x86: Update x86inc.asm

Make things up-to-date with upstream.

https://code.videolan.org/videolan/x86inc.asm

1 files changed, 462 insertions, 210 deletions

diff --git a/libavutil/x86/x86inc.asm b/libavutil/x86/x86inc.asm
index e099ee4b10..e61d924bc1 100644
--- a/libavutil/x86/x86inc.asm
+++ b/libavutil/x86/x86inc.asm
@@ -1,7 +1,7 @@
 ;*****************************************************************************
-;* x86inc.asm: x264asm abstraction layer
+;* x86inc.asm: x86 abstraction layer
 ;*****************************************************************************
-;* Copyright (C) 2005-2018 x264 project
+;* Copyright (C) 2005-2024 x264 project
 ;*
 ;* Authors: Loren Merritt <lorenm@u.washington.edu>
 ;*          Henrik Gramner <henrik@gramner.com>
@@ -21,21 +21,14 @@
 ;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 ;*****************************************************************************
 
-; This is a header file for the x264ASM assembly language, which uses
+; This is a header file for the x86inc.asm assembly language, which uses
 ; NASM/YASM syntax combined with a large number of macros to provide easy
 ; abstraction between different calling conventions (x86_32, win64, linux64).
 ; It also has various other useful features to simplify writing the kind of
-; DSP functions that are most often used in x264.
-
-; Unlike the rest of x264, this file is available under an ISC license, as it
-; has significant usefulness outside of x264 and we want it to be available
-; to the largest audience possible.  Of course, if you modify it for your own
-; purposes to add a new feature, we strongly encourage contributing a patch
-; as this feature might be useful for others as well.  Send patches or ideas
-; to x264-devel@videolan.org .
+; DSP functions that are most often used.
 
 %ifndef private_prefix
-    %define private_prefix x264
+    %error private_prefix not defined
 %endif
 
 %ifndef public_prefix
@@ -68,12 +61,19 @@
 %endif
 
 %define FORMAT_ELF 0
+%define FORMAT_MACHO 0
 %ifidn __OUTPUT_FORMAT__,elf
     %define FORMAT_ELF 1
 %elifidn __OUTPUT_FORMAT__,elf32
     %define FORMAT_ELF 1
 %elifidn __OUTPUT_FORMAT__,elf64
     %define FORMAT_ELF 1
+%elifidn __OUTPUT_FORMAT__,macho
+    %define FORMAT_MACHO 1
+%elifidn __OUTPUT_FORMAT__,macho32
+    %define FORMAT_MACHO 1
+%elifidn __OUTPUT_FORMAT__,macho64
+    %define FORMAT_MACHO 1
 %endif
 
 %ifdef PREFIX
@@ -82,6 +82,11 @@
     %define mangle(x) x
 %endif
 
+; Use VEX-encoding even in non-AVX functions
+%ifndef FORCE_VEX_ENCODING
+    %define FORCE_VEX_ENCODING 0
+%endif
+
 ; aout does not support align=
 ; NOTE: This section is out of sync with x264, in order to
 ; keep supporting OS/2.
@@ -99,28 +104,27 @@
     %endif
 %endmacro
 
-%if WIN64
-    %define PIC
-%elif ARCH_X86_64 == 0
-; x86_32 doesn't require PIC.
-; Some distros prefer shared objects to be PIC, but nothing breaks if
-; the code contains a few textrels, so we'll skip that complexity.
-    %undef PIC
-%endif
-%ifdef PIC
+%if ARCH_X86_64
+    %define PIC 1 ; always use PIC on x86-64
     default rel
+%elifidn __OUTPUT_FORMAT__,win32
+    %define PIC 0 ; PIC isn't used on 32-bit Windows
+%elifndef PIC
+    %define PIC 0
 %endif
 
-%macro CPUNOP 1
-    %if HAVE_CPUNOP
-        CPU %1
+%define HAVE_PRIVATE_EXTERN 1
+%ifdef __NASM_VERSION_ID__
+    %use smartalign
+    %if __NASM_VERSION_ID__ < 0x020e0000 ; 2.14
+        %define HAVE_PRIVATE_EXTERN 0
     %endif
-%endmacro
+%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
-; covers most of x264's asm.
+; covers most use cases.
 
 ; PROLOGUE:
 ; %1 = number of arguments. loads them from stack if needed.
@@ -232,6 +236,18 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
     %define gprsize 4
 %endif
 
+%macro LEA 2
+%if ARCH_X86_64
+    lea %1, [%2]
+%elif PIC
+    call $+5 ; special-cased to not affect the RSB on most CPU:s
+    pop %1
+    add %1, (%2)-$+1
+%else
+    mov %1, %2
+%endif
+%endmacro
+
 ; Repeats an instruction/operation for multiple arguments.
 ; Example usage: "REPX {psrlw x, 8}, m0, m1, m2, m3"
 %macro REPX 2-* ; operation, args
@@ -303,6 +319,10 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
     %endif
 %endmacro
 
+%if ARCH_X86_64 == 0
+    %define movsxd movifnidn
+%endif
+
 %macro movsxdifnidn 2
     %ifnidn %1, %2
         movsxd %1, %2
@@ -354,7 +374,46 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
 %define vzeroupper_required (mmsize > 16 && (ARCH_X86_64 == 0 || xmm_regs_used > 16 || notcpuflag(avx512)))
 %define high_mm_regs (16*cpuflag(avx512))
 
-%macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
+; Large stack allocations on Windows need to use stack probing in order
+; to guarantee that all stack memory is committed before accessing it.
+; This is done by ensuring that the guard page(s) at the end of the
+; currently committed pages are touched prior to any pages beyond that.
+%if WIN64
+    %assign STACK_PROBE_SIZE 8192
+%elifidn __OUTPUT_FORMAT__, win32
+    %assign STACK_PROBE_SIZE 4096
+%else
+    %assign STACK_PROBE_SIZE 0
+%endif
+
+%macro PROBE_STACK 1 ; stack_size
+    %if STACK_PROBE_SIZE
+        %assign %%i STACK_PROBE_SIZE
+        %rep %1 / STACK_PROBE_SIZE
+            mov eax, [rsp-%%i]
+            %assign %%i %%i+STACK_PROBE_SIZE
+        %endrep
+    %endif
+%endmacro
+
+%macro RESET_STACK_STATE 0
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset - stack_size_padded
+    %else
+        %xdefine rstk rsp
+    %endif
+    %assign stack_size 0
+    %assign stack_size_padded 0
+    %assign xmm_regs_used 0
+%endmacro
+
+%macro ALLOC_STACK 0-2 0, 0 ; stack_size, n_xmm_regs
+    RESET_STACK_STATE
+    %ifnum %2
+        %if mmsize != 8
+            %assign xmm_regs_used %2
+        %endif
+    %endif
     %ifnum %1
         %if %1 != 0
             %assign %%pad 0
@@ -364,16 +423,14 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
             %endif
             %if WIN64
                 %assign %%pad %%pad + 32 ; shadow space
-                %if mmsize != 8
-                    %assign xmm_regs_used %2
-                    %if xmm_regs_used > 8
-                        %assign %%pad %%pad + (xmm_regs_used-8)*16 ; callee-saved xmm registers
-                    %endif
+                %if xmm_regs_used > 8
+                    %assign %%pad %%pad + (xmm_regs_used-8)*16 ; callee-saved xmm registers
                 %endif
             %endif
             %if required_stack_alignment <= STACK_ALIGNMENT
                 ; maintain the current stack alignment
                 %assign stack_size_padded stack_size + %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+                PROBE_STACK stack_size_padded
                 SUB rsp, stack_size_padded
             %else
                 %assign %%reg_num (regs_used - 1)
@@ -389,6 +446,7 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
                     %xdefine rstkm rstk
                 %endif
                 %assign stack_size_padded stack_size + ((%%pad + required_stack_alignment-1) & ~(required_stack_alignment-1))
+                PROBE_STACK stack_size_padded
                 mov rstk, rsp
                 and rsp, ~(required_stack_alignment-1)
                 sub rsp, stack_size_padded
@@ -399,7 +457,7 @@ DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
     %endif
 %endmacro
 
-%macro SETUP_STACK_POINTER 1
+%macro SETUP_STACK_POINTER 0-1 0
     %ifnum %1
         %if %1 != 0 && required_stack_alignment > STACK_ALIGNMENT
             %if %1 > 0
@@ -462,35 +520,62 @@ DECLARE_REG 14, R13, 120
     %endif
 %endmacro
 
-%macro WIN64_PUSH_XMM 0
-    ; Use the shadow space to store XMM6 and XMM7, the rest needs stack space allocated.
-    %if xmm_regs_used > 6 + high_mm_regs
-        movaps [rstk + stack_offset +  8], xmm6
-    %endif
-    %if xmm_regs_used > 7 + high_mm_regs
-        movaps [rstk + stack_offset + 24], xmm7
-    %endif
-    %assign %%xmm_regs_on_stack xmm_regs_used - high_mm_regs - 8
-    %if %%xmm_regs_on_stack > 0
-        %assign %%i 8
-        %rep %%xmm_regs_on_stack
-            movaps [rsp + (%%i-8)*16 + stack_size + 32], xmm %+ %%i
-            %assign %%i %%i+1
-        %endrep
+; Push XMM registers to the stack. If no argument is specified all used register
+; will be pushed, otherwise only push previously unpushed registers.
+%macro WIN64_PUSH_XMM 0-2 ; new_xmm_regs_used, xmm_regs_pushed
+    %if mmsize != 8
+        %if %0 == 2
+            %assign %%pushed %2
+            %assign xmm_regs_used %1
+        %elif %0 == 1
+            %assign %%pushed xmm_regs_used
+            %assign xmm_regs_used %1
+        %else
+            %assign %%pushed 0
+        %endif
+        ; Use the shadow space to store XMM6 and XMM7, the rest needs stack space allocated.
+        %if %%pushed <= 6 + high_mm_regs && xmm_regs_used > 6 + high_mm_regs
+            movaps [rstk + stack_offset +  8], xmm6
+        %endif
+        %if %%pushed <= 7 + high_mm_regs && xmm_regs_used > 7 + high_mm_regs
+            movaps [rstk + stack_offset + 24], xmm7
+        %endif
+        %assign %%pushed %%pushed - high_mm_regs - 8
+        %if %%pushed < 0
+            %assign %%pushed 0
+        %endif
+        %assign %%regs_to_push xmm_regs_used - %%pushed - high_mm_regs - 8
+        %if %%regs_to_push > 0
+            ASSERT (%%regs_to_push + %%pushed) * 16 <= stack_size_padded - stack_size - 32
+            %assign %%i %%pushed + 8
+            %rep %%regs_to_push
+                movaps [rsp + (%%i-8)*16 + stack_size + 32], xmm %+ %%i
+                %assign %%i %%i+1
+            %endrep
+        %endif
     %endif
 %endmacro
 
-%macro WIN64_SPILL_XMM 1
-    %assign xmm_regs_used %1
-    ASSERT xmm_regs_used <= 16 + high_mm_regs
-    %assign %%xmm_regs_on_stack xmm_regs_used - high_mm_regs - 8
-    %if %%xmm_regs_on_stack > 0
-        ; Allocate stack space for callee-saved xmm registers plus shadow space and align the stack.
-        %assign %%pad %%xmm_regs_on_stack*16 + 32
-        %assign stack_size_padded %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
-        SUB rsp, stack_size_padded
+; Allocated stack space for XMM registers and push all, or a subset, of those
+%macro WIN64_SPILL_XMM 1-2 ; xmm_regs_used, xmm_regs_reserved
+    RESET_STACK_STATE
+    %if mmsize != 8
+        %assign xmm_regs_used %1
+        ASSERT xmm_regs_used <= 16 + high_mm_regs
+        %if %0 == 2
+            ASSERT %2 >= %1
+            %assign %%xmm_regs_on_stack %2 - high_mm_regs - 8
+        %else
+            %assign %%xmm_regs_on_stack %1 - high_mm_regs - 8
+        %endif
+        %if %%xmm_regs_on_stack > 0
+            ; Allocate stack space for callee-saved xmm registers plus shadow space and align the stack.
+            %assign %%pad %%xmm_regs_on_stack*16 + 32
+            %assign stack_size_padded %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+            SUB rsp, stack_size_padded
+        %endif
+        WIN64_PUSH_XMM
     %endif
-    WIN64_PUSH_XMM
 %endmacro
 
 %macro WIN64_RESTORE_XMM_INTERNAL 0
@@ -521,9 +606,7 @@ DECLARE_REG 14, R13, 120
 
 %macro WIN64_RESTORE_XMM 0
     WIN64_RESTORE_XMM_INTERNAL
-    %assign stack_offset (stack_offset-stack_size_padded)
-    %assign stack_size_padded 0
-    %assign xmm_regs_used 0
+    RESET_STACK_STATE
 %endmacro
 
 %define has_epilogue regs_used > 7 || stack_size > 0 || vzeroupper_required || xmm_regs_used > 6+high_mm_regs
@@ -558,12 +641,11 @@ DECLARE_REG 14, R13, 72
 %macro PROLOGUE 2-5+ 0, 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
     %assign num_args %1
     %assign regs_used %2
-    %assign xmm_regs_used %3
     ASSERT regs_used >= num_args
     SETUP_STACK_POINTER %4
     ASSERT regs_used <= 15
     PUSH_IF_USED 9, 10, 11, 12, 13, 14
-    ALLOC_STACK %4
+    ALLOC_STACK %4, %3
     LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
     %if %0 > 4
         %ifnum %4
@@ -627,7 +709,7 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
     SETUP_STACK_POINTER %4
     ASSERT regs_used <= 7
     PUSH_IF_USED 3, 4, 5, 6
-    ALLOC_STACK %4
+    ALLOC_STACK %4, %3
     LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
     %if %0 > 4
         %ifnum %4
@@ -660,11 +742,19 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 %endif ;======================================================================
 
 %if WIN64 == 0
-    %macro WIN64_SPILL_XMM 1
+    %macro WIN64_SPILL_XMM 1-2
+        RESET_STACK_STATE
+        %if mmsize != 8
+            %assign xmm_regs_used %1
+        %endif
     %endmacro
     %macro WIN64_RESTORE_XMM 0
+        RESET_STACK_STATE
     %endmacro
-    %macro WIN64_PUSH_XMM 0
+    %macro WIN64_PUSH_XMM 0-2
+        %if mmsize != 8 && %0 >= 1
+            %assign xmm_regs_used %1
+        %endif
     %endmacro
 %endif
 
@@ -705,7 +795,7 @@ DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
 
 BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae, jna, jnae, jb, jbe, jnb, jnbe, jc, jnc, js, jns, jo, jno, jp, jnp
 
-%macro TAIL_CALL 2 ; callee, is_nonadjacent
+%macro TAIL_CALL 1-2 1 ; callee, is_nonadjacent
     %if has_epilogue
         call %1
         RET
@@ -735,22 +825,25 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
 %endmacro
 %macro cglobal_internal 2-3+
     annotate_function_size
-    %if %1
-        %xdefine %%FUNCTION_PREFIX private_prefix
-        %xdefine %%VISIBILITY hidden
-    %else
-        %xdefine %%FUNCTION_PREFIX public_prefix
-        %xdefine %%VISIBILITY
-    %endif
     %ifndef cglobaled_%2
-        %xdefine %2 mangle(%%FUNCTION_PREFIX %+ _ %+ %2)
+        %if %1
+            %xdefine %2 mangle(private_prefix %+ _ %+ %2)
+        %else
+            %xdefine %2 mangle(public_prefix %+ _ %+ %2)
+        %endif
         %xdefine %2.skip_prologue %2 %+ .skip_prologue
         CAT_XDEFINE cglobaled_, %2, 1
     %endif
     %xdefine current_function %2
     %xdefine current_function_section __SECT__
     %if FORMAT_ELF
-        global %2:function %%VISIBILITY
+        %if %1
+            global %2:function hidden
+        %else
+            global %2:function
+        %endif
+    %elif FORMAT_MACHO && HAVE_PRIVATE_EXTERN && %1
+        global %2:private_extern
     %else
         global %2
     %endif
@@ -771,6 +864,8 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
 %macro cglobal_label 1
     %if FORMAT_ELF
         global current_function %+ %1:function hidden
+    %elif FORMAT_MACHO && HAVE_PRIVATE_EXTERN
+        global current_function %+ %1:private_extern
     %else
         global current_function %+ %1
     %endif
@@ -796,15 +891,34 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
     %xdefine %1 mangle(private_prefix %+ _ %+ %1)
     %if FORMAT_ELF
         global %1:data hidden
+    %elif FORMAT_MACHO && HAVE_PRIVATE_EXTERN
+        global %1:private_extern
     %else
         global %1
     %endif
     %1: %2
 %endmacro
 
-; This is needed for ELF, otherwise the GNU linker assumes the stack is executable by default.
 %if FORMAT_ELF
+    ; The GNU linker assumes the stack is executable by default.
     [SECTION .note.GNU-stack noalloc noexec nowrite progbits]
+
+    %ifdef __NASM_VERSION_ID__
+        %if __NASM_VERSION_ID__ >= 0x020e0300 ; 2.14.03
+            %if ARCH_X86_64
+                ; Control-flow Enforcement Technology (CET) properties.
+                [SECTION .note.gnu.property alloc noexec nowrite note align=gprsize]
+                dd 0x00000004  ; n_namesz
+                dd gprsize + 8 ; n_descsz
+                dd 0x00000005  ; n_type = NT_GNU_PROPERTY_TYPE_0
+                db "GNU",0     ; n_name
+                dd 0xc0000002  ; pr_type = GNU_PROPERTY_X86_FEATURE_1_AND
+                dd 0x00000004  ; pr_datasz
+                dd 0x00000002  ; pr_data = GNU_PROPERTY_X86_FEATURE_1_SHSTK
+                dd 0x00000000  ; pr_padding
+            %endif
+        %endif
+    %endif
 %endif
 
 ; Tell debuggers how large the function was.
@@ -828,32 +942,34 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
 ; cpuflags
 
 %assign cpuflags_mmx       (1<<0)
-%assign cpuflags_mmx2      (1<<1) | cpuflags_mmx
-%assign cpuflags_3dnow     (1<<2) | cpuflags_mmx
-%assign cpuflags_3dnowext  (1<<3) | cpuflags_3dnow
-%assign cpuflags_sse       (1<<4) | cpuflags_mmx2
-%assign cpuflags_sse2      (1<<5) | cpuflags_sse
-%assign cpuflags_sse2slow  (1<<6) | cpuflags_sse2
-%assign cpuflags_lzcnt     (1<<7) | cpuflags_sse2
-%assign cpuflags_sse3      (1<<8) | cpuflags_sse2
-%assign cpuflags_ssse3     (1<<9) | cpuflags_sse3
-%assign cpuflags_sse4      (1<<10)| cpuflags_ssse3
-%assign cpuflags_sse42     (1<<11)| cpuflags_sse4
-%assign cpuflags_aesni     (1<<12)| cpuflags_sse42
-%assign cpuflags_avx       (1<<13)| cpuflags_sse42
-%assign cpuflags_xop       (1<<14)| cpuflags_avx
-%assign cpuflags_fma4      (1<<15)| cpuflags_avx
-%assign cpuflags_fma3      (1<<16)| cpuflags_avx
-%assign cpuflags_bmi1      (1<<17)| cpuflags_avx|cpuflags_lzcnt
-%assign cpuflags_bmi2      (1<<18)| cpuflags_bmi1
-%assign cpuflags_avx2      (1<<19)| cpuflags_fma3|cpuflags_bmi2
-%assign cpuflags_avx512    (1<<20)| cpuflags_avx2 ; F, CD, BW, DQ, VL
-%assign cpuflags_avx512icl (1<<25)| cpuflags_avx512
-
-%assign cpuflags_cache32   (1<<21)
-%assign cpuflags_cache64   (1<<22)
-%assign cpuflags_aligned   (1<<23) ; not a cpu feature, but a function variant
-%assign cpuflags_atom      (1<<24)
+%assign cpuflags_mmx2      (1<<1)  | cpuflags_mmx
+%assign cpuflags_3dnow     (1<<2)  | cpuflags_mmx
+%assign cpuflags_3dnowext  (1<<3)  | cpuflags_3dnow
+%assign cpuflags_sse       (1<<4)  | cpuflags_mmx2
+%assign cpuflags_sse2      (1<<5)  | cpuflags_sse
+%assign cpuflags_sse2slow  (1<<6)  | cpuflags_sse2
+%assign cpuflags_lzcnt     (1<<7)  | cpuflags_sse2
+%assign cpuflags_sse3      (1<<8)  | cpuflags_sse2
+%assign cpuflags_ssse3     (1<<9)  | cpuflags_sse3
+%assign cpuflags_sse4      (1<<10) | cpuflags_ssse3
+%assign cpuflags_sse42     (1<<11) | cpuflags_sse4
+%assign cpuflags_aesni     (1<<12) | cpuflags_sse42
+%assign cpuflags_clmul     (1<<13) | cpuflags_sse42
+%assign cpuflags_gfni      (1<<14) | cpuflags_aesni|cpuflags_clmul
+%assign cpuflags_avx       (1<<15) | cpuflags_sse42
+%assign cpuflags_xop       (1<<16) | cpuflags_avx
+%assign cpuflags_fma4      (1<<17) | cpuflags_avx
+%assign cpuflags_fma3      (1<<18) | cpuflags_avx
+%assign cpuflags_bmi1      (1<<19) | cpuflags_avx|cpuflags_lzcnt
+%assign cpuflags_bmi2      (1<<20) | cpuflags_bmi1
+%assign cpuflags_avx2      (1<<21) | cpuflags_fma3|cpuflags_bmi2
+%assign cpuflags_avx512    (1<<22) | cpuflags_avx2 ; F, CD, BW, DQ, VL
+%assign cpuflags_avx512icl (1<<23) | cpuflags_avx512|cpuflags_gfni ; VNNI, IFMA, VBMI, VBMI2, VPOPCNTDQ, BITALG, VAES, VPCLMULQDQ
+
+%assign cpuflags_cache32   (1<<24)
+%assign cpuflags_cache64   (1<<25)
+%assign cpuflags_aligned   (1<<26) ; not a cpu feature, but a function variant
+%assign cpuflags_atom      (1<<27)
 
 ; Returns a boolean value expressing whether or not the specified cpuflag is enabled.
 %define    cpuflag(x) (((((cpuflags & (cpuflags_ %+ x)) ^ (cpuflags_ %+ x)) - 1) >> 31) & 1)
@@ -895,9 +1011,17 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
     %endif
 
     %if ARCH_X86_64 || cpuflag(sse2)
-        CPUNOP amdnop
+        %ifdef __NASM_VERSION_ID__
+            ALIGNMODE p6
+        %else
+            CPU amdnop
+        %endif
     %else
-        CPUNOP basicnop
+        %ifdef __NASM_VERSION_ID__
+            ALIGNMODE nop
+        %else
+            CPU basicnop
+        %endif
     %endif
 %endmacro
 
@@ -971,7 +1095,7 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
 %endmacro
 
 %macro INIT_XMM 0-1+
-    %assign avx_enabled 0
+    %assign avx_enabled FORCE_VEX_ENCODING
     %define RESET_MM_PERMUTATION INIT_XMM %1
     %define mmsize 16
     %define mova movdqa
@@ -983,6 +1107,9 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
     %if WIN64
         AVX512_MM_PERMUTATION 6 ; Swap callee-saved registers with volatile registers
     %endif
+    %xdefine bcstw 1to8
+    %xdefine bcstd 1to4
+    %xdefine bcstq 1to2
 %endmacro
 
 %macro INIT_YMM 0-1+
@@ -996,6 +1123,9 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
     INIT_CPUFLAGS %1
     DEFINE_MMREGS ymm
     AVX512_MM_PERMUTATION
+    %xdefine bcstw 1to16
+    %xdefine bcstd 1to8
+    %xdefine bcstq 1to4
 %endmacro
 
 %macro INIT_ZMM 0-1+
@@ -1009,6 +1139,9 @@ BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae,
     INIT_CPUFLAGS %1
     DEFINE_MMREGS zmm
     AVX512_MM_PERMUTATION
+    %xdefine bcstw 1to32
+    %xdefine bcstd 1to16
+    %xdefine bcstq 1to8
 %endmacro
 
 INIT_XMM
@@ -1106,19 +1239,32 @@ INIT_XMM
     %endif
     %assign %%i 0
     %rep num_mmregs
-        CAT_XDEFINE %%f, %%i, m %+ %%i
+        %xdefine %%tmp m %+ %%i
+        CAT_XDEFINE %%f, %%i, regnumof %+ %%tmp
         %assign %%i %%i+1
     %endrep
 %endmacro
 
-%macro LOAD_MM_PERMUTATION 1 ; name to load from
-    %ifdef %1_m0
+%macro LOAD_MM_PERMUTATION 0-1 ; name to load from
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %xdefine %%tmp %%f %+ 0
+    %ifnum %%tmp
+        DEFINE_MMREGS mmtype
         %assign %%i 0
         %rep num_mmregs
-            CAT_XDEFINE m, %%i, %1_m %+ %%i
-            CAT_XDEFINE nn, m %+ %%i, %%i
+            %xdefine %%tmp %%f %+ %%i
+            CAT_XDEFINE %%m, %%i, m %+ %%tmp
             %assign %%i %%i+1
         %endrep
+        %rep num_mmregs
+            %assign %%i %%i-1
+            CAT_XDEFINE m, %%i, %%m %+ %%i
+            CAT_XDEFINE nn, m %+ %%i, %%i
+        %endrep
     %endif
 %endmacro
 
@@ -1224,8 +1370,22 @@ INIT_XMM
         %ifdef cpuname
             %if notcpuflag(%2)
                 %error use of ``%1'' %2 instruction in cpuname function: current_function
-            %elif cpuflags_%2 < cpuflags_sse && notcpuflag(sse2) && __sizeofreg > 8
+            %elif %3 == 0 && __sizeofreg == 16 && notcpuflag(sse2)
                 %error use of ``%1'' sse2 instruction in cpuname function: current_function
+            %elif %3 == 0 && __sizeofreg == 32 && notcpuflag(avx2)
+                %error use of ``%1'' avx2 instruction in cpuname function: current_function
+            %elif __sizeofreg == 16 && notcpuflag(sse)
+                %error use of ``%1'' sse instruction in cpuname function: current_function
+            %elif __sizeofreg == 32 && notcpuflag(avx)
+                %error use of ``%1'' avx instruction in cpuname function: current_function
+            %elif __sizeofreg == 64 && notcpuflag(avx512)
+                %error use of ``%1'' avx512 instruction in cpuname function: current_function
+            %elifidn %1, pextrw ; special case because the base instruction is mmx2,
+                %ifnid %6       ; but sse4 is required for memory operands
+                    %if notcpuflag(sse4)
+                        %error use of ``%1'' sse4 instruction in cpuname function: current_function
+                    %endif
+                %endif
             %endif
         %endif
     %endif
@@ -1267,11 +1427,79 @@ INIT_XMM
             %1 %6, __src2
         %endif
     %elif %0 >= 9
-        __instr %6, %7, %8, %9
+        %if avx_enabled && __sizeofreg >= 16 && %4 == 1
+            %ifnnum regnumof%7
+                %if %3
+                    vmovaps %6, %7
+                %else
+                    vmovdqa %6, %7
+                %endif
+                __instr %6, %6, %8, %9
+            %else
+                __instr %6, %7, %8, %9
+            %endif
+        %else
+            __instr %6, %7, %8, %9
+        %endif
     %elif %0 == 8
-        __instr %6, %7, %8
+        %if avx_enabled && __sizeofreg >= 16 && %4 == 0
+            %xdefine __src1 %7
+            %xdefine __src2 %8
+            %if %5
+                %ifnum regnumof%7
+                    %ifnum regnumof%8
+                        %if regnumof%7 < 8 && regnumof%8 >= 8 && regnumof%8 < 16 && sizeof%8 <= 32
+                            ; Most VEX-encoded instructions require an additional byte to encode when
+                            ; src2 is a high register (e.g. m8..15). If the instruction is commutative
+                            ; we can swap src1 and src2 when doing so reduces the instruction length.
+                            %xdefine __src1 %8
+                            %xdefine __src2 %7
+                        %endif
+                    %endif
+                %elifnum regnumof%8 ; put memory operands in src2 when possible
+                    %xdefine __src1 %8
+                    %xdefine __src2 %7
+                %else
+                    %assign __emulate_avx 1
+                %endif
+            %elifnnum regnumof%7
+                ; EVEX allows imm8 shift instructions to be used with memory operands,
+                ; but VEX does not. This handles those special cases.
+                %ifnnum %8
+                    %assign __emulate_avx 1
+                %elif notcpuflag(avx512)
+                    %assign __emulate_avx 1
+                %endif
+            %endif
+            %if __emulate_avx ; a separate load is required
+                %if %3
+                    vmovaps %6, %7
+                %else
+                    vmovdqa %6, %7
+                %endif
+                __instr %6, %6, %8
+            %else
+                __instr %6, __src1, __src2
+            %endif
+        %else
+            __instr %6, %7, %8
+        %endif
     %elif %0 == 7
-        __instr %6, %7
+        %if avx_enabled && __sizeofreg >= 16 && %5
+            %xdefine __src1 %6
+            %xdefine __src2 %7
+            %ifnum regnumof%6
+                %ifnum regnumof%7
+                    %if regnumof%6 < 8 && regnumof%7 >= 8 && regnumof%7 < 16 && sizeof%7 <= 32
+                        %xdefine __src1 %7
+                        %xdefine __src2 %6
+                    %endif
+                %endif
+            %endif
+            __instr %6, __src1, __src2
+        %else
+            __instr %6, %7
+        %endif
     %else
         __instr %6
     %endif
@@ -1318,8 +1546,8 @@ AVX_INSTR andpd, sse2, 1, 0, 1
 AVX_INSTR andps, sse, 1, 0, 1
 AVX_INSTR blendpd, sse4, 1, 1, 0
 AVX_INSTR blendps, sse4, 1, 1, 0
-AVX_INSTR blendvpd, sse4 ; can't be emulated
-AVX_INSTR blendvps, sse4 ; can't be emulated
+AVX_INSTR blendvpd, sse4, 1, 1, 0 ; last operand must be xmm0 with legacy encoding
+AVX_INSTR blendvps, sse4, 1, 1, 0 ; last operand must be xmm0 with legacy encoding
 AVX_INSTR cmpeqpd, sse2, 1, 0, 1
 AVX_INSTR cmpeqps, sse, 1, 0, 1
 AVX_INSTR cmpeqsd, sse2, 1, 0, 0
@@ -1356,38 +1584,41 @@ AVX_INSTR cmpunordpd, sse2, 1, 0, 1
 AVX_INSTR cmpunordps, sse, 1, 0, 1
 AVX_INSTR cmpunordsd, sse2, 1, 0, 0
 AVX_INSTR cmpunordss, sse, 1, 0, 0
-AVX_INSTR comisd, sse2
-AVX_INSTR comiss, sse
-AVX_INSTR cvtdq2pd, sse2
-AVX_INSTR cvtdq2ps, sse2
-AVX_INSTR cvtpd2dq, sse2
-AVX_INSTR cvtpd2ps, sse2
-AVX_INSTR cvtps2dq, sse2
-AVX_INSTR cvtps2pd, sse2
-AVX_INSTR cvtsd2si, sse2
+AVX_INSTR comisd, sse2, 1
+AVX_INSTR comiss, sse, 1
+AVX_INSTR cvtdq2pd, sse2, 1
+AVX_INSTR cvtdq2ps, sse2, 1
+AVX_INSTR cvtpd2dq, sse2, 1
+AVX_INSTR cvtpd2ps, sse2, 1
+AVX_INSTR cvtps2dq, sse2, 1
+AVX_INSTR cvtps2pd, sse2, 1
+AVX_INSTR cvtsd2si, sse2, 1
 AVX_INSTR cvtsd2ss, sse2, 1, 0, 0
 AVX_INSTR cvtsi2sd, sse2, 1, 0, 0
 AVX_INSTR cvtsi2ss, sse, 1, 0, 0
 AVX_INSTR cvtss2sd, sse2, 1, 0, 0
-AVX_INSTR cvtss2si, sse
-AVX_INSTR cvttpd2dq, sse2
-AVX_INSTR cvttps2dq, sse2
-AVX_INSTR cvttsd2si, sse2
-AVX_INSTR cvttss2si, sse
+AVX_INSTR cvtss2si, sse, 1
+AVX_INSTR cvttpd2dq, sse2, 1
+AVX_INSTR cvttps2dq, sse2, 1
+AVX_INSTR cvttsd2si, sse2, 1
+AVX_INSTR cvttss2si, sse, 1
 AVX_INSTR divpd, sse2, 1, 0, 0
 AVX_INSTR divps, sse, 1, 0, 0
 AVX_INSTR divsd, sse2, 1, 0, 0
 AVX_INSTR divss, sse, 1, 0, 0
 AVX_INSTR dppd, sse4, 1, 1, 0
 AVX_INSTR dpps, sse4, 1, 1, 0
-AVX_INSTR extractps, sse4
+AVX_INSTR extractps, sse4, 1
+AVX_INSTR gf2p8affineinvqb, gfni, 0, 1, 0
+AVX_INSTR gf2p8affineqb, gfni, 0, 1, 0
+AVX_INSTR gf2p8mulb, gfni, 0, 0, 0
 AVX_INSTR haddpd, sse3, 1, 0, 0
 AVX_INSTR haddps, sse3, 1, 0, 0
 AVX_INSTR hsubpd, sse3, 1, 0, 0
 AVX_INSTR hsubps, sse3, 1, 0, 0
 AVX_INSTR insertps, sse4, 1, 1, 0
 AVX_INSTR lddqu, sse3
-AVX_INSTR ldmxcsr, sse
+AVX_INSTR ldmxcsr, sse, 1
 AVX_INSTR maskmovdqu, sse2
 AVX_INSTR maxpd, sse2, 1, 0, 1
 AVX_INSTR maxps, sse, 1, 0, 1
@@ -1397,10 +1628,10 @@ AVX_INSTR minpd, sse2, 1, 0, 1
 AVX_INSTR minps, sse, 1, 0, 1
 AVX_INSTR minsd, sse2, 1, 0, 0
 AVX_INSTR minss, sse, 1, 0, 0
-AVX_INSTR movapd, sse2
-AVX_INSTR movaps, sse
+AVX_INSTR movapd, sse2, 1
+AVX_INSTR movaps, sse, 1
 AVX_INSTR movd, mmx
-AVX_INSTR movddup, sse3
+AVX_INSTR movddup, sse3, 1
 AVX_INSTR movdqa, sse2
 AVX_INSTR movdqu, sse2
 AVX_INSTR movhlps, sse, 1, 0, 0
@@ -1409,19 +1640,19 @@ AVX_INSTR movhps, sse, 1, 0, 0
 AVX_INSTR movlhps, sse, 1, 0, 0
 AVX_INSTR movlpd, sse2, 1, 0, 0
 AVX_INSTR movlps, sse, 1, 0, 0
-AVX_INSTR movmskpd, sse2
-AVX_INSTR movmskps, sse
+AVX_INSTR movmskpd, sse2, 1
+AVX_INSTR movmskps, sse, 1
 AVX_INSTR movntdq, sse2
 AVX_INSTR movntdqa, sse4
-AVX_INSTR movntpd, sse2
-AVX_INSTR movntps, sse
+AVX_INSTR movntpd, sse2, 1
+AVX_INSTR movntps, sse, 1
 AVX_INSTR movq, mmx
 AVX_INSTR movsd, sse2, 1, 0, 0
-AVX_INSTR movshdup, sse3
-AVX_INSTR movsldup, sse3
+AVX_INSTR movshdup, sse3, 1
+AVX_INSTR movsldup, sse3, 1
 AVX_INSTR movss, sse, 1, 0, 0
-AVX_INSTR movupd, sse2
-AVX_INSTR movups, sse
+AVX_INSTR movupd, sse2, 1
+AVX_INSTR movups, sse, 1
 AVX_INSTR mpsadbw, sse4, 0, 1, 0
 AVX_INSTR mulpd, sse2, 1, 0, 1
 AVX_INSTR mulps, sse, 1, 0, 1
@@ -1432,90 +1663,90 @@ AVX_INSTR orps, sse, 1, 0, 1
 AVX_INSTR pabsb, ssse3
 AVX_INSTR pabsd, ssse3
 AVX_INSTR pabsw, ssse3
-AVX_INSTR packsswb, mmx, 0, 0, 0
 AVX_INSTR packssdw, mmx, 0, 0, 0
-AVX_INSTR packuswb, mmx, 0, 0, 0
+AVX_INSTR packsswb, mmx, 0, 0, 0
 AVX_INSTR packusdw, sse4, 0, 0, 0
+AVX_INSTR packuswb, mmx, 0, 0, 0
 AVX_INSTR paddb, mmx, 0, 0, 1
-AVX_INSTR paddw, mmx, 0, 0, 1
 AVX_INSTR paddd, mmx, 0, 0, 1
 AVX_INSTR paddq, sse2, 0, 0, 1
 AVX_INSTR paddsb, mmx, 0, 0, 1
 AVX_INSTR paddsw, mmx, 0, 0, 1
 AVX_INSTR paddusb, mmx, 0, 0, 1
 AVX_INSTR paddusw, mmx, 0, 0, 1
+AVX_INSTR paddw, mmx, 0, 0, 1
 AVX_INSTR palignr, ssse3, 0, 1, 0
 AVX_INSTR pand, mmx, 0, 0, 1
 AVX_INSTR pandn, mmx, 0, 0, 0
 AVX_INSTR pavgb, mmx2, 0, 0, 1
 AVX_INSTR pavgw, mmx2, 0, 0, 1
-AVX_INSTR pblendvb, sse4 ; can't be emulated
+AVX_INSTR pblendvb, sse4, 0, 1, 0 ; last operand must be xmm0 with legacy encoding
 AVX_INSTR pblendw, sse4, 0, 1, 0
-AVX_INSTR pclmulqdq, fnord, 0, 1, 0
-AVX_INSTR pclmulhqhqdq, fnord, 0, 0, 0
-AVX_INSTR pclmulhqlqdq, fnord, 0, 0, 0
-AVX_INSTR pclmullqhqdq, fnord, 0, 0, 0
-AVX_INSTR pclmullqlqdq, fnord, 0, 0, 0
-AVX_INSTR pcmpestri, sse42
-AVX_INSTR pcmpestrm, sse42
-AVX_INSTR pcmpistri, sse42
-AVX_INSTR pcmpistrm, sse42
+AVX_INSTR pclmulhqhqdq, clmul, 0, 0, 0
+AVX_INSTR pclmulhqlqdq, clmul, 0, 0, 0
+AVX_INSTR pclmullqhqdq, clmul, 0, 0, 0
+AVX_INSTR pclmullqlqdq, clmul, 0, 0, 0
+AVX_INSTR pclmulqdq, clmul, 0, 1, 0
 AVX_INSTR pcmpeqb, mmx, 0, 0, 1
-AVX_INSTR pcmpeqw, mmx, 0, 0, 1
 AVX_INSTR pcmpeqd, mmx, 0, 0, 1
 AVX_INSTR pcmpeqq, sse4, 0, 0, 1
+AVX_INSTR pcmpeqw, mmx, 0, 0, 1
+AVX_INSTR pcmpestri, sse42
+AVX_INSTR pcmpestrm, sse42
 AVX_INSTR pcmpgtb, mmx, 0, 0, 0
-AVX_INSTR pcmpgtw, mmx, 0, 0, 0
 AVX_INSTR pcmpgtd, mmx, 0, 0, 0
 AVX_INSTR pcmpgtq, sse42, 0, 0, 0
+AVX_INSTR pcmpgtw, mmx, 0, 0, 0
+AVX_INSTR pcmpistri, sse42
+AVX_INSTR pcmpistrm, sse42
 AVX_INSTR pextrb, sse4
 AVX_INSTR pextrd, sse4
 AVX_INSTR pextrq, sse4
 AVX_INSTR pextrw, mmx2
-AVX_INSTR phaddw, ssse3, 0, 0, 0
 AVX_INSTR phaddd, ssse3, 0, 0, 0
 AVX_INSTR phaddsw, ssse3, 0, 0, 0
+AVX_INSTR phaddw, ssse3, 0, 0, 0
 AVX_INSTR phminposuw, sse4
-AVX_INSTR phsubw, ssse3, 0, 0, 0
 AVX_INSTR phsubd, ssse3, 0, 0, 0
 AVX_INSTR phsubsw, ssse3, 0, 0, 0
+AVX_INSTR phsubw, ssse3, 0, 0, 0
 AVX_INSTR pinsrb, sse4, 0, 1, 0
 AVX_INSTR pinsrd, sse4, 0, 1, 0
 AVX_INSTR pinsrq, sse4, 0, 1, 0
 AVX_INSTR pinsrw, mmx2, 0, 1, 0
-AVX_INSTR pmaddwd, mmx, 0, 0, 1
 AVX_INSTR pmaddubsw, ssse3, 0, 0, 0
+AVX_INSTR pmaddwd, mmx, 0, 0, 1
 AVX_INSTR pmaxsb, sse4, 0, 0, 1
-AVX_INSTR pmaxsw, mmx2, 0, 0, 1
 AVX_INSTR pmaxsd, sse4, 0, 0, 1
+AVX_INSTR pmaxsw, mmx2, 0, 0, 1
 AVX_INSTR pmaxub, mmx2, 0, 0, 1
-AVX_INSTR pmaxuw, sse4, 0, 0, 1
 AVX_INSTR pmaxud, sse4, 0, 0, 1
+AVX_INSTR pmaxuw, sse4, 0, 0, 1
 AVX_INSTR pminsb, sse4, 0, 0, 1
-AVX_INSTR pminsw, mmx2, 0, 0, 1
 AVX_INSTR pminsd, sse4, 0, 0, 1
+AVX_INSTR pminsw, mmx2, 0, 0, 1
 AVX_INSTR pminub, mmx2, 0, 0, 1
-AVX_INSTR pminuw, sse4, 0, 0, 1
 AVX_INSTR pminud, sse4, 0, 0, 1
+AVX_INSTR pminuw, sse4, 0, 0, 1
 AVX_INSTR pmovmskb, mmx2
-AVX_INSTR pmovsxbw, sse4
 AVX_INSTR pmovsxbd, sse4
 AVX_INSTR pmovsxbq, sse4
+AVX_INSTR pmovsxbw, sse4
+AVX_INSTR pmovsxdq, sse4
 AVX_INSTR pmovsxwd, sse4
 AVX_INSTR pmovsxwq, sse4
-AVX_INSTR pmovsxdq, sse4
-AVX_INSTR pmovzxbw, sse4
 AVX_INSTR pmovzxbd, sse4
 AVX_INSTR pmovzxbq, sse4
+AVX_INSTR pmovzxbw, sse4
+AVX_INSTR pmovzxdq, sse4
 AVX_INSTR pmovzxwd, sse4
 AVX_INSTR pmovzxwq, sse4
-AVX_INSTR pmovzxdq, sse4
 AVX_INSTR pmuldq, sse4, 0, 0, 1
 AVX_INSTR pmulhrsw, ssse3, 0, 0, 1
 AVX_INSTR pmulhuw, mmx2, 0, 0, 1
 AVX_INSTR pmulhw, mmx, 0, 0, 1
-AVX_INSTR pmullw, mmx, 0, 0, 1
 AVX_INSTR pmulld, sse4, 0, 0, 1
+AVX_INSTR pmullw, mmx, 0, 0, 1
 AVX_INSTR pmuludq, sse2, 0, 0, 1
 AVX_INSTR por, mmx, 0, 0, 1
 AVX_INSTR psadbw, mmx2, 0, 0, 1
@@ -1524,57 +1755,57 @@ AVX_INSTR pshufd, sse2
 AVX_INSTR pshufhw, sse2
 AVX_INSTR pshuflw, sse2
 AVX_INSTR psignb, ssse3, 0, 0, 0
-AVX_INSTR psignw, ssse3, 0, 0, 0
 AVX_INSTR psignd, ssse3, 0, 0, 0
-AVX_INSTR psllw, mmx, 0, 0, 0
+AVX_INSTR psignw, ssse3, 0, 0, 0
 AVX_INSTR pslld, mmx, 0, 0, 0
-AVX_INSTR psllq, mmx, 0, 0, 0
 AVX_INSTR pslldq, sse2, 0, 0, 0
-AVX_INSTR psraw, mmx, 0, 0, 0
+AVX_INSTR psllq, mmx, 0, 0, 0
+AVX_INSTR psllw, mmx, 0, 0, 0
 AVX_INSTR psrad, mmx, 0, 0, 0
-AVX_INSTR psrlw, mmx, 0, 0, 0
+AVX_INSTR psraw, mmx, 0, 0, 0
 AVX_INSTR psrld, mmx, 0, 0, 0
-AVX_INSTR psrlq, mmx, 0, 0, 0
 AVX_INSTR psrldq, sse2, 0, 0, 0
+AVX_INSTR psrlq, mmx, 0, 0, 0
+AVX_INSTR psrlw, mmx, 0, 0, 0
 AVX_INSTR psubb, mmx, 0, 0, 0
-AVX_INSTR psubw, mmx, 0, 0, 0
 AVX_INSTR psubd, mmx, 0, 0, 0
 AVX_INSTR psubq, sse2, 0, 0, 0
 AVX_INSTR psubsb, mmx, 0, 0, 0
 AVX_INSTR psubsw, mmx, 0, 0, 0
 AVX_INSTR psubusb, mmx, 0, 0, 0
 AVX_INSTR psubusw, mmx, 0, 0, 0
+AVX_INSTR psubw, mmx, 0, 0, 0
 AVX_INSTR ptest, sse4
 AVX_INSTR punpckhbw, mmx, 0, 0, 0
-AVX_INSTR punpckhwd, mmx, 0, 0, 0
 AVX_INSTR punpckhdq, mmx, 0, 0, 0
 AVX_INSTR punpckhqdq, sse2, 0, 0, 0
+AVX_INSTR punpckhwd, mmx, 0, 0, 0
 AVX_INSTR punpcklbw, mmx, 0, 0, 0
-AVX_INSTR punpcklwd, mmx, 0, 0, 0
 AVX_INSTR punpckldq, mmx, 0, 0, 0
 AVX_INSTR punpcklqdq, sse2, 0, 0, 0
+AVX_INSTR punpcklwd, mmx, 0, 0, 0
 AVX_INSTR pxor, mmx, 0, 0, 1
-AVX_INSTR rcpps, sse
+AVX_INSTR rcpps, sse, 1
 AVX_INSTR rcpss, sse, 1, 0, 0
-AVX_INSTR roundpd, sse4
-AVX_INSTR roundps, sse4
+AVX_INSTR roundpd, sse4, 1
+AVX_INSTR roundps, sse4, 1
 AVX_INSTR roundsd, sse4, 1, 1, 0
 AVX_INSTR roundss, sse4, 1, 1, 0
-AVX_INSTR rsqrtps, sse
+AVX_INSTR rsqrtps, sse, 1
 AVX_INSTR rsqrtss, sse, 1, 0, 0
 AVX_INSTR shufpd, sse2, 1, 1, 0
 AVX_INSTR shufps, sse, 1, 1, 0
-AVX_INSTR sqrtpd, sse2
-AVX_INSTR sqrtps, sse
+AVX_INSTR sqrtpd, sse2, 1
+AVX_INSTR sqrtps, sse, 1
 AVX_INSTR sqrtsd, sse2, 1, 0, 0
 AVX_INSTR sqrtss, sse, 1, 0, 0
-AVX_INSTR stmxcsr, sse
+AVX_INSTR stmxcsr, sse, 1
 AVX_INSTR subpd, sse2, 1, 0, 0
 AVX_INSTR subps, sse, 1, 0, 0
 AVX_INSTR subsd, sse2, 1, 0, 0
 AVX_INSTR subss, sse, 1, 0, 0
-AVX_INSTR ucomisd, sse2
-AVX_INSTR ucomiss, sse
+AVX_INSTR ucomisd, sse2, 1
+AVX_INSTR ucomiss, sse, 1
 AVX_INSTR unpckhpd, sse2, 1, 0, 0
 AVX_INSTR unpckhps, sse, 1, 0, 0
 AVX_INSTR unpcklpd, sse2, 1, 0, 0
@@ -1584,8 +1815,41 @@ AVX_INSTR xorps, sse, 1, 0, 1
 
 ; 3DNow instructions, for sharing code between AVX, SSE and 3DN
 AVX_INSTR pfadd, 3dnow, 1, 0, 1
-AVX_INSTR pfsub, 3dnow, 1, 0, 0
 AVX_INSTR pfmul, 3dnow, 1, 0, 1
+AVX_INSTR pfsub, 3dnow, 1, 0, 0
+
+;%1 == instruction
+;%2 == minimal instruction set
+%macro GPR_INSTR 2
+    %macro %1 2-5 fnord, %1, %2
+        %ifdef cpuname
+            %if notcpuflag(%5)
+                %error use of ``%4'' %5 instruction in cpuname function: current_function
+            %endif
+        %endif
+        %ifidn %3, fnord
+            %4 %1, %2
+        %else
+            %4 %1, %2, %3
+        %endif
+    %endmacro
+%endmacro
+
+GPR_INSTR andn, bmi1
+GPR_INSTR bextr, bmi1
+GPR_INSTR blsi, bmi1
+GPR_INSTR blsmsk, bmi1
+GPR_INSTR blsr, bmi1
+GPR_INSTR bzhi, bmi2
+GPR_INSTR crc32, sse42
+GPR_INSTR mulx, bmi2
+GPR_INSTR pdep, bmi2
+GPR_INSTR pext, bmi2
+GPR_INSTR popcnt, sse42
+GPR_INSTR rorx, bmi2
+GPR_INSTR sarx, bmi2
+GPR_INSTR shlx, bmi2
+GPR_INSTR shrx, bmi2
 
 ; base-4 constants for shuffles
 %assign i 0
@@ -1618,15 +1882,11 @@ AVX_INSTR pfmul, 3dnow, 1, 0, 1
     %endmacro
 %endmacro
 
-FMA_INSTR  pmacsww,  pmullw, paddw
-FMA_INSTR  pmacsdd,  pmulld, paddd ; sse4 emulation
-FMA_INSTR pmacsdql,  pmuldq, paddq ; sse4 emulation
+FMA_INSTR pmacsdd,  pmulld,  paddd ; sse4 emulation
+FMA_INSTR pmacsdql, pmuldq,  paddq ; sse4 emulation
+FMA_INSTR pmacsww,  pmullw,  paddw
 FMA_INSTR pmadcswd, pmaddwd, paddd
 
-; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
-; This lets us use tzcnt without bumping the yasm version requirement yet.
-%define tzcnt rep bsf
-
 ; 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.
@@ -1687,6 +1947,11 @@ FMA4_INSTR fnmsub,   pd, ps, sd, ss
                 %assign %%evex_required 1
             %endif
         %endif
+        %ifnum regnumof%3
+            %if regnumof%3 >= 16 || sizeof%3 > 32
+                %assign %%evex_required 1
+            %endif
+        %endif
         %if %%evex_required
             %6 %%args
         %else
@@ -1711,16 +1976,3 @@ EVEX_INSTR vrcpps,         vrcp14ps,   1 ; EVEX versions have higher precision
 EVEX_INSTR vrcpss,         vrcp14ss,   1
 EVEX_INSTR vrsqrtps,       vrsqrt14ps, 1
 EVEX_INSTR vrsqrtss,       vrsqrt14ss, 1
-
-; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug (fixed in 1.3.0)
-%ifdef __YASM_VER__
-    %if __YASM_VERSION_ID__ < 0x01030000 && ARCH_X86_64 == 0
-        %macro vpbroadcastq 2
-            %if sizeof%1 == 16
-                movddup %1, %2
-            %else
-                vbroadcastsd %1, %2
-            %endif
-        %endmacro
-    %endif
-%endif