X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=common%2Fx86%2Fx86inc.asm;h=a4d81398b6306e78c3e657a8aaa2b9703683a92d;hb=366fa85885053c7b836a4272a4fbec1852103979;hp=c7ac95b70c34eaef31168fab746921ff4a48883e;hpb=23d4434de9ab5ef32ebb03401d971b8579a65fc6;p=x264

diff --git a/common/x86/x86inc.asm b/common/x86/x86inc.asm
index c7ac95b7..a4d81398 100644
--- a/common/x86/x86inc.asm
+++ b/common/x86/x86inc.asm
@@ -64,6 +64,15 @@
     %endif
 %endif
 
+%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 FORMAT_ELF 1
+%endif
+
 %ifdef PREFIX
     %define mangle(x) _ %+ x
 %else
@@ -74,10 +83,6 @@
     SECTION .rodata align=%1
 %endmacro
 
-%macro SECTION_TEXT 0-1 16
-    SECTION .text align=%1
-%endmacro
-
 %if WIN64
     %define PIC
 %elif ARCH_X86_64 == 0
@@ -90,6 +95,10 @@
     default rel
 %endif
 
+%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
@@ -137,6 +146,7 @@
     %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
@@ -161,9 +171,9 @@
     %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
@@ -273,7 +283,7 @@ 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
 
@@ -364,8 +374,10 @@ 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
@@ -479,9 +491,9 @@ 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
 
@@ -518,17 +530,17 @@ 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
 
@@ -574,29 +586,29 @@ 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
@@ -609,24 +621,26 @@ 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
@@ -641,6 +655,7 @@ 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
 
 ;=============================================================================
@@ -662,6 +677,7 @@ 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
         %xdefine %%VISIBILITY hidden
@@ -675,7 +691,8 @@ 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__,elf
+    %xdefine current_function_section __SECT__
+    %if FORMAT_ELF
         global %2:function %%VISIBILITY
     %else
         global %2
@@ -701,14 +718,16 @@ 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__,elf
+    %if FORMAT_ELF
         global %1:data hidden
     %else
         global %1
@@ -716,12 +735,29 @@ 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__,elf
-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)
@@ -750,8 +786,9 @@ 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))
+; 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.
@@ -789,9 +826,17 @@ SECTION .note.GNU-stack noalloc noexec nowrite progbits
     %endif
 
     %if ARCH_X86_64 || cpuflag(sse2)
-        CPU amdnop
+        %ifdef __NASM_VER__
+            ALIGNMODE k8
+        %else
+            CPU amdnop
+        %endif
     %else
-        CPU basicnop
+        %ifdef __NASM_VER__
+            ALIGNMODE nop
+        %else
+            CPU basicnop
+        %endif
     %endif
 %endmacro
 
@@ -820,14 +865,14 @@ 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
@@ -838,7 +883,7 @@ 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
@@ -846,9 +891,9 @@ 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
@@ -859,7 +904,7 @@ 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
@@ -867,9 +912,9 @@ 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
@@ -893,7 +938,7 @@ 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
@@ -911,23 +956,23 @@ 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-*
@@ -937,7 +982,7 @@ INIT_XMM
         %xdefine m%2 %%tmp
         CAT_XDEFINE nn, m%1, %1
         CAT_XDEFINE nn, m%2, %2
-    %rotate 1
+        %rotate 1
     %endrep
 %endmacro
 
@@ -945,7 +990,7 @@ INIT_XMM
     %xdefine %%args nn %+ %1
     %rep %0-1
         %xdefine %%args %%args, nn %+ %2
-    %rotate 1
+        %rotate 1
     %endrep
     SWAP_INTERNAL_NUM %%args
 %endmacro
@@ -962,7 +1007,7 @@ INIT_XMM
     %assign %%i 0
     %rep num_mmregs
         CAT_XDEFINE %%f, %%i, m %+ %%i
-    %assign %%i %%i+1
+        %assign %%i %%i+1
     %endrep
 %endmacro
 
@@ -972,20 +1017,20 @@ 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
@@ -1028,7 +1073,7 @@ INIT_XMM
     %endif
     CAT_XDEFINE sizeofxmm, i, 16
     CAT_XDEFINE sizeofymm, i, 32
-%assign i i+1
+    %assign i i+1
 %endrep
 %undef i
 
@@ -1070,6 +1115,8 @@ 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
+                %error use of ``%1'' sse2 instruction in cpuname function: current_function
             %endif
         %endif
     %endif
@@ -1206,7 +1253,7 @@ AVX_INSTR minsd, sse2, 1, 0, 1
 AVX_INSTR minss, sse, 1, 0, 1
 AVX_INSTR movapd, sse2
 AVX_INSTR movaps, sse
-AVX_INSTR movd
+AVX_INSTR movd, mmx
 AVX_INSTR movddup, sse3
 AVX_INSTR movdqa, sse2
 AVX_INSTR movdqu, sse2
@@ -1222,7 +1269,7 @@ AVX_INSTR movntdq, sse2
 AVX_INSTR movntdqa, sse4
 AVX_INSTR movntpd, sse2
 AVX_INSTR movntps, sse
-AVX_INSTR movq
+AVX_INSTR movq, mmx
 AVX_INSTR movsd, sse2, 1, 0, 0
 AVX_INSTR movshdup, sse3
 AVX_INSTR movsldup, sse3
@@ -1403,7 +1450,7 @@ 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
@@ -1426,55 +1473,54 @@ 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)
+%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