexec nowrite progbits |
| %endif |
| jmp _x86return |