michael@0: ;
michael@0: ;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
michael@0: ;
michael@0: ;  Use of this source code is governed by a BSD-style license
michael@0: ;  that can be found in the LICENSE file in the root of the source
michael@0: ;  tree. An additional intellectual property rights grant can be found
michael@0: ;  in the file PATENTS.  All contributing project authors may
michael@0: ;  be found in the AUTHORS file in the root of the source tree.
michael@0: ;
michael@0: 
michael@0: 
michael@0: %include "vpx_config.asm"
michael@0: 
michael@0: ; 32/64 bit compatibility macros
michael@0: ;
michael@0: ; In general, we make the source use 64 bit syntax, then twiddle with it using
michael@0: ; the preprocessor to get the 32 bit syntax on 32 bit platforms.
michael@0: ;
michael@0: %ifidn __OUTPUT_FORMAT__,elf32
michael@0: %define ABI_IS_32BIT 1
michael@0: %elifidn __OUTPUT_FORMAT__,macho32
michael@0: %define ABI_IS_32BIT 1
michael@0: %elifidn __OUTPUT_FORMAT__,win32
michael@0: %define ABI_IS_32BIT 1
michael@0: %elifidn __OUTPUT_FORMAT__,aout
michael@0: %define ABI_IS_32BIT 1
michael@0: %else
michael@0: %define ABI_IS_32BIT 0
michael@0: %endif
michael@0: 
michael@0: %if ABI_IS_32BIT
michael@0: %define rax eax
michael@0: %define rbx ebx
michael@0: %define rcx ecx
michael@0: %define rdx edx
michael@0: %define rsi esi
michael@0: %define rdi edi
michael@0: %define rsp esp
michael@0: %define rbp ebp
michael@0: %define movsxd mov
michael@0: %macro movq 2
michael@0:   %ifidn %1,eax
michael@0:     movd %1,%2
michael@0:   %elifidn %2,eax
michael@0:     movd %1,%2
michael@0:   %elifidn %1,ebx
michael@0:     movd %1,%2
michael@0:   %elifidn %2,ebx
michael@0:     movd %1,%2
michael@0:   %elifidn %1,ecx
michael@0:     movd %1,%2
michael@0:   %elifidn %2,ecx
michael@0:     movd %1,%2
michael@0:   %elifidn %1,edx
michael@0:     movd %1,%2
michael@0:   %elifidn %2,edx
michael@0:     movd %1,%2
michael@0:   %elifidn %1,esi
michael@0:     movd %1,%2
michael@0:   %elifidn %2,esi
michael@0:     movd %1,%2
michael@0:   %elifidn %1,edi
michael@0:     movd %1,%2
michael@0:   %elifidn %2,edi
michael@0:     movd %1,%2
michael@0:   %elifidn %1,esp
michael@0:     movd %1,%2
michael@0:   %elifidn %2,esp
michael@0:     movd %1,%2
michael@0:   %elifidn %1,ebp
michael@0:     movd %1,%2
michael@0:   %elifidn %2,ebp
michael@0:     movd %1,%2
michael@0:   %else
michael@0:     movq %1,%2
michael@0:   %endif
michael@0: %endmacro
michael@0: %endif
michael@0: 
michael@0: 
michael@0: ; LIBVPX_YASM_WIN64
michael@0: ; Set LIBVPX_YASM_WIN64 if output is Windows 64bit so the code will work if x64
michael@0: ; or win64 is defined on the Yasm command line.
michael@0: %ifidn __OUTPUT_FORMAT__,win64
michael@0: %define LIBVPX_YASM_WIN64 1
michael@0: %elifidn __OUTPUT_FORMAT__,x64
michael@0: %define LIBVPX_YASM_WIN64 1
michael@0: %else
michael@0: %define LIBVPX_YASM_WIN64 0
michael@0: %endif
michael@0: 
michael@0: ; sym()
michael@0: ; Return the proper symbol name for the target ABI.
michael@0: ;
michael@0: ; Certain ABIs, notably MS COFF and Darwin MACH-O, require that symbols
michael@0: ; with C linkage be prefixed with an underscore.
michael@0: ;
michael@0: %ifidn   __OUTPUT_FORMAT__,elf32
michael@0: %define sym(x) x
michael@0: %elifidn __OUTPUT_FORMAT__,elf64
michael@0: %define sym(x) x
michael@0: %elifidn __OUTPUT_FORMAT__,elfx32
michael@0: %define sym(x) x
michael@0: %elif LIBVPX_YASM_WIN64
michael@0: %define sym(x) x
michael@0: %else
michael@0: %define sym(x) _ %+ x
michael@0: %endif
michael@0: 
michael@0: ;  PRIVATE
michael@0: ;  Macro for the attribute to hide a global symbol for the target ABI.
michael@0: ;  This is only active if CHROMIUM is defined.
michael@0: ;
michael@0: ;  Chromium doesn't like exported global symbols due to symbol clashing with
michael@0: ;  plugins among other things.
michael@0: ;
michael@0: ;  Requires Chromium's patched copy of yasm:
michael@0: ;    http://src.chromium.org/viewvc/chrome?view=rev&revision=73761
michael@0: ;    http://www.tortall.net/projects/yasm/ticket/236
michael@0: ;
michael@0: %ifdef CHROMIUM
michael@0:   %ifidn   __OUTPUT_FORMAT__,elf32
michael@0:     %define PRIVATE :hidden
michael@0:   %elifidn __OUTPUT_FORMAT__,elf64
michael@0:     %define PRIVATE :hidden
michael@0:   %elifidn __OUTPUT_FORMAT__,elfx32
michael@0:     %define PRIVATE :hidden
michael@0:   %elif LIBVPX_YASM_WIN64
michael@0:     %define PRIVATE
michael@0:   %else
michael@0:     %define PRIVATE :private_extern
michael@0:   %endif
michael@0: %else
michael@0:   %define PRIVATE
michael@0: %endif
michael@0: 
michael@0: ; arg()
michael@0: ; Return the address specification of the given argument
michael@0: ;
michael@0: %if ABI_IS_32BIT
michael@0:   %define arg(x) [ebp+8+4*x]
michael@0: %else
michael@0:   ; 64 bit ABI passes arguments in registers. This is a workaround to get up
michael@0:   ; and running quickly. Relies on SHADOW_ARGS_TO_STACK
michael@0:   %if LIBVPX_YASM_WIN64
michael@0:     %define arg(x) [rbp+16+8*x]
michael@0:   %else
michael@0:     %define arg(x) [rbp-8-8*x]
michael@0:   %endif
michael@0: %endif
michael@0: 
michael@0: ; REG_SZ_BYTES, REG_SZ_BITS
michael@0: ; Size of a register
michael@0: %if ABI_IS_32BIT
michael@0: %define REG_SZ_BYTES 4
michael@0: %define REG_SZ_BITS  32
michael@0: %else
michael@0: %define REG_SZ_BYTES 8
michael@0: %define REG_SZ_BITS  64
michael@0: %endif
michael@0: 
michael@0: 
michael@0: ; ALIGN_STACK <alignment> <register>
michael@0: ; This macro aligns the stack to the given alignment (in bytes). The stack
michael@0: ; is left such that the previous value of the stack pointer is the first
michael@0: ; argument on the stack (ie, the inverse of this macro is 'pop rsp.')
michael@0: ; This macro uses one temporary register, which is not preserved, and thus
michael@0: ; must be specified as an argument.
michael@0: %macro ALIGN_STACK 2
michael@0:     mov         %2, rsp
michael@0:     and         rsp, -%1
michael@0:     lea         rsp, [rsp - (%1 - REG_SZ_BYTES)]
michael@0:     push        %2
michael@0: %endmacro
michael@0: 
michael@0: 
michael@0: ;
michael@0: ; The Microsoft assembler tries to impose a certain amount of type safety in
michael@0: ; its register usage. YASM doesn't recognize these directives, so we just
michael@0: ; %define them away to maintain as much compatibility as possible with the
michael@0: ; original inline assembler we're porting from.
michael@0: ;
michael@0: %idefine PTR
michael@0: %idefine XMMWORD
michael@0: %idefine MMWORD
michael@0: 
michael@0: ; PIC macros
michael@0: ;
michael@0: %if ABI_IS_32BIT
michael@0:   %if CONFIG_PIC=1
michael@0:   %ifidn __OUTPUT_FORMAT__,elf32
michael@0:     %define GET_GOT_SAVE_ARG 1
michael@0:     %define WRT_PLT wrt ..plt
michael@0:     %macro GET_GOT 1
michael@0:       extern _GLOBAL_OFFSET_TABLE_
michael@0:       push %1
michael@0:       call %%get_got
michael@0:       %%sub_offset:
michael@0:       jmp %%exitGG
michael@0:       %%get_got:
michael@0:       mov %1, [esp]
michael@0:       add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
michael@0:       ret
michael@0:       %%exitGG:
michael@0:       %undef GLOBAL
michael@0:       %define GLOBAL(x) x + %1 wrt ..gotoff
michael@0:       %undef RESTORE_GOT
michael@0:       %define RESTORE_GOT pop %1
michael@0:     %endmacro
michael@0:   %elifidn __OUTPUT_FORMAT__,macho32
michael@0:     %define GET_GOT_SAVE_ARG 1
michael@0:     %macro GET_GOT 1
michael@0:       push %1
michael@0:       call %%get_got
michael@0:       %%get_got:
michael@0:       pop  %1
michael@0:       %undef GLOBAL
michael@0:       %define GLOBAL(x) x + %1 - %%get_got
michael@0:       %undef RESTORE_GOT
michael@0:       %define RESTORE_GOT pop %1
michael@0:     %endmacro
michael@0:   %endif
michael@0:   %endif
michael@0: 
michael@0:   %ifdef CHROMIUM
michael@0:     %ifidn __OUTPUT_FORMAT__,macho32
michael@0:       %define HIDDEN_DATA(x) x:private_extern
michael@0:     %else
michael@0:       %define HIDDEN_DATA(x) x
michael@0:     %endif
michael@0:   %else
michael@0:     %define HIDDEN_DATA(x) x
michael@0:   %endif
michael@0: %else
michael@0:   %macro GET_GOT 1
michael@0:   %endmacro
michael@0:   %define GLOBAL(x) rel x
michael@0:   %ifidn __OUTPUT_FORMAT__,elf64
michael@0:     %define WRT_PLT wrt ..plt
michael@0:     %define HIDDEN_DATA(x) x:data hidden
michael@0:   %elifidn __OUTPUT_FORMAT__,elfx32
michael@0:     %define WRT_PLT wrt ..plt
michael@0:     %define HIDDEN_DATA(x) x:data hidden
michael@0:   %elifidn __OUTPUT_FORMAT__,macho64
michael@0:     %ifdef CHROMIUM
michael@0:       %define HIDDEN_DATA(x) x:private_extern
michael@0:     %else
michael@0:       %define HIDDEN_DATA(x) x
michael@0:     %endif
michael@0:   %else
michael@0:     %define HIDDEN_DATA(x) x
michael@0:   %endif
michael@0: %endif
michael@0: %ifnmacro GET_GOT
michael@0:     %macro GET_GOT 1
michael@0:     %endmacro
michael@0:     %define GLOBAL(x) x
michael@0: %endif
michael@0: %ifndef RESTORE_GOT
michael@0: %define RESTORE_GOT
michael@0: %endif
michael@0: %ifndef WRT_PLT
michael@0: %define WRT_PLT
michael@0: %endif
michael@0: 
michael@0: %if ABI_IS_32BIT
michael@0:   %macro SHADOW_ARGS_TO_STACK 1
michael@0:   %endm
michael@0:   %define UNSHADOW_ARGS
michael@0: %else
michael@0: %if LIBVPX_YASM_WIN64
michael@0:   %macro SHADOW_ARGS_TO_STACK 1 ; argc
michael@0:     %if %1 > 0
michael@0:         mov arg(0),rcx
michael@0:     %endif
michael@0:     %if %1 > 1
michael@0:         mov arg(1),rdx
michael@0:     %endif
michael@0:     %if %1 > 2
michael@0:         mov arg(2),r8
michael@0:     %endif
michael@0:     %if %1 > 3
michael@0:         mov arg(3),r9
michael@0:     %endif
michael@0:   %endm
michael@0: %else
michael@0:   %macro SHADOW_ARGS_TO_STACK 1 ; argc
michael@0:     %if %1 > 0
michael@0:         push rdi
michael@0:     %endif
michael@0:     %if %1 > 1
michael@0:         push rsi
michael@0:     %endif
michael@0:     %if %1 > 2
michael@0:         push rdx
michael@0:     %endif
michael@0:     %if %1 > 3
michael@0:         push rcx
michael@0:     %endif
michael@0:     %if %1 > 4
michael@0:         push r8
michael@0:     %endif
michael@0:     %if %1 > 5
michael@0:         push r9
michael@0:     %endif
michael@0:     %if %1 > 6
michael@0:       %assign i %1-6
michael@0:       %assign off 16
michael@0:       %rep i
michael@0:         mov rax,[rbp+off]
michael@0:         push rax
michael@0:         %assign off off+8
michael@0:       %endrep
michael@0:     %endif
michael@0:   %endm
michael@0: %endif
michael@0:   %define UNSHADOW_ARGS mov rsp, rbp
michael@0: %endif
michael@0: 
michael@0: ; Win64 ABI requires that XMM6:XMM15 are callee saved
michael@0: ; SAVE_XMM n, [u]
michael@0: ; store registers 6-n on the stack
michael@0: ; if u is specified, use unaligned movs.
michael@0: ; Win64 ABI requires 16 byte stack alignment, but then pushes an 8 byte return
michael@0: ; value. Typically we follow this up with 'push rbp' - re-aligning the stack -
michael@0: ; but in some cases this is not done and unaligned movs must be used.
michael@0: %if LIBVPX_YASM_WIN64
michael@0: %macro SAVE_XMM 1-2 a
michael@0:   %if %1 < 6
michael@0:     %error Only xmm registers 6-15 must be preserved
michael@0:   %else
michael@0:     %assign last_xmm %1
michael@0:     %define movxmm movdq %+ %2
michael@0:     %assign xmm_stack_space ((last_xmm - 5) * 16)
michael@0:     sub rsp, xmm_stack_space
michael@0:     %assign i 6
michael@0:     %rep (last_xmm - 5)
michael@0:       movxmm [rsp + ((i - 6) * 16)], xmm %+ i
michael@0:       %assign i i+1
michael@0:     %endrep
michael@0:   %endif
michael@0: %endmacro
michael@0: %macro RESTORE_XMM 0
michael@0:   %ifndef last_xmm
michael@0:     %error RESTORE_XMM must be paired with SAVE_XMM n
michael@0:   %else
michael@0:     %assign i last_xmm
michael@0:     %rep (last_xmm - 5)
michael@0:       movxmm xmm %+ i, [rsp +((i - 6) * 16)]
michael@0:       %assign i i-1
michael@0:     %endrep
michael@0:     add rsp, xmm_stack_space
michael@0:     ; there are a couple functions which return from multiple places.
michael@0:     ; otherwise, we could uncomment these:
michael@0:     ; %undef last_xmm
michael@0:     ; %undef xmm_stack_space
michael@0:     ; %undef movxmm
michael@0:   %endif
michael@0: %endmacro
michael@0: %else
michael@0: %macro SAVE_XMM 1-2
michael@0: %endmacro
michael@0: %macro RESTORE_XMM 0
michael@0: %endmacro
michael@0: %endif
michael@0: 
michael@0: ; Name of the rodata section
michael@0: ;
michael@0: ; .rodata seems to be an elf-ism, as it doesn't work on OSX.
michael@0: ;
michael@0: %ifidn __OUTPUT_FORMAT__,macho64
michael@0: %define SECTION_RODATA section .text
michael@0: %elifidn __OUTPUT_FORMAT__,macho32
michael@0: %macro SECTION_RODATA 0
michael@0: section .text
michael@0: %endmacro
michael@0: %elifidn __OUTPUT_FORMAT__,aout
michael@0: %define SECTION_RODATA section .data
michael@0: %else
michael@0: %define SECTION_RODATA section .rodata
michael@0: %endif
michael@0: 
michael@0: 
michael@0: ; Tell GNU ld that we don't require an executable stack.
michael@0: %ifidn __OUTPUT_FORMAT__,elf32
michael@0: section .note.GNU-stack noalloc noexec nowrite progbits
michael@0: section .text
michael@0: %elifidn __OUTPUT_FORMAT__,elf64
michael@0: section .note.GNU-stack noalloc noexec nowrite progbits
michael@0: section .text
michael@0: %elifidn __OUTPUT_FORMAT__,elfx32
michael@0: section .note.GNU-stack noalloc noexec nowrite progbits
michael@0: section .text
michael@0: %endif
michael@0: