1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vpx_ports/x86_abi_support.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,395 @@
1.4 +;
1.5 +; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 +;
1.7 +; Use of this source code is governed by a BSD-style license
1.8 +; that can be found in the LICENSE file in the root of the source
1.9 +; tree. An additional intellectual property rights grant can be found
1.10 +; in the file PATENTS. All contributing project authors may
1.11 +; be found in the AUTHORS file in the root of the source tree.
1.12 +;
1.13 +
1.14 +
1.15 +%include "vpx_config.asm"
1.16 +
1.17 +; 32/64 bit compatibility macros
1.18 +;
1.19 +; In general, we make the source use 64 bit syntax, then twiddle with it using
1.20 +; the preprocessor to get the 32 bit syntax on 32 bit platforms.
1.21 +;
1.22 +%ifidn __OUTPUT_FORMAT__,elf32
1.23 +%define ABI_IS_32BIT 1
1.24 +%elifidn __OUTPUT_FORMAT__,macho32
1.25 +%define ABI_IS_32BIT 1
1.26 +%elifidn __OUTPUT_FORMAT__,win32
1.27 +%define ABI_IS_32BIT 1
1.28 +%elifidn __OUTPUT_FORMAT__,aout
1.29 +%define ABI_IS_32BIT 1
1.30 +%else
1.31 +%define ABI_IS_32BIT 0
1.32 +%endif
1.33 +
1.34 +%if ABI_IS_32BIT
1.35 +%define rax eax
1.36 +%define rbx ebx
1.37 +%define rcx ecx
1.38 +%define rdx edx
1.39 +%define rsi esi
1.40 +%define rdi edi
1.41 +%define rsp esp
1.42 +%define rbp ebp
1.43 +%define movsxd mov
1.44 +%macro movq 2
1.45 + %ifidn %1,eax
1.46 + movd %1,%2
1.47 + %elifidn %2,eax
1.48 + movd %1,%2
1.49 + %elifidn %1,ebx
1.50 + movd %1,%2
1.51 + %elifidn %2,ebx
1.52 + movd %1,%2
1.53 + %elifidn %1,ecx
1.54 + movd %1,%2
1.55 + %elifidn %2,ecx
1.56 + movd %1,%2
1.57 + %elifidn %1,edx
1.58 + movd %1,%2
1.59 + %elifidn %2,edx
1.60 + movd %1,%2
1.61 + %elifidn %1,esi
1.62 + movd %1,%2
1.63 + %elifidn %2,esi
1.64 + movd %1,%2
1.65 + %elifidn %1,edi
1.66 + movd %1,%2
1.67 + %elifidn %2,edi
1.68 + movd %1,%2
1.69 + %elifidn %1,esp
1.70 + movd %1,%2
1.71 + %elifidn %2,esp
1.72 + movd %1,%2
1.73 + %elifidn %1,ebp
1.74 + movd %1,%2
1.75 + %elifidn %2,ebp
1.76 + movd %1,%2
1.77 + %else
1.78 + movq %1,%2
1.79 + %endif
1.80 +%endmacro
1.81 +%endif
1.82 +
1.83 +
1.84 +; LIBVPX_YASM_WIN64
1.85 +; Set LIBVPX_YASM_WIN64 if output is Windows 64bit so the code will work if x64
1.86 +; or win64 is defined on the Yasm command line.
1.87 +%ifidn __OUTPUT_FORMAT__,win64
1.88 +%define LIBVPX_YASM_WIN64 1
1.89 +%elifidn __OUTPUT_FORMAT__,x64
1.90 +%define LIBVPX_YASM_WIN64 1
1.91 +%else
1.92 +%define LIBVPX_YASM_WIN64 0
1.93 +%endif
1.94 +
1.95 +; sym()
1.96 +; Return the proper symbol name for the target ABI.
1.97 +;
1.98 +; Certain ABIs, notably MS COFF and Darwin MACH-O, require that symbols
1.99 +; with C linkage be prefixed with an underscore.
1.100 +;
1.101 +%ifidn __OUTPUT_FORMAT__,elf32
1.102 +%define sym(x) x
1.103 +%elifidn __OUTPUT_FORMAT__,elf64
1.104 +%define sym(x) x
1.105 +%elifidn __OUTPUT_FORMAT__,elfx32
1.106 +%define sym(x) x
1.107 +%elif LIBVPX_YASM_WIN64
1.108 +%define sym(x) x
1.109 +%else
1.110 +%define sym(x) _ %+ x
1.111 +%endif
1.112 +
1.113 +; PRIVATE
1.114 +; Macro for the attribute to hide a global symbol for the target ABI.
1.115 +; This is only active if CHROMIUM is defined.
1.116 +;
1.117 +; Chromium doesn't like exported global symbols due to symbol clashing with
1.118 +; plugins among other things.
1.119 +;
1.120 +; Requires Chromium's patched copy of yasm:
1.121 +; http://src.chromium.org/viewvc/chrome?view=rev&revision=73761
1.122 +; http://www.tortall.net/projects/yasm/ticket/236
1.123 +;
1.124 +%ifdef CHROMIUM
1.125 + %ifidn __OUTPUT_FORMAT__,elf32
1.126 + %define PRIVATE :hidden
1.127 + %elifidn __OUTPUT_FORMAT__,elf64
1.128 + %define PRIVATE :hidden
1.129 + %elifidn __OUTPUT_FORMAT__,elfx32
1.130 + %define PRIVATE :hidden
1.131 + %elif LIBVPX_YASM_WIN64
1.132 + %define PRIVATE
1.133 + %else
1.134 + %define PRIVATE :private_extern
1.135 + %endif
1.136 +%else
1.137 + %define PRIVATE
1.138 +%endif
1.139 +
1.140 +; arg()
1.141 +; Return the address specification of the given argument
1.142 +;
1.143 +%if ABI_IS_32BIT
1.144 + %define arg(x) [ebp+8+4*x]
1.145 +%else
1.146 + ; 64 bit ABI passes arguments in registers. This is a workaround to get up
1.147 + ; and running quickly. Relies on SHADOW_ARGS_TO_STACK
1.148 + %if LIBVPX_YASM_WIN64
1.149 + %define arg(x) [rbp+16+8*x]
1.150 + %else
1.151 + %define arg(x) [rbp-8-8*x]
1.152 + %endif
1.153 +%endif
1.154 +
1.155 +; REG_SZ_BYTES, REG_SZ_BITS
1.156 +; Size of a register
1.157 +%if ABI_IS_32BIT
1.158 +%define REG_SZ_BYTES 4
1.159 +%define REG_SZ_BITS 32
1.160 +%else
1.161 +%define REG_SZ_BYTES 8
1.162 +%define REG_SZ_BITS 64
1.163 +%endif
1.164 +
1.165 +
1.166 +; ALIGN_STACK <alignment> <register>
1.167 +; This macro aligns the stack to the given alignment (in bytes). The stack
1.168 +; is left such that the previous value of the stack pointer is the first
1.169 +; argument on the stack (ie, the inverse of this macro is 'pop rsp.')
1.170 +; This macro uses one temporary register, which is not preserved, and thus
1.171 +; must be specified as an argument.
1.172 +%macro ALIGN_STACK 2
1.173 + mov %2, rsp
1.174 + and rsp, -%1
1.175 + lea rsp, [rsp - (%1 - REG_SZ_BYTES)]
1.176 + push %2
1.177 +%endmacro
1.178 +
1.179 +
1.180 +;
1.181 +; The Microsoft assembler tries to impose a certain amount of type safety in
1.182 +; its register usage. YASM doesn't recognize these directives, so we just
1.183 +; %define them away to maintain as much compatibility as possible with the
1.184 +; original inline assembler we're porting from.
1.185 +;
1.186 +%idefine PTR
1.187 +%idefine XMMWORD
1.188 +%idefine MMWORD
1.189 +
1.190 +; PIC macros
1.191 +;
1.192 +%if ABI_IS_32BIT
1.193 + %if CONFIG_PIC=1
1.194 + %ifidn __OUTPUT_FORMAT__,elf32
1.195 + %define GET_GOT_SAVE_ARG 1
1.196 + %define WRT_PLT wrt ..plt
1.197 + %macro GET_GOT 1
1.198 + extern _GLOBAL_OFFSET_TABLE_
1.199 + push %1
1.200 + call %%get_got
1.201 + %%sub_offset:
1.202 + jmp %%exitGG
1.203 + %%get_got:
1.204 + mov %1, [esp]
1.205 + add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
1.206 + ret
1.207 + %%exitGG:
1.208 + %undef GLOBAL
1.209 + %define GLOBAL(x) x + %1 wrt ..gotoff
1.210 + %undef RESTORE_GOT
1.211 + %define RESTORE_GOT pop %1
1.212 + %endmacro
1.213 + %elifidn __OUTPUT_FORMAT__,macho32
1.214 + %define GET_GOT_SAVE_ARG 1
1.215 + %macro GET_GOT 1
1.216 + push %1
1.217 + call %%get_got
1.218 + %%get_got:
1.219 + pop %1
1.220 + %undef GLOBAL
1.221 + %define GLOBAL(x) x + %1 - %%get_got
1.222 + %undef RESTORE_GOT
1.223 + %define RESTORE_GOT pop %1
1.224 + %endmacro
1.225 + %endif
1.226 + %endif
1.227 +
1.228 + %ifdef CHROMIUM
1.229 + %ifidn __OUTPUT_FORMAT__,macho32
1.230 + %define HIDDEN_DATA(x) x:private_extern
1.231 + %else
1.232 + %define HIDDEN_DATA(x) x
1.233 + %endif
1.234 + %else
1.235 + %define HIDDEN_DATA(x) x
1.236 + %endif
1.237 +%else
1.238 + %macro GET_GOT 1
1.239 + %endmacro
1.240 + %define GLOBAL(x) rel x
1.241 + %ifidn __OUTPUT_FORMAT__,elf64
1.242 + %define WRT_PLT wrt ..plt
1.243 + %define HIDDEN_DATA(x) x:data hidden
1.244 + %elifidn __OUTPUT_FORMAT__,elfx32
1.245 + %define WRT_PLT wrt ..plt
1.246 + %define HIDDEN_DATA(x) x:data hidden
1.247 + %elifidn __OUTPUT_FORMAT__,macho64
1.248 + %ifdef CHROMIUM
1.249 + %define HIDDEN_DATA(x) x:private_extern
1.250 + %else
1.251 + %define HIDDEN_DATA(x) x
1.252 + %endif
1.253 + %else
1.254 + %define HIDDEN_DATA(x) x
1.255 + %endif
1.256 +%endif
1.257 +%ifnmacro GET_GOT
1.258 + %macro GET_GOT 1
1.259 + %endmacro
1.260 + %define GLOBAL(x) x
1.261 +%endif
1.262 +%ifndef RESTORE_GOT
1.263 +%define RESTORE_GOT
1.264 +%endif
1.265 +%ifndef WRT_PLT
1.266 +%define WRT_PLT
1.267 +%endif
1.268 +
1.269 +%if ABI_IS_32BIT
1.270 + %macro SHADOW_ARGS_TO_STACK 1
1.271 + %endm
1.272 + %define UNSHADOW_ARGS
1.273 +%else
1.274 +%if LIBVPX_YASM_WIN64
1.275 + %macro SHADOW_ARGS_TO_STACK 1 ; argc
1.276 + %if %1 > 0
1.277 + mov arg(0),rcx
1.278 + %endif
1.279 + %if %1 > 1
1.280 + mov arg(1),rdx
1.281 + %endif
1.282 + %if %1 > 2
1.283 + mov arg(2),r8
1.284 + %endif
1.285 + %if %1 > 3
1.286 + mov arg(3),r9
1.287 + %endif
1.288 + %endm
1.289 +%else
1.290 + %macro SHADOW_ARGS_TO_STACK 1 ; argc
1.291 + %if %1 > 0
1.292 + push rdi
1.293 + %endif
1.294 + %if %1 > 1
1.295 + push rsi
1.296 + %endif
1.297 + %if %1 > 2
1.298 + push rdx
1.299 + %endif
1.300 + %if %1 > 3
1.301 + push rcx
1.302 + %endif
1.303 + %if %1 > 4
1.304 + push r8
1.305 + %endif
1.306 + %if %1 > 5
1.307 + push r9
1.308 + %endif
1.309 + %if %1 > 6
1.310 + %assign i %1-6
1.311 + %assign off 16
1.312 + %rep i
1.313 + mov rax,[rbp+off]
1.314 + push rax
1.315 + %assign off off+8
1.316 + %endrep
1.317 + %endif
1.318 + %endm
1.319 +%endif
1.320 + %define UNSHADOW_ARGS mov rsp, rbp
1.321 +%endif
1.322 +
1.323 +; Win64 ABI requires that XMM6:XMM15 are callee saved
1.324 +; SAVE_XMM n, [u]
1.325 +; store registers 6-n on the stack
1.326 +; if u is specified, use unaligned movs.
1.327 +; Win64 ABI requires 16 byte stack alignment, but then pushes an 8 byte return
1.328 +; value. Typically we follow this up with 'push rbp' - re-aligning the stack -
1.329 +; but in some cases this is not done and unaligned movs must be used.
1.330 +%if LIBVPX_YASM_WIN64
1.331 +%macro SAVE_XMM 1-2 a
1.332 + %if %1 < 6
1.333 + %error Only xmm registers 6-15 must be preserved
1.334 + %else
1.335 + %assign last_xmm %1
1.336 + %define movxmm movdq %+ %2
1.337 + %assign xmm_stack_space ((last_xmm - 5) * 16)
1.338 + sub rsp, xmm_stack_space
1.339 + %assign i 6
1.340 + %rep (last_xmm - 5)
1.341 + movxmm [rsp + ((i - 6) * 16)], xmm %+ i
1.342 + %assign i i+1
1.343 + %endrep
1.344 + %endif
1.345 +%endmacro
1.346 +%macro RESTORE_XMM 0
1.347 + %ifndef last_xmm
1.348 + %error RESTORE_XMM must be paired with SAVE_XMM n
1.349 + %else
1.350 + %assign i last_xmm
1.351 + %rep (last_xmm - 5)
1.352 + movxmm xmm %+ i, [rsp +((i - 6) * 16)]
1.353 + %assign i i-1
1.354 + %endrep
1.355 + add rsp, xmm_stack_space
1.356 + ; there are a couple functions which return from multiple places.
1.357 + ; otherwise, we could uncomment these:
1.358 + ; %undef last_xmm
1.359 + ; %undef xmm_stack_space
1.360 + ; %undef movxmm
1.361 + %endif
1.362 +%endmacro
1.363 +%else
1.364 +%macro SAVE_XMM 1-2
1.365 +%endmacro
1.366 +%macro RESTORE_XMM 0
1.367 +%endmacro
1.368 +%endif
1.369 +
1.370 +; Name of the rodata section
1.371 +;
1.372 +; .rodata seems to be an elf-ism, as it doesn't work on OSX.
1.373 +;
1.374 +%ifidn __OUTPUT_FORMAT__,macho64
1.375 +%define SECTION_RODATA section .text
1.376 +%elifidn __OUTPUT_FORMAT__,macho32
1.377 +%macro SECTION_RODATA 0
1.378 +section .text
1.379 +%endmacro
1.380 +%elifidn __OUTPUT_FORMAT__,aout
1.381 +%define SECTION_RODATA section .data
1.382 +%else
1.383 +%define SECTION_RODATA section .rodata
1.384 +%endif
1.385 +
1.386 +
1.387 +; Tell GNU ld that we don't require an executable stack.
1.388 +%ifidn __OUTPUT_FORMAT__,elf32
1.389 +section .note.GNU-stack noalloc noexec nowrite progbits
1.390 +section .text
1.391 +%elifidn __OUTPUT_FORMAT__,elf64
1.392 +section .note.GNU-stack noalloc noexec nowrite progbits
1.393 +section .text
1.394 +%elifidn __OUTPUT_FORMAT__,elfx32
1.395 +section .note.GNU-stack noalloc noexec nowrite progbits
1.396 +section .text
1.397 +%endif
1.398 +
