combine1: pushq %r12 movq %rdi, %r12 pushq %rbp movq %rsi, %rbp pushq %rbx xorl %ebx, %ebx subq $16, %rsp movl $0x3f800000, (%rsi) jmp .L2 .L3: leaq 12(%rsp), %rdx movq %rbx, %rsi movq %r12, %rdi addq $1, %rbx call get_vec_element movss 0(%rbp), %xmm0 mulss 12(%rsp), %xmm0 movss %xmm0, 0(%rbp) .L2: movq %r12, %rdi call vec_length cmpq %rax, %rbx jl .L3 addq $16, %rsp popq %rbx popq %rbp popq %r12 ret combine2: pushq %r13 movq %rdi, %r13 pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx xorl %ebx, %ebx subq $24, %rsp call vec_length testq %rax, %rax movq %rax, %r12 movl $0x3f800000, 0(%rbp) jle .L6 .L10: leaq 12(%rsp), %rdx movq %rbx, %rsi movq %r13, %rdi addq $1, %rbx call get_vec_element movss 0(%rbp), %xmm0 cmpq %r12, %rbx mulss 12(%rsp), %xmm0 movss %xmm0, 0(%rbp) jne .L10 .L6: addq $24, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret combine4b: pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx subq $8, %rsp call vec_length testq %rax, %rax movss .LC0(%rip), %xmm0 jle .L13 xorl %edx, %edx .L15: cmpq %rdx, (%rbx) jle .L14 movq 8(%rbx), %rcx mulss (%rcx,%rdx,4), %xmm0 .L14: addq $1, %rdx cmpq %rax, %rdx jne .L15 .L13: movss %xmm0, 0(%rbp) addq $8, %rsp popq %rbx popq %rbp ret combine3: pushq %r12 pushq %rbp movq %rdi, %rbp pushq %rbx movq %rsi, %rbx call vec_length movq %rbp, %rdi movq %rax, %r12 call get_vec_start movss .LC0(%rip), %xmm0 testq %r12, %r12 movss %xmm0, (%rbx) jle .L18 movq %rax, %rdx leaq (%rax,%r12,4), %rax .L21: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rax, %rdx movss %xmm0, (%rbx) jne .L21 .L18: popq %rbx popq %rbp popq %r12 ret combine3w: pushq %r12 pushq %rbp movq %rdi, %rbp pushq %rbx movq %rsi, %rbx call vec_length movq %rbp, %rdi movq %rax, %r12 call get_vec_start movss .LC0(%rip), %xmm0 testq %r12, %r12 movss %xmm0, (%rbx) jle .L23 movq %rax, %rdx leaq (%rax,%r12,4), %rax .L26: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rax, %rdx movss %xmm0, (%rbx) jne .L26 .L23: popq %rbx popq %rbp popq %r12 ret combine4: pushq %r12 movq %rsi, %r12 pushq %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %rbp call get_vec_start testq %rbp, %rbp jle .L31 movq %rax, %rdx leaq (%rax,%rbp,4), %rax movss .LC0(%rip), %xmm0 .L30: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rax, %rdx jne .L30 .L29: popq %rbx popq %rbp movss %xmm0, (%r12) popq %r12 ret .L31: movss .LC0(%rip), %xmm0 jmp .L29 combine4p: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rdx leaq (%rax,%r12,4), %rax movss .LC0(%rip), %xmm0 cmpq %rax, %rdx jnb .L34 .L35: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L35 .L34: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret combine5: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -1(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm0 jle .L39 .L41: mulss (%rcx,%rdx,4), %xmm0 mulss 4(%rcx,%rdx,4), %xmm0 addq $2, %rdx cmpq %rdx, %rbp jg .L41 leaq -2(%rbx), %rax shrq %rax leaq 2(%rax,%rax), %rdx .L39: cmpq %rdx, %rbx jle .L42 .L43: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L43 .L42: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll3a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -2(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movss .LC0(%rip), %xmm0 jle .L47 .L48: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm0 mulss 8(%rax,%rdx,4), %xmm0 addq $3, %rdx cmpq %rdx, %rbp jg .L48 .L47: cmpq %rdx, %rbx jle .L49 .L50: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L50 .L49: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret combine5p: pushq %r12 movq %rsi, %r12 pushq %rbp movq %rdi, %rbp pushq %rbx call get_vec_start movq %rbp, %rdi movq %rax, %rbx call vec_length leaq (%rbx,%rax,4), %rax leaq -4(%rax), %rcx cmpq %rcx, %rbx jnb .L59 movq %rbx, %rdx movss .LC0(%rip), %xmm0 .L56: mulss (%rdx), %xmm0 addq $8, %rdx mulss -4(%rdx), %xmm0 cmpq %rdx, %rcx ja .L56 movq %rax, %rdx subq %rbx, %rdx subq $5, %rdx shrq $3, %rdx leaq 8(%rbx,%rdx,8), %rbx jmp .L61 .L58: mulss (%rbx), %xmm0 addq $4, %rbx .L61: cmpq %rbx, %rax ja .L58 popq %rbx popq %rbp movss %xmm0, (%r12) popq %r12 ret .L59: movss .LC0(%rip), %xmm0 jmp .L61 unroll2aw_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -1(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm0 jle .L63 .L65: mulss (%rcx,%rdx,4), %xmm0 addq $2, %rdx cmpq %rdx, %rbp mulss -4(%rcx,%rdx,4), %xmm0 jg .L65 leaq -2(%rbx), %rax shrq %rax leaq 2(%rax,%rax), %rdx .L63: cmpq %rdx, %rbx jle .L66 .L67: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L67 .L66: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -3(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm0 jle .L71 .L73: mulss (%rcx,%rdx,4), %xmm0 mulss 4(%rcx,%rdx,4), %xmm0 mulss 8(%rcx,%rdx,4), %xmm0 mulss 12(%rcx,%rdx,4), %xmm0 addq $4, %rdx cmpq %rdx, %rbp jg .L73 leaq -4(%rbx), %rax shrq $2, %rax leaq 4(,%rax,4), %rdx .L71: cmpq %rdx, %rbx jle .L74 .L75: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L75 .L74: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll5a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -4(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L83 movq %rax, %rcx movss .LC0(%rip), %xmm0 xorl %edx, %edx .L80: mulss (%rcx), %xmm0 addq $5, %rdx addq $20, %rcx mulss -16(%rcx), %xmm0 mulss -12(%rcx), %xmm0 mulss -8(%rcx), %xmm0 mulss -4(%rcx), %xmm0 cmpq %rdx, %rbp jg .L80 .L79: cmpq %rdx, %rbx jle .L81 .L82: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L82 .L81: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L83: movss .LC0(%rip), %xmm0 xorl %edx, %edx jmp .L79 unroll6a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -5(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L90 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L87: mulss (%rdx), %xmm0 addq $6, %rcx addq $24, %rdx mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L87 .L86: cmpq %rcx, %rbx jle .L88 .L89: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L89 .L88: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L90: movss .LC0(%rip), %xmm0 xorl %ecx, %ecx jmp .L86 unroll7a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -6(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L97 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L94: mulss (%rdx), %xmm0 addq $7, %rcx addq $28, %rdx mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L94 .L93: cmpq %rcx, %rbx jle .L95 .L96: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L96 .L95: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L97: movss .LC0(%rip), %xmm0 xorl %ecx, %ecx jmp .L93 unroll8a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -7(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L105 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L102: mulss (%rdx), %xmm0 addq $8, %rcx addq $32, %rdx mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L102 leaq -8(%rbx), %rdx shrq $3, %rdx leaq 8(,%rdx,8), %rdx .L100: cmpq %rdx, %rbx jle .L103 .L104: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L104 .L103: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L105: movss .LC0(%rip), %xmm0 xorl %edx, %edx jmp .L100 unroll9a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -8(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L112 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L109: mulss (%rdx), %xmm0 addq $9, %rcx addq $36, %rdx mulss -32(%rdx), %xmm0 mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L109 .L108: cmpq %rcx, %rbx jle .L110 .L111: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L111 .L110: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L112: movss .LC0(%rip), %xmm0 xorl %ecx, %ecx jmp .L108 unroll10a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -9(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L119 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L116: mulss (%rdx), %xmm0 addq $10, %rcx addq $40, %rdx mulss -36(%rdx), %xmm0 mulss -32(%rdx), %xmm0 mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L116 .L115: cmpq %rcx, %rbx jle .L117 .L118: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L118 .L117: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L119: movss .LC0(%rip), %xmm0 xorl %ecx, %ecx jmp .L115 unroll16a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -15(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L127 movq %rax, %rdx movss .LC0(%rip), %xmm0 xorl %ecx, %ecx .L124: mulss (%rdx), %xmm0 addq $16, %rcx addq $64, %rdx mulss -60(%rdx), %xmm0 mulss -56(%rdx), %xmm0 mulss -52(%rdx), %xmm0 mulss -48(%rdx), %xmm0 mulss -44(%rdx), %xmm0 mulss -40(%rdx), %xmm0 mulss -36(%rdx), %xmm0 mulss -32(%rdx), %xmm0 mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rcx, %rbp jg .L124 leaq -16(%rbx), %rdx andq $-16, %rdx addq $16, %rdx .L122: cmpq %rdx, %rbx jle .L125 .L126: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L126 .L125: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L127: movss .LC0(%rip), %xmm0 xorl %edx, %edx jmp .L122 unroll2_combine: pushq %r12 movq %rsi, %r12 pushq %rbp movq %rdi, %rbp pushq %rbx call vec_length movq %rbp, %rdi movq %rax, %rbx call get_vec_start movq %rax, %rcx movq %rbx, %rax shrq $63, %rax leaq (%rbx,%rax), %rdi andl $1, %edi subq %rax, %rdi subq %rdi, %rbx leaq (%rcx,%rbx,4), %rax cmpq %rax, %rcx jnb .L135 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L132: mulss (%rdx), %xmm0 addq $8, %rdx mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L132 movq %rcx, %rdx notq %rdx addq %rax, %rdx shrq $3, %rdx leaq 8(%rcx,%rdx,8), %rcx .L130: leaq (%rax,%rdi,4), %rax cmpq %rcx, %rax jbe .L133 .L134: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L134 .L133: popq %rbx popq %rbp movss %xmm0, (%r12) popq %r12 ret .L135: movss .LC0(%rip), %xmm0 jmp .L130 unroll3_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rdx leaq -8(%rax,%r12,4), %rax movss .LC0(%rip), %xmm0 cmpq %rax, %rdx jnb .L138 .L139: mulss (%rdx), %xmm0 addq $12, %rdx mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L139 .L138: addq $8, %rax cmpq %rdx, %rax jbe .L140 .L141: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L141 .L140: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll4_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rcx leaq -12(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L150 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L147: mulss (%rdx), %xmm0 addq $16, %rdx mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L147 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-16, %rdx leaq 16(%rcx,%rdx), %rcx .L145: addq $12, %rax cmpq %rcx, %rax jbe .L148 .L149: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L149 .L148: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret .L150: movss .LC0(%rip), %xmm0 jmp .L145 unroll8_combine: pushq %r12 movq %rsi, %r12 pushq %rbp movq %rdi, %rbp pushq %rbx call vec_length movq %rbp, %rdi movq %rax, %rbx call get_vec_start movq %rbx, %rdx movq %rax, %rcx sarq $63, %rdx shrq $61, %rdx leaq (%rbx,%rdx), %rdi andl $7, %edi subq %rdx, %rdi subq %rdi, %rbx leaq (%rax,%rbx,4), %rax cmpq %rax, %rcx jnb .L158 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L155: mulss (%rdx), %xmm0 addq $32, %rdx mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L155 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx .L153: leaq (%rax,%rdi,4), %rax cmpq %rcx, %rax jbe .L156 .L157: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L157 .L156: popq %rbx popq %rbp movss %xmm0, (%r12) popq %r12 ret .L158: movss .LC0(%rip), %xmm0 jmp .L153 unroll16_combine: pushq %r12 movq %rsi, %r12 pushq %rbp movq %rdi, %rbp pushq %rbx call vec_length movq %rbp, %rdi movq %rax, %rbx call get_vec_start movq %rbx, %rdx movq %rax, %rcx sarq $63, %rdx shrq $60, %rdx leaq (%rbx,%rdx), %rdi andl $15, %edi subq %rdx, %rdi subq %rdi, %rbx leaq (%rax,%rbx,4), %rax cmpq %rax, %rcx jnb .L166 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L163: mulss (%rdx), %xmm0 addq $64, %rdx mulss -60(%rdx), %xmm0 mulss -56(%rdx), %xmm0 mulss -52(%rdx), %xmm0 mulss -48(%rdx), %xmm0 mulss -44(%rdx), %xmm0 mulss -40(%rdx), %xmm0 mulss -36(%rdx), %xmm0 mulss -32(%rdx), %xmm0 mulss -28(%rdx), %xmm0 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L163 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-64, %rdx leaq 64(%rcx,%rdx), %rcx .L161: leaq (%rax,%rdi,4), %rax cmpq %rcx, %rax jbe .L164 .L165: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L165 .L164: popq %rbx popq %rbp movss %xmm0, (%r12) popq %r12 ret .L166: movss .LC0(%rip), %xmm0 jmp .L161 combine6: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -1(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start movss .LC0(%rip), %xmm1 xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movaps %xmm1, %xmm0 jle .L169 .L171: mulss (%rcx,%rdx,4), %xmm0 mulss 4(%rcx,%rdx,4), %xmm1 addq $2, %rdx cmpq %rdx, %rbp jg .L171 leaq -2(%rbx), %rax shrq %rax leaq 2(%rax,%rax), %rdx .L169: cmpq %rdx, %rbx jle .L172 .L173: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L173 .L172: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4x2a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -3(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start movss .LC0(%rip), %xmm1 xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movaps %xmm1, %xmm0 jle .L177 .L179: mulss (%rcx,%rdx,4), %xmm0 mulss 4(%rcx,%rdx,4), %xmm1 mulss 8(%rcx,%rdx,4), %xmm0 mulss 12(%rcx,%rdx,4), %xmm1 addq $4, %rdx cmpq %rdx, %rbp jg .L179 leaq -4(%rbx), %rax shrq $2, %rax leaq 4(,%rax,4), %rdx .L177: cmpq %rdx, %rbx jle .L180 .L181: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L181 .L180: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8x2a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -7(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L190 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm0 .L187: mulss (%rdx), %xmm0 addq $8, %rcx addq $32, %rdx mulss -28(%rdx), %xmm1 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm1 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm1 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L187 leaq -8(%rbx), %rdx shrq $3, %rdx leaq 8(,%rdx,8), %rdx .L185: cmpq %rdx, %rbx jle .L188 .L189: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L189 .L188: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L190: movss .LC0(%rip), %xmm1 xorl %edx, %edx movaps %xmm1, %xmm0 jmp .L185 unroll3x3a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -2(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start movss .LC0(%rip), %xmm1 xorl %edx, %edx testq %rbp, %rbp movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jle .L193 .L194: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm2 mulss 8(%rax,%rdx,4), %xmm1 addq $3, %rdx cmpq %rdx, %rbp jg .L194 .L193: cmpq %rdx, %rbx jle .L195 .L196: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L196 .L195: mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4x4a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -3(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm1 jle .L205 movaps %xmm1, %xmm3 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 xorl %edx, %edx .L202: mulss (%rcx,%rdx,4), %xmm0 mulss 4(%rcx,%rdx,4), %xmm2 mulss 8(%rcx,%rdx,4), %xmm3 mulss 12(%rcx,%rdx,4), %xmm1 addq $4, %rdx cmpq %rdx, %rbp jg .L202 leaq -4(%rbx), %rax mulss %xmm3, %xmm1 shrq $2, %rax leaq 4(,%rax,4), %rdx .L200: cmpq %rdx, %rbx jle .L203 .L204: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L204 .L203: mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L205: movaps %xmm1, %xmm2 xorl %edx, %edx movaps %xmm1, %xmm0 jmp .L200 unroll8x4a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -7(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L213 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 .L210: mulss (%rdx), %xmm0 addq $8, %rcx addq $32, %rdx mulss -28(%rdx), %xmm3 mulss -24(%rdx), %xmm2 mulss -20(%rdx), %xmm1 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm3 mulss -8(%rdx), %xmm2 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L210 leaq -8(%rbx), %rdx shrq $3, %rdx leaq 8(,%rdx,8), %rdx .L208: cmpq %rdx, %rbx jle .L211 .L212: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L212 .L211: mulss %xmm3, %xmm0 mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L213: movss .LC0(%rip), %xmm1 xorl %edx, %edx movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 jmp .L208 unroll12x6a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -11(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L221 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L218: mulss (%rdx), %xmm0 addq $12, %rcx addq $48, %rdx mulss -44(%rdx), %xmm2 mulss -40(%rdx), %xmm5 mulss -36(%rdx), %xmm4 mulss -32(%rdx), %xmm3 mulss -28(%rdx), %xmm1 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm2 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L218 mulss %xmm5, %xmm4 mulss %xmm3, %xmm1 .L216: cmpq %rcx, %rbx jle .L219 .L220: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L220 .L219: mulss %xmm2, %xmm0 mulss %xmm4, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L221: movss .LC0(%rip), %xmm1 xorl %ecx, %ecx movaps %xmm1, %xmm4 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jmp .L216 unroll12x12a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -11(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L229 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm10 movaps %xmm1, %xmm11 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L226: addq $12, %rcx mulss (%rdx), %xmm0 addq $48, %rdx mulss -24(%rdx), %xmm7 mulss -44(%rdx), %xmm2 mulss -20(%rdx), %xmm6 mulss -40(%rdx), %xmm11 mulss -16(%rdx), %xmm5 mulss -36(%rdx), %xmm10 mulss -12(%rdx), %xmm4 mulss -32(%rdx), %xmm9 mulss -8(%rdx), %xmm3 mulss -28(%rdx), %xmm8 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L226 mulss %xmm11, %xmm10 mulss %xmm9, %xmm8 mulss %xmm7, %xmm6 mulss %xmm5, %xmm4 mulss %xmm3, %xmm1 .L224: cmpq %rcx, %rbx jle .L227 .L228: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L228 .L227: mulss %xmm2, %xmm0 mulss %xmm10, %xmm0 mulss %xmm8, %xmm0 mulss %xmm6, %xmm0 mulss %xmm4, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L229: movss .LC0(%rip), %xmm1 xorl %ecx, %ecx movaps %xmm1, %xmm4 movaps %xmm1, %xmm6 movaps %xmm1, %xmm8 movaps %xmm1, %xmm10 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jmp .L224 unroll16x16a_combine: pushq %r12 movq %rsi, %r12 pushq %rbp movq %rdi, %rbp pushq %rbx call vec_length movq %rbp, %rdi movq %rax, %rbx call get_vec_start cmpq $15, %rbx jle .L237 leaq -16(%rbx), %rcx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm3 movq %rax, %rdx shrq $4, %rcx movaps %xmm1, %xmm4 leaq 1(%rcx), %r8 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 xorl %ecx, %ecx movq %r8, %rdi movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 salq $6, %rdi movaps %xmm1, %xmm9 movaps %xmm1, %xmm10 movaps %xmm1, %xmm11 movaps %xmm1, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L234: addq $64, %rcx mulss (%rdx), %xmm0 addq $64, %rdx mulss -40(%rdx), %xmm11 mulss -60(%rdx), %xmm2 mulss -36(%rdx), %xmm10 mulss -56(%rdx), %xmm15 mulss -32(%rdx), %xmm9 mulss -52(%rdx), %xmm14 mulss -28(%rdx), %xmm8 mulss -48(%rdx), %xmm13 mulss -24(%rdx), %xmm7 mulss -44(%rdx), %xmm12 mulss -20(%rdx), %xmm6 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rdi, %rcx jne .L234 mulss %xmm11, %xmm10 movq %r8, %rdx mulss %xmm9, %xmm8 salq $4, %rdx mulss %xmm7, %xmm6 mulss %xmm5, %xmm4 mulss %xmm3, %xmm1 mulss %xmm8, %xmm10 mulss %xmm15, %xmm14 mulss %xmm13, %xmm12 mulss %xmm1, %xmm4 mulss %xmm6, %xmm10 .L232: cmpq %rdx, %rbx jle .L235 .L236: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L236 .L235: mulss %xmm2, %xmm0 popq %rbx popq %rbp mulss %xmm14, %xmm0 mulss %xmm12, %xmm0 mulss %xmm10, %xmm0 mulss %xmm4, %xmm0 movss %xmm0, (%r12) popq %r12 ret .L237: movss .LC0(%rip), %xmm4 xorl %edx, %edx movaps %xmm4, %xmm10 movaps %xmm4, %xmm12 movaps %xmm4, %xmm14 movaps %xmm4, %xmm2 movaps %xmm4, %xmm0 jmp .L232 unroll20x20a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $40, %rsp call vec_length leaq -19(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L245 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movss %xmm1, 24(%rsp) movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movss %xmm1, 20(%rsp) movaps %xmm1, %xmm5 movss %xmm1, 16(%rsp) movaps %xmm1, %xmm6 movss %xmm1, 12(%rsp) movaps %xmm1, %xmm7 movss %xmm1, 28(%rsp) movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm10 movaps %xmm1, %xmm11 movaps %xmm1, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movaps %xmm1, %xmm2 .L242: movss 28(%rsp), %xmm0 addq $20, %rcx mulss (%rdx), %xmm0 addq $80, %rdx mulss -76(%rdx), %xmm2 mulss -72(%rdx), %xmm15 mulss -68(%rdx), %xmm14 mulss -44(%rdx), %xmm11 mulss -64(%rdx), %xmm13 movss %xmm0, 28(%rsp) mulss -40(%rdx), %xmm10 movss 12(%rsp), %xmm0 mulss -60(%rdx), %xmm12 mulss -56(%rdx), %xmm0 mulss -36(%rdx), %xmm9 mulss -32(%rdx), %xmm8 mulss -28(%rdx), %xmm7 mulss -24(%rdx), %xmm6 mulss -20(%rdx), %xmm5 movss %xmm0, 12(%rsp) mulss -16(%rdx), %xmm4 movss 16(%rsp), %xmm0 mulss -12(%rdx), %xmm3 mulss -52(%rdx), %xmm0 mulss -4(%rdx), %xmm1 movss %xmm0, 16(%rsp) movss 20(%rsp), %xmm0 mulss -48(%rdx), %xmm0 movss %xmm0, 20(%rsp) movss 24(%rsp), %xmm0 mulss -8(%rdx), %xmm0 cmpq %rcx, %rbp movss %xmm0, 24(%rsp) jg .L242 mulss %xmm15, %xmm14 movss 12(%rsp), %xmm15 mulss 16(%rsp), %xmm15 movss 28(%rsp), %xmm0 mulss 20(%rsp), %xmm11 mulss 24(%rsp), %xmm1 mulss %xmm8, %xmm7 mulss %xmm6, %xmm5 mulss %xmm4, %xmm3 mulss %xmm10, %xmm9 mulss %xmm11, %xmm15 mulss %xmm5, %xmm7 mulss %xmm1, %xmm3 mulss %xmm13, %xmm12 mulss %xmm9, %xmm15 mulss %xmm3, %xmm7 .L240: cmpq %rcx, %rbx jle .L243 .L244: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L244 .L243: mulss %xmm2, %xmm0 mulss %xmm14, %xmm0 mulss %xmm12, %xmm0 mulss %xmm15, %xmm0 mulss %xmm7, %xmm0 movss %xmm0, 0(%r13) addq $40, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L245: movss .LC0(%rip), %xmm7 xorl %ecx, %ecx movaps %xmm7, %xmm15 movaps %xmm7, %xmm12 movaps %xmm7, %xmm14 movaps %xmm7, %xmm2 movaps %xmm7, %xmm0 jmp .L240 unroll5x5a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -4(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L253 movss .LC0(%rip), %xmm1 movq %rax, %rcx xorl %edx, %edx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L250: addq $5, %rdx mulss (%rcx), %xmm0 addq $20, %rcx mulss -16(%rcx), %xmm2 mulss -12(%rcx), %xmm4 mulss -8(%rcx), %xmm3 mulss -4(%rcx), %xmm1 cmpq %rdx, %rbp jg .L250 mulss %xmm4, %xmm3 mulss %xmm1, %xmm3 .L248: cmpq %rdx, %rbx jle .L251 .L252: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L252 .L251: mulss %xmm2, %xmm0 mulss %xmm3, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L253: movss .LC0(%rip), %xmm3 xorl %edx, %edx movaps %xmm3, %xmm2 movaps %xmm3, %xmm0 jmp .L248 unroll6x6a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -5(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L261 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L258: addq $6, %rcx mulss (%rdx), %xmm0 addq $24, %rdx mulss -20(%rdx), %xmm2 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L258 mulss %xmm5, %xmm4 mulss %xmm3, %xmm1 .L256: cmpq %rcx, %rbx jle .L259 .L260: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L260 .L259: mulss %xmm2, %xmm0 mulss %xmm4, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L261: movss .LC0(%rip), %xmm1 xorl %ecx, %ecx movaps %xmm1, %xmm4 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jmp .L256 unroll7x7a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -6(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L269 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L266: addq $7, %rcx mulss (%rdx), %xmm0 addq $28, %rdx mulss -24(%rdx), %xmm2 mulss -20(%rdx), %xmm6 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L266 mulss %xmm4, %xmm3 mulss %xmm6, %xmm5 mulss %xmm1, %xmm3 .L264: cmpq %rcx, %rbx jle .L267 .L268: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L268 .L267: mulss %xmm2, %xmm0 mulss %xmm5, %xmm0 mulss %xmm3, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L269: movss .LC0(%rip), %xmm3 xorl %ecx, %ecx movaps %xmm3, %xmm5 movaps %xmm3, %xmm2 movaps %xmm3, %xmm0 jmp .L264 unroll8x8a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -7(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L277 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L274: addq $8, %rcx mulss (%rdx), %xmm0 addq $32, %rdx mulss -28(%rdx), %xmm2 mulss -24(%rdx), %xmm7 mulss -20(%rdx), %xmm6 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L274 mulss %xmm5, %xmm4 leaq -8(%rbx), %rdx mulss %xmm3, %xmm1 mulss %xmm7, %xmm6 shrq $3, %rdx leaq 8(,%rdx,8), %rdx mulss %xmm1, %xmm4 .L272: cmpq %rdx, %rbx jle .L275 .L276: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L276 .L275: mulss %xmm2, %xmm0 mulss %xmm6, %xmm0 mulss %xmm4, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L277: movss .LC0(%rip), %xmm4 xorl %edx, %edx movaps %xmm4, %xmm6 movaps %xmm4, %xmm2 movaps %xmm4, %xmm0 jmp .L272 unroll9x9a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -8(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L285 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L282: addq $9, %rcx mulss (%rdx), %xmm0 addq $36, %rdx mulss -32(%rdx), %xmm2 mulss -28(%rdx), %xmm8 mulss -24(%rdx), %xmm7 mulss -20(%rdx), %xmm6 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L282 mulss %xmm6, %xmm5 mulss %xmm4, %xmm3 mulss %xmm8, %xmm7 mulss %xmm3, %xmm5 mulss %xmm1, %xmm5 .L280: cmpq %rcx, %rbx jle .L283 .L284: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L284 .L283: mulss %xmm2, %xmm0 mulss %xmm7, %xmm0 mulss %xmm5, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L285: movss .LC0(%rip), %xmm5 xorl %ecx, %ecx movaps %xmm5, %xmm7 movaps %xmm5, %xmm2 movaps %xmm5, %xmm0 jmp .L280 unroll10x10a_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -9(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L293 movss .LC0(%rip), %xmm1 movq %rax, %rdx xorl %ecx, %ecx movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L290: addq $10, %rcx mulss (%rdx), %xmm0 addq $40, %rdx mulss -36(%rdx), %xmm2 mulss -32(%rdx), %xmm9 mulss -28(%rdx), %xmm8 mulss -24(%rdx), %xmm7 mulss -20(%rdx), %xmm6 mulss -16(%rdx), %xmm5 mulss -12(%rdx), %xmm4 mulss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp jg .L290 mulss %xmm7, %xmm6 mulss %xmm5, %xmm4 mulss %xmm9, %xmm8 mulss %xmm3, %xmm1 mulss %xmm4, %xmm6 .L288: cmpq %rcx, %rbx jle .L291 .L292: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L292 .L291: mulss %xmm2, %xmm0 mulss %xmm8, %xmm0 mulss %xmm6, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L293: movss .LC0(%rip), %xmm1 xorl %ecx, %ecx movaps %xmm1, %xmm6 movaps %xmm1, %xmm8 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jmp .L288 unrollx2as_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length movq %rax, %rbx movq %r12, %rdi movq %rax, %rbp shrq $63, %rbx addq %rax, %rbx sarq %rbx call get_vec_start movss .LC0(%rip), %xmm0 testq %rbx, %rbx movq %rax, %rcx movaps %xmm0, %xmm1 leaq (%rax,%rbx,4), %rax jle .L296 xorl %edx, %edx .L297: mulss (%rcx,%rdx,4), %xmm1 mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L297 .L296: leaq (%rbx,%rbx), %rdx cmpq %rdx, %rbp jle .L298 .L299: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbp, %rdx jne .L299 .L298: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8x2_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L308 movss .LC0(%rip), %xmm1 movq %rcx, %rdx movaps %xmm1, %xmm0 .L305: mulss (%rdx), %xmm0 addq $32, %rdx mulss -28(%rdx), %xmm1 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm1 mulss -16(%rdx), %xmm0 mulss -12(%rdx), %xmm1 mulss -8(%rdx), %xmm0 mulss -4(%rdx), %xmm1 cmpq %rdx, %rax ja .L305 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx .L303: addq $28, %rax cmpq %rcx, %rax jbe .L306 .L307: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L307 .L306: mulss %xmm1, %xmm0 movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret .L308: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 jmp .L303 unroll9x3_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rdx leaq -32(%rax,%r12,4), %rax movss .LC0(%rip), %xmm1 cmpq %rax, %rdx movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 jnb .L311 .L312: mulss (%rdx), %xmm0 addq $36, %rdx mulss -32(%rdx), %xmm2 mulss -28(%rdx), %xmm1 mulss -24(%rdx), %xmm0 mulss -20(%rdx), %xmm2 mulss -16(%rdx), %xmm1 mulss -12(%rdx), %xmm0 mulss -8(%rdx), %xmm2 mulss -4(%rdx), %xmm1 cmpq %rdx, %rax ja .L312 .L311: addq $32, %rax cmpq %rdx, %rax jbe .L313 .L314: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L314 .L313: mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll8x4_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L323 movss .LC0(%rip), %xmm0 movq %rcx, %rdx movaps %xmm0, %xmm1 movaps %xmm0, %xmm2 movaps %xmm0, %xmm3 .L320: mulss (%rdx), %xmm3 addq $32, %rdx mulss -28(%rdx), %xmm2 mulss -24(%rdx), %xmm1 mulss -20(%rdx), %xmm0 mulss -16(%rdx), %xmm3 mulss -12(%rdx), %xmm2 mulss -8(%rdx), %xmm1 mulss -4(%rdx), %xmm0 cmpq %rdx, %rax ja .L320 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx .L318: addq $28, %rax cmpq %rcx, %rax jbe .L321 .L322: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L322 .L321: mulss %xmm3, %xmm0 mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret .L323: movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 movaps %xmm0, %xmm2 movaps %xmm0, %xmm3 jmp .L318 unroll8x8_combine: pushq %r12 pushq %rbp movq %rsi, %rbp pushq %rbx movq %rdi, %rbx call vec_length movq %rbx, %rdi movq %rax, %r12 call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L331 movss .LC0(%rip), %xmm1 movq %rcx, %rdx movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm0 .L328: mulss (%rdx), %xmm0 addq $32, %rdx mulss -28(%rdx), %xmm7 mulss -24(%rdx), %xmm6 mulss -20(%rdx), %xmm5 mulss -16(%rdx), %xmm4 mulss -12(%rdx), %xmm3 mulss -8(%rdx), %xmm2 mulss -4(%rdx), %xmm1 cmpq %rdx, %rax ja .L328 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx .L326: addq $28, %rax cmpq %rcx, %rax jbe .L329 .L330: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L330 .L329: mulss %xmm7, %xmm0 mulss %xmm6, %xmm0 mulss %xmm5, %xmm0 mulss %xmm4, %xmm0 mulss %xmm3, %xmm0 mulss %xmm2, %xmm0 mulss %xmm1, %xmm0 movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret .L331: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm0 jmp .L326 combine7: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -1(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm0 jle .L334 .L336: movss (%rcx,%rdx,4), %xmm1 mulss 4(%rcx,%rdx,4), %xmm1 addq $2, %rdx cmpq %rdx, %rbp mulss %xmm1, %xmm0 jg .L336 leaq -2(%rbx), %rax shrq %rax leaq 2(%rax,%rax), %rdx .L334: cmpq %rdx, %rbx jle .L337 .L338: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L338 .L337: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll3aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -2(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movss .LC0(%rip), %xmm1 jle .L342 .L343: movss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm0 mulss 8(%rax,%rdx,4), %xmm0 addq $3, %rdx cmpq %rdx, %rbp mulss %xmm0, %xmm1 jg .L343 .L342: cmpq %rdx, %rbx jle .L344 .L345: mulss (%rax,%rdx,4), %xmm1 addq $1, %rdx cmpq %rbx, %rdx jne .L345 .L344: movss %xmm1, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -3(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start xorl %edx, %edx testq %rbp, %rbp movq %rax, %rcx movss .LC0(%rip), %xmm1 jle .L349 .L351: movss (%rcx,%rdx,4), %xmm0 movss 8(%rcx,%rdx,4), %xmm2 mulss 4(%rcx,%rdx,4), %xmm0 mulss 12(%rcx,%rdx,4), %xmm2 addq $4, %rdx cmpq %rdx, %rbp mulss %xmm2, %xmm0 mulss %xmm0, %xmm1 jg .L351 leaq -4(%rbx), %rax shrq $2, %rax leaq 4(,%rax,4), %rdx .L349: cmpq %rdx, %rbx jle .L352 .L353: mulss (%rcx,%rdx,4), %xmm1 addq $1, %rdx cmpq %rbx, %rdx jne .L353 .L352: movss %xmm1, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll5aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -4(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L361 movq %rax, %rcx movss .LC0(%rip), %xmm1 xorl %edx, %edx .L358: movss (%rcx), %xmm0 addq $5, %rdx addq $20, %rcx movss -12(%rcx), %xmm2 mulss -16(%rcx), %xmm0 mulss -8(%rcx), %xmm2 mulss %xmm2, %xmm0 mulss -4(%rcx), %xmm0 cmpq %rdx, %rbp mulss %xmm0, %xmm1 jg .L358 .L357: cmpq %rdx, %rbx jle .L359 .L360: mulss (%rax,%rdx,4), %xmm1 addq $1, %rdx cmpq %rbx, %rdx jne .L360 .L359: movss %xmm1, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L361: movss .LC0(%rip), %xmm1 xorl %edx, %edx jmp .L357 unroll6aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -5(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L368 movq %rax, %rdx movss .LC0(%rip), %xmm1 xorl %ecx, %ecx .L365: movss (%rdx), %xmm0 addq $6, %rcx addq $24, %rdx movss -16(%rdx), %xmm2 mulss -20(%rdx), %xmm0 mulss -12(%rdx), %xmm2 mulss %xmm2, %xmm0 movss -8(%rdx), %xmm2 mulss -4(%rdx), %xmm2 cmpq %rcx, %rbp mulss %xmm2, %xmm0 mulss %xmm0, %xmm1 jg .L365 .L364: cmpq %rcx, %rbx jle .L366 .L367: mulss (%rax,%rcx,4), %xmm1 addq $1, %rcx cmpq %rbx, %rcx jne .L367 .L366: movss %xmm1, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L368: movss .LC0(%rip), %xmm1 xorl %ecx, %ecx jmp .L364 unroll7aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -6(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L375 movq %rax, %rdx movss .LC0(%rip), %xmm2 xorl %ecx, %ecx .L372: movss (%rdx), %xmm0 addq $7, %rcx addq $28, %rdx movss -20(%rdx), %xmm1 mulss -24(%rdx), %xmm0 mulss -16(%rdx), %xmm1 mulss %xmm1, %xmm0 movss -12(%rdx), %xmm1 mulss -8(%rdx), %xmm1 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp mulss %xmm1, %xmm0 mulss %xmm0, %xmm2 jg .L372 .L371: cmpq %rcx, %rbx jle .L373 .L374: mulss (%rax,%rcx,4), %xmm2 addq $1, %rcx cmpq %rbx, %rcx jne .L374 .L373: movss %xmm2, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L375: movss .LC0(%rip), %xmm2 xorl %ecx, %ecx jmp .L371 unroll8aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -7(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L383 movq %rax, %rdx movss .LC0(%rip), %xmm2 xorl %ecx, %ecx .L380: movss (%rdx), %xmm0 addq $8, %rcx addq $32, %rdx movss -24(%rdx), %xmm1 mulss -28(%rdx), %xmm0 mulss -20(%rdx), %xmm1 movss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm3 mulss %xmm1, %xmm0 movss -16(%rdx), %xmm1 mulss -12(%rdx), %xmm1 cmpq %rcx, %rbp mulss %xmm3, %xmm1 mulss %xmm1, %xmm0 mulss %xmm0, %xmm2 jg .L380 leaq -8(%rbx), %rdx shrq $3, %rdx leaq 8(,%rdx,8), %rdx .L378: cmpq %rdx, %rbx jle .L381 .L382: mulss (%rax,%rdx,4), %xmm2 addq $1, %rdx cmpq %rbx, %rdx jne .L382 .L381: movss %xmm2, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L383: movss .LC0(%rip), %xmm2 xorl %edx, %edx jmp .L378 unroll9aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -8(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L390 movq %rax, %rdx movss .LC0(%rip), %xmm2 xorl %ecx, %ecx .L387: movss (%rdx), %xmm0 addq $9, %rcx addq $36, %rdx movss -28(%rdx), %xmm1 mulss -32(%rdx), %xmm0 mulss -24(%rdx), %xmm1 movss -12(%rdx), %xmm3 mulss -8(%rdx), %xmm3 mulss %xmm1, %xmm0 movss -20(%rdx), %xmm1 mulss -16(%rdx), %xmm1 mulss %xmm3, %xmm1 mulss -4(%rdx), %xmm1 cmpq %rcx, %rbp mulss %xmm1, %xmm0 mulss %xmm0, %xmm2 jg .L387 .L386: cmpq %rcx, %rbx jle .L388 .L389: mulss (%rax,%rcx,4), %xmm2 addq $1, %rcx cmpq %rbx, %rcx jne .L389 .L388: movss %xmm2, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L390: movss .LC0(%rip), %xmm2 xorl %ecx, %ecx jmp .L386 unroll10aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -9(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L397 movq %rax, %rdx movss .LC0(%rip), %xmm2 xorl %ecx, %ecx .L394: movss (%rdx), %xmm0 addq $10, %rcx addq $40, %rdx movss -32(%rdx), %xmm1 mulss -36(%rdx), %xmm0 mulss -28(%rdx), %xmm1 movss -16(%rdx), %xmm3 mulss -12(%rdx), %xmm3 mulss %xmm1, %xmm0 movss -24(%rdx), %xmm1 mulss -20(%rdx), %xmm1 mulss %xmm3, %xmm1 movss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm3 cmpq %rcx, %rbp mulss %xmm3, %xmm1 mulss %xmm1, %xmm0 mulss %xmm0, %xmm2 jg .L394 .L393: cmpq %rcx, %rbx jle .L395 .L396: mulss (%rax,%rcx,4), %xmm2 addq $1, %rcx cmpq %rbx, %rcx jne .L396 .L395: movss %xmm2, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L397: movss .LC0(%rip), %xmm2 xorl %ecx, %ecx jmp .L393 unroll12aa_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length leaq -11(%rax), %rbp movq %r12, %rdi movq %rax, %rbx call get_vec_start testq %rbp, %rbp jle .L404 movq %rax, %rdx movss .LC0(%rip), %xmm2 xorl %ecx, %ecx .L401: movss (%rdx), %xmm0 addq $12, %rcx addq $48, %rdx movss -40(%rdx), %xmm1 mulss -44(%rdx), %xmm0 mulss -36(%rdx), %xmm1 movss -24(%rdx), %xmm3 mulss -20(%rdx), %xmm3 mulss %xmm1, %xmm0 movss -32(%rdx), %xmm1 mulss -28(%rdx), %xmm1 mulss %xmm3, %xmm1 movss -8(%rdx), %xmm3 mulss -4(%rdx), %xmm3 mulss %xmm1, %xmm0 movss -16(%rdx), %xmm1 mulss -12(%rdx), %xmm1 cmpq %rcx, %rbp mulss %xmm3, %xmm1 mulss %xmm1, %xmm0 mulss %xmm0, %xmm2 jg .L401 .L400: cmpq %rcx, %rbx jle .L402 .L403: mulss (%rax,%rcx,4), %xmm2 addq $1, %rcx cmpq %rbx, %rcx jne .L403 .L402: movss %xmm2, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret .L404: movss .LC0(%rip), %xmm2 xorl %ecx, %ecx jmp .L400 simd_v1_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $208, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm1 testb $31, %bl movl %eax, %edx movss %xmm1, 48(%rsp) movss %xmm1, 52(%rsp) movaps %xmm1, %xmm0 movss %xmm1, 56(%rsp) movss %xmm1, 60(%rsp) movss %xmm1, 64(%rsp) movss %xmm1, 68(%rsp) movss %xmm1, 72(%rsp) movss %xmm1, 76(%rsp) je .L407 testl %eax, %eax jne .L412 jmp .L415 .L409: testl %edx, %edx je .L415 .L412: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L409 .L407: cmpl $7, %edx jbe .L411 movq 48(%rsp), %rax movl %edx, %r10d movq %rax, 176(%rsp) movq 56(%rsp), %rax movq %rax, 184(%rsp) movq 64(%rsp), %rax movq %rax, 192(%rsp) movq 72(%rsp), %rax movq %rax, 200(%rsp) movq %rbx, %rax jmp .L414 .L432: movq %r9, 176(%rsp) movq %r8, 184(%rsp) movq %rdi, 192(%rsp) movq %rcx, 200(%rsp) .L414: movq (%rax), %rcx subl $8, %r10d addq $32, %rax movaps 176(%rsp), %xmm4 movq %rcx, 112(%rsp) movq -24(%rax), %rcx movaps 192(%rsp), %xmm6 movq %rcx, 120(%rsp) movq -16(%rax), %rcx movaps 112(%rsp), %xmm2 movq %rcx, 128(%rsp) movq -8(%rax), %rcx cmpl $7, %r10d mulps %xmm4, %xmm2 movq %rcx, 136(%rsp) movaps 128(%rsp), %xmm3 mulps %xmm6, %xmm3 movaps %xmm2, (%rsp) movq (%rsp), %r9 movaps %xmm2, 16(%rsp) movaps %xmm3, (%rsp) movaps %xmm3, 32(%rsp) movq (%rsp), %rdi movq 24(%rsp), %r8 movq 40(%rsp), %rcx movq %r9, 144(%rsp) movq %r8, 152(%rsp) movq %rdi, 160(%rsp) movq %rcx, 168(%rsp) ja .L432 subl $8, %edx movq %r9, 48(%rsp) movq %r9, 80(%rsp) movl %edx, %esi movq %r8, 88(%rsp) movq %rdi, 96(%rsp) shrl $3, %esi movq %rcx, 104(%rsp) movq %r8, 56(%rsp) movl %esi, %eax negl %esi movq %rdi, 64(%rsp) addq $1, %rax leal (%rdx,%rsi,8), %edx movq %rcx, 72(%rsp) salq $5, %rax movss 48(%rsp), %xmm1 addq %rax, %rbx .L411: testl %edx, %edx je .L415 .L416: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 jne .L416 .L415: leaq 52(%rsp), %rax leaq 80(%rsp), %rdx jmp .L418 .L417: movss (%rax), %xmm1 addq $4, %rax .L418: cmpq %rdx, %rax mulss %xmm1, %xmm0 jne .L417 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v2_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $320, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm0 testb $31, %bl movss %xmm0, 64(%rsp) movl %eax, %ecx movss %xmm0, 68(%rsp) movq 64(%rsp), %rdx movss %xmm0, 72(%rsp) movss %xmm0, 76(%rsp) movq %rdx, 256(%rsp) movss %xmm0, 80(%rsp) movq 72(%rsp), %rdx movss %xmm0, 84(%rsp) movss %xmm0, 88(%rsp) movss %xmm0, 92(%rsp) movq %rdx, 264(%rsp) movq 80(%rsp), %rdx movq %rdx, 272(%rsp) movq 88(%rsp), %rdx movq %rdx, 280(%rsp) je .L434 testl %eax, %eax jne .L439 jmp .L435 .L436: testl %ecx, %ecx je .L435 .L439: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L436 .L434: cmpl $15, %ecx movq 256(%rsp), %rax jbe .L459 movq %rax, 96(%rsp) movq 264(%rsp), %rax movl %ecx, %r10d movq %rbx, %rdx movq %rax, 104(%rsp) movq 272(%rsp), %rax movq %rax, 112(%rsp) movq 280(%rsp), %rax movq %rax, 120(%rsp) jmp .L441 .L460: movq %r9, 96(%rsp) movq %r8, 104(%rsp) movq %rdi, 112(%rsp) movq %rax, 120(%rsp) .L441: movq (%rdx), %rax subl $16, %r10d addq $64, %rdx movaps 96(%rsp), %xmm5 movq %rax, 160(%rsp) movq -56(%rdx), %rax movaps 112(%rsp), %xmm2 movq %rax, 168(%rsp) movq -48(%rdx), %rax movaps 160(%rsp), %xmm1 movq %rax, 176(%rsp) movq -40(%rdx), %rax mulps %xmm5, %xmm1 movaps 256(%rsp), %xmm3 movq %rax, 184(%rsp) movq -32(%rdx), %rax movaps 176(%rsp), %xmm7 movq %rax, 192(%rsp) movq -24(%rdx), %rax mulps %xmm7, %xmm2 movaps %xmm1, 16(%rsp) movq %rax, 200(%rsp) movq -16(%rdx), %rax movaps 192(%rsp), %xmm6 movq %rax, 208(%rsp) movq -8(%rdx), %rax cmpl $15, %r10d mulps %xmm6, %xmm3 movq 16(%rsp), %r9 movaps %xmm1, 32(%rsp) movq %rax, 216(%rsp) movq 40(%rsp), %r8 movaps %xmm2, 16(%rsp) movaps %xmm2, 48(%rsp) movq 16(%rsp), %rdi movq 56(%rsp), %rax movq %r9, 224(%rsp) movq %r8, 232(%rsp) movq %rdi, 240(%rsp) movq %rax, 248(%rsp) movaps %xmm3, (%rsp) movq (%rsp), %rsi movaps %xmm3, 32(%rsp) movaps 272(%rsp), %xmm4 movq %rsi, 256(%rsp) movq 40(%rsp), %rsi movaps 208(%rsp), %xmm7 mulps %xmm7, %xmm4 movq %rsi, 264(%rsp) movaps %xmm4, 16(%rsp) movq 16(%rsp), %r11 movaps %xmm4, 48(%rsp) movq %r11, 272(%rsp) movq 56(%rsp), %r11 movq %r11, 280(%rsp) ja .L460 subl $16, %ecx movq %r9, 128(%rsp) movq %r8, 136(%rsp) movl %ecx, %esi movq %rdi, 144(%rsp) movq %rax, 152(%rsp) shrl $4, %esi movl %esi, %edx sall $4, %esi addq $1, %rdx subl %esi, %ecx salq $6, %rdx addq %rdx, %rbx .L438: testl %ecx, %ecx je .L442 .L443: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L443 .L442: movaps 256(%rsp), %xmm3 mulps 128(%rsp), %xmm3 movaps 144(%rsp), %xmm4 movaps %xmm3, 16(%rsp) movq 16(%rsp), %rdx movaps %xmm3, 32(%rsp) movaps 272(%rsp), %xmm3 movq %rdx, 64(%rsp) movq 40(%rsp), %rdx mulps %xmm3, %xmm4 movq %rdx, 72(%rsp) xorl %edx, %edx movaps %xmm4, 16(%rsp) movq 16(%rsp), %rax movaps %xmm4, 48(%rsp) movq %rax, 80(%rsp) movq 56(%rsp), %rax movq %rax, 88(%rsp) xorl %eax, %eax .L445: addq $1, %rax mulss 64(%rsp,%rdx,4), %xmm0 cmpq $8, %rax movq %rax, %rdx jne .L445 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret .L435: movq 256(%rsp), %rax movq %rax, 128(%rsp) movq 264(%rsp), %rax movq %rax, 136(%rsp) movq 272(%rsp), %rax movq %rax, 144(%rsp) movq 280(%rsp), %rax movq %rax, 152(%rsp) jmp .L442 .L459: movq %rax, 128(%rsp) movq 264(%rsp), %rax movq %rax, 136(%rsp) movq 272(%rsp), %rax movq %rax, 144(%rsp) movq 280(%rsp), %rax movq %rax, 152(%rsp) jmp .L438 simd_v4_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $544, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm0 testb $31, %bl movss %xmm0, 96(%rsp) movl %eax, %ecx movss %xmm0, 100(%rsp) movq 96(%rsp), %rdx movss %xmm0, 104(%rsp) movss %xmm0, 108(%rsp) movq %rdx, 128(%rsp) movss %xmm0, 112(%rsp) movq 104(%rsp), %rdx movss %xmm0, 116(%rsp) movss %xmm0, 120(%rsp) movss %xmm0, 124(%rsp) movq %rdx, 136(%rsp) movq 112(%rsp), %rdx movq %rdx, 144(%rsp) movq 120(%rsp), %rdx movq %rdx, 152(%rsp) je .L462 testl %eax, %eax jne .L467 jmp .L463 .L464: testl %ecx, %ecx je .L463 .L467: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L464 .L462: cmpl $31, %ecx movq 128(%rsp), %rdi jbe .L487 movq 144(%rsp), %rdx movq 136(%rsp), %rsi movl %ecx, %r10d movq 152(%rsp), %rax movq %rdi, 480(%rsp) movq %rdi, 192(%rsp) movq %rdi, 256(%rsp) movq %rdx, 496(%rsp) movq %rdx, 208(%rsp) movq %rdx, 272(%rsp) movq %rsi, 488(%rsp) movq %rbx, %rdx movq %rax, 504(%rsp) movq %rsi, 200(%rsp) movq %rax, 216(%rsp) movq %rsi, 264(%rsp) movq %rax, 280(%rsp) jmp .L469 .L488: movq %r9, 480(%rsp) movq %r8, 488(%rsp) movq %rdi, 496(%rsp) movq %rax, 504(%rsp) .L469: movq (%rdx), %rax subl $32, %r10d subq $-128, %rdx movaps 480(%rsp), %xmm2 movq %rax, 288(%rsp) movq -120(%rdx), %rax movq %rax, 296(%rsp) movq -112(%rdx), %rax movaps 288(%rsp), %xmm1 movq %rax, 304(%rsp) movq -104(%rdx), %rax mulps %xmm2, %xmm1 movq %rax, 312(%rsp) movq -96(%rdx), %rax movq %rax, 320(%rsp) movq -88(%rdx), %rax movq %rax, 328(%rsp) movq -80(%rdx), %rax movq %rax, 336(%rsp) movq -72(%rdx), %rax movq %rax, 344(%rsp) movq -64(%rdx), %rax movq %rax, 352(%rsp) movq -56(%rdx), %rax movq %rax, 360(%rsp) movq -48(%rdx), %rax movq %rax, 368(%rsp) movq -40(%rdx), %rax movq %rax, 376(%rsp) movq -32(%rdx), %rax movq %rax, 384(%rsp) movq -24(%rdx), %rax movq %rax, 392(%rsp) movq -16(%rdx), %rax movq %rax, 400(%rsp) movq -8(%rdx), %rax cmpl $31, %r10d movq %rax, 408(%rsp) movaps %xmm1, 48(%rsp) movq 48(%rsp), %r9 movaps %xmm1, 64(%rsp) movaps 496(%rsp), %xmm4 movq 72(%rsp), %r8 movq %r9, 416(%rsp) movaps 304(%rsp), %xmm2 movaps 336(%rsp), %xmm7 mulps %xmm4, %xmm2 movq %r8, 424(%rsp) movaps 272(%rsp), %xmm4 mulps %xmm7, %xmm4 movaps 192(%rsp), %xmm5 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rdi movaps %xmm2, 80(%rsp) movaps %xmm4, 48(%rsp) movq 88(%rsp), %rax movq 48(%rsp), %rsi movaps %xmm4, 80(%rsp) movq %rdi, 432(%rsp) movaps 352(%rsp), %xmm4 movq %rsi, 272(%rsp) movq 88(%rsp), %rsi movaps 256(%rsp), %xmm3 mulps %xmm4, %xmm5 movq %rax, 440(%rsp) movaps 320(%rsp), %xmm6 movq %rsi, 280(%rsp) mulps %xmm6, %xmm3 movaps 208(%rsp), %xmm6 movaps %xmm5, 48(%rsp) movaps %xmm3, 32(%rsp) movq 32(%rsp), %r11 movaps %xmm3, 64(%rsp) movaps 368(%rsp), %xmm5 movq %r11, 256(%rsp) movq 72(%rsp), %r11 mulps %xmm5, %xmm6 movq %r11, 264(%rsp) movaps %xmm6, 32(%rsp) movq 48(%rsp), %r11 movq 32(%rsp), %rsi movaps 32(%rsp), %xmm7 movq %r11, 192(%rsp) movq %rsi, 208(%rsp) movaps %xmm7, 80(%rsp) movq 88(%rsp), %rsi movaps 128(%rsp), %xmm7 movq %rsi, 216(%rsp) movaps 384(%rsp), %xmm1 movaps 144(%rsp), %xmm2 mulps %xmm1, %xmm7 movaps 400(%rsp), %xmm1 movaps 48(%rsp), %xmm6 mulps %xmm2, %xmm1 movaps %xmm6, 64(%rsp) movq 72(%rsp), %r11 movaps %xmm7, 48(%rsp) movaps %xmm1, 32(%rsp) movaps %xmm7, 64(%rsp) movaps %xmm1, 80(%rsp) movq 48(%rsp), %rsi movq %r11, 200(%rsp) movq 32(%rsp), %r11 movq %rsi, 128(%rsp) movq 72(%rsp), %rsi movq %r11, 144(%rsp) movq 88(%rsp), %r11 movq %rsi, 136(%rsp) movq %r11, 152(%rsp) ja .L488 subl $32, %ecx movq %r9, 512(%rsp) movq %r8, 520(%rsp) movl %ecx, %esi movq %rdi, 528(%rsp) movq %rax, 536(%rsp) shrl $5, %esi movl %esi, %edx sall $5, %esi addq $1, %rdx subl %esi, %ecx salq $7, %rdx addq %rdx, %rbx .L466: testl %ecx, %ecx je .L470 .L471: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L471 .L470: movaps 528(%rsp), %xmm6 mulps 272(%rsp), %xmm6 movaps 512(%rsp), %xmm5 movaps 128(%rsp), %xmm7 mulps 256(%rsp), %xmm5 movaps %xmm6, 80(%rsp) movaps %xmm6, 32(%rsp) movaps %xmm5, 64(%rsp) movaps %xmm5, 48(%rsp) movaps 192(%rsp), %xmm6 movq 48(%rsp), %rdx movq 32(%rsp), %rax mulps %xmm6, %xmm7 movaps 144(%rsp), %xmm5 movq %rdx, 448(%rsp) movq %rax, 464(%rsp) movq 72(%rsp), %rdx movq 88(%rsp), %rax movaps %xmm7, 16(%rsp) movq %rdx, 456(%rsp) movq %rax, 472(%rsp) movaps 208(%rsp), %xmm7 movq 16(%rsp), %rdx mulps %xmm7, %xmm5 movq %rdx, 224(%rsp) movaps %xmm5, (%rsp) movq (%rsp), %rax movaps 16(%rsp), %xmm5 movq %rax, 240(%rsp) movaps (%rsp), %xmm6 movaps %xmm5, 64(%rsp) movq 72(%rsp), %rdx movaps %xmm6, 80(%rsp) movq 88(%rsp), %rax movq %rdx, 232(%rsp) movq %rax, 248(%rsp) movaps 32(%rsp), %xmm7 movaps 48(%rsp), %xmm6 mulps (%rsp), %xmm7 mulps %xmm5, %xmm6 movaps %xmm6, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm7, 48(%rsp) movaps %xmm6, 64(%rsp) movaps %xmm7, 80(%rsp) movq 48(%rsp), %rax movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rdx, 104(%rsp) xorl %edx, %edx movq %rax, 120(%rsp) xorl %eax, %eax .L473: addq $1, %rax mulss 96(%rsp,%rdx,4), %xmm0 cmpq $8, %rax movq %rax, %rdx jne .L473 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret .L463: movq 128(%rsp), %rsi movq 136(%rsp), %rcx movq 144(%rsp), %rdx movq 152(%rsp), %rax movq %rsi, 192(%rsp) movq %rcx, 200(%rsp) movq %rdx, 208(%rsp) movq %rax, 216(%rsp) movq %rsi, 256(%rsp) movq %rcx, 264(%rsp) movq %rdx, 272(%rsp) movq %rax, 280(%rsp) movq %rsi, 512(%rsp) movq %rcx, 520(%rsp) movq %rdx, 528(%rsp) movq %rax, 536(%rsp) jmp .L470 .L487: movq 136(%rsp), %rsi movq 144(%rsp), %rdx movq 152(%rsp), %rax movq %rdi, 512(%rsp) movq %rdi, 192(%rsp) movq %rdi, 256(%rsp) movq %rsi, 520(%rsp) movq %rdx, 528(%rsp) movq %rax, 536(%rsp) movq %rsi, 200(%rsp) movq %rdx, 208(%rsp) movq %rax, 216(%rsp) movq %rsi, 264(%rsp) movq %rdx, 272(%rsp) movq %rax, 280(%rsp) jmp .L466 simd_v8_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $928, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm0 testb $31, %bl movss %xmm0, 96(%rsp) movl %eax, %ecx movss %xmm0, 100(%rsp) movq 96(%rsp), %rdx movss %xmm0, 104(%rsp) movss %xmm0, 108(%rsp) movq %rdx, 128(%rsp) movss %xmm0, 112(%rsp) movq 104(%rsp), %rdx movss %xmm0, 116(%rsp) movss %xmm0, 120(%rsp) movss %xmm0, 124(%rsp) movq %rdx, 136(%rsp) movq 112(%rsp), %rdx movq %rdx, 144(%rsp) movq 120(%rsp), %rdx movq %rdx, 152(%rsp) je .L490 testl %eax, %eax jne .L495 jmp .L491 .L492: testl %ecx, %ecx je .L491 .L495: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L492 .L490: cmpl $63, %ecx jbe .L515 movq 128(%rsp), %r8 movq 136(%rsp), %rdi movl %ecx, %r10d movq 144(%rsp), %rsi movq 152(%rsp), %rax movq %rbx, %rdx movq %r8, 352(%rsp) movq %rdi, 360(%rsp) movq %rsi, 368(%rsp) movq %rax, 376(%rsp) movq %r8, 832(%rsp) movq %rdi, 840(%rsp) movq %rsi, 848(%rsp) movq %rax, 856(%rsp) movq %r8, 800(%rsp) movq %rdi, 808(%rsp) movq %rsi, 816(%rsp) movq %rax, 824(%rsp) movq %r8, 768(%rsp) movq %rdi, 776(%rsp) movq %rsi, 784(%rsp) movq %rax, 792(%rsp) movq %r8, 736(%rsp) movq %rdi, 744(%rsp) movq %rsi, 752(%rsp) movq %rax, 760(%rsp) movq %r8, 384(%rsp) movq %rdi, 392(%rsp) movq %rsi, 400(%rsp) movq %rax, 408(%rsp) movq %r8, 416(%rsp) movq %rdi, 424(%rsp) movq %rsi, 432(%rsp) movq %rax, 440(%rsp) jmp .L497 .L516: movq %r9, 352(%rsp) movq %r8, 360(%rsp) movq %rdi, 368(%rsp) movq %rax, 376(%rsp) .L497: movq (%rdx), %rax subl $64, %r10d addq $256, %rdx movq %rax, 448(%rsp) movq -248(%rdx), %rax movq %rax, 456(%rsp) movq -240(%rdx), %rax movq %rax, 464(%rsp) movq -232(%rdx), %rax movq %rax, 472(%rsp) movq -224(%rdx), %rax movq %rax, 480(%rsp) movq -216(%rdx), %rax movq %rax, 488(%rsp) movq -208(%rdx), %rax movq %rax, 496(%rsp) movq -200(%rdx), %rax movq %rax, 504(%rsp) movq -192(%rdx), %rax movq %rax, 512(%rsp) movq -184(%rdx), %rax movq %rax, 520(%rsp) movq -176(%rdx), %rax movq %rax, 528(%rsp) movq -168(%rdx), %rax movq %rax, 536(%rsp) movq -160(%rdx), %rax movq %rax, 544(%rsp) movq -152(%rdx), %rax movq %rax, 552(%rsp) movq -144(%rdx), %rax movq %rax, 560(%rsp) movq -136(%rdx), %rax movq %rax, 568(%rsp) movq -128(%rdx), %rax movq %rax, 576(%rsp) movq -120(%rdx), %rax movaps 352(%rsp), %xmm3 movq %rax, 584(%rsp) movq -112(%rdx), %rax movaps 448(%rsp), %xmm1 movq %rax, 592(%rsp) movq -104(%rdx), %rax mulps %xmm3, %xmm1 movq %rax, 600(%rsp) movq -96(%rdx), %rax movq %rax, 608(%rsp) movq -88(%rdx), %rax movaps %xmm1, 48(%rsp) movq 48(%rsp), %r9 movq %rax, 616(%rsp) movq -80(%rdx), %rax movq %rax, 624(%rsp) movq -72(%rdx), %rax movq %rax, 632(%rsp) movq -64(%rdx), %rax movq %rax, 640(%rsp) movq -56(%rdx), %rax movq %rax, 648(%rsp) movq -48(%rdx), %rax movq %rax, 656(%rsp) movq -40(%rdx), %rax movq %rax, 664(%rsp) movq -32(%rdx), %rax movq %rax, 672(%rsp) movq -24(%rdx), %rax movq %rax, 680(%rsp) movq -16(%rdx), %rax movq %rax, 688(%rsp) movq -8(%rdx), %rax cmpl $63, %r10d movq %rax, 696(%rsp) movaps %xmm1, 64(%rsp) movq 72(%rsp), %r8 movq %r9, 704(%rsp) movaps 416(%rsp), %xmm3 movq %r8, 712(%rsp) movaps 480(%rsp), %xmm7 mulps %xmm7, %xmm3 movaps 432(%rsp), %xmm4 movaps 368(%rsp), %xmm5 movaps %xmm3, 32(%rsp) movq 32(%rsp), %r11 movaps 496(%rsp), %xmm3 movq %r11, 416(%rsp) movaps 464(%rsp), %xmm2 mulps %xmm3, %xmm4 mulps %xmm5, %xmm2 movaps 512(%rsp), %xmm6 movaps 528(%rsp), %xmm7 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rdi movaps %xmm4, 48(%rsp) movaps %xmm2, 80(%rsp) movaps 48(%rsp), %xmm5 movq 88(%rsp), %rax movq 48(%rsp), %rsi movaps %xmm5, 80(%rsp) movq %rdi, 720(%rsp) movaps 384(%rsp), %xmm5 movq %rsi, 432(%rsp) movq 88(%rsp), %rsi mulps %xmm6, %xmm5 movaps 400(%rsp), %xmm6 movq %rax, 728(%rsp) movaps 32(%rsp), %xmm4 mulps %xmm7, %xmm6 movq %rsi, 440(%rsp) movaps %xmm4, 64(%rsp) movq 72(%rsp), %r11 movaps %xmm5, 32(%rsp) movq %r11, 424(%rsp) movaps %xmm6, 48(%rsp) movq 32(%rsp), %r11 movq 48(%rsp), %rsi movaps %xmm5, 64(%rsp) movaps %xmm6, 80(%rsp) movaps 128(%rsp), %xmm7 movq %r11, 384(%rsp) movq %rsi, 400(%rsp) movaps 144(%rsp), %xmm1 movq 72(%rsp), %r11 movq 88(%rsp), %rsi movaps 544(%rsp), %xmm3 movaps 560(%rsp), %xmm4 mulps %xmm3, %xmm7 movq %r11, 392(%rsp) movq %rsi, 408(%rsp) movaps 736(%rsp), %xmm2 mulps %xmm4, %xmm1 movaps 752(%rsp), %xmm3 movaps %xmm7, 32(%rsp) movq 32(%rsp), %r11 movaps %xmm1, 48(%rsp) movaps %xmm7, 64(%rsp) movaps %xmm1, 80(%rsp) movaps 576(%rsp), %xmm7 movq 48(%rsp), %rsi movq %r11, 128(%rsp) movaps 592(%rsp), %xmm1 mulps %xmm7, %xmm2 movq 72(%rsp), %r11 mulps %xmm1, %xmm3 movq %rsi, 144(%rsp) movq 88(%rsp), %rsi movq %r11, 136(%rsp) movaps %xmm2, 32(%rsp) movq %rsi, 152(%rsp) movq 32(%rsp), %r11 movaps %xmm3, 48(%rsp) movaps %xmm2, 64(%rsp) movaps %xmm3, 80(%rsp) movaps 608(%rsp), %xmm4 movq 48(%rsp), %rsi movq %r11, 736(%rsp) movaps 624(%rsp), %xmm5 mulps 768(%rsp), %xmm4 movq 72(%rsp), %r11 movq %rsi, 752(%rsp) mulps 784(%rsp), %xmm5 movq 88(%rsp), %rsi movaps 640(%rsp), %xmm1 movq %r11, 744(%rsp) movaps 656(%rsp), %xmm2 movq %rsi, 760(%rsp) movaps 800(%rsp), %xmm6 movaps 816(%rsp), %xmm7 mulps %xmm1, %xmm6 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rsi mulps %xmm2, %xmm7 movaps %xmm4, 32(%rsp) movaps %xmm4, 64(%rsp) movaps %xmm5, 80(%rsp) movaps %xmm7, 48(%rsp) movq 32(%rsp), %r11 movq %rsi, 784(%rsp) movq 88(%rsp), %rsi movaps %xmm6, 32(%rsp) movq %r11, 768(%rsp) movq 72(%rsp), %r11 movaps %xmm6, 64(%rsp) movq %r11, 776(%rsp) movq %rsi, 792(%rsp) movq 32(%rsp), %r11 movq 48(%rsp), %rsi movaps %xmm7, 80(%rsp) movaps 832(%rsp), %xmm5 movq %r11, 800(%rsp) movq %rsi, 816(%rsp) movaps 848(%rsp), %xmm2 movq 72(%rsp), %r11 movq 88(%rsp), %rsi movaps 672(%rsp), %xmm1 movaps 688(%rsp), %xmm6 mulps %xmm5, %xmm1 movq %r11, 808(%rsp) movq %rsi, 824(%rsp) mulps %xmm6, %xmm2 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rsi movaps %xmm2, 32(%rsp) movaps %xmm1, 64(%rsp) movaps %xmm2, 80(%rsp) movq 32(%rsp), %r11 movq %rsi, 832(%rsp) movq 72(%rsp), %rsi movq %r11, 848(%rsp) movq 88(%rsp), %r11 movq %rsi, 840(%rsp) movq %r11, 856(%rsp) ja .L516 subl $64, %ecx movq %r9, 896(%rsp) movq %r8, 904(%rsp) movl %ecx, %esi movq %rdi, 912(%rsp) movq %rax, 920(%rsp) shrl $6, %esi movl %esi, %edx sall $6, %esi addq $1, %rdx subl %esi, %ecx salq $8, %rdx addq %rdx, %rbx .L494: testl %ecx, %ecx je .L498 .L499: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L499 .L498: movaps 416(%rsp), %xmm7 movaps 896(%rsp), %xmm2 mulps %xmm7, %xmm2 movaps 384(%rsp), %xmm4 movaps 400(%rsp), %xmm5 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rdx movaps 432(%rsp), %xmm2 movq %rdx, 320(%rsp) mulps 912(%rsp), %xmm2 movaps 48(%rsp), %xmm7 movaps %xmm7, 64(%rsp) movq 72(%rsp), %rdx movaps %xmm7, %xmm3 movaps %xmm2, 80(%rsp) movq %rdx, 328(%rsp) movaps %xmm2, 32(%rsp) movaps 128(%rsp), %xmm2 movq 32(%rsp), %rax mulps %xmm4, %xmm2 movq %rax, 336(%rsp) movq 88(%rsp), %rax movaps %xmm2, 16(%rsp) movq %rax, 344(%rsp) movq 16(%rsp), %rdx movaps 144(%rsp), %xmm2 movq %rdx, 288(%rsp) mulps %xmm5, %xmm2 movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 72(%rsp), %rdx mulps %xmm6, %xmm3 movaps %xmm2, (%rsp) movq (%rsp), %rax movq %rdx, 296(%rsp) movaps %xmm2, 80(%rsp) movaps %xmm3, %xmm7 movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rax, 312(%rsp) movaps 32(%rsp), %xmm4 mulps %xmm2, %xmm4 movaps %xmm3, 64(%rsp) movaps %xmm3, 16(%rsp) movaps %xmm4, 80(%rsp) movaps %xmm4, %xmm3 movaps %xmm4, (%rsp) movaps 768(%rsp), %xmm5 movq 16(%rsp), %rdx movq (%rsp), %rax movaps 736(%rsp), %xmm4 mulps %xmm4, %xmm5 movq %rdx, 256(%rsp) movaps 784(%rsp), %xmm6 movq 72(%rsp), %rdx movq %rax, 272(%rsp) movq 88(%rsp), %rax movaps %xmm5, 48(%rsp) movq %rdx, 264(%rsp) movq %rax, 280(%rsp) movaps 752(%rsp), %xmm5 movq 48(%rsp), %rdx mulps %xmm5, %xmm6 movq %rdx, 224(%rsp) movaps %xmm6, 32(%rsp) movq 32(%rsp), %rax movaps 32(%rsp), %xmm2 movq %rax, 240(%rsp) movaps 48(%rsp), %xmm6 movaps %xmm2, 80(%rsp) mulps %xmm6, %xmm7 movq 88(%rsp), %rax movaps %xmm6, 64(%rsp) movaps 32(%rsp), %xmm2 movq 72(%rsp), %rdx movq %rax, 248(%rsp) mulps %xmm3, %xmm2 movaps %xmm7, 16(%rsp) movq %rdx, 232(%rsp) movaps %xmm2, (%rsp) movq 16(%rsp), %rdx movq (%rsp), %rax movaps %xmm7, 64(%rsp) movaps %xmm2, 80(%rsp) movaps 800(%rsp), %xmm4 movq %rdx, 192(%rsp) movq %rax, 208(%rsp) movaps 832(%rsp), %xmm3 movq 72(%rsp), %rdx movq 88(%rsp), %rax mulps %xmm4, %xmm3 movaps 816(%rsp), %xmm5 movaps 848(%rsp), %xmm4 movq %rdx, 200(%rsp) movq %rax, 216(%rsp) mulps %xmm5, %xmm4 movaps (%rsp), %xmm6 movaps %xmm7, %xmm5 movaps %xmm3, 48(%rsp) movq 48(%rsp), %rdx mulps %xmm3, %xmm5 movaps %xmm3, 64(%rsp) mulps %xmm4, %xmm6 movaps %xmm4, 32(%rsp) movaps %xmm4, 80(%rsp) movaps %xmm5, 48(%rsp) movq 32(%rsp), %rax movq %rdx, 160(%rsp) movq 72(%rsp), %rdx movaps %xmm5, 64(%rsp) movq %rax, 176(%rsp) movq 88(%rsp), %rax movaps %xmm6, 80(%rsp) movq %rdx, 168(%rsp) movq %rax, 184(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 48(%rsp) movq 48(%rsp), %rax movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rdx, 104(%rsp) xorl %edx, %edx movq %rax, 120(%rsp) xorl %eax, %eax .L501: addq $1, %rax mulss 96(%rsp,%rdx,4), %xmm0 cmpq $8, %rax movq %rax, %rdx jne .L501 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret .L491: movq 128(%rsp), %rsi movq 136(%rsp), %rcx movq 144(%rsp), %rdx movq 152(%rsp), %rax movq %rsi, 896(%rsp) movq %rcx, 904(%rsp) movq %rdx, 912(%rsp) movq %rax, 920(%rsp) movq %rsi, 832(%rsp) movq %rcx, 840(%rsp) movq %rdx, 848(%rsp) movq %rax, 856(%rsp) movq %rsi, 800(%rsp) movq %rcx, 808(%rsp) movq %rdx, 816(%rsp) movq %rax, 824(%rsp) movq %rsi, 768(%rsp) movq %rcx, 776(%rsp) movq %rdx, 784(%rsp) movq %rax, 792(%rsp) movq %rsi, 736(%rsp) movq %rcx, 744(%rsp) movq %rdx, 752(%rsp) movq %rax, 760(%rsp) movq %rsi, 384(%rsp) movq %rcx, 392(%rsp) movq %rdx, 400(%rsp) movq %rax, 408(%rsp) movq %rsi, 416(%rsp) movq %rcx, 424(%rsp) movq %rdx, 432(%rsp) movq %rax, 440(%rsp) jmp .L498 .L515: movq 128(%rsp), %rdi movq 136(%rsp), %rsi movq 144(%rsp), %rdx movq 152(%rsp), %rax movq %rdi, 896(%rsp) movq %rsi, 904(%rsp) movq %rdx, 912(%rsp) movq %rax, 920(%rsp) movq %rdi, 832(%rsp) movq %rsi, 840(%rsp) movq %rdx, 848(%rsp) movq %rax, 856(%rsp) movq %rdi, 800(%rsp) movq %rsi, 808(%rsp) movq %rdx, 816(%rsp) movq %rax, 824(%rsp) movq %rdi, 768(%rsp) movq %rsi, 776(%rsp) movq %rdx, 784(%rsp) movq %rax, 792(%rsp) movq %rdi, 736(%rsp) movq %rsi, 744(%rsp) movq %rdx, 752(%rsp) movq %rax, 760(%rsp) movq %rdi, 384(%rsp) movq %rsi, 392(%rsp) movq %rdx, 400(%rsp) movq %rax, 408(%rsp) movq %rdi, 416(%rsp) movq %rsi, 424(%rsp) movq %rdx, 432(%rsp) movq %rax, 440(%rsp) jmp .L494 simd_v10_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $1120, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm0 testb $31, %bl movss %xmm0, 96(%rsp) movl %eax, %edx movss %xmm0, 100(%rsp) movq 96(%rsp), %rcx movss %xmm0, 104(%rsp) movss %xmm0, 108(%rsp) movq %rcx, 736(%rsp) movss %xmm0, 112(%rsp) movq 104(%rsp), %rcx movss %xmm0, 116(%rsp) movss %xmm0, 120(%rsp) movss %xmm0, 124(%rsp) movq %rcx, 744(%rsp) movq 112(%rsp), %rcx movq %rcx, 752(%rsp) movq 120(%rsp), %rcx movq %rcx, 760(%rsp) je .L518 testl %eax, %eax jne .L523 jmp .L519 .L520: testl %edx, %edx je .L519 .L523: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L520 .L518: cmpl $79, %edx movl %edx, %r11d jbe .L543 movq 736(%rsp), %rsi movq 752(%rsp), %rdx movq 760(%rsp), %rax movq 744(%rsp), %rcx movq %rsi, 1056(%rsp) movq %rdx, 1072(%rsp) movq %rcx, 1064(%rsp) movq %rax, 1080(%rsp) movq %rsi, 704(%rsp) movq %rcx, 712(%rsp) movq %rdx, 720(%rsp) movq %rax, 728(%rsp) movq %rsi, 672(%rsp) movq %rcx, 680(%rsp) movq %rdx, 688(%rsp) movq %rax, 696(%rsp) movq %rsi, 640(%rsp) movq %rcx, 648(%rsp) movq %rdx, 656(%rsp) movq %rax, 664(%rsp) movq %rsi, 608(%rsp) movq %rcx, 616(%rsp) movq %rdx, 624(%rsp) movq %rax, 632(%rsp) movq %rsi, 576(%rsp) movq %rcx, 584(%rsp) movq %rdx, 592(%rsp) movq %rax, 600(%rsp) movq %rsi, 544(%rsp) movq %rcx, 552(%rsp) movq %rdx, 560(%rsp) movq %rax, 568(%rsp) movq %rsi, 512(%rsp) movq %rcx, 520(%rsp) movq %rdx, 528(%rsp) movq %rax, 536(%rsp) movq %rsi, 480(%rsp) movq %rcx, 488(%rsp) movq %rdx, 496(%rsp) movq %rax, 504(%rsp) jmp .L525 .L544: movl %edx, %r11d movq %r8, 1056(%rsp) movq %rdi, 1064(%rsp) movq %rcx, 1072(%rsp) movq %rax, 1080(%rsp) .L525: movq (%rbx), %rax addq $320, %rbx movq %rax, 128(%rsp) movq -312(%rbx), %rax movq %rax, 136(%rsp) movq -304(%rbx), %rax movq %rax, 144(%rsp) movq -296(%rbx), %rax movq %rax, 152(%rsp) movq -288(%rbx), %rax movq %rax, 160(%rsp) movq -280(%rbx), %rax movq %rax, 168(%rsp) movq -272(%rbx), %rax movq %rax, 176(%rsp) movq -264(%rbx), %rax movq %rax, 184(%rsp) movq -256(%rbx), %rax movq %rax, 192(%rsp) movq -248(%rbx), %rax movq %rax, 200(%rsp) movq -240(%rbx), %rax movq %rax, 208(%rsp) movq -232(%rbx), %rax movq %rax, 216(%rsp) movq -224(%rbx), %rax movq %rax, 224(%rsp) movq -216(%rbx), %rax movq %rax, 232(%rsp) movq -208(%rbx), %rax movq %rax, 240(%rsp) movq -200(%rbx), %rax movq %rax, 248(%rsp) movq -192(%rbx), %rax movq %rax, 256(%rsp) movq -184(%rbx), %rax movq %rax, 264(%rsp) movq -176(%rbx), %rax movq %rax, 272(%rsp) movq -168(%rbx), %rax movq %rax, 280(%rsp) movq -160(%rbx), %rax movq %rax, 288(%rsp) movq -152(%rbx), %rax movq %rax, 296(%rsp) movq -144(%rbx), %rax movq %rax, 304(%rsp) movq -136(%rbx), %rax movq %rax, 312(%rsp) movq -128(%rbx), %rax movq %rax, 320(%rsp) movq -120(%rbx), %rax movq %rax, 328(%rsp) movq -112(%rbx), %rax movq %rax, 336(%rsp) movq -104(%rbx), %rax movq %rax, 344(%rsp) movq -96(%rbx), %rax movq %rax, 352(%rsp) movq -88(%rbx), %rax movq %rax, 360(%rsp) movq -80(%rbx), %rax movq %rax, 368(%rsp) movq -72(%rbx), %rax movq %rax, 376(%rsp) movq -64(%rbx), %rax movq %rax, 384(%rsp) movq -56(%rbx), %rax movq %rax, 392(%rsp) movq -48(%rbx), %rax movaps 1056(%rsp), %xmm7 movq %rax, 400(%rsp) movq -40(%rbx), %rax movaps 128(%rsp), %xmm1 movq %rax, 408(%rsp) movq -32(%rbx), %rax mulps %xmm7, %xmm1 movaps 1072(%rsp), %xmm2 movq %rax, 416(%rsp) movq -24(%rbx), %rax movaps 160(%rsp), %xmm3 movq %rax, 424(%rsp) movq -16(%rbx), %rax movaps %xmm1, 48(%rsp) mulps 480(%rsp), %xmm3 movq %rax, 432(%rsp) movq -8(%rbx), %rax movaps %xmm1, 64(%rsp) movq 48(%rsp), %r8 movq 72(%rsp), %rdi movaps %xmm3, 32(%rsp) movq %rax, 440(%rsp) movq %r8, 448(%rsp) movaps 144(%rsp), %xmm1 movq %rdi, 456(%rsp) movaps 176(%rsp), %xmm4 mulps %xmm1, %xmm2 mulps 496(%rsp), %xmm4 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rcx movaps %xmm2, 80(%rsp) movq 88(%rsp), %rax movq %rcx, 464(%rsp) movq %rax, 472(%rsp) movaps %xmm4, 48(%rsp) movq 32(%rsp), %rsi movq 48(%rsp), %rdx movaps %xmm3, 64(%rsp) movaps %xmm4, 80(%rsp) movaps 512(%rsp), %xmm5 movq %rsi, 480(%rsp) movq %rdx, 496(%rsp) movaps 192(%rsp), %xmm7 movq 72(%rsp), %rsi movq 88(%rsp), %rdx movaps 528(%rsp), %xmm6 mulps %xmm7, %xmm5 movaps 208(%rsp), %xmm1 movq %rsi, 488(%rsp) movq %rdx, 504(%rsp) mulps %xmm1, %xmm6 movaps 224(%rsp), %xmm4 movaps %xmm5, 32(%rsp) movq 32(%rsp), %rsi movaps %xmm5, 64(%rsp) movaps %xmm6, 48(%rsp) movaps %xmm6, 80(%rsp) movaps 240(%rsp), %xmm5 movq 48(%rsp), %rdx movq %rsi, 512(%rsp) movaps 544(%rsp), %xmm7 movq 72(%rsp), %rsi movaps 560(%rsp), %xmm1 mulps %xmm4, %xmm7 movq %rdx, 528(%rsp) movq 88(%rsp), %rdx mulps %xmm5, %xmm1 movq %rsi, 520(%rsp) movq %rdx, 536(%rsp) movaps %xmm7, 32(%rsp) movq 32(%rsp), %rsi movaps %xmm1, 48(%rsp) movaps %xmm7, 64(%rsp) movaps %xmm1, 80(%rsp) movaps 576(%rsp), %xmm2 movq 48(%rsp), %rdx movq %rsi, 544(%rsp) movaps 256(%rsp), %xmm1 movq 72(%rsp), %rsi mulps %xmm1, %xmm2 movaps 592(%rsp), %xmm3 movq %rdx, 560(%rsp) movq 88(%rsp), %rdx movaps 288(%rsp), %xmm5 movq %rsi, 552(%rsp) movaps 304(%rsp), %xmm6 movq %rdx, 568(%rsp) movaps %xmm2, 32(%rsp) movq 32(%rsp), %rsi movaps 272(%rsp), %xmm2 movq %rsi, 576(%rsp) mulps %xmm2, %xmm3 movaps %xmm3, 48(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm4 movq %rdx, 592(%rsp) movaps %xmm4, 80(%rsp) movq 88(%rsp), %rdx movaps 608(%rsp), %xmm4 movq %rdx, 600(%rsp) mulps %xmm5, %xmm4 movaps 624(%rsp), %xmm5 movaps 32(%rsp), %xmm3 mulps %xmm6, %xmm5 movaps %xmm3, 64(%rsp) movq 72(%rsp), %rsi movaps %xmm4, 32(%rsp) movaps %xmm5, 48(%rsp) movq %rsi, 584(%rsp) movq 32(%rsp), %rsi movaps %xmm4, 64(%rsp) movq 48(%rsp), %rdx movaps %xmm5, 80(%rsp) movaps 640(%rsp), %xmm6 movq %rsi, 608(%rsp) movq %rdx, 624(%rsp) movaps 656(%rsp), %xmm7 movq 72(%rsp), %rsi movq 88(%rsp), %rdx movaps 320(%rsp), %xmm2 movaps 336(%rsp), %xmm3 mulps %xmm2, %xmm6 movq %rsi, 616(%rsp) movq %rdx, 632(%rsp) movaps 672(%rsp), %xmm1 mulps %xmm3, %xmm7 movaps 688(%rsp), %xmm2 movaps %xmm6, 32(%rsp) movq 32(%rsp), %rsi movaps %xmm7, 48(%rsp) movaps %xmm6, 64(%rsp) movaps %xmm7, 80(%rsp) movaps 352(%rsp), %xmm6 movq 48(%rsp), %rdx movq %rsi, 640(%rsp) movaps 368(%rsp), %xmm7 mulps %xmm6, %xmm1 movq 72(%rsp), %rsi mulps %xmm7, %xmm2 movq %rdx, 656(%rsp) movq 88(%rsp), %rdx movq %rsi, 648(%rsp) movaps %xmm1, 32(%rsp) movq %rdx, 664(%rsp) movq 32(%rsp), %rsi movaps %xmm2, 48(%rsp) movaps %xmm1, 64(%rsp) movq 48(%rsp), %rdx movaps %xmm2, 80(%rsp) movq %rsi, 672(%rsp) movq 72(%rsp), %rsi movaps 400(%rsp), %xmm4 movq %rsi, 680(%rsp) movq %rdx, 688(%rsp) mulps 720(%rsp), %xmm4 movaps 384(%rsp), %xmm3 movq 88(%rsp), %rdx mulps 704(%rsp), %xmm3 movaps 736(%rsp), %xmm5 movq %rdx, 696(%rsp) movaps 416(%rsp), %xmm7 movaps 432(%rsp), %xmm1 mulps %xmm7, %xmm5 movaps 752(%rsp), %xmm6 movaps %xmm4, 48(%rsp) movq 48(%rsp), %rsi mulps %xmm1, %xmm6 movaps %xmm3, 32(%rsp) movaps %xmm4, 80(%rsp) movaps %xmm3, 64(%rsp) movaps %xmm6, 48(%rsp) movq 32(%rsp), %rdx movq %rsi, 720(%rsp) movq 88(%rsp), %rsi movq 48(%rsp), %r9 movaps %xmm5, 32(%rsp) movq %rdx, 704(%rsp) movq 72(%rsp), %rdx movaps %xmm5, 64(%rsp) movaps %xmm6, 80(%rsp) movq %rsi, 728(%rsp) movq 72(%rsp), %r10 movq 32(%rsp), %rsi movq %rdx, 712(%rsp) leal -80(%r11), %edx cmpl $79, %edx movq %rsi, 736(%rsp) movq %r10, 744(%rsp) movq %r9, 752(%rsp) movq 88(%rsp), %r9 movq %r9, 760(%rsp) ja .L544 movq %r8, 1088(%rsp) movq %rdi, 1096(%rsp) movq %rcx, 1104(%rsp) movq %rax, 1112(%rsp) .L522: testl %edx, %edx je .L526 .L527: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 jne .L527 .L526: movaps 1088(%rsp), %xmm5 movaps 480(%rsp), %xmm4 mulps %xmm5, %xmm4 movaps 1104(%rsp), %xmm6 movaps 496(%rsp), %xmm5 mulps %xmm6, %xmm5 movaps %xmm4, 64(%rsp) movaps %xmm4, 32(%rsp) movaps %xmm5, 48(%rsp) movaps %xmm5, 80(%rsp) movaps 512(%rsp), %xmm6 movq 32(%rsp), %rdx movq 48(%rsp), %rax movaps 560(%rsp), %xmm4 mulps 544(%rsp), %xmm6 movq %rdx, 768(%rsp) movq %rax, 784(%rsp) mulps 528(%rsp), %xmm4 movq 72(%rsp), %rdx movq 88(%rsp), %rax movaps 32(%rsp), %xmm5 movq %rdx, 776(%rsp) movq %rax, 792(%rsp) movaps %xmm6, 16(%rsp) movq 16(%rsp), %rdx mulps %xmm6, %xmm5 movaps %xmm4, (%rsp) movaps %xmm6, 64(%rsp) movaps %xmm4, 80(%rsp) movq (%rsp), %rax movq %rdx, 800(%rsp) movq 72(%rsp), %rdx movq %rax, 816(%rsp) movq 88(%rsp), %rax movq %rdx, 808(%rsp) movq %rax, 824(%rsp) movaps %xmm5, 16(%rsp) movq 16(%rsp), %rdx movaps %xmm5, 64(%rsp) movaps 48(%rsp), %xmm6 movq %rdx, 832(%rsp) movq 72(%rsp), %rdx mulps %xmm4, %xmm6 movaps 576(%rsp), %xmm7 movaps 592(%rsp), %xmm1 movq %rdx, 840(%rsp) movaps %xmm6, (%rsp) movaps %xmm5, %xmm6 movq (%rsp), %rax movaps (%rsp), %xmm4 movq %rax, 848(%rsp) movaps %xmm4, 80(%rsp) movq 88(%rsp), %rax movaps 608(%rsp), %xmm4 movq %rax, 856(%rsp) movaps 624(%rsp), %xmm5 mulps %xmm7, %xmm4 mulps %xmm1, %xmm5 mulps %xmm4, %xmm6 movaps %xmm4, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 64(%rsp) movaps %xmm5, 32(%rsp) movaps %xmm5, 80(%rsp) movaps %xmm6, 16(%rsp) movaps (%rsp), %xmm4 movq 32(%rsp), %rax movq %rdx, 864(%rsp) mulps %xmm5, %xmm4 movq 72(%rsp), %rdx movq %rax, 880(%rsp) movq 88(%rsp), %rax movq %rdx, 872(%rsp) movq 16(%rsp), %rdx movq %rax, 888(%rsp) movaps %xmm4, (%rsp) movaps %xmm6, 64(%rsp) movaps %xmm4, 80(%rsp) movaps 672(%rsp), %xmm5 movq %rdx, 896(%rsp) movq 72(%rsp), %rdx movaps 640(%rsp), %xmm4 movq (%rsp), %rax mulps %xmm4, %xmm5 movaps 656(%rsp), %xmm7 movq %rdx, 904(%rsp) movaps 688(%rsp), %xmm6 movq %rax, 912(%rsp) movq 88(%rsp), %rax mulps %xmm7, %xmm6 movaps 16(%rsp), %xmm7 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rdx mulps %xmm5, %xmm7 movaps %xmm5, 64(%rsp) movaps %xmm6, 32(%rsp) movaps %xmm7, 16(%rsp) movaps %xmm7, %xmm3 movaps %xmm6, 80(%rsp) movq %rdx, 928(%rsp) movq 72(%rsp), %rdx movaps (%rsp), %xmm4 movq %rax, 920(%rsp) movq 32(%rsp), %rax movq %rdx, 936(%rsp) movq 16(%rsp), %rdx mulps %xmm6, %xmm4 movaps %xmm7, 64(%rsp) movq %rax, 944(%rsp) movq %rdx, 960(%rsp) movaps %xmm3, %xmm7 movq 88(%rsp), %rax movq 72(%rsp), %rdx movaps %xmm4, (%rsp) movaps %xmm4, 80(%rsp) movq %rax, 952(%rsp) movq (%rsp), %rax movq %rdx, 968(%rsp) movaps 704(%rsp), %xmm6 movq %rax, 976(%rsp) movq 88(%rsp), %rax movaps 736(%rsp), %xmm5 movaps 720(%rsp), %xmm4 mulps %xmm6, %xmm5 movq %rax, 984(%rsp) movaps 752(%rsp), %xmm6 mulps %xmm4, %xmm6 movaps (%rsp), %xmm4 mulps %xmm5, %xmm7 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 32(%rsp) mulps %xmm6, %xmm4 movaps %xmm5, 64(%rsp) movaps %xmm6, 80(%rsp) movaps %xmm7, 48(%rsp) movq 32(%rsp), %rax movq %rdx, 992(%rsp) movq 72(%rsp), %rdx movaps %xmm7, 64(%rsp) movq %rax, 1008(%rsp) movq 88(%rsp), %rax movaps %xmm4, 80(%rsp) movq %rdx, 1000(%rsp) movq %rax, 1016(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 48(%rsp) movq 48(%rsp), %rax movq %rdx, 96(%rsp) movq %rax, 112(%rsp) movq 72(%rsp), %rdx movq 88(%rsp), %rax movq %rdx, 104(%rsp) movq %rax, 120(%rsp) xorl %edx, %edx xorl %eax, %eax .L529: addq $1, %rax mulss 96(%rsp,%rdx,4), %xmm0 cmpq $8, %rax movq %rax, %rdx jne .L529 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret .L519: movq 736(%rsp), %rsi movq 752(%rsp), %rdx movq 760(%rsp), %rax movq 744(%rsp), %rcx movq %rsi, 1088(%rsp) movq %rdx, 1104(%rsp) movq %rcx, 1096(%rsp) movq %rax, 1112(%rsp) movq %rsi, 704(%rsp) movq %rcx, 712(%rsp) movq %rdx, 720(%rsp) movq %rax, 728(%rsp) movq %rsi, 672(%rsp) movq %rcx, 680(%rsp) movq %rdx, 688(%rsp) movq %rax, 696(%rsp) movq %rsi, 640(%rsp) movq %rcx, 648(%rsp) movq %rdx, 656(%rsp) movq %rax, 664(%rsp) movq %rsi, 608(%rsp) movq %rcx, 616(%rsp) movq %rdx, 624(%rsp) movq %rax, 632(%rsp) movq %rsi, 576(%rsp) movq %rcx, 584(%rsp) movq %rdx, 592(%rsp) movq %rax, 600(%rsp) movq %rsi, 544(%rsp) movq %rcx, 552(%rsp) movq %rdx, 560(%rsp) movq %rax, 568(%rsp) movq %rsi, 512(%rsp) movq %rcx, 520(%rsp) movq %rdx, 528(%rsp) movq %rax, 536(%rsp) movq %rsi, 480(%rsp) movq %rcx, 488(%rsp) movq %rdx, 496(%rsp) movq %rax, 504(%rsp) jmp .L526 .L543: movq 736(%rsp), %rdi movq 752(%rsp), %rcx movq 760(%rsp), %rax movq 744(%rsp), %rsi movq %rdi, 1088(%rsp) movq %rcx, 1104(%rsp) movq %rsi, 1096(%rsp) movq %rax, 1112(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 576(%rsp) movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 544(%rsp) movq %rsi, 552(%rsp) movq %rcx, 560(%rsp) movq %rax, 568(%rsp) movq %rdi, 512(%rsp) movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq %rax, 536(%rsp) movq %rdi, 480(%rsp) movq %rsi, 488(%rsp) movq %rcx, 496(%rsp) movq %rax, 504(%rsp) jmp .L522 simd_v12_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $1312, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm0 testb $31, %bl movss %xmm0, 96(%rsp) movl %eax, %edx movss %xmm0, 100(%rsp) movq 96(%rsp), %rcx movss %xmm0, 104(%rsp) movss %xmm0, 108(%rsp) movq %rcx, 128(%rsp) movss %xmm0, 112(%rsp) movq 104(%rsp), %rcx movss %xmm0, 116(%rsp) movss %xmm0, 120(%rsp) movss %xmm0, 124(%rsp) movq %rcx, 136(%rsp) movq 112(%rsp), %rcx movq %rcx, 144(%rsp) movq 120(%rsp), %rcx movq %rcx, 152(%rsp) je .L547 testl %eax, %eax jne .L552 jmp .L548 .L549: testl %edx, %edx je .L548 .L552: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L549 movl %edx, %eax .L547: cmpl $95, %eax jbe .L571 movq 128(%rsp), %r8 movq 144(%rsp), %rsi movl %eax, %edx movq 152(%rsp), %rcx movq 136(%rsp), %rdi movq %r8, 1248(%rsp) movq %rsi, 1264(%rsp) movq %rdi, 1256(%rsp) movq %rcx, 1272(%rsp) movq %r8, 160(%rsp) movq %rdi, 168(%rsp) movq %rsi, 176(%rsp) movq %rcx, 184(%rsp) movq %r8, 864(%rsp) movq %rdi, 872(%rsp) movq %rsi, 880(%rsp) movq %rcx, 888(%rsp) movq %r8, 832(%rsp) movq %rdi, 840(%rsp) movq %rsi, 848(%rsp) movq %rcx, 856(%rsp) movq %r8, 800(%rsp) movq %rdi, 808(%rsp) movq %rsi, 816(%rsp) movq %rcx, 824(%rsp) movq %r8, 768(%rsp) movq %rdi, 776(%rsp) movq %rsi, 784(%rsp) movq %rcx, 792(%rsp) movq %r8, 736(%rsp) movq %rdi, 744(%rsp) movq %rsi, 752(%rsp) movq %rcx, 760(%rsp) movq %r8, 704(%rsp) movq %rdi, 712(%rsp) movq %rsi, 720(%rsp) movq %rcx, 728(%rsp) movq %r8, 672(%rsp) movq %rdi, 680(%rsp) movq %rsi, 688(%rsp) movq %rcx, 696(%rsp) movq %r8, 640(%rsp) movq %rdi, 648(%rsp) movq %rsi, 656(%rsp) movq %rcx, 664(%rsp) movq %r8, 608(%rsp) movq %rdi, 616(%rsp) movq %rsi, 624(%rsp) movq %rcx, 632(%rsp) jmp .L554 .L572: movq %r8, 1248(%rsp) movq %rdi, 1256(%rsp) movq %rcx, 1264(%rsp) movq %rax, 1272(%rsp) .L554: movq (%rbx), %rax subl $96, %edx addq $384, %rbx movq %rax, 192(%rsp) movq -376(%rbx), %rax movq %rax, 200(%rsp) movq -368(%rbx), %rax movq %rax, 208(%rsp) movq -360(%rbx), %rax movq %rax, 216(%rsp) movq -352(%rbx), %rax movq %rax, 224(%rsp) movq -344(%rbx), %rax movq %rax, 232(%rsp) movq -336(%rbx), %rax movq %rax, 240(%rsp) movq -328(%rbx), %rax movq %rax, 248(%rsp) movq -320(%rbx), %rax movq %rax, 256(%rsp) movq -312(%rbx), %rax movq %rax, 264(%rsp) movq -304(%rbx), %rax movq %rax, 272(%rsp) movq -296(%rbx), %rax movq %rax, 280(%rsp) movq -288(%rbx), %rax movq %rax, 288(%rsp) movq -280(%rbx), %rax movq %rax, 296(%rsp) movq -272(%rbx), %rax movq %rax, 304(%rsp) movq -264(%rbx), %rax movq %rax, 312(%rsp) movq -256(%rbx), %rax movq %rax, 320(%rsp) movq -248(%rbx), %rax movq %rax, 328(%rsp) movq -240(%rbx), %rax movq %rax, 336(%rsp) movq -232(%rbx), %rax movq %rax, 344(%rsp) movq -224(%rbx), %rax movq %rax, 352(%rsp) movq -216(%rbx), %rax movq %rax, 360(%rsp) movq -208(%rbx), %rax movq %rax, 368(%rsp) movq -200(%rbx), %rax movq %rax, 376(%rsp) movq -192(%rbx), %rax movq %rax, 384(%rsp) movq -184(%rbx), %rax movq %rax, 392(%rsp) movq -176(%rbx), %rax movq %rax, 400(%rsp) movq -168(%rbx), %rax movq %rax, 408(%rsp) movq -160(%rbx), %rax movq %rax, 416(%rsp) movq -152(%rbx), %rax movq %rax, 424(%rsp) movq -144(%rbx), %rax movq %rax, 432(%rsp) movq -136(%rbx), %rax movq %rax, 440(%rsp) movq -128(%rbx), %rax movq %rax, 448(%rsp) movq -120(%rbx), %rax movq %rax, 456(%rsp) movq -112(%rbx), %rax movaps 192(%rsp), %xmm4 movq %rax, 464(%rsp) movq -104(%rbx), %rax movaps 208(%rsp), %xmm6 movq %rax, 472(%rsp) movq -96(%rbx), %rax movaps 1248(%rsp), %xmm1 movq %rax, 480(%rsp) movq -88(%rbx), %rax movaps 1264(%rsp), %xmm2 mulps %xmm4, %xmm1 movq %rax, 488(%rsp) movq -80(%rbx), %rax mulps %xmm6, %xmm2 movq %rax, 496(%rsp) movq -72(%rbx), %rax movaps %xmm1, 48(%rsp) movq 48(%rsp), %r8 movq %rax, 504(%rsp) movq -64(%rbx), %rax movq %rax, 512(%rsp) movq -56(%rbx), %rax movq %rax, 520(%rsp) movq -48(%rbx), %rax movq %rax, 528(%rsp) movq -40(%rbx), %rax movq %rax, 536(%rsp) movq -32(%rbx), %rax movq %rax, 544(%rsp) movq -24(%rbx), %rax movq %rax, 552(%rsp) movq -16(%rbx), %rax movq %rax, 560(%rsp) movq -8(%rbx), %rax cmpl $95, %edx movaps %xmm1, 64(%rsp) movq %rax, 568(%rsp) movaps %xmm2, 48(%rsp) movq 48(%rsp), %rcx movq 72(%rsp), %rdi movaps %xmm2, 80(%rsp) movaps 608(%rsp), %xmm3 movq 88(%rsp), %rax movq %r8, 576(%rsp) movaps 224(%rsp), %xmm4 movq %rdi, 584(%rsp) movq %rcx, 592(%rsp) mulps %xmm4, %xmm3 movaps 240(%rsp), %xmm5 movq %rax, 600(%rsp) movaps 624(%rsp), %xmm4 mulps %xmm5, %xmm4 movaps 256(%rsp), %xmm1 movaps 272(%rsp), %xmm2 movaps 640(%rsp), %xmm5 movaps 656(%rsp), %xmm6 mulps %xmm1, %xmm5 movaps %xmm3, 32(%rsp) movq 32(%rsp), %r9 mulps %xmm2, %xmm6 movaps %xmm4, 48(%rsp) movaps %xmm3, 64(%rsp) movaps %xmm4, 80(%rsp) movaps %xmm5, 32(%rsp) movq 48(%rsp), %rsi movq %r9, 608(%rsp) movq 72(%rsp), %r9 movaps %xmm6, 48(%rsp) movq %rsi, 624(%rsp) movq 88(%rsp), %rsi movaps %xmm5, 64(%rsp) movq %r9, 616(%rsp) movq %rsi, 632(%rsp) movq 32(%rsp), %r9 movq 48(%rsp), %rsi movaps %xmm6, 80(%rsp) movaps 672(%rsp), %xmm7 movq %r9, 640(%rsp) movq 72(%rsp), %r9 movaps 288(%rsp), %xmm5 movq %rsi, 656(%rsp) movq 88(%rsp), %rsi mulps %xmm5, %xmm7 movaps 320(%rsp), %xmm2 movq %r9, 648(%rsp) mulps 704(%rsp), %xmm2 movaps 688(%rsp), %xmm1 movq %rsi, 664(%rsp) movaps 304(%rsp), %xmm6 movaps %xmm7, 32(%rsp) movq 32(%rsp), %r9 mulps %xmm6, %xmm1 movaps %xmm7, 64(%rsp) movaps %xmm2, 32(%rsp) movaps %xmm1, 48(%rsp) movaps %xmm1, 80(%rsp) movq %r9, 672(%rsp) movq 72(%rsp), %r9 movaps 336(%rsp), %xmm3 movq 48(%rsp), %rsi mulps 720(%rsp), %xmm3 movq %r9, 680(%rsp) movq 32(%rsp), %r9 movaps %xmm2, 64(%rsp) movq %rsi, 688(%rsp) movq 88(%rsp), %rsi movq %r9, 704(%rsp) movq 72(%rsp), %r9 movaps %xmm3, 48(%rsp) movq %rsi, 696(%rsp) movq 48(%rsp), %rsi movaps %xmm3, 80(%rsp) movq %r9, 712(%rsp) movq %rsi, 720(%rsp) movaps 736(%rsp), %xmm4 movq 88(%rsp), %rsi movaps 352(%rsp), %xmm6 movaps 752(%rsp), %xmm5 mulps %xmm6, %xmm4 movq %rsi, 728(%rsp) movaps 368(%rsp), %xmm7 mulps %xmm7, %xmm5 movaps 768(%rsp), %xmm6 movaps %xmm4, 32(%rsp) movq 32(%rsp), %r9 movaps %xmm4, 64(%rsp) movaps %xmm5, 48(%rsp) movaps %xmm5, 80(%rsp) movaps 384(%rsp), %xmm3 movq 48(%rsp), %rsi movq %r9, 736(%rsp) movaps 400(%rsp), %xmm4 mulps %xmm3, %xmm6 movq 72(%rsp), %r9 movaps 784(%rsp), %xmm7 movq %rsi, 752(%rsp) movq 88(%rsp), %rsi mulps %xmm4, %xmm7 movq %r9, 744(%rsp) movaps %xmm6, 32(%rsp) movq %rsi, 760(%rsp) movq 32(%rsp), %r9 movaps %xmm6, 64(%rsp) movaps %xmm7, 48(%rsp) movaps %xmm7, 80(%rsp) movq 48(%rsp), %rsi movq %r9, 768(%rsp) movq 72(%rsp), %r9 movq %rsi, 784(%rsp) movq 88(%rsp), %rsi movq %r9, 776(%rsp) movq %rsi, 792(%rsp) movaps 800(%rsp), %xmm1 movaps 416(%rsp), %xmm7 mulps %xmm7, %xmm1 movaps 816(%rsp), %xmm2 movaps 448(%rsp), %xmm4 movaps %xmm1, 32(%rsp) movq 32(%rsp), %r9 movaps 432(%rsp), %xmm1 movq %r9, 800(%rsp) mulps %xmm1, %xmm2 movaps 464(%rsp), %xmm5 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm3 movq %rsi, 816(%rsp) movaps %xmm3, 80(%rsp) movq 88(%rsp), %rsi movaps 832(%rsp), %xmm3 movq %rsi, 824(%rsp) mulps %xmm4, %xmm3 movaps 848(%rsp), %xmm4 movaps 32(%rsp), %xmm2 mulps %xmm5, %xmm4 movaps %xmm2, 64(%rsp) movq 72(%rsp), %r9 movaps %xmm3, 32(%rsp) movaps %xmm4, 48(%rsp) movaps %xmm3, 64(%rsp) movaps %xmm4, 80(%rsp) movq 48(%rsp), %rsi movq %r9, 808(%rsp) movq 32(%rsp), %r9 movq %rsi, 848(%rsp) movq 88(%rsp), %rsi movq %r9, 832(%rsp) movq 72(%rsp), %r9 movq %r9, 840(%rsp) movq %rsi, 856(%rsp) movaps 864(%rsp), %xmm5 movaps 880(%rsp), %xmm6 movaps 480(%rsp), %xmm1 movaps 496(%rsp), %xmm2 mulps %xmm1, %xmm5 movaps 128(%rsp), %xmm7 mulps %xmm2, %xmm6 movaps 144(%rsp), %xmm1 movaps %xmm5, 32(%rsp) movq 32(%rsp), %r9 movaps %xmm6, 48(%rsp) movaps %xmm5, 64(%rsp) movaps %xmm6, 80(%rsp) movaps 512(%rsp), %xmm5 movq 48(%rsp), %rsi movq %r9, 864(%rsp) movaps 528(%rsp), %xmm6 mulps %xmm5, %xmm7 movq 72(%rsp), %r9 mulps %xmm6, %xmm1 movq %rsi, 880(%rsp) movq 88(%rsp), %rsi movaps 544(%rsp), %xmm2 movq %r9, 872(%rsp) mulps 160(%rsp), %xmm2 movq %rsi, 888(%rsp) movaps %xmm7, 32(%rsp) movq 32(%rsp), %r9 movaps %xmm1, 48(%rsp) movaps %xmm7, 64(%rsp) movaps %xmm1, 80(%rsp) movq 48(%rsp), %rsi movq %r9, 128(%rsp) movq 72(%rsp), %r9 movq %rsi, 144(%rsp) movq 88(%rsp), %rsi movq %r9, 136(%rsp) movq %rsi, 152(%rsp) movaps %xmm2, 32(%rsp) movq 32(%rsp), %rsi movaps %xmm2, 64(%rsp) movaps 560(%rsp), %xmm3 movq %rsi, 160(%rsp) movq 72(%rsp), %rsi mulps 176(%rsp), %xmm3 movq %rsi, 168(%rsp) movaps %xmm3, 48(%rsp) movq 48(%rsp), %r9 movaps %xmm3, 80(%rsp) movq %r9, 176(%rsp) movq 88(%rsp), %r9 movq %r9, 184(%rsp) ja .L572 movq %r8, 1280(%rsp) movq %rdi, 1288(%rsp) movq %rcx, 1296(%rsp) movq %rax, 1304(%rsp) .L551: testl %edx, %edx je .L555 .L556: addq $4, %rbx subl $1, %edx mulss -4(%rbx), %xmm0 jne .L556 .L555: movaps 1280(%rsp), %xmm7 movaps 608(%rsp), %xmm6 movaps 1296(%rsp), %xmm1 mulps %xmm7, %xmm6 movaps 624(%rsp), %xmm7 mulps %xmm1, %xmm7 movaps 688(%rsp), %xmm4 movaps %xmm6, %xmm2 movaps %xmm6, 64(%rsp) movaps %xmm7, 80(%rsp) movaps %xmm7, %xmm3 movaps %xmm6, 16(%rsp) movaps %xmm7, 32(%rsp) movaps 640(%rsp), %xmm6 movq 16(%rsp), %rdx movq 32(%rsp), %rax movaps 672(%rsp), %xmm7 mulps %xmm7, %xmm6 movq %rdx, 896(%rsp) movaps 656(%rsp), %xmm7 movq 72(%rsp), %rdx movq %rax, 912(%rsp) mulps %xmm4, %xmm7 movq 88(%rsp), %rax movaps %xmm6, 48(%rsp) movaps %xmm6, %xmm5 movq %rdx, 904(%rsp) movaps %xmm6, 64(%rsp) movaps %xmm2, %xmm6 movaps %xmm7, (%rsp) mulps %xmm5, %xmm6 movaps %xmm7, 80(%rsp) movaps (%rsp), %xmm7 movq 48(%rsp), %rdx movq %rax, 920(%rsp) movq (%rsp), %rax mulps %xmm3, %xmm7 movaps %xmm6, 16(%rsp) movq %rdx, 928(%rsp) movq 72(%rsp), %rdx movaps %xmm6, %xmm1 movq %rax, 944(%rsp) movq 88(%rsp), %rax movq %rdx, 936(%rsp) movaps %xmm7, %xmm2 movq %rax, 952(%rsp) movaps %xmm7, (%rsp) movq 16(%rsp), %rdx movq (%rsp), %rax movaps %xmm6, 64(%rsp) movaps %xmm7, 80(%rsp) movaps 704(%rsp), %xmm3 movq %rdx, 960(%rsp) movq 72(%rsp), %rdx movaps 736(%rsp), %xmm6 movq %rax, 976(%rsp) movq 88(%rsp), %rax mulps %xmm3, %xmm6 movaps 720(%rsp), %xmm7 movq %rdx, 968(%rsp) mulps 752(%rsp), %xmm7 movq %rax, 984(%rsp) movaps %xmm6, 48(%rsp) movq 48(%rsp), %rdx mulps %xmm6, %xmm1 movaps %xmm6, 64(%rsp) movaps %xmm7, 32(%rsp) mulps %xmm7, %xmm2 movaps %xmm1, 16(%rsp) movaps %xmm7, 80(%rsp) movaps %xmm2, (%rsp) movq %rdx, 992(%rsp) movq 72(%rsp), %rdx movq 32(%rsp), %rax movaps %xmm1, 64(%rsp) movq %rdx, 1000(%rsp) movq 16(%rsp), %rdx movaps %xmm2, %xmm1 movq %rax, 1008(%rsp) movq %rdx, 1024(%rsp) movq 88(%rsp), %rax movq 72(%rsp), %rdx movaps %xmm2, 80(%rsp) movq %rax, 1016(%rsp) movq (%rsp), %rax movq %rdx, 1032(%rsp) movaps 768(%rsp), %xmm2 movq %rax, 1040(%rsp) movq 88(%rsp), %rax movaps 800(%rsp), %xmm3 mulps %xmm2, %xmm3 movq %rax, 1048(%rsp) movaps 816(%rsp), %xmm6 movaps %xmm3, 48(%rsp) movq 48(%rsp), %rdx movaps 784(%rsp), %xmm3 movq %rdx, 1056(%rsp) mulps %xmm3, %xmm6 movaps 48(%rsp), %xmm7 movaps %xmm7, 64(%rsp) movq 72(%rsp), %rdx movaps %xmm6, 32(%rsp) movq 32(%rsp), %rax mulps %xmm6, %xmm1 movaps %xmm6, 80(%rsp) movaps 16(%rsp), %xmm7 movq %rax, 1072(%rsp) movq 88(%rsp), %rax mulps 48(%rsp), %xmm7 movaps %xmm1, (%rsp) movq %rdx, 1064(%rsp) movq %rax, 1080(%rsp) movaps %xmm1, 80(%rsp) movaps %xmm7, 16(%rsp) movq (%rsp), %rax movq 16(%rsp), %rdx movaps %xmm7, 64(%rsp) movaps %xmm7, %xmm5 movaps %xmm5, %xmm6 movq %rdx, 1088(%rsp) movq %rax, 1104(%rsp) movq 72(%rsp), %rdx movq 88(%rsp), %rax movq %rdx, 1096(%rsp) movq %rax, 1112(%rsp) movaps 832(%rsp), %xmm1 movaps 864(%rsp), %xmm2 mulps %xmm1, %xmm2 movaps 880(%rsp), %xmm3 movaps 160(%rsp), %xmm1 movaps %xmm2, 48(%rsp) movq 48(%rsp), %rdx movaps 848(%rsp), %xmm2 movq %rdx, 1120(%rsp) mulps %xmm2, %xmm3 movaps %xmm3, 32(%rsp) movq 32(%rsp), %rax movaps 32(%rsp), %xmm7 movq %rax, 1136(%rsp) movaps %xmm7, 80(%rsp) movq 88(%rsp), %rax movaps (%rsp), %xmm7 movq %rax, 1144(%rsp) movaps 48(%rsp), %xmm3 mulps 32(%rsp), %xmm7 movaps %xmm3, 64(%rsp) mulps %xmm3, %xmm6 movq 72(%rsp), %rdx movq %rdx, 1128(%rsp) movaps %xmm6, 16(%rsp) movaps %xmm6, %xmm4 movq 16(%rsp), %rdx movaps %xmm6, 64(%rsp) movaps %xmm7, %xmm5 movaps %xmm7, (%rsp) movaps %xmm7, 80(%rsp) movaps 128(%rsp), %xmm6 movq (%rsp), %rax movq %rdx, 1152(%rsp) mulps %xmm6, %xmm1 movq 72(%rsp), %rdx movq %rax, 1168(%rsp) movq 88(%rsp), %rax movq %rdx, 1160(%rsp) movq %rax, 1176(%rsp) movaps %xmm1, 48(%rsp) movq 48(%rsp), %rdx movaps 144(%rsp), %xmm1 movq %rdx, 1184(%rsp) movaps 176(%rsp), %xmm2 mulps %xmm1, %xmm2 movaps %xmm2, 32(%rsp) movq 32(%rsp), %rax movaps 32(%rsp), %xmm3 movq %rax, 1200(%rsp) movaps 48(%rsp), %xmm2 movaps %xmm3, 80(%rsp) movaps %xmm4, %xmm3 movq 88(%rsp), %rax movaps %xmm2, 64(%rsp) mulps %xmm2, %xmm3 movaps 32(%rsp), %xmm6 movq 72(%rsp), %rdx movq %rax, 1208(%rsp) mulps %xmm5, %xmm6 movaps %xmm3, 48(%rsp) movq %rdx, 1192(%rsp) movq 48(%rsp), %rdx movaps %xmm3, 64(%rsp) movaps %xmm6, 48(%rsp) movaps %xmm6, 80(%rsp) movq 48(%rsp), %rax movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rdx, 104(%rsp) xorl %edx, %edx movq %rax, 120(%rsp) xorl %eax, %eax .L558: addq $1, %rax mulss 96(%rsp,%rdx,4), %xmm0 cmpq $8, %rax movq %rax, %rdx jne .L558 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret .L548: movq 128(%rsp), %rsi movq 144(%rsp), %rdx movq 152(%rsp), %rax movq 136(%rsp), %rcx movq %rsi, 1280(%rsp) movq %rdx, 1296(%rsp) movq %rcx, 1288(%rsp) movq %rax, 1304(%rsp) movq %rsi, 160(%rsp) movq %rcx, 168(%rsp) movq %rdx, 176(%rsp) movq %rax, 184(%rsp) movq %rsi, 864(%rsp) movq %rcx, 872(%rsp) movq %rdx, 880(%rsp) movq %rax, 888(%rsp) movq %rsi, 832(%rsp) movq %rcx, 840(%rsp) movq %rdx, 848(%rsp) movq %rax, 856(%rsp) movq %rsi, 800(%rsp) movq %rcx, 808(%rsp) movq %rdx, 816(%rsp) movq %rax, 824(%rsp) movq %rsi, 768(%rsp) movq %rcx, 776(%rsp) movq %rdx, 784(%rsp) movq %rax, 792(%rsp) movq %rsi, 736(%rsp) movq %rcx, 744(%rsp) movq %rdx, 752(%rsp) movq %rax, 760(%rsp) movq %rsi, 704(%rsp) movq %rcx, 712(%rsp) movq %rdx, 720(%rsp) movq %rax, 728(%rsp) movq %rsi, 672(%rsp) movq %rcx, 680(%rsp) movq %rdx, 688(%rsp) movq %rax, 696(%rsp) movq %rsi, 640(%rsp) movq %rcx, 648(%rsp) movq %rdx, 656(%rsp) movq %rax, 664(%rsp) movq %rsi, 608(%rsp) movq %rcx, 616(%rsp) movq %rdx, 624(%rsp) movq %rax, 632(%rsp) jmp .L555 .L571: movq 128(%rsp), %rdi movq 144(%rsp), %rcx movq 152(%rsp), %rax movq 136(%rsp), %rsi movq %rdi, 1280(%rsp) movq %rcx, 1296(%rsp) movq %rsi, 1288(%rsp) movq %rax, 1304(%rsp) movq %rdi, 160(%rsp) movq %rsi, 168(%rsp) movq %rcx, 176(%rsp) movq %rax, 184(%rsp) movq %rdi, 864(%rsp) movq %rsi, 872(%rsp) movq %rcx, 880(%rsp) movq %rax, 888(%rsp) movq %rdi, 832(%rsp) movq %rsi, 840(%rsp) movq %rcx, 848(%rsp) movq %rax, 856(%rsp) movq %rdi, 800(%rsp) movq %rsi, 808(%rsp) movq %rcx, 816(%rsp) movq %rax, 824(%rsp) movq %rdi, 768(%rsp) movq %rsi, 776(%rsp) movq %rcx, 784(%rsp) movq %rax, 792(%rsp) movq %rdi, 736(%rsp) movq %rsi, 744(%rsp) movq %rcx, 752(%rsp) movq %rax, 760(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) jmp .L551 simd_v2a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $288, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm1 testb $31, %bl movl %eax, %ecx movss %xmm1, 64(%rsp) movss %xmm1, 68(%rsp) movaps %xmm1, %xmm0 movss %xmm1, 72(%rsp) movss %xmm1, 76(%rsp) movss %xmm1, 80(%rsp) movss %xmm1, 84(%rsp) movss %xmm1, 88(%rsp) movss %xmm1, 92(%rsp) je .L574 testl %eax, %eax jne .L579 jmp .L582 .L576: testl %ecx, %ecx je .L582 .L579: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L576 .L574: cmpl $15, %ecx jbe .L578 movq 64(%rsp), %rax movl %ecx, %r10d movq %rbx, %rdx movq %rax, 256(%rsp) movq 72(%rsp), %rax movq %rax, 264(%rsp) movq 80(%rsp), %rax movq %rax, 272(%rsp) movq 88(%rsp), %rax movq %rax, 280(%rsp) jmp .L581 .L599: movq %r9, 256(%rsp) movq %r8, 264(%rsp) movq %rdi, 272(%rsp) movq %rax, 280(%rsp) .L581: movq (%rdx), %rax subl $16, %r10d addq $64, %rdx movaps 256(%rsp), %xmm4 movq %rax, 96(%rsp) movq -56(%rdx), %rax movq %rax, 104(%rsp) movq -48(%rdx), %rax movaps 96(%rsp), %xmm2 movq %rax, 112(%rsp) movq -40(%rdx), %rax movq %rax, 120(%rsp) movq -32(%rdx), %rax movaps 112(%rsp), %xmm3 movq %rax, 128(%rsp) movq -24(%rdx), %rax movq %rax, 136(%rsp) movq -16(%rdx), %rax movaps 128(%rsp), %xmm6 movq %rax, 144(%rsp) movq -8(%rdx), %rax cmpl $15, %r10d mulps %xmm6, %xmm2 movq %rax, 152(%rsp) movaps 144(%rsp), %xmm7 mulps %xmm7, %xmm3 movaps %xmm2, 16(%rsp) movq 16(%rsp), %rsi mulps %xmm2, %xmm4 movaps %xmm2, 32(%rsp) movaps %xmm3, (%rsp) movaps %xmm3, 48(%rsp) movq (%rsp), %rax movq %rsi, 160(%rsp) movq 40(%rsp), %rsi movq %rax, 176(%rsp) movq 56(%rsp), %rax movq %rsi, 168(%rsp) movq %rax, 184(%rsp) movaps %xmm4, 16(%rsp) movq 16(%rsp), %r9 movaps %xmm4, 32(%rsp) movaps 272(%rsp), %xmm5 movq 40(%rsp), %r8 movq %r9, 192(%rsp) mulps %xmm3, %xmm5 movq %r8, 200(%rsp) movaps %xmm5, 16(%rsp) movq 16(%rsp), %rdi movaps %xmm5, 48(%rsp) movq 56(%rsp), %rax movq %rdi, 208(%rsp) movq %rax, 216(%rsp) ja .L599 subl $16, %ecx movq %r9, 64(%rsp) movq %r9, 224(%rsp) movl %ecx, %esi movq %r8, 232(%rsp) movq %rdi, 240(%rsp) shrl $4, %esi movq %rax, 248(%rsp) movq %r8, 72(%rsp) movl %esi, %edx sall $4, %esi movq %rdi, 80(%rsp) addq $1, %rdx movq %rax, 88(%rsp) subl %esi, %ecx salq $6, %rdx movss 64(%rsp), %xmm1 addq %rdx, %rbx .L578: testl %ecx, %ecx je .L582 .L583: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L583 .L582: leaq 68(%rsp), %rax leaq 96(%rsp), %rdx jmp .L585 .L584: movss (%rax), %xmm1 addq $4, %rax .L585: cmpq %rdx, %rax mulss %xmm1, %xmm0 jne .L584 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v4a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $448, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm1 testb $31, %bl movl %eax, %ecx movss %xmm1, 96(%rsp) movss %xmm1, 100(%rsp) movaps %xmm1, %xmm0 movss %xmm1, 104(%rsp) movss %xmm1, 108(%rsp) movss %xmm1, 112(%rsp) movss %xmm1, 116(%rsp) movss %xmm1, 120(%rsp) movss %xmm1, 124(%rsp) je .L601 testl %eax, %eax jne .L606 jmp .L609 .L603: testl %ecx, %ecx je .L609 .L606: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L603 .L601: cmpl $31, %ecx jbe .L605 movq 96(%rsp), %rax movl %ecx, %r10d movq %rbx, %rdx movq %rax, 384(%rsp) movq 104(%rsp), %rax movq %rax, 392(%rsp) movq 112(%rsp), %rax movq %rax, 400(%rsp) movq 120(%rsp), %rax movq %rax, 408(%rsp) jmp .L608 .L626: movq %r9, 384(%rsp) movq %r8, 392(%rsp) movq %rdi, 400(%rsp) movq %rax, 408(%rsp) .L608: movq (%rdx), %rax subl $32, %r10d subq $-128, %rdx movq %rax, 128(%rsp) movq -120(%rdx), %rax movq %rax, 136(%rsp) movq -112(%rdx), %rax movaps 128(%rsp), %xmm2 movq %rax, 144(%rsp) movq -104(%rdx), %rax movq %rax, 152(%rsp) movq -96(%rdx), %rax movq %rax, 160(%rsp) movq -88(%rdx), %rax movq %rax, 168(%rsp) movq -80(%rdx), %rax movaps 160(%rsp), %xmm3 movq %rax, 176(%rsp) movq -72(%rdx), %rax mulps %xmm3, %xmm2 movq %rax, 184(%rsp) movq -64(%rdx), %rax movq %rax, 192(%rsp) movq -56(%rdx), %rax movq %rax, 200(%rsp) movq -48(%rdx), %rax movq %rax, 208(%rsp) movq -40(%rdx), %rax movq %rax, 216(%rsp) movq -32(%rdx), %rax movq %rax, 224(%rsp) movq -24(%rdx), %rax movq %rax, 232(%rsp) movq -16(%rdx), %rax movq %rax, 240(%rsp) movq -8(%rdx), %rax cmpl $31, %r10d movq %rax, 248(%rsp) movaps %xmm2, (%rsp) movq (%rsp), %rsi movaps %xmm2, 64(%rsp) movaps 176(%rsp), %xmm4 movq %rsi, 256(%rsp) movq 72(%rsp), %rsi movaps 144(%rsp), %xmm3 mulps %xmm4, %xmm3 movq %rsi, 264(%rsp) movaps 224(%rsp), %xmm7 movaps 192(%rsp), %xmm4 movaps %xmm3, 80(%rsp) mulps %xmm7, %xmm4 movaps %xmm3, 32(%rsp) movaps 208(%rsp), %xmm5 movq 32(%rsp), %rax movaps %xmm4, %xmm6 movaps 240(%rsp), %xmm3 mulps %xmm2, %xmm6 movq %rax, 272(%rsp) movq 88(%rsp), %rax mulps %xmm3, %xmm5 movaps 32(%rsp), %xmm7 movaps %xmm4, 16(%rsp) movq %rax, 280(%rsp) movq 16(%rsp), %rsi movaps %xmm4, 64(%rsp) mulps %xmm5, %xmm7 movaps %xmm5, 48(%rsp) movaps %xmm5, 80(%rsp) movaps %xmm6, 16(%rsp) movq 48(%rsp), %rax movq %rsi, 288(%rsp) movq 72(%rsp), %rsi movaps %xmm7, 48(%rsp) movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rsi, 296(%rsp) movq %rax, 312(%rsp) movq 16(%rsp), %rsi movaps %xmm6, 64(%rsp) movq 48(%rsp), %rax movaps %xmm7, 80(%rsp) movaps 384(%rsp), %xmm1 movq %rax, 336(%rsp) movq 88(%rsp), %rax movaps 400(%rsp), %xmm2 mulps %xmm6, %xmm1 movq %rsi, 320(%rsp) movq 72(%rsp), %rsi mulps %xmm7, %xmm2 movq %rax, 344(%rsp) movq %rsi, 328(%rsp) movaps %xmm1, 32(%rsp) movq 32(%rsp), %r9 movaps %xmm1, 64(%rsp) movaps %xmm2, 48(%rsp) movaps %xmm2, 80(%rsp) movq 48(%rsp), %rdi movq 72(%rsp), %r8 movq 88(%rsp), %rax movq %r9, 352(%rsp) movq %r8, 360(%rsp) movq %rdi, 368(%rsp) movq %rax, 376(%rsp) ja .L626 subl $32, %ecx movq %r9, 96(%rsp) movq %r9, 416(%rsp) movl %ecx, %esi movq %r8, 424(%rsp) movq %rdi, 432(%rsp) shrl $5, %esi movq %rax, 440(%rsp) movq %r8, 104(%rsp) movl %esi, %edx sall $5, %esi movq %rdi, 112(%rsp) addq $1, %rdx movq %rax, 120(%rsp) subl %esi, %ecx salq $7, %rdx movss 96(%rsp), %xmm1 addq %rdx, %rbx .L605: testl %ecx, %ecx je .L609 .L610: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L610 .L609: leaq 100(%rsp), %rax leaq 128(%rsp), %rdx jmp .L612 .L611: movss (%rax), %xmm1 addq $4, %rax .L612: cmpq %rdx, %rax mulss %xmm1, %xmm0 jne .L611 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v8a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 movq %rdi, %r13 pushq %r12 movq %rsi, %r12 pushq %rbx andq $-32, %rsp subq $736, %rsp call get_vec_start movq %r13, %rdi movq %rax, %rbx call vec_length movss .LC0(%rip), %xmm1 testb $31, %bl movl %eax, %ecx movss %xmm1, 128(%rsp) movss %xmm1, 132(%rsp) movaps %xmm1, %xmm0 movss %xmm1, 136(%rsp) movss %xmm1, 140(%rsp) movss %xmm1, 144(%rsp) movss %xmm1, 148(%rsp) movss %xmm1, 152(%rsp) movss %xmm1, 156(%rsp) je .L628 testl %eax, %eax jne .L633 jmp .L636 .L630: testl %ecx, %ecx je .L636 .L633: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 testb $31, %bl jne .L630 .L628: cmpl $63, %ecx jbe .L632 movq 128(%rsp), %rax movl %ecx, %r10d movq %rbx, %rdx movq %rax, 704(%rsp) movq 136(%rsp), %rax movq %rax, 712(%rsp) movq 144(%rsp), %rax movq %rax, 720(%rsp) movq 152(%rsp), %rax movq %rax, 728(%rsp) jmp .L635 .L653: movq %r9, 704(%rsp) movq %r8, 712(%rsp) movq %rdi, 720(%rsp) movq %rax, 728(%rsp) .L635: movq (%rdx), %rax subl $64, %r10d addq $256, %rdx movq %rax, 160(%rsp) movq -248(%rdx), %rax movq %rax, 168(%rsp) movq -240(%rdx), %rax movq %rax, 176(%rsp) movq -232(%rdx), %rax movq %rax, 184(%rsp) movq -224(%rdx), %rax movq %rax, 192(%rsp) movq -216(%rdx), %rax movq %rax, 200(%rsp) movq -208(%rdx), %rax movq %rax, 208(%rsp) movq -200(%rdx), %rax movq %rax, 216(%rsp) movq -192(%rdx), %rax movq %rax, 224(%rsp) movq -184(%rdx), %rax movq %rax, 232(%rsp) movq -176(%rdx), %rax movq %rax, 240(%rsp) movq -168(%rdx), %rax movq %rax, 248(%rsp) movq -160(%rdx), %rax movq %rax, 256(%rsp) movq -152(%rdx), %rax movq %rax, 264(%rsp) movq -144(%rdx), %rax movq %rax, 272(%rsp) movq -136(%rdx), %rax movq %rax, 280(%rsp) movq -128(%rdx), %rax movq %rax, 288(%rsp) movq -120(%rdx), %rax movaps 192(%rsp), %xmm4 movq %rax, 296(%rsp) movq -112(%rdx), %rax movaps 176(%rsp), %xmm5 movq %rax, 304(%rsp) movq -104(%rdx), %rax movaps 160(%rsp), %xmm2 movq %rax, 312(%rsp) movq -96(%rdx), %rax movaps 208(%rsp), %xmm3 mulps %xmm4, %xmm2 movq %rax, 320(%rsp) movq -88(%rdx), %rax mulps %xmm5, %xmm3 movq %rax, 328(%rsp) movq -80(%rdx), %rax movaps %xmm2, 32(%rsp) movq %rax, 336(%rsp) movq -72(%rdx), %rax movq %rax, 344(%rsp) movq -64(%rdx), %rax movq %rax, 352(%rsp) movq -56(%rdx), %rax movq %rax, 360(%rsp) movq -48(%rdx), %rax movq %rax, 368(%rsp) movq -40(%rdx), %rax movq %rax, 376(%rsp) movq -32(%rdx), %rax movq %rax, 384(%rsp) movq -24(%rdx), %rax movq %rax, 392(%rsp) movq -16(%rdx), %rax movq %rax, 400(%rsp) movq -8(%rdx), %rax cmpl $63, %r10d movq %rax, 408(%rsp) movaps %xmm3, 64(%rsp) movq 32(%rsp), %rsi movq 64(%rsp), %rax movaps %xmm2, 96(%rsp) movaps %xmm3, 112(%rsp) movaps 256(%rsp), %xmm4 movq %rsi, 416(%rsp) movq 104(%rsp), %rsi mulps 224(%rsp), %xmm4 movaps 272(%rsp), %xmm5 movq %rsi, 424(%rsp) movq %rax, 432(%rsp) mulps 240(%rsp), %xmm5 movq 120(%rsp), %rax movaps %xmm4, %xmm6 movq %rax, 440(%rsp) movaps %xmm4, 48(%rsp) movq 48(%rsp), %rsi mulps %xmm2, %xmm6 movaps %xmm4, 96(%rsp) movaps %xmm5, %xmm7 movaps %xmm5, 80(%rsp) mulps %xmm3, %xmm7 movaps %xmm6, (%rsp) movaps %xmm5, 112(%rsp) movaps %xmm7, 16(%rsp) movq %rsi, 448(%rsp) movq 104(%rsp), %rsi movq 80(%rsp), %rax movaps %xmm6, 96(%rsp) movq %rsi, 456(%rsp) movq (%rsp), %rsi movq %rax, 464(%rsp) movq %rsi, 480(%rsp) movq 120(%rsp), %rax movq 104(%rsp), %rsi movaps %xmm7, 112(%rsp) movq %rax, 472(%rsp) movq 16(%rsp), %rax movq %rsi, 488(%rsp) movaps 320(%rsp), %xmm3 movq %rax, 496(%rsp) movq 120(%rsp), %rax movaps 288(%rsp), %xmm1 movaps 336(%rsp), %xmm4 mulps %xmm3, %xmm1 movq %rax, 504(%rsp) movaps 304(%rsp), %xmm2 mulps %xmm4, %xmm2 movaps 384(%rsp), %xmm7 movaps %xmm1, 96(%rsp) movaps %xmm1, 64(%rsp) movaps %xmm2, 32(%rsp) movaps %xmm2, 112(%rsp) movaps 400(%rsp), %xmm1 movq 64(%rsp), %rsi movq 32(%rsp), %rax movaps 352(%rsp), %xmm3 movaps 368(%rsp), %xmm4 mulps %xmm7, %xmm3 movq %rsi, 512(%rsp) movq %rax, 528(%rsp) movq 104(%rsp), %rsi movq 120(%rsp), %rax mulps %xmm1, %xmm4 movq %rsi, 520(%rsp) movq %rax, 536(%rsp) movaps %xmm3, 80(%rsp) movq 80(%rsp), %rsi movaps %xmm3, %xmm2 movaps %xmm4, 48(%rsp) movaps %xmm3, 96(%rsp) movaps %xmm4, 112(%rsp) movq 48(%rsp), %rax movq %rsi, 544(%rsp) movq 104(%rsp), %rsi movq %rax, 560(%rsp) movq 120(%rsp), %rax movq %rsi, 552(%rsp) movq %rax, 568(%rsp) movaps 64(%rsp), %xmm5 movaps 32(%rsp), %xmm6 mulps %xmm2, %xmm5 movaps (%rsp), %xmm7 mulps %xmm4, %xmm6 movaps 16(%rsp), %xmm1 movaps %xmm5, %xmm4 movaps %xmm5, 80(%rsp) mulps %xmm4, %xmm7 movq 80(%rsp), %rsi movaps %xmm6, 64(%rsp) mulps %xmm6, %xmm1 movaps %xmm5, 96(%rsp) movaps %xmm6, 112(%rsp) movaps %xmm7, %xmm6 movaps %xmm7, 80(%rsp) movq 64(%rsp), %rax movq %rsi, 576(%rsp) movq 104(%rsp), %rsi movaps 704(%rsp), %xmm2 movq %rax, 592(%rsp) movq 120(%rsp), %rax movaps %xmm7, 96(%rsp) movaps %xmm1, %xmm7 movq %rsi, 584(%rsp) movaps %xmm1, 64(%rsp) mulps %xmm6, %xmm2 movaps %xmm1, 112(%rsp) movaps 720(%rsp), %xmm3 movq 80(%rsp), %rsi movq %rax, 600(%rsp) movq 64(%rsp), %rax mulps %xmm7, %xmm3 movaps %xmm2, 80(%rsp) movq %rsi, 608(%rsp) movq 104(%rsp), %rsi movaps %xmm2, 96(%rsp) movq %rax, 624(%rsp) movq 120(%rsp), %rax movq 80(%rsp), %r9 movq %rsi, 616(%rsp) movq %rax, 632(%rsp) movaps %xmm3, 80(%rsp) movq 80(%rsp), %rdi movq 104(%rsp), %r8 movaps %xmm3, 112(%rsp) movq 120(%rsp), %rax movq %r9, 640(%rsp) movq %r8, 648(%rsp) movq %rdi, 656(%rsp) movq %rax, 664(%rsp) ja .L653 subl $64, %ecx movq %r9, 128(%rsp) movq %r9, 672(%rsp) movl %ecx, %esi movq %r8, 680(%rsp) movq %rdi, 688(%rsp) shrl $6, %esi movq %rax, 696(%rsp) movq %r8, 136(%rsp) movl %esi, %edx sall $6, %esi movq %rdi, 144(%rsp) addq $1, %rdx movq %rax, 152(%rsp) subl %esi, %ecx salq $8, %rdx movss 128(%rsp), %xmm1 addq %rdx, %rbx .L632: testl %ecx, %ecx je .L636 .L637: addq $4, %rbx subl $1, %ecx mulss -4(%rbx), %xmm0 jne .L637 .L636: leaq 132(%rsp), %rax leaq 160(%rsp), %rdx jmp .L639 .L638: movss (%rax), %xmm1 addq $4, %rax .L639: cmpq %rdx, %rax mulss %xmm1, %xmm0 jne .L638 movss %xmm0, (%r12) leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret unroll4x2as_combine: pushq %r13 movq %rsi, %r13 pushq %r12 movq %rdi, %r12 pushq %rbp pushq %rbx subq $8, %rsp call vec_length movq %rax, %rbx movq %r12, %rdi movq %rax, %rbp shrq $63, %rbx addq %rax, %rbx sarq %rbx call get_vec_start movss .LC0(%rip), %xmm0 testq %rbx, %rbx movq %rax, %rcx movaps %xmm0, %xmm1 leaq (%rax,%rbx,4), %rax jle .L655 xorl %edx, %edx .L656: mulss (%rcx,%rdx,4), %xmm1 mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L656 .L655: leaq (%rbx,%rbx), %rdx cmpq %rdx, %rbp jle .L657 .L658: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbp, %rdx jne .L658 .L657: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret register_combiners: movl $combine1, %esi subq $8, %rsp movl $combine1_descr, %edx movq %rsi, %rdi call add_combiner movl $combine2_descr, %edx movl $combine1, %esi movl $combine2, %edi call add_combiner movl $combine3_descr, %edx movl $combine1, %esi movl $combine3, %edi call add_combiner movl $combine3w_descr, %edx movl $combine1, %esi movl $combine3w, %edi call add_combiner movl $combine4_descr, %edx movl $combine1, %esi movl $combine4, %edi call add_combiner movl $combine4b_descr, %edx movl $combine1, %esi movl $combine4b, %edi call add_combiner movl $combine4p_descr, %edx movl $combine1, %esi movl $combine4p, %edi call add_combiner movl $combine5_descr, %edx movl $combine1, %esi movl $combine5, %edi call add_combiner movl $combine5p_descr, %edx movl $combine1, %esi movl $combine5p, %edi call add_combiner movl $unroll2aw_descr, %edx movl $combine1, %esi movl $unroll2aw_combine, %edi call add_combiner movl $unroll3a_descr, %edx movl $combine1, %esi movl $unroll3a_combine, %edi call add_combiner movl $unroll4a_descr, %edx movl $combine1, %esi movl $unroll4a_combine, %edi call add_combiner movl $unroll5a_descr, %edx movl $combine1, %esi movl $unroll5a_combine, %edi call add_combiner movl $unroll6a_descr, %edx movl $combine1, %esi movl $unroll6a_combine, %edi call add_combiner movl $unroll7a_descr, %edx movl $combine1, %esi movl $unroll7a_combine, %edi call add_combiner movl $unroll8a_descr, %edx movl $combine1, %esi movl $unroll8a_combine, %edi call add_combiner movl $unroll9a_descr, %edx movl $combine1, %esi movl $unroll9a_combine, %edi call add_combiner movl $unroll10a_descr, %edx movl $combine1, %esi movl $unroll10a_combine, %edi call add_combiner movl $unroll16a_descr, %edx movl $combine1, %esi movl $unroll16a_combine, %edi call add_combiner movl $unroll2_descr, %edx movl $combine1, %esi movl $unroll2_combine, %edi call add_combiner movl $unroll3_descr, %edx movl $combine1, %esi movl $unroll3_combine, %edi call add_combiner movl $unroll4_descr, %edx movl $combine1, %esi movl $unroll4_combine, %edi call add_combiner movl $unroll8_descr, %edx movl $combine1, %esi movl $unroll8_combine, %edi call add_combiner movl $unroll16_descr, %edx movl $combine1, %esi movl $unroll16_combine, %edi call add_combiner movl $combine6_descr, %edx movl $combine1, %esi movl $combine6, %edi call add_combiner movl $unroll4x2a_descr, %edx movl $combine1, %esi movl $unroll4x2a_combine, %edi call add_combiner movl $unroll8x2a_descr, %edx movl $combine1, %esi movl $unroll8x2a_combine, %edi call add_combiner movl $unroll3x3a_descr, %edx movl $combine1, %esi movl $unroll3x3a_combine, %edi call add_combiner movl $unroll4x4a_descr, %edx movl $combine1, %esi movl $unroll4x4a_combine, %edi call add_combiner movl $unroll5x5a_descr, %edx movl $combine1, %esi movl $unroll5x5a_combine, %edi call add_combiner movl $unroll6x6a_descr, %edx movl $combine1, %esi movl $unroll6x6a_combine, %edi call add_combiner movl $unroll7x7a_descr, %edx movl $combine1, %esi movl $unroll7x7a_combine, %edi call add_combiner movl $unroll8x4a_descr, %edx movl $combine1, %esi movl $unroll8x4a_combine, %edi call add_combiner movl $unroll8x8a_descr, %edx movl $combine1, %esi movl $unroll8x8a_combine, %edi call add_combiner movl $unroll9x9a_descr, %edx movl $combine1, %esi movl $unroll9x9a_combine, %edi call add_combiner movl $unroll10x10a_descr, %edx movl $combine1, %esi movl $unroll10x10a_combine, %edi call add_combiner movl $unroll12x6a_descr, %edx movl $combine1, %esi movl $unroll12x6a_combine, %edi call add_combiner movl $unroll12x12a_descr, %edx movl $combine1, %esi movl $unroll12x12a_combine, %edi call add_combiner movl $unroll16x16a_descr, %edx movl $combine1, %esi movl $unroll16x16a_combine, %edi call add_combiner movl $unroll20x20a_descr, %edx movl $combine1, %esi movl $unroll20x20a_combine, %edi call add_combiner movl $unroll8x2_descr, %edx movl $combine1, %esi movl $unroll8x2_combine, %edi call add_combiner movl $unroll8x4_descr, %edx movl $combine1, %esi movl $unroll8x4_combine, %edi call add_combiner movl $unroll8x8_descr, %edx movl $combine1, %esi movl $unroll8x8_combine, %edi call add_combiner movl $unroll9x3_descr, %edx movl $combine1, %esi movl $unroll9x3_combine, %edi call add_combiner movl $unrollx2as_descr, %edx movl $combine1, %esi movl $unrollx2as_combine, %edi call add_combiner movl $combine7_descr, %edx movl $combine1, %esi movl $combine7, %edi call add_combiner movl $unroll3aa_descr, %edx movl $combine1, %esi movl $unroll3aa_combine, %edi call add_combiner movl $unroll4aa_descr, %edx movl $combine1, %esi movl $unroll4aa_combine, %edi call add_combiner movl $unroll5aa_descr, %edx movl $combine1, %esi movl $unroll5aa_combine, %edi call add_combiner movl $unroll6aa_descr, %edx movl $combine1, %esi movl $unroll6aa_combine, %edi call add_combiner movl $unroll7aa_descr, %edx movl $combine1, %esi movl $unroll7aa_combine, %edi call add_combiner movl $unroll8aa_descr, %edx movl $combine1, %esi movl $unroll8aa_combine, %edi call add_combiner movl $unroll9aa_descr, %edx movl $combine1, %esi movl $unroll9aa_combine, %edi call add_combiner movl $unroll10aa_descr, %edx movl $combine1, %esi movl $unroll10aa_combine, %edi call add_combiner movl $unroll12aa_descr, %edx movl $combine1, %esi movl $unroll12aa_combine, %edi call add_combiner movl $simd_v1_descr, %edx movl $combine1, %esi movl $simd_v1_combine, %edi call add_combiner movl $simd_v2_descr, %edx movl $combine1, %esi movl $simd_v2_combine, %edi call add_combiner movl $simd_v4_descr, %edx movl $combine1, %esi movl $simd_v4_combine, %edi call add_combiner movl $simd_v8_descr, %edx movl $combine1, %esi movl $simd_v8_combine, %edi call add_combiner movl $simd_v10_descr, %edx movl $combine1, %esi movl $simd_v10_combine, %edi call add_combiner movl $simd_v12_descr, %edx movl $combine1, %esi movl $simd_v12_combine, %edi call add_combiner movl $simd_v2a_descr, %edx movl $combine1, %esi movl $simd_v2a_combine, %edi call add_combiner movl $simd_v4a_descr, %edx movl $combine1, %esi movl $simd_v4a_combine, %edi call add_combiner movl $simd_v8a_combine, %edi movl $simd_v8a_descr, %edx movl $combine1, %esi call add_combiner movsd .LC1(%rip), %xmm1 movl $simd_v8a_combine, %edi movsd .LC2(%rip), %xmm0 addq $8, %rsp jmp log_combiner simd_v8a_descr: simd_v4a_descr: simd_v2a_descr: simd_v12_descr: simd_v10_descr: simd_v8_descr: simd_v4_descr: simd_v2_descr: simd_v1_descr: unroll12aa_descr: unroll10aa_descr: unroll9aa_descr: unroll8aa_descr: unroll7aa_descr: unroll6aa_descr: unroll5aa_descr: unroll4aa_descr: unroll3aa_descr: combine7_descr: unroll8x8_descr: unroll8x4_descr: unroll9x3_descr: unroll8x2_descr: unroll4x2as_descr: unrollx2as_descr: unroll10x10a_descr: unroll9x9a_descr: unroll8x8a_descr: unroll7x7a_descr: unroll6x6a_descr: unroll5x5a_descr: unroll20x20a_descr: unroll16x16a_descr: unroll12x12a_descr: unroll12x6a_descr: unroll8x4a_descr: unroll4x4a_descr: unroll3x3a_descr: unroll8x2a_descr: unroll4x2a_descr: combine6_descr: unroll16_descr: unroll8_descr: unroll4_descr: unroll3_descr: unroll2_descr: unroll16a_descr: unroll10a_descr: unroll9a_descr: unroll8a_descr: unroll7a_descr: unroll6a_descr: unroll5a_descr: unroll4a_descr: unroll2aw_descr: combine5p_descr: unroll3a_descr: combine5_descr: combine4p_descr: combine4b_descr: combine4_descr: combine3w_descr: combine3_descr: combine2_descr: combine1_descr: