Commit 7d8e2b13 authored by Yaowu Xu's avatar Yaowu Xu

x86inc.asm: port changes from libvpx

Cherry-picked the following commits:
e56f9639 define GET_GOT_DEFINED if not previously defined
a4b47b89 x86inc.asm: only set visibility for chromium builds
a97c83f7 Only use .text sections for aout
9679be4b Use .text instead of .rodata on macho
7e065cd5 Copy PIC handling code from x86_abi_support
71ccd103 Set 'private_extern' visibility for macho targets
1be46ef6 Expand PIC default to macho64 and respect CONFIG_PIC from libvpx
4de9641f Use libvpx defines to set name mangling rules
2d6393a9 Customize x86inc.asm for libvpx
2497356d Update x86inc.asm from x264

Change-Id: I2fe309bf7b5cea866177cec954024ced59aed789
parent e8f0478c
URL: http://git.videolan.org/?p=x264.git
Version: a95584945dd9ce3acc66c6cd8f6796bc4404d40d
URL: https://git.videolan.org/git/x264.git
Version: d23d18655249944c1ca894b451e2c82c7a584c62
License: ISC
License File: LICENSE
......@@ -13,12 +13,8 @@ Prefix functions with aom by default.
Manage name mangling (prefixing with '_') manually because 'PREFIX' does not
exist in libaom.
Expand PIC default to macho64 and respect CONFIG_PIC from libaom
Catch all elf formats for 'hidden' status and SECTION notes.
Avoid 'amdnop' when building with nasm.
Set 'private_extern' visibility for macho targets.
Copy PIC 'GLOBAL' macros from x86_abi_support.asm
Use .text instead of .rodata on macho to avoid broken tables in PIC mode.
Use .text with no alignment for aout
Only use 'hidden' visibility with Chromium
Move '%use smartalign' for nasm out of 'INIT_CPUFLAGS' and before
'ALIGNMODE'.
;*****************************************************************************
;* x86inc.asm: x264asm abstraction layer
;*****************************************************************************
;* Copyright (C) 2005-2015 x264 project
;* Copyright (C) 2005-2016 x264 project
;*
;* Authors: Loren Merritt <lorenm@u.washington.edu>
;* Anton Mitrofanov <BugMaster@narod.ru>
......@@ -66,16 +66,35 @@
%endif
%endif
%ifidn __OUTPUT_FORMAT__,elf32
%define mangle(x) x
%define FORMAT_ELF 0
%ifidn __OUTPUT_FORMAT__,elf
%define FORMAT_ELF 1
%elifidn __OUTPUT_FORMAT__,elf32
%define FORMAT_ELF 1
%elifidn __OUTPUT_FORMAT__,elf64
%define mangle(x) x
%elifidn __OUTPUT_FORMAT__,x64
%define mangle(x) x
%elifidn __OUTPUT_FORMAT__,win64
%define mangle(x) x
%define FORMAT_ELF 1
%endif
%define FORMAT_MACHO 0
%ifidn __OUTPUT_FORMAT__,macho32
%define FORMAT_MACHO 1
%elifidn __OUTPUT_FORMAT__,macho64
%define FORMAT_MACHO 1
%endif
; Set PREFIX for libaom builds.
%if FORMAT_ELF
%undef PREFIX
%elif WIN64
%undef PREFIX
%else
%define PREFIX
%endif
%ifdef PREFIX
%define mangle(x) _ %+ x
%else
%define mangle(x) x
%endif
; In some instances macho32 tables get misaligned when using .rodata.
......@@ -94,14 +113,6 @@
%endif
%endmacro
%macro SECTION_TEXT 0-1 16
%ifidn __OUTPUT_FORMAT__,aout
SECTION .text
%else
SECTION .text align=%1
%endif
%endmacro
; PIC macros are copied from aom_ports/x86_abi_support.asm. The "define PIC"
; from original code is added in for 64bit.
%ifidn __OUTPUT_FORMAT__,elf32
......@@ -188,8 +199,16 @@
%ifdef PIC
default rel
%endif
%ifndef GET_GOT_DEFINED
%define GET_GOT_DEFINED 0
%endif
; Done with PIC macros
%ifdef __NASM_VER__
%use smartalign
%endif
; Macros to eliminate most code duplication between x86_32 and x86_64:
; Currently this works only for leaf functions which load all their arguments
; into registers at the start, and make no other use of the stack. Luckily that
......@@ -237,6 +256,7 @@
%define r%1w %2w
%define r%1b %2b
%define r%1h %2h
%define %2q %2
%if %0 == 2
%define r%1m %2d
%define r%1mp %2
......@@ -261,9 +281,9 @@
%define e%1h %3
%define r%1b %2
%define e%1b %2
%if ARCH_X86_64 == 0
%define r%1 e%1
%endif
%if ARCH_X86_64 == 0
%define r%1 e%1
%endif
%endmacro
DECLARE_REG_SIZE ax, al, ah
......@@ -373,7 +393,7 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
%macro ASSERT 1
%if (%1) == 0
%error assert failed
%error assertion ``%1'' failed
%endif
%endmacro
......@@ -464,8 +484,10 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
%if %1 != 0 && required_stack_alignment > STACK_ALIGNMENT
%if %1 > 0
%assign regs_used (regs_used + 1)
%elif ARCH_X86_64 && regs_used == num_args && num_args <= 4 + UNIX64 * 2
%warning "Stack pointer will overwrite register argument"
%endif
%if ARCH_X86_64 && regs_used < 5 + UNIX64 * 3
; Ensure that we don't clobber any registers containing arguments
%assign regs_used 5 + UNIX64 * 3
%endif
%endif
%endif
......@@ -579,9 +601,9 @@ DECLARE_REG 14, R15, 120
%macro RET 0
WIN64_RESTORE_XMM_INTERNAL rsp
POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
%if mmsize == 32
vzeroupper
%endif
%if mmsize == 32
vzeroupper
%endif
AUTO_REP_RET
%endmacro
......@@ -618,17 +640,17 @@ DECLARE_REG 14, R15, 72
%define has_epilogue regs_used > 9 || mmsize == 32 || stack_size > 0
%macro RET 0
%if stack_size_padded > 0
%if required_stack_alignment > STACK_ALIGNMENT
mov rsp, rstkm
%else
add rsp, stack_size_padded
%endif
%endif
%if stack_size_padded > 0
%if required_stack_alignment > STACK_ALIGNMENT
mov rsp, rstkm
%else
add rsp, stack_size_padded
%endif
%endif
POP_IF_USED 14, 13, 12, 11, 10, 9
%if mmsize == 32
vzeroupper
%endif
%if mmsize == 32
vzeroupper
%endif
AUTO_REP_RET
%endmacro
......@@ -674,29 +696,29 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
%define has_epilogue regs_used > 3 || mmsize == 32 || stack_size > 0
%macro RET 0
%if stack_size_padded > 0
%if required_stack_alignment > STACK_ALIGNMENT
mov rsp, rstkm
%else
add rsp, stack_size_padded
%endif
%endif
%if stack_size_padded > 0
%if required_stack_alignment > STACK_ALIGNMENT
mov rsp, rstkm
%else
add rsp, stack_size_padded
%endif
%endif
POP_IF_USED 6, 5, 4, 3
%if mmsize == 32
vzeroupper
%endif
%if mmsize == 32
vzeroupper
%endif
AUTO_REP_RET
%endmacro
%endif ;======================================================================
%if WIN64 == 0
%macro WIN64_SPILL_XMM 1
%endmacro
%macro WIN64_RESTORE_XMM 1
%endmacro
%macro WIN64_PUSH_XMM 0
%endmacro
%macro WIN64_SPILL_XMM 1
%endmacro
%macro WIN64_RESTORE_XMM 1
%endmacro
%macro WIN64_PUSH_XMM 0
%endmacro
%endif
; On AMD cpus <=K10, an ordinary ret is slow if it immediately follows either
......@@ -709,24 +731,26 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
%else
rep ret
%endif
annotate_function_size
%endmacro
%define last_branch_adr $$
%macro AUTO_REP_RET 0
%ifndef cpuflags
times ((last_branch_adr-$)>>31)+1 rep ; times 1 iff $ != last_branch_adr.
%elif notcpuflag(ssse3)
times ((last_branch_adr-$)>>31)+1 rep
%if notcpuflag(ssse3)
times ((last_branch_adr-$)>>31)+1 rep ; times 1 iff $ == last_branch_adr.
%endif
ret
annotate_function_size
%endmacro
%macro BRANCH_INSTR 0-*
%rep %0
%macro %1 1-2 %1
%2 %1
%%branch_instr:
%xdefine last_branch_adr %%branch_instr
%if notcpuflag(ssse3)
%%branch_instr equ $
%xdefine last_branch_adr %%branch_instr
%endif
%endmacro
%rotate 1
%endrep
......@@ -741,6 +765,7 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
%elif %2
jmp %1
%endif
annotate_function_size
%endmacro
;=============================================================================
......@@ -762,6 +787,7 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
cglobal_internal 0, %1 %+ SUFFIX, %2
%endmacro
%macro cglobal_internal 2-3+
annotate_function_size
%if %1
%xdefine %%FUNCTION_PREFIX private_prefix
; libaom explicitly sets visibility in shared object builds. Avoid
......@@ -782,17 +808,10 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
CAT_XDEFINE cglobaled_, %2, 1
%endif
%xdefine current_function %2
%ifidn __OUTPUT_FORMAT__,elf32
global %2:function %%VISIBILITY
%elifidn __OUTPUT_FORMAT__,elf64
%xdefine current_function_section __SECT__
%if FORMAT_ELF
global %2:function %%VISIBILITY
%elifidn __OUTPUT_FORMAT__,macho32
%ifdef __NASM_VER__
global %2
%else
global %2:private_extern
%endif
%elifidn __OUTPUT_FORMAT__,macho64
%elif FORMAT_MACHO
%ifdef __NASM_VER__
global %2
%else
......@@ -822,16 +841,16 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
; like cextern, but without the prefix
%macro cextern_naked 1
%xdefine %1 mangle(%1)
%ifdef PREFIX
%xdefine %1 mangle(%1)
%endif
CAT_XDEFINE cglobaled_, %1, 1
extern %1
%endmacro
%macro const 1-2+
%xdefine %1 mangle(private_prefix %+ _ %+ %1)
%ifidn __OUTPUT_FORMAT__,elf32
global %1:data hidden
%elifidn __OUTPUT_FORMAT__,elf64
%if FORMAT_ELF
global %1:data hidden
%else
global %1
......@@ -839,14 +858,29 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
%1: %2
%endmacro
; This is needed for ELF, otherwise the GNU linker assumes the stack is
; executable by default.
%ifidn __OUTPUT_FORMAT__,elf32
SECTION .note.GNU-stack noalloc noexec nowrite progbits
%elifidn __OUTPUT_FORMAT__,elf64
SECTION .note.GNU-stack noalloc noexec nowrite progbits
; This is needed for ELF, otherwise the GNU linker assumes the stack is executable by default.
%if FORMAT_ELF
[SECTION .note.GNU-stack noalloc noexec nowrite progbits]
%endif
; Tell debuggers how large the function was.
; This may be invoked multiple times per function; we rely on later instances overriding earlier ones.
; This is invoked by RET and similar macros, and also cglobal does it for the previous function,
; but if the last function in a source file doesn't use any of the standard macros for its epilogue,
; then its size might be unspecified.
%macro annotate_function_size 0
%ifdef __YASM_VER__
%ifdef current_function
%if FORMAT_ELF
current_function_section
%%ecf equ $
size current_function %%ecf - current_function
__SECT__
%endif
%endif
%endif
%endmacro
; cpuflags
%assign cpuflags_mmx (1<<0)
......@@ -875,12 +909,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%assign cpuflags_bmi1 (1<<22)|cpuflags_lzcnt
%assign cpuflags_bmi2 (1<<23)|cpuflags_bmi1
%define cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
%ifdef __NASM_VER__
%use smartalign
%endif
; Returns a boolean value expressing whether or not the specified cpuflag is enabled.
%define cpuflag(x) (((((cpuflags & (cpuflags_ %+ x)) ^ (cpuflags_ %+ x)) - 1) >> 31) & 1)
%define notcpuflag(x) (cpuflag(x) ^ 1)
; Takes an arbitrary number of cpuflags from the above list.
; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
......@@ -917,12 +948,18 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%endif
%endif
%ifdef __NASM_VER__
ALIGNMODE k7
%elif ARCH_X86_64 || cpuflag(sse2)
CPU amdnop
%if ARCH_X86_64 || cpuflag(sse2)
%ifdef __NASM_VER__
ALIGNMODE k8
%else
CPU amdnop
%endif
%else
CPU basicnop
%ifdef __NASM_VER__
ALIGNMODE nop
%else
CPU basicnop
%endif
%endif
%endmacro
......@@ -951,14 +988,14 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define movnta movntq
%assign %%i 0
%rep 8
CAT_XDEFINE m, %%i, mm %+ %%i
CAT_XDEFINE nnmm, %%i, %%i
%assign %%i %%i+1
CAT_XDEFINE m, %%i, mm %+ %%i
CAT_XDEFINE nnmm, %%i, %%i
%assign %%i %%i+1
%endrep
%rep 8
CAT_UNDEF m, %%i
CAT_UNDEF nnmm, %%i
%assign %%i %%i+1
CAT_UNDEF m, %%i
CAT_UNDEF nnmm, %%i
%assign %%i %%i+1
%endrep
INIT_CPUFLAGS %1
%endmacro
......@@ -969,7 +1006,7 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define mmsize 16
%define num_mmregs 8
%if ARCH_X86_64
%define num_mmregs 16
%define num_mmregs 16
%endif
%define mova movdqa
%define movu movdqu
......@@ -977,9 +1014,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define movnta movntdq
%assign %%i 0
%rep num_mmregs
CAT_XDEFINE m, %%i, xmm %+ %%i
CAT_XDEFINE nnxmm, %%i, %%i
%assign %%i %%i+1
CAT_XDEFINE m, %%i, xmm %+ %%i
CAT_XDEFINE nnxmm, %%i, %%i
%assign %%i %%i+1
%endrep
INIT_CPUFLAGS %1
%endmacro
......@@ -990,7 +1027,7 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define mmsize 32
%define num_mmregs 8
%if ARCH_X86_64
%define num_mmregs 16
%define num_mmregs 16
%endif
%define mova movdqa
%define movu movdqu
......@@ -998,9 +1035,9 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
%define movnta movntdq
%assign %%i 0
%rep num_mmregs
CAT_XDEFINE m, %%i, ymm %+ %%i
CAT_XDEFINE nnymm, %%i, %%i
%assign %%i %%i+1
CAT_XDEFINE m, %%i, ymm %+ %%i
CAT_XDEFINE nnymm, %%i, %%i
%assign %%i %%i+1
%endrep
INIT_CPUFLAGS %1
%endmacro
......@@ -1024,7 +1061,7 @@ INIT_XMM
%assign i 0
%rep 16
DECLARE_MMCAST i
%assign i i+1
%assign i i+1
%endrep
; I often want to use macros that permute their arguments. e.g. there's no
......@@ -1042,23 +1079,23 @@ INIT_XMM
; doesn't cost any cycles.
%macro PERMUTE 2-* ; takes a list of pairs to swap
%rep %0/2
%xdefine %%tmp%2 m%2
%rotate 2
%endrep
%rep %0/2
%xdefine m%1 %%tmp%2
CAT_XDEFINE nn, m%1, %1
%rotate 2
%endrep
%rep %0/2
%xdefine %%tmp%2 m%2
%rotate 2
%endrep
%rep %0/2
%xdefine m%1 %%tmp%2
CAT_XDEFINE nn, m%1, %1
%rotate 2
%endrep
%endmacro
%macro SWAP 2+ ; swaps a single chain (sometimes more concise than pairs)
%ifnum %1 ; SWAP 0, 1, ...
SWAP_INTERNAL_NUM %1, %2
%else ; SWAP m0, m1, ...
SWAP_INTERNAL_NAME %1, %2
%endif
%ifnum %1 ; SWAP 0, 1, ...
SWAP_INTERNAL_NUM %1, %2
%else ; SWAP m0, m1, ...
SWAP_INTERNAL_NAME %1, %2
%endif
%endmacro
%macro SWAP_INTERNAL_NUM 2-*
......@@ -1068,7 +1105,7 @@ INIT_XMM
%xdefine m%2 %%tmp
CAT_XDEFINE nn, m%1, %1
CAT_XDEFINE nn, m%2, %2
%rotate 1
%rotate 1
%endrep
%endmacro
......@@ -1076,7 +1113,7 @@ INIT_XMM
%xdefine %%args nn %+ %1
%rep %0-1
%xdefine %%args %%args, nn %+ %2
%rotate 1
%rotate 1
%endrep
SWAP_INTERNAL_NUM %%args
%endmacro
......@@ -1093,7 +1130,7 @@ INIT_XMM
%assign %%i 0
%rep num_mmregs
CAT_XDEFINE %%f, %%i, m %+ %%i
%assign %%i %%i+1
%assign %%i %%i+1
%endrep
%endmacro
......@@ -1103,20 +1140,20 @@ INIT_XMM
%rep num_mmregs
CAT_XDEFINE m, %%i, %1_m %+ %%i
CAT_XDEFINE nn, m %+ %%i, %%i
%assign %%i %%i+1
%assign %%i %%i+1
%endrep
%endif
%endmacro
; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
%macro call 1
call_internal %1, %1 %+ SUFFIX
call_internal %1 %+ SUFFIX, %1
%endmacro
%macro call_internal 2
%xdefine %%i %1
%ifndef cglobaled_%1
%ifdef cglobaled_%2
%xdefine %%i %2
%xdefine %%i %2
%ifndef cglobaled_%2
%ifdef cglobaled_%1
%xdefine %%i %1
%endif
%endif
call %%i
......@@ -1159,7 +1196,7 @@ INIT_XMM
%endif
CAT_XDEFINE sizeofxmm, i, 16
CAT_XDEFINE sizeofymm, i, 32
%assign i i+1
%assign i i+1
%endrep
%undef i
......@@ -1536,7 +1573,7 @@ AVX_INSTR pfmul, 3dnow, 1, 0, 1
%else
CAT_XDEFINE q, j, i
%endif
%assign i i+1
%assign i i+1
%endrep
%undef i
%undef j
......@@ -1559,55 +1596,54 @@ FMA_INSTR pmacsdd, pmulld, paddd ; sse4 emulation
FMA_INSTR pmacsdql, pmuldq, paddq ; sse4 emulation
FMA_INSTR pmadcswd, pmaddwd, paddd
; convert FMA4 to FMA3 if possible
%macro FMA4_INSTR 4
%macro %1 4-8 %1, %2, %3, %4
%if cpuflag(fma4)
v%5 %1, %2, %3, %4
%elifidn %1, %2
v%6 %1, %4, %3 ; %1 = %1 * %3 + %4
%elifidn %1, %3
v%7 %1, %2, %4 ; %1 = %2 * %1 + %4
%elifidn %1, %4
v%8 %1, %2, %3 ; %1 = %2 * %3 + %1
%else
%error fma3 emulation of ``%5 %1, %2, %3, %4'' is not supported
%endif
%endmacro
%endmacro
FMA4_INSTR fmaddpd, fmadd132pd, fmadd213pd, fmadd231pd
FMA4_INSTR fmaddps, fmadd132ps, fmadd213ps, fmadd231ps
FMA4_INSTR fmaddsd, fmadd132sd, fmadd213sd, fmadd231sd
FMA4_INSTR fmaddss, fmadd132ss, fmadd213ss, fmadd231ss
FMA4_INSTR fmaddsubpd, fmaddsub132pd, fmaddsub213pd, fmaddsub231pd
FMA4_INSTR fmaddsubps, fmaddsub132ps, fmaddsub213ps, fmaddsub231ps
FMA4_INSTR fmsubaddpd, fmsubadd132pd, fmsubadd213pd, fmsubadd231pd
FMA4_INSTR fmsubaddps, fmsubadd132ps, fmsubadd213ps, fmsubadd231ps
FMA4_INSTR fmsubpd, fmsub132pd, fmsub213pd, fmsub231pd
FMA4_INSTR fmsubps, fmsub132ps, fmsub213ps, fmsub231ps
FMA4_INSTR fmsubsd, fmsub132sd, fmsub213sd, fmsub231sd
FMA4_INSTR fmsubss, fmsub132ss, fmsub213ss, fmsub231ss
FMA4_INSTR fnmaddpd, fnmadd132pd, fnmadd213pd, fnmadd231pd
FMA4_INSTR fnmaddps, fnmadd132ps, fnmadd213ps, fnmadd231ps
FMA4_INSTR fnmaddsd, fnmadd132sd, fnmadd213sd, fnmadd231sd
FMA4_INSTR fnmaddss, fnmadd132ss, fnmadd213ss, fnmadd231ss
FMA4_INSTR fnmsubpd, fnmsub132pd, fnmsub213pd, fnmsub231pd
FMA4_INSTR fnmsubps, fnmsub132ps, fnmsub213ps, fnmsub231ps
FMA4_INSTR fnmsubsd, fnmsub132sd, fnmsub213sd, fnmsub231sd
FMA4_INSTR fnmsubss, fnmsub132ss, fnmsub213ss, fnmsub231ss
; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug
%if ARCH_X86_64 == 0
%macro vpbroadcastq 2
%if sizeof%1 == 16
movddup %1, %2
%else
vbroadcastsd %1, %2
%endif
%endmacro
; Macros for consolidating FMA3 and FMA4 using 4-operand (dst, src1, src2, src3) syntax.
; FMA3 is only possible if dst is the same as one of the src registers.
; Either src2 or src3 can be a memory operand.
%macro FMA4_INSTR 2-*
%push fma4_instr
%xdefine %$prefix %1
%rep %0 - 1
%macro %$prefix%2 4-6 %$prefix, %2
%if notcpuflag(fma3) && notcpuflag(fma4)
%error use of ``%5%6'' fma instruction in cpuname function: current_function
%elif cpuflag(fma4)
v%5%6 %1, %2, %3, %4
%elifidn %1, %2
; If %3 or %4 is a memory operand it needs to be encoded as the last operand.
%ifid %3
v%{5}213%6 %2, %3, %4
%else
v%{5}132%6 %2, %4, %3
%endif
%elifidn %1, %3
v%{5}213%6 %3, %2, %4
%elifidn %1, %4
v%{5}231%6 %4, %2, %3
%else
%error fma3 emulation of ``%5%6 %1, %2, %3, %4'' is not supported
%endif
%endmacro
%rotate 1
%endrep
%pop
%endmacro
FMA4_INSTR fmadd, pd, ps, sd, ss
FMA4_INSTR fmaddsub, pd, ps
FMA4_INSTR fmsub, pd, ps, sd, ss
FMA4_INSTR fmsubadd, pd, ps
FMA4_INSTR fnmadd, pd, ps, sd, ss
FMA4_INSTR fnmsub, pd, ps, sd, ss
; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug (fixed in 1.3.0)