combine1: pushq %r12 pushq %rbp pushq %rbx subq $16, %rsp movq %rdi, %r12 movq %rsi, %rbp movl $0x3f800000, (%rsi) movl $0, %ebx jmp .L2 .L3: leaq 12(%rsp), %rdx movq %rbx, %rsi movq %r12, %rdi call get_vec_element movss 0(%rbp), %xmm0 mulss 12(%rsp), %xmm0 movss %xmm0, 0(%rbp) addq $1, %rbx .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 pushq %r12 pushq %rbp pushq %rbx subq $24, %rsp movq %rdi, %r13 movq %rsi, %rbp call vec_length movq %rax, %r12 movl $0x3f800000, 0(%rbp) testq %rax, %rax jle .L5 movl $0, %ebx .L7: leaq 12(%rsp), %rdx movq %rbx, %rsi movq %r13, %rdi call get_vec_element movss 0(%rbp), %xmm0 mulss 12(%rsp), %xmm0 movss %xmm0, 0(%rbp) addq $1, %rbx cmpq %r12, %rbx jne .L7 .L5: addq $24, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret combine4b: pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %rbx movq %rsi, %rbp call vec_length testq %rax, %rax jle .L13 movss .LC0(%rip), %xmm0 movl $0, %edx .L12: testq %rdx, %rdx js .L11 cmpq %rdx, (%rbx) jle .L11 movq 8(%rbx), %rcx mulss (%rcx,%rdx,4), %xmm0 .L11: addq $1, %rdx cmpq %rax, %rdx jne .L12 jmp .L10 .L13: movss .LC0(%rip), %xmm0 .L10: movss %xmm0, 0(%rbp) addq $8, %rsp popq %rbx popq %rbp ret combine3: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbp movq %rsi, %rbx call vec_length movq %rax, %r12 movq %rbp, %rdi call get_vec_start movl $0x3f800000, (%rbx) testq %r12, %r12 jle .L15 movq %rax, %rdx leaq (%rax,%r12,4), %rax .L17: movss (%rbx), %xmm0 mulss (%rdx), %xmm0 movss %xmm0, (%rbx) addq $4, %rdx cmpq %rax, %rdx jne .L17 .L15: popq %rbx popq %rbp popq %r12 ret combine3w: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbp movq %rsi, %rbx call vec_length movq %rax, %r12 movq %rbp, %rdi call get_vec_start movl $0x3f800000, (%rbx) testq %r12, %r12 jle .L19 movq %rax, %rdx leaq (%rax,%r12,4), %rax movss .LC0(%rip), %xmm0 .L21: mulss (%rdx), %xmm0 movss %xmm0, (%rbx) addq $4, %rdx cmpq %rax, %rdx jne .L21 .L19: popq %rbx popq %rbp popq %r12 ret combine4: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %r12 call vec_length movq %rax, %rbp movq %rbx, %rdi call get_vec_start testq %rbp, %rbp jle .L26 movq %rax, %rdx leaq (%rax,%rbp,4), %rax movss .LC0(%rip), %xmm0 .L25: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rax, %rdx jne .L25 jmp .L24 .L26: movss .LC0(%rip), %xmm0 .L24: movss %xmm0, (%r12) popq %rbx popq %rbp popq %r12 ret combine4p: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rdx leaq (%rax,%r12,4), %rax cmpq %rax, %rdx jnb .L31 movss .LC0(%rip), %xmm0 .L30: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L30 jmp .L29 .L31: movss .LC0(%rip), %xmm0 .L29: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret combine5: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -1(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L38 movss .LC0(%rip), %xmm0 movl $0, %edx .L35: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm0 addq $2, %rdx cmpq %rdx, %rbp jg .L35 jmp .L34 .L38: movss .LC0(%rip), %xmm0 movl $0, %edx .L34: cmpq %rdx, %rbx jle .L36 .L37: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L37 .L36: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll3a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -2(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L45 movss .LC0(%rip), %xmm0 movl $0, %edx .L42: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm0 mulss 8(%rax,%rdx,4), %xmm0 addq $3, %rdx cmpq %rdx, %rbp jg .L42 jmp .L41 .L45: movss .LC0(%rip), %xmm0 movl $0, %edx .L41: cmpq %rdx, %rbx jle .L43 .L44: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L44 .L43: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret combine5p: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbp movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %rbp, %rdi call vec_length leaq (%rbx,%rax,4), %rax leaq -4(%rax), %rcx cmpq %rcx, %rbx jnb .L53 movq %rbx, %rdx movss .LC0(%rip), %xmm0 .L50: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 addq $8, %rdx cmpq %rdx, %rcx ja .L50 movq %rax, %rdx subq %rbx, %rdx leaq -5(%rdx), %rdx shrq $3, %rdx leaq 8(%rbx,%rdx,8), %rbx jmp .L48 .L53: movss .LC0(%rip), %xmm0 .L48: cmpq %rbx, %rax jbe .L51 .L52: mulss (%rbx), %xmm0 addq $4, %rbx cmpq %rbx, %rax ja .L52 .L51: movss %xmm0, (%r12) popq %rbx popq %rbp popq %r12 ret unroll2aw_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -1(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L60 movss .LC0(%rip), %xmm0 movl $0, %edx .L57: mulss (%rax,%rdx,4), %xmm0 addq $2, %rdx mulss -4(%rax,%rdx,4), %xmm0 cmpq %rdx, %rbp jg .L57 jmp .L56 .L60: movss .LC0(%rip), %xmm0 movl $0, %edx .L56: cmpq %rdx, %rbx jle .L58 .L59: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L59 .L58: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -3(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L67 movss .LC0(%rip), %xmm0 movl $0, %edx .L64: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm0 mulss 8(%rax,%rdx,4), %xmm0 mulss 12(%rax,%rdx,4), %xmm0 addq $4, %rdx cmpq %rdx, %rbp jg .L64 jmp .L63 .L67: movss .LC0(%rip), %xmm0 movl $0, %edx .L63: cmpq %rdx, %rbx jle .L65 .L66: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L66 .L65: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll5a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -4(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L74 movq %rax, %rcx movss .LC0(%rip), %xmm0 movl $0, %edx .L71: mulss (%rcx), %xmm0 mulss 4(%rcx), %xmm0 mulss 8(%rcx), %xmm0 mulss 12(%rcx), %xmm0 mulss 16(%rcx), %xmm0 addq $5, %rdx addq $20, %rcx cmpq %rdx, %rbp jg .L71 jmp .L70 .L74: movss .LC0(%rip), %xmm0 movl $0, %edx .L70: cmpq %rdx, %rbx jle .L72 .L73: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L73 .L72: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll6a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -5(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L81 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L78: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 addq $6, %rcx addq $24, %rdx cmpq %rcx, %rbp jg .L78 jmp .L77 .L81: movss .LC0(%rip), %xmm0 movl $0, %ecx .L77: cmpq %rcx, %rbx jle .L79 .L80: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L80 .L79: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll7a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -6(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L88 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L85: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 addq $7, %rcx addq $28, %rdx cmpq %rcx, %rbp jg .L85 jmp .L84 .L88: movss .LC0(%rip), %xmm0 movl $0, %ecx .L84: cmpq %rcx, %rbx jle .L86 .L87: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L87 .L86: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -7(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L95 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L92: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 addq $8, %rcx addq $32, %rdx cmpq %rcx, %rbp jg .L92 jmp .L91 .L95: movss .LC0(%rip), %xmm0 movl $0, %ecx .L91: cmpq %rcx, %rbx jle .L93 .L94: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L94 .L93: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll9a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -8(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L102 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L99: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 mulss 32(%rdx), %xmm0 addq $9, %rcx addq $36, %rdx cmpq %rcx, %rbp jg .L99 jmp .L98 .L102: movss .LC0(%rip), %xmm0 movl $0, %ecx .L98: cmpq %rcx, %rbx jle .L100 .L101: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L101 .L100: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll10a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -9(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L109 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L106: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 mulss 32(%rdx), %xmm0 mulss 36(%rdx), %xmm0 addq $10, %rcx addq $40, %rdx cmpq %rcx, %rbp jg .L106 jmp .L105 .L109: movss .LC0(%rip), %xmm0 movl $0, %ecx .L105: cmpq %rcx, %rbx jle .L107 .L108: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L108 .L107: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll16a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -15(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L116 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L113: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 mulss 32(%rdx), %xmm0 mulss 36(%rdx), %xmm0 mulss 40(%rdx), %xmm0 mulss 44(%rdx), %xmm0 mulss 48(%rdx), %xmm0 mulss 52(%rdx), %xmm0 mulss 56(%rdx), %xmm0 mulss 60(%rdx), %xmm0 addq $16, %rcx addq $64, %rdx cmpq %rcx, %rbp jg .L113 jmp .L112 .L116: movss .LC0(%rip), %xmm0 movl $0, %ecx .L112: cmpq %rcx, %rbx jle .L114 .L115: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L115 .L114: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll2_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %r12 movq %rsi, %rbp call vec_length movq %rax, %rbx movq %r12, %rdi call get_vec_start movq %rax, %rcx movq %rbx, %rax shrq $63, %rax leaq (%rbx,%rax), %rsi andl $1, %esi subq %rax, %rsi movslq %esi, %rsi subq %rsi, %rbx leaq (%rcx,%rbx,4), %rax cmpq %rax, %rcx jnb .L124 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L121: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 addq $8, %rdx cmpq %rdx, %rax ja .L121 movq %rcx, %rdx notq %rdx addq %rax, %rdx shrq $3, %rdx leaq 8(%rcx,%rdx,8), %rcx jmp .L119 .L124: movss .LC0(%rip), %xmm0 .L119: leaq (%rax,%rsi,4), %rax cmpq %rcx, %rax jbe .L122 .L123: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L123 .L122: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll3_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rdx leaq -8(%rax,%r12,4), %rax cmpq %rax, %rdx jnb .L131 movss .LC0(%rip), %xmm0 .L128: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 addq $12, %rdx cmpq %rdx, %rax ja .L128 jmp .L127 .L131: movss .LC0(%rip), %xmm0 .L127: addq $8, %rax cmpq %rdx, %rax jbe .L129 .L130: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L130 .L129: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll4_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rcx leaq -12(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L139 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L136: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 addq $16, %rdx cmpq %rdx, %rax ja .L136 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-16, %rdx leaq 16(%rcx,%rdx), %rcx jmp .L134 .L139: movss .LC0(%rip), %xmm0 .L134: addq $12, %rax cmpq %rcx, %rax jbe .L137 .L138: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L138 .L137: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll8_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %r12 movq %rsi, %rbp call vec_length movq %rax, %rbx movq %r12, %rdi call get_vec_start movq %rax, %rcx movq %rbx, %rax sarq $63, %rax shrq $61, %rax leaq (%rbx,%rax), %rsi andl $7, %esi subq %rax, %rsi movslq %esi, %rsi subq %rsi, %rbx leaq (%rcx,%rbx,4), %rax cmpq %rax, %rcx jnb .L147 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L144: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 addq $32, %rdx cmpq %rdx, %rax ja .L144 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx jmp .L142 .L147: movss .LC0(%rip), %xmm0 .L142: leaq (%rax,%rsi,4), %rax cmpq %rcx, %rax jbe .L145 .L146: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L146 .L145: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll16_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %r12 movq %rsi, %rbp call vec_length movq %rax, %rbx movq %r12, %rdi call get_vec_start movq %rax, %rcx movq %rbx, %rax sarq $63, %rax shrq $60, %rax leaq (%rbx,%rax), %rsi andl $15, %esi subq %rax, %rsi movslq %esi, %rsi subq %rsi, %rbx leaq (%rcx,%rbx,4), %rax cmpq %rax, %rcx jnb .L155 movq %rcx, %rdx movss .LC0(%rip), %xmm0 .L152: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm0 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm0 mulss 32(%rdx), %xmm0 mulss 36(%rdx), %xmm0 mulss 40(%rdx), %xmm0 mulss 44(%rdx), %xmm0 mulss 48(%rdx), %xmm0 mulss 52(%rdx), %xmm0 mulss 56(%rdx), %xmm0 mulss 60(%rdx), %xmm0 addq $64, %rdx cmpq %rdx, %rax ja .L152 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-64, %rdx leaq 64(%rcx,%rdx), %rcx jmp .L150 .L155: movss .LC0(%rip), %xmm0 .L150: leaq (%rax,%rsi,4), %rax cmpq %rcx, %rax jbe .L153 .L154: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L154 .L153: movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret combine6: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -1(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L162 movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %edx .L159: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm1 addq $2, %rdx cmpq %rdx, %rbp jg .L159 jmp .L158 .L162: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %edx .L158: cmpq %rdx, %rbx jle .L160 .L161: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L161 .L160: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4x2a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -3(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L169 movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %edx .L166: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm1 mulss 8(%rax,%rdx,4), %xmm0 mulss 12(%rax,%rdx,4), %xmm1 addq $4, %rdx cmpq %rdx, %rbp jg .L166 jmp .L165 .L169: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %edx .L165: cmpq %rdx, %rbx jle .L167 .L168: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L168 .L167: mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8x2a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -7(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L176 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %ecx .L173: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm1 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm1 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm1 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm1 addq $8, %rcx addq $32, %rdx cmpq %rcx, %rbp jg .L173 jmp .L172 .L176: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 movl $0, %ecx .L172: cmpq %rcx, %rbx jle .L174 .L175: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L175 .L174: mulss %xmm1, %xmm0 movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll3x3a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -2(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L183 movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 movl $0, %edx .L180: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm2 mulss 8(%rax,%rdx,4), %xmm1 addq $3, %rdx cmpq %rdx, %rbp jg .L180 jmp .L179 .L183: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 movl $0, %edx .L179: cmpq %rdx, %rbx jle .L181 .L182: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L182 .L181: mulss %xmm2, %xmm0 mulss %xmm0, %xmm1 movss %xmm1, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4x4a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -3(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L190 movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 movl $0, %edx .L187: mulss (%rax,%rdx,4), %xmm0 mulss 4(%rax,%rdx,4), %xmm3 mulss 8(%rax,%rdx,4), %xmm2 mulss 12(%rax,%rdx,4), %xmm1 addq $4, %rdx cmpq %rdx, %rbp jg .L187 jmp .L186 .L190: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 movl $0, %edx .L186: cmpq %rdx, %rbx jle .L188 .L189: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L189 .L188: mulss %xmm3, %xmm0 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8x4a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -7(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L197 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 movl $0, %ecx .L194: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm3 mulss 8(%rdx), %xmm2 mulss 12(%rdx), %xmm1 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm3 mulss 24(%rdx), %xmm2 mulss 28(%rdx), %xmm1 addq $8, %rcx addq $32, %rdx cmpq %rcx, %rbp jg .L194 jmp .L193 .L197: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm0 movl $0, %ecx .L193: cmpq %rcx, %rbx jle .L195 .L196: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L196 .L195: mulss %xmm3, %xmm0 mulss %xmm0, %xmm2 mulss %xmm2, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll12x6a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -11(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L204 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm0 movl $0, %ecx .L201: mulss (%rdx), %xmm0 mulss 24(%rdx), %xmm0 mulss 4(%rdx), %xmm5 mulss 28(%rdx), %xmm5 mulss 8(%rdx), %xmm4 mulss 32(%rdx), %xmm4 mulss 12(%rdx), %xmm3 mulss 36(%rdx), %xmm3 mulss 16(%rdx), %xmm2 mulss 40(%rdx), %xmm2 mulss 20(%rdx), %xmm1 mulss 44(%rdx), %xmm1 addq $12, %rcx addq $48, %rdx cmpq %rcx, %rbp jg .L201 jmp .L200 .L204: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm0 movl $0, %ecx .L200: cmpq %rcx, %rbx jle .L202 .L203: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L203 .L202: mulss %xmm5, %xmm0 mulss %xmm4, %xmm3 mulss %xmm3, %xmm0 mulss %xmm1, %xmm2 movaps %xmm0, %xmm1 mulss %xmm2, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll12x12a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -11(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L211 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 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, %xmm0 movl $0, %ecx .L208: mulss (%rdx), %xmm0 mulss 24(%rdx), %xmm6 mulss 4(%rdx), %xmm11 mulss 28(%rdx), %xmm5 mulss 8(%rdx), %xmm10 mulss 32(%rdx), %xmm4 mulss 12(%rdx), %xmm9 mulss 36(%rdx), %xmm3 mulss 16(%rdx), %xmm8 mulss 40(%rdx), %xmm2 mulss 20(%rdx), %xmm7 mulss 44(%rdx), %xmm1 addq $12, %rcx addq $48, %rdx cmpq %rcx, %rbp jg .L208 jmp .L207 .L211: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 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, %xmm0 movl $0, %ecx .L207: cmpq %rcx, %rbx jle .L209 .L210: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L210 .L209: mulss %xmm11, %xmm0 mulss %xmm10, %xmm9 mulss %xmm9, %xmm0 mulss %xmm7, %xmm8 movaps %xmm0, %xmm7 mulss %xmm8, %xmm7 mulss %xmm5, %xmm6 movaps %xmm7, %xmm5 mulss %xmm6, %xmm5 mulss %xmm3, %xmm4 movaps %xmm5, %xmm3 mulss %xmm4, %xmm3 mulss %xmm1, %xmm2 mulss %xmm2, %xmm3 movss %xmm3, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll16x16a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -15(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L218 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 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, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movaps %xmm1, %xmm0 movl $0, %ecx .L215: mulss (%rdx), %xmm0 mulss 24(%rdx), %xmm10 mulss 4(%rdx), %xmm15 mulss 28(%rdx), %xmm9 mulss 8(%rdx), %xmm14 mulss 32(%rdx), %xmm8 mulss 12(%rdx), %xmm13 mulss 36(%rdx), %xmm7 mulss 16(%rdx), %xmm12 mulss 40(%rdx), %xmm6 mulss 20(%rdx), %xmm11 mulss 44(%rdx), %xmm5 mulss 48(%rdx), %xmm4 mulss 52(%rdx), %xmm3 mulss 56(%rdx), %xmm2 mulss 60(%rdx), %xmm1 addq $16, %rcx addq $64, %rdx cmpq %rcx, %rbp jg .L215 jmp .L214 .L218: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 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, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movaps %xmm1, %xmm0 movl $0, %ecx .L214: cmpq %rcx, %rbx jle .L216 .L217: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L217 .L216: mulss %xmm15, %xmm0 mulss %xmm14, %xmm13 mulss %xmm13, %xmm0 mulss %xmm11, %xmm12 mulss %xmm12, %xmm0 mulss %xmm10, %xmm9 mulss %xmm8, %xmm7 mulss %xmm9, %xmm7 mulss %xmm5, %xmm6 movaps %xmm7, %xmm5 mulss %xmm6, %xmm5 mulss %xmm0, %xmm5 mulss %xmm4, %xmm3 mulss %xmm2, %xmm1 mulss %xmm1, %xmm3 mulss %xmm3, %xmm5 movss %xmm5, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll20x20a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $40, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -19(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L225 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm10 movss %xmm1, 24(%rsp) movss %xmm1, 20(%rsp) movss %xmm1, 16(%rsp) movss %xmm1, 12(%rsp) movaps %xmm1, %xmm11 movaps %xmm1, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movl $0, %ecx movss %xmm1, 28(%rsp) .L222: movss 28(%rsp), %xmm0 mulss (%rdx), %xmm0 movss %xmm0, 28(%rsp) movss 12(%rsp), %xmm0 mulss 24(%rdx), %xmm0 movss %xmm0, 12(%rsp) mulss 4(%rdx), %xmm15 movss 16(%rsp), %xmm0 mulss 28(%rdx), %xmm0 movss %xmm0, 16(%rsp) mulss 8(%rdx), %xmm14 movss 20(%rsp), %xmm0 mulss 32(%rdx), %xmm0 movss %xmm0, 20(%rsp) mulss 12(%rdx), %xmm13 movss 24(%rsp), %xmm0 mulss 36(%rdx), %xmm0 movss %xmm0, 24(%rsp) mulss 16(%rdx), %xmm12 mulss 40(%rdx), %xmm10 mulss 20(%rdx), %xmm11 mulss 44(%rdx), %xmm9 mulss 48(%rdx), %xmm8 mulss 52(%rdx), %xmm7 mulss 56(%rdx), %xmm6 mulss 60(%rdx), %xmm5 mulss 64(%rdx), %xmm4 mulss 68(%rdx), %xmm3 mulss 72(%rdx), %xmm2 mulss 76(%rdx), %xmm1 addq $20, %rcx addq $80, %rdx cmpq %rcx, %rbp jg .L222 movss 28(%rsp), %xmm0 jmp .L221 .L225: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm10 movss %xmm1, 24(%rsp) movss %xmm1, 20(%rsp) movss %xmm1, 16(%rsp) movss %xmm1, 12(%rsp) movaps %xmm1, %xmm11 movaps %xmm1, %xmm12 movaps %xmm1, %xmm13 movaps %xmm1, %xmm14 movaps %xmm1, %xmm15 movaps %xmm1, %xmm0 movl $0, %ecx .L221: cmpq %rcx, %rbx jle .L223 .L224: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L224 .L223: mulss %xmm15, %xmm0 mulss %xmm14, %xmm13 mulss %xmm13, %xmm0 mulss %xmm12, %xmm11 mulss %xmm11, %xmm0 movss 12(%rsp), %xmm12 mulss 16(%rsp), %xmm12 movss 20(%rsp), %xmm11 mulss 24(%rsp), %xmm11 mulss %xmm12, %xmm11 mulss %xmm9, %xmm10 movaps %xmm11, %xmm9 mulss %xmm10, %xmm9 mulss %xmm9, %xmm0 mulss %xmm8, %xmm7 mulss %xmm6, %xmm5 mulss %xmm7, %xmm5 mulss %xmm4, %xmm3 mulss %xmm2, %xmm1 mulss %xmm3, %xmm1 mulss %xmm5, %xmm1 mulss %xmm1, %xmm0 movss %xmm0, (%r12) addq $40, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll5x5a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -4(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L232 movq %rax, %rcx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm0 movl $0, %edx .L229: mulss (%rcx), %xmm0 mulss 4(%rcx), %xmm4 mulss 8(%rcx), %xmm3 mulss 12(%rcx), %xmm2 mulss 16(%rcx), %xmm1 addq $5, %rdx addq $20, %rcx cmpq %rdx, %rbp jg .L229 jmp .L228 .L232: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm0 movl $0, %edx .L228: cmpq %rdx, %rbx jle .L230 .L231: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L231 .L230: mulss %xmm4, %xmm0 mulss %xmm3, %xmm2 mulss %xmm2, %xmm1 mulss %xmm0, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll6x6a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -5(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L239 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm0 movl $0, %ecx .L236: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm5 mulss 8(%rdx), %xmm4 mulss 12(%rdx), %xmm3 mulss 16(%rdx), %xmm2 mulss 20(%rdx), %xmm1 addq $6, %rcx addq $24, %rdx cmpq %rcx, %rbp jg .L236 jmp .L235 .L239: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm0 movl $0, %ecx .L235: cmpq %rcx, %rbx jle .L237 .L238: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L238 .L237: mulss %xmm5, %xmm0 mulss %xmm4, %xmm3 mulss %xmm3, %xmm0 mulss %xmm1, %xmm2 movaps %xmm0, %xmm1 mulss %xmm2, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll7x7a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -6(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L246 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm0 movl $0, %ecx .L243: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm6 mulss 8(%rdx), %xmm5 mulss 12(%rdx), %xmm4 mulss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm2 mulss 24(%rdx), %xmm1 addq $7, %rcx addq $28, %rdx cmpq %rcx, %rbp jg .L243 jmp .L242 .L246: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm0 movl $0, %ecx .L242: cmpq %rcx, %rbx jle .L244 .L245: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L245 .L244: mulss %xmm6, %xmm0 mulss %xmm5, %xmm4 mulss %xmm4, %xmm0 mulss %xmm3, %xmm2 mulss %xmm2, %xmm1 mulss %xmm0, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8x8a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -7(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L253 movq %rax, %rdx 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 movl $0, %ecx .L250: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm7 mulss 8(%rdx), %xmm6 mulss 12(%rdx), %xmm5 mulss 16(%rdx), %xmm4 mulss 20(%rdx), %xmm3 mulss 24(%rdx), %xmm2 mulss 28(%rdx), %xmm1 addq $8, %rcx addq $32, %rdx cmpq %rcx, %rbp jg .L250 jmp .L249 .L253: 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 movl $0, %ecx .L249: cmpq %rcx, %rbx jle .L251 .L252: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L252 .L251: mulss %xmm7, %xmm0 mulss %xmm6, %xmm5 mulss %xmm5, %xmm0 mulss %xmm4, %xmm3 mulss %xmm2, %xmm1 mulss %xmm3, %xmm1 mulss %xmm1, %xmm0 movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll9x9a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -8(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L260 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm0 movl $0, %ecx .L257: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm8 mulss 8(%rdx), %xmm7 mulss 12(%rdx), %xmm6 mulss 16(%rdx), %xmm5 mulss 20(%rdx), %xmm4 mulss 24(%rdx), %xmm3 mulss 28(%rdx), %xmm2 mulss 32(%rdx), %xmm1 addq $9, %rcx addq $36, %rdx cmpq %rcx, %rbp jg .L257 jmp .L256 .L260: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm0 movl $0, %ecx .L256: cmpq %rcx, %rbx jle .L258 .L259: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L259 .L258: mulss %xmm8, %xmm0 mulss %xmm7, %xmm6 mulss %xmm6, %xmm0 mulss %xmm5, %xmm4 mulss %xmm3, %xmm2 mulss %xmm4, %xmm2 mulss %xmm2, %xmm1 mulss %xmm0, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll10x10a_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -9(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L267 movq %rax, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm0 movl $0, %ecx .L264: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm9 mulss 8(%rdx), %xmm8 mulss 12(%rdx), %xmm7 mulss 16(%rdx), %xmm6 mulss 20(%rdx), %xmm5 mulss 24(%rdx), %xmm4 mulss 28(%rdx), %xmm3 mulss 32(%rdx), %xmm2 mulss 36(%rdx), %xmm1 addq $10, %rcx addq $40, %rdx cmpq %rcx, %rbp jg .L264 jmp .L263 .L267: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm3 movaps %xmm1, %xmm4 movaps %xmm1, %xmm5 movaps %xmm1, %xmm6 movaps %xmm1, %xmm7 movaps %xmm1, %xmm8 movaps %xmm1, %xmm9 movaps %xmm1, %xmm0 movl $0, %ecx .L263: cmpq %rcx, %rbx jle .L265 .L266: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L266 .L265: mulss %xmm9, %xmm0 mulss %xmm8, %xmm7 mulss %xmm7, %xmm0 mulss %xmm6, %xmm5 mulss %xmm4, %xmm3 mulss %xmm5, %xmm3 mulss %xmm3, %xmm0 mulss %xmm1, %xmm2 movaps %xmm0, %xmm1 mulss %xmm2, %xmm1 movss %xmm1, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unrollx2as_combine: pushq %r14 pushq %r13 pushq %r12 pushq %rbp pushq %rbx movq %rdi, %r14 movq %rsi, %r12 call vec_length movq %rax, %rbx movq %rax, %r13 shrq $63, %r13 addq %rax, %r13 sarq %r13 movq %r14, %rdi call get_vec_start movq %rax, %rcx leaq (%rax,%r13,4), %rax testq %r13, %r13 jle .L274 movq %r13, %rbp movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 movl $0, %edx .L271: mulss (%rcx,%rdx,4), %xmm1 mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbp, %rdx jne .L271 jmp .L270 .L274: movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 .L270: leaq (%r13,%r13), %rdx cmpq %rdx, %rbx jle .L272 .L273: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L273 .L272: mulss %xmm1, %xmm0 movss %xmm0, (%r12) popq %rbx popq %rbp popq %r12 popq %r13 popq %r14 ret unroll8x2_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L282 movq %rcx, %rdx movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 .L279: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm1 mulss 8(%rdx), %xmm0 mulss 12(%rdx), %xmm1 mulss 16(%rdx), %xmm0 mulss 20(%rdx), %xmm1 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm1 addq $32, %rdx cmpq %rdx, %rax ja .L279 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx jmp .L277 .L282: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm0 .L277: addq $28, %rax cmpq %rcx, %rax jbe .L280 .L281: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L281 .L280: mulss %xmm1, %xmm0 movss %xmm0, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll9x3_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rdx leaq -32(%rax,%r12,4), %rax cmpq %rax, %rdx jnb .L289 movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L286: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm2 mulss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm2 mulss 20(%rdx), %xmm1 mulss 24(%rdx), %xmm0 mulss 28(%rdx), %xmm2 mulss 32(%rdx), %xmm1 addq $36, %rdx cmpq %rdx, %rax ja .L286 jmp .L285 .L289: movss .LC0(%rip), %xmm1 movaps %xmm1, %xmm2 movaps %xmm1, %xmm0 .L285: addq $32, %rax cmpq %rdx, %rax jbe .L287 .L288: mulss (%rdx), %xmm0 addq $4, %rdx cmpq %rdx, %rax ja .L288 .L287: mulss %xmm2, %xmm0 mulss %xmm0, %xmm1 movss %xmm1, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll8x4_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L297 movq %rcx, %rdx movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 movaps %xmm0, %xmm2 movaps %xmm0, %xmm3 .L294: mulss (%rdx), %xmm3 mulss 4(%rdx), %xmm2 mulss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm0 mulss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm2 mulss 24(%rdx), %xmm1 mulss 28(%rdx), %xmm0 addq $32, %rdx cmpq %rdx, %rax ja .L294 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx jmp .L292 .L297: movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 movaps %xmm0, %xmm2 movaps %xmm0, %xmm3 .L292: addq $28, %rax cmpq %rcx, %rax jbe .L295 .L296: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L296 .L295: mulss %xmm3, %xmm0 mulss %xmm0, %xmm2 mulss %xmm2, %xmm1 movss %xmm1, 0(%rbp) popq %rbx popq %rbp popq %r12 ret unroll8x8_combine: pushq %r12 pushq %rbp pushq %rbx movq %rdi, %rbx movq %rsi, %rbp call vec_length movq %rax, %r12 movq %rbx, %rdi call get_vec_start movq %rax, %rcx leaq -28(%rax,%r12,4), %rax cmpq %rax, %rcx jnb .L305 movq %rcx, %rdx 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 .L302: mulss (%rdx), %xmm0 mulss 4(%rdx), %xmm7 mulss 8(%rdx), %xmm6 mulss 12(%rdx), %xmm5 mulss 16(%rdx), %xmm4 mulss 20(%rdx), %xmm3 mulss 24(%rdx), %xmm2 mulss 28(%rdx), %xmm1 addq $32, %rdx cmpq %rdx, %rax ja .L302 movq %rcx, %rdx notq %rdx addq %rax, %rdx andq $-32, %rdx leaq 32(%rcx,%rdx), %rcx jmp .L300 .L305: 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 .L300: addq $28, %rax cmpq %rcx, %rax jbe .L303 .L304: mulss (%rcx), %xmm0 addq $4, %rcx cmpq %rcx, %rax ja .L304 .L303: mulss %xmm7, %xmm0 mulss %xmm0, %xmm6 mulss %xmm6, %xmm5 mulss %xmm5, %xmm4 mulss %xmm4, %xmm3 mulss %xmm3, %xmm2 mulss %xmm2, %xmm1 movss %xmm1, 0(%rbp) popq %rbx popq %rbp popq %r12 ret combine7: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -1(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L312 movss .LC0(%rip), %xmm0 movl $0, %edx .L309: movss (%rax,%rdx,4), %xmm1 mulss 4(%rax,%rdx,4), %xmm1 mulss %xmm1, %xmm0 addq $2, %rdx cmpq %rdx, %rbp jg .L309 jmp .L308 .L312: movss .LC0(%rip), %xmm0 movl $0, %edx .L308: cmpq %rdx, %rbx jle .L310 .L311: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L311 .L310: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll3aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -2(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L319 movss .LC0(%rip), %xmm0 movl $0, %edx .L316: movss (%rax,%rdx,4), %xmm1 mulss 4(%rax,%rdx,4), %xmm1 mulss 8(%rax,%rdx,4), %xmm1 mulss %xmm1, %xmm0 addq $3, %rdx cmpq %rdx, %rbp jg .L316 jmp .L315 .L319: movss .LC0(%rip), %xmm0 movl $0, %edx .L315: cmpq %rdx, %rbx jle .L317 .L318: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L318 .L317: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll4aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r12 movq %rsi, %r13 call vec_length movq %rax, %rbx leaq -3(%rax), %rbp movq %r12, %rdi call get_vec_start testq %rbp, %rbp jle .L326 movss .LC0(%rip), %xmm0 movl $0, %edx .L323: movss (%rax,%rdx,4), %xmm2 mulss 4(%rax,%rdx,4), %xmm2 movss 8(%rax,%rdx,4), %xmm1 mulss 12(%rax,%rdx,4), %xmm1 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $4, %rdx cmpq %rdx, %rbp jg .L323 jmp .L322 .L326: movss .LC0(%rip), %xmm0 movl $0, %edx .L322: cmpq %rdx, %rbx jle .L324 .L325: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L325 .L324: movss %xmm0, 0(%r13) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll5aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -4(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L333 movq %rax, %rcx movss .LC0(%rip), %xmm0 movl $0, %edx .L330: movss (%rcx), %xmm2 mulss 4(%rcx), %xmm2 movss 8(%rcx), %xmm1 mulss 12(%rcx), %xmm1 mulss %xmm2, %xmm1 mulss 16(%rcx), %xmm1 mulss %xmm1, %xmm0 addq $5, %rdx addq $20, %rcx cmpq %rdx, %rbp jg .L330 jmp .L329 .L333: movss .LC0(%rip), %xmm0 movl $0, %edx .L329: cmpq %rdx, %rbx jle .L331 .L332: mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L332 .L331: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll6aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -5(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L340 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L337: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm2, %xmm1 movss 16(%rdx), %xmm2 mulss 20(%rdx), %xmm2 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $6, %rcx addq $24, %rdx cmpq %rcx, %rbp jg .L337 jmp .L336 .L340: movss .LC0(%rip), %xmm0 movl $0, %ecx .L336: cmpq %rcx, %rbx jle .L338 .L339: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L339 .L338: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll7aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -6(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L347 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L344: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm2, %xmm1 movss 16(%rdx), %xmm2 mulss 20(%rdx), %xmm2 mulss 24(%rdx), %xmm2 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $7, %rcx addq $28, %rdx cmpq %rcx, %rbp jg .L344 jmp .L343 .L347: movss .LC0(%rip), %xmm0 movl $0, %ecx .L343: cmpq %rcx, %rbx jle .L345 .L346: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L346 .L345: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll8aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -7(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L354 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L351: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm1, %xmm2 movss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm3 movss 24(%rdx), %xmm1 mulss 28(%rdx), %xmm1 mulss %xmm3, %xmm1 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $8, %rcx addq $32, %rdx cmpq %rcx, %rbp jg .L351 jmp .L350 .L354: movss .LC0(%rip), %xmm0 movl $0, %ecx .L350: cmpq %rcx, %rbx jle .L352 .L353: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L353 .L352: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll9aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -8(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L361 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L358: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm2, %xmm1 movss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm3 movss 24(%rdx), %xmm2 mulss 28(%rdx), %xmm2 mulss %xmm3, %xmm2 mulss 32(%rdx), %xmm2 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $9, %rcx addq $36, %rdx cmpq %rcx, %rbp jg .L358 jmp .L357 .L361: movss .LC0(%rip), %xmm0 movl $0, %ecx .L357: cmpq %rcx, %rbx jle .L359 .L360: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L360 .L359: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll10aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -9(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L368 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L365: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm1, %xmm2 movss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm3 movss 24(%rdx), %xmm1 mulss 28(%rdx), %xmm1 mulss %xmm3, %xmm1 movss 32(%rdx), %xmm3 mulss 36(%rdx), %xmm3 mulss %xmm3, %xmm1 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $10, %rcx addq $40, %rdx cmpq %rcx, %rbp jg .L365 jmp .L364 .L368: movss .LC0(%rip), %xmm0 movl $0, %ecx .L364: cmpq %rcx, %rbx jle .L366 .L367: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L367 .L366: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret unroll12aa_combine: pushq %r13 pushq %r12 pushq %rbp pushq %rbx subq $8, %rsp movq %rdi, %r13 movq %rsi, %r12 call vec_length movq %rax, %rbx leaq -11(%rax), %rbp movq %r13, %rdi call get_vec_start testq %rbp, %rbp jle .L375 movq %rax, %rdx movss .LC0(%rip), %xmm0 movl $0, %ecx .L372: movss (%rdx), %xmm2 mulss 4(%rdx), %xmm2 movss 8(%rdx), %xmm1 mulss 12(%rdx), %xmm1 mulss %xmm1, %xmm2 movss 16(%rdx), %xmm3 mulss 20(%rdx), %xmm3 movss 24(%rdx), %xmm1 mulss 28(%rdx), %xmm1 mulss %xmm3, %xmm1 mulss %xmm2, %xmm1 movss 32(%rdx), %xmm3 mulss 36(%rdx), %xmm3 movss 40(%rdx), %xmm2 mulss 44(%rdx), %xmm2 mulss %xmm3, %xmm2 mulss %xmm2, %xmm1 mulss %xmm1, %xmm0 addq $12, %rcx addq $48, %rdx cmpq %rcx, %rbp jg .L372 jmp .L371 .L375: movss .LC0(%rip), %xmm0 movl $0, %ecx .L371: cmpq %rcx, %rbx jle .L373 .L374: mulss (%rax,%rcx,4), %xmm0 addq $1, %rcx cmpq %rbx, %rcx jne .L374 .L373: movss %xmm0, (%r12) addq $8, %rsp popq %rbx popq %rbp popq %r12 popq %r13 ret simd_v1_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $208, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L379: movss %xmm0, 48(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L379 testb $31, %bl je .L392 testl %eax, %eax je .L393 movss .LC0(%rip), %xmm0 .L385: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L382 jmp .L380 .L392: movss .LC0(%rip), %xmm0 .L380: movl %edx, %edi cmpl $7, %edx ja .L383 jmp .L384 .L382: testl %edx, %edx jne .L385 jmp .L384 .L383: movq 48(%rsp), %rax movq %rax, 112(%rsp) movq 56(%rsp), %rax movq %rax, 120(%rsp) movq 64(%rsp), %rax movq %rax, 128(%rsp) movq 72(%rsp), %rax movq %rax, 136(%rsp) movq %rbx, %rax .L387: movq (%rax), %rcx movq %rcx, 176(%rsp) movq 8(%rax), %rcx movq %rcx, 184(%rsp) movq 16(%rax), %rcx movq %rcx, 192(%rsp) movq 24(%rax), %rcx movq %rcx, 200(%rsp) movaps 112(%rsp), %xmm1 mulps 176(%rsp), %xmm1 movaps %xmm1, (%rsp) movq (%rsp), %rsi movaps %xmm1, 16(%rsp) movaps 128(%rsp), %xmm1 mulps 192(%rsp), %xmm1 movaps %xmm1, (%rsp) movq (%rsp), %rcx movaps %xmm1, 32(%rsp) movq %rsi, 144(%rsp) movq 24(%rsp), %rsi movq %rsi, 152(%rsp) movq %rcx, 160(%rsp) movq 40(%rsp), %rcx movq %rcx, 168(%rsp) addq $32, %rax subl $8, %edx cmpl $7, %edx jbe .L386 movq 144(%rsp), %rcx movq %rcx, 112(%rsp) movq %rsi, 120(%rsp) movq 160(%rsp), %rcx movq %rcx, 128(%rsp) movq 168(%rsp), %rcx movq %rcx, 136(%rsp) jmp .L387 .L386: movq 144(%rsp), %r8 movq %r8, 80(%rsp) movq 152(%rsp), %rsi movq %rsi, 88(%rsp) movq 160(%rsp), %rcx movq %rcx, 96(%rsp) movq 168(%rsp), %rdx movq %rdx, 104(%rsp) subl $8, %edi movl %edi, %eax shrl $3, %eax movl %eax, %r9d addq $1, %r9 salq $5, %r9 addq %r9, %rbx movq %r8, 48(%rsp) movq %rsi, 56(%rsp) movq %rcx, 64(%rsp) movq %rdx, 72(%rsp) negl %eax leal (%rdi,%rax,8), %edx .L384: testl %edx, %edx je .L388 .L389: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L389 jmp .L388 .L391: mulss 48(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L391 movss %xmm0, (%r12) jmp .L394 .L393: movss .LC0(%rip), %xmm0 jmp .L384 .L388: movl $0, %edx movl $0, %eax jmp .L391 .L394: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v2_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $352, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L398: movss %xmm0, 64(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L398 movq 64(%rsp), %rcx movq %rcx, 128(%rsp) movq 72(%rsp), %rcx movq %rcx, 136(%rsp) movq 80(%rsp), %rcx movq %rcx, 144(%rsp) movq 88(%rsp), %rcx movq %rcx, 152(%rsp) testb $31, %bl je .L411 testl %eax, %eax je .L412 movss .LC0(%rip), %xmm0 .L404: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L401 jmp .L399 .L411: movss .LC0(%rip), %xmm0 .L399: movl %edx, %edi cmpl $15, %edx ja .L402 jmp .L403 .L401: testl %edx, %edx jne .L404 jmp .L403 .L402: movq 128(%rsp), %rax movq %rax, 192(%rsp) movq 136(%rsp), %rax movq %rax, 200(%rsp) movq 144(%rsp), %rax movq %rax, 208(%rsp) movq 152(%rsp), %rax movq %rax, 216(%rsp) movl %edx, %eax movq %rbx, %rdx .L406: movq (%rdx), %rcx movq %rcx, 256(%rsp) movq 8(%rdx), %rcx movq %rcx, 264(%rsp) movq 16(%rdx), %rcx movq %rcx, 272(%rsp) movq 24(%rdx), %rcx movq %rcx, 280(%rsp) movq 32(%rdx), %rcx movq %rcx, 224(%rsp) movq 40(%rdx), %rcx movq %rcx, 232(%rsp) movq 48(%rdx), %rcx movq %rcx, 240(%rsp) movq 56(%rdx), %rcx movq %rcx, 248(%rsp) movaps 192(%rsp), %xmm1 mulps 256(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movq 16(%rsp), %rsi movaps %xmm1, 32(%rsp) movaps 208(%rsp), %xmm1 mulps 272(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movq 16(%rsp), %rcx movaps %xmm1, 48(%rsp) movq %rsi, 160(%rsp) movq 40(%rsp), %rsi movq %rsi, 168(%rsp) movq %rcx, 176(%rsp) movq 56(%rsp), %rcx movq %rcx, 184(%rsp) movaps 224(%rsp), %xmm1 mulps 128(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 144(%rsp), %xmm1 mulps 240(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rsi movaps 16(%rsp), %xmm4 movaps %xmm4, 32(%rsp) movq (%rsp), %rcx movaps %xmm1, 48(%rsp) movq %rsi, 128(%rsp) movq 40(%rsp), %rsi movq %rsi, 136(%rsp) movq %rcx, 144(%rsp) movq 56(%rsp), %rcx movq %rcx, 152(%rsp) addq $64, %rdx subl $16, %eax cmpl $15, %eax jbe .L405 movq 160(%rsp), %rcx movq %rcx, 192(%rsp) movq 168(%rsp), %rcx movq %rcx, 200(%rsp) movq 176(%rsp), %rcx movq %rcx, 208(%rsp) movq 184(%rsp), %rcx movq %rcx, 216(%rsp) jmp .L406 .L405: movq 160(%rsp), %r9 movq %r9, 96(%rsp) movq 168(%rsp), %r8 movq %r8, 104(%rsp) movq 176(%rsp), %rsi movq %rsi, 112(%rsp) movq 184(%rsp), %rcx movq %rcx, 120(%rsp) leal -16(%rdi), %edx movl %edx, %eax shrl $4, %eax movl %eax, %edi addq $1, %rdi salq $6, %rdi addq %rdi, %rbx movq %r9, 64(%rsp) movq %r8, 72(%rsp) movq %rsi, 80(%rsp) movq %rcx, 88(%rsp) sall $4, %eax subl %eax, %edx .L403: testl %edx, %edx je .L407 .L408: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L408 jmp .L407 .L410: mulss 64(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L410 movss %xmm0, (%r12) jmp .L413 .L412: movss .LC0(%rip), %xmm0 jmp .L403 .L407: movq 64(%rsp), %rax movq %rax, 288(%rsp) movq 72(%rsp), %rax movq %rax, 296(%rsp) movq 80(%rsp), %rax movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rax, 312(%rsp) movaps 288(%rsp), %xmm1 mulps 128(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movq 16(%rsp), %rdx movaps %xmm1, 32(%rsp) movaps 304(%rsp), %xmm1 mulps 144(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movq 16(%rsp), %rax movaps %xmm1, 48(%rsp) movq %rdx, 64(%rsp) movq 40(%rsp), %rdx movq %rdx, 72(%rsp) movq %rax, 80(%rsp) movq 56(%rsp), %rax movq %rax, 88(%rsp) movl $0, %edx movl $0, %eax jmp .L410 .L413: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v4_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $576, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L417: movss %xmm0, 96(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L417 movq 96(%rsp), %rcx movq %rcx, 320(%rsp) movq 104(%rsp), %rcx movq %rcx, 328(%rsp) movq 112(%rsp), %rcx movq %rcx, 336(%rsp) movq 120(%rsp), %rcx movq %rcx, 344(%rsp) testb $31, %bl je .L430 testl %eax, %eax je .L431 movss .LC0(%rip), %xmm0 .L423: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L420 jmp .L418 .L430: movss .LC0(%rip), %xmm0 .L418: movl %edx, %edi cmpl $31, %edx ja .L421 movq 320(%rsp), %rdi movq %rdi, 256(%rsp) movq 328(%rsp), %rsi movq %rsi, 264(%rsp) movq 336(%rsp), %rcx movq %rcx, 272(%rsp) movq 344(%rsp), %rax movq %rax, 280(%rsp) movq %rdi, 288(%rsp) movq %rsi, 296(%rsp) movq %rcx, 304(%rsp) movq %rax, 312(%rsp) jmp .L422 .L420: testl %edx, %edx jne .L423 jmp .L419 .L421: movq 320(%rsp), %r9 movq %r9, 384(%rsp) movq 328(%rsp), %r8 movq %r8, 392(%rsp) movq 336(%rsp), %rsi movq %rsi, 400(%rsp) movq 344(%rsp), %rcx movq %rcx, 408(%rsp) movl %edx, %eax movq %r9, 256(%rsp) movq %r8, 264(%rsp) movq %rsi, 272(%rsp) movq %rcx, 280(%rsp) movq %r9, 288(%rsp) movq %r8, 296(%rsp) movq %rsi, 304(%rsp) movq %rcx, 312(%rsp) movq %rbx, %rdx .L425: movq (%rdx), %rcx movq %rcx, 512(%rsp) movq 8(%rdx), %rcx movq %rcx, 520(%rsp) movq 16(%rdx), %rcx movq %rcx, 528(%rsp) movq 24(%rdx), %rcx movq %rcx, 536(%rsp) movq 32(%rdx), %rcx movq %rcx, 480(%rsp) movq 40(%rdx), %rcx movq %rcx, 488(%rsp) movq 48(%rdx), %rcx movq %rcx, 496(%rsp) movq 56(%rdx), %rcx movq %rcx, 504(%rsp) movq 64(%rdx), %rcx movq %rcx, 448(%rsp) movq 72(%rdx), %rcx movq %rcx, 456(%rsp) movq 80(%rdx), %rcx movq %rcx, 464(%rsp) movq 88(%rdx), %rcx movq %rcx, 472(%rsp) movq 96(%rdx), %rcx movq %rcx, 416(%rsp) movq 104(%rdx), %rcx movq %rcx, 424(%rsp) movq 112(%rdx), %rcx movq %rcx, 432(%rsp) movq 120(%rdx), %rcx movq %rcx, 440(%rsp) movaps 384(%rsp), %xmm1 mulps 512(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rsi movaps %xmm1, 64(%rsp) movaps 400(%rsp), %xmm1 mulps 528(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 352(%rsp) movq 72(%rsp), %rsi movq %rsi, 360(%rsp) movq %rcx, 368(%rsp) movq 88(%rsp), %rcx movq %rcx, 376(%rsp) movaps 480(%rsp), %xmm1 mulps 320(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 336(%rsp), %xmm1 mulps 496(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 320(%rsp) movq 72(%rsp), %rsi movq %rsi, 328(%rsp) movq %rcx, 336(%rsp) movq 88(%rsp), %rcx movq %rcx, 344(%rsp) movaps 288(%rsp), %xmm1 mulps 448(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 304(%rsp), %xmm1 mulps 464(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 288(%rsp) movq 72(%rsp), %rsi movq %rsi, 296(%rsp) movq %rcx, 304(%rsp) movq 88(%rsp), %rcx movq %rcx, 312(%rsp) movaps 256(%rsp), %xmm1 mulps 416(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 272(%rsp), %xmm1 mulps 432(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 256(%rsp) movq 72(%rsp), %rsi movq %rsi, 264(%rsp) movq %rcx, 272(%rsp) movq 88(%rsp), %rcx movq %rcx, 280(%rsp) subq $-128, %rdx subl $32, %eax cmpl $31, %eax jbe .L424 movq 352(%rsp), %rcx movq %rcx, 384(%rsp) movq 360(%rsp), %rcx movq %rcx, 392(%rsp) movq 368(%rsp), %rcx movq %rcx, 400(%rsp) movq 376(%rsp), %rcx movq %rcx, 408(%rsp) jmp .L425 .L424: movq 352(%rsp), %r9 movq %r9, 224(%rsp) movq 360(%rsp), %r8 movq %r8, 232(%rsp) movq 368(%rsp), %rsi movq %rsi, 240(%rsp) movq 376(%rsp), %rcx movq %rcx, 248(%rsp) leal -32(%rdi), %edx movl %edx, %eax shrl $5, %eax movl %eax, %edi addq $1, %rdi salq $7, %rdi addq %rdi, %rbx movq %r9, 96(%rsp) movq %r8, 104(%rsp) movq %rsi, 112(%rsp) movq %rcx, 120(%rsp) sall $5, %eax subl %eax, %edx .L422: testl %edx, %edx je .L426 .L427: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L427 jmp .L426 .L429: mulss 96(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L429 movss %xmm0, (%r12) jmp .L432 .L431: movss .LC0(%rip), %xmm0 .L419: movq 320(%rsp), %rdi movq %rdi, 256(%rsp) movq 328(%rsp), %rsi movq %rsi, 264(%rsp) movq 336(%rsp), %rcx movq %rcx, 272(%rsp) movq 344(%rsp), %rax movq %rax, 280(%rsp) movq %rdi, 288(%rsp) movq %rsi, 296(%rsp) movq %rcx, 304(%rsp) movq %rax, 312(%rsp) jmp .L422 .L426: movq 96(%rsp), %rax movq %rax, 544(%rsp) movq 104(%rsp), %rax movq %rax, 552(%rsp) movq 112(%rsp), %rax movq %rax, 560(%rsp) movq 120(%rsp), %rax movq %rax, 568(%rsp) movaps 544(%rsp), %xmm1 mulps 320(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 560(%rsp), %xmm1 mulps 336(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 128(%rsp) movq 72(%rsp), %rdx movq %rdx, 136(%rsp) movq %rax, 144(%rsp) movq 88(%rsp), %rax movq %rax, 152(%rsp) movaps 288(%rsp), %xmm1 mulps 256(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 304(%rsp), %xmm1 mulps 272(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 160(%rsp) movq 72(%rsp), %rdx movq %rdx, 168(%rsp) movq %rax, 176(%rsp) movq 88(%rsp), %rax movq %rax, 184(%rsp) movaps 48(%rsp), %xmm4 mulps %xmm6, %xmm4 movaps %xmm4, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 64(%rsp) movaps 32(%rsp), %xmm5 mulps %xmm1, %xmm5 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rax movaps %xmm5, 80(%rsp) movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rdx, 104(%rsp) movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rax, 120(%rsp) movl $0, %edx movl $0, %eax jmp .L429 .L432: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v8_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $960, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L436: movss %xmm0, 96(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L436 movq 96(%rsp), %rcx movq %rcx, 544(%rsp) movq 104(%rsp), %rcx movq %rcx, 552(%rsp) movq 112(%rsp), %rcx movq %rcx, 560(%rsp) movq 120(%rsp), %rcx movq %rcx, 568(%rsp) testb $31, %bl je .L449 testl %eax, %eax je .L450 movss .LC0(%rip), %xmm0 .L442: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L439 jmp .L437 .L449: movss .LC0(%rip), %xmm0 .L437: movl %edx, %edi cmpl $63, %edx ja .L440 movq 544(%rsp), %rdi movq %rdi, 352(%rsp) movq 552(%rsp), %rsi movq %rsi, 360(%rsp) movq 560(%rsp), %rcx movq %rcx, 368(%rsp) movq 568(%rsp), %rax movq %rax, 376(%rsp) movq %rdi, 384(%rsp) movq %rsi, 392(%rsp) movq %rcx, 400(%rsp) movq %rax, 408(%rsp) movq %rdi, 416(%rsp) movq %rsi, 424(%rsp) movq %rcx, 432(%rsp) movq %rax, 440(%rsp) movq %rdi, 448(%rsp) movq %rsi, 456(%rsp) movq %rcx, 464(%rsp) movq %rax, 472(%rsp) movq %rdi, 480(%rsp) movq %rsi, 488(%rsp) movq %rcx, 496(%rsp) movq %rax, 504(%rsp) movq %rdi, 512(%rsp) movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq %rax, 536(%rsp) jmp .L441 .L439: testl %edx, %edx jne .L442 jmp .L438 .L440: movq 544(%rsp), %r9 movq %r9, 608(%rsp) movq 552(%rsp), %r8 movq %r8, 616(%rsp) movq 560(%rsp), %rsi movq %rsi, 624(%rsp) movq 568(%rsp), %rcx movq %rcx, 632(%rsp) movl %edx, %eax movq %rbx, %rdx movq %r9, 352(%rsp) movq %r8, 360(%rsp) movq %rsi, 368(%rsp) movq %rcx, 376(%rsp) movq %r9, 384(%rsp) movq %r8, 392(%rsp) movq %rsi, 400(%rsp) movq %rcx, 408(%rsp) movq %r9, 416(%rsp) movq %r8, 424(%rsp) movq %rsi, 432(%rsp) movq %rcx, 440(%rsp) movq %r9, 448(%rsp) movq %r8, 456(%rsp) movq %rsi, 464(%rsp) movq %rcx, 472(%rsp) movq %r9, 480(%rsp) movq %r8, 488(%rsp) movq %rsi, 496(%rsp) movq %rcx, 504(%rsp) movq %r9, 512(%rsp) movq %r8, 520(%rsp) movq %rsi, 528(%rsp) movq %rcx, 536(%rsp) .L444: movq (%rdx), %rcx movq %rcx, 864(%rsp) movq 8(%rdx), %rcx movq %rcx, 872(%rsp) movq 16(%rdx), %rcx movq %rcx, 880(%rsp) movq 24(%rdx), %rcx movq %rcx, 888(%rsp) movq 32(%rdx), %rcx movq %rcx, 832(%rsp) movq 40(%rdx), %rcx movq %rcx, 840(%rsp) movq 48(%rdx), %rcx movq %rcx, 848(%rsp) movq 56(%rdx), %rcx movq %rcx, 856(%rsp) movq 64(%rdx), %rcx movq %rcx, 800(%rsp) movq 72(%rdx), %rcx movq %rcx, 808(%rsp) movq 80(%rdx), %rcx movq %rcx, 816(%rsp) movq 88(%rdx), %rcx movq %rcx, 824(%rsp) movq 96(%rdx), %rcx movq %rcx, 768(%rsp) movq 104(%rdx), %rcx movq %rcx, 776(%rsp) movq 112(%rdx), %rcx movq %rcx, 784(%rsp) movq 120(%rdx), %rcx movq %rcx, 792(%rsp) movq 128(%rdx), %rcx movq %rcx, 736(%rsp) movq 136(%rdx), %rcx movq %rcx, 744(%rsp) movq 144(%rdx), %rcx movq %rcx, 752(%rsp) movq 152(%rdx), %rcx movq %rcx, 760(%rsp) movq 160(%rdx), %rcx movq %rcx, 704(%rsp) movq 168(%rdx), %rcx movq %rcx, 712(%rsp) movq 176(%rdx), %rcx movq %rcx, 720(%rsp) movq 184(%rdx), %rcx movq %rcx, 728(%rsp) movq 192(%rdx), %rcx movq %rcx, 672(%rsp) movq 200(%rdx), %rcx movq %rcx, 680(%rsp) movq 208(%rdx), %rcx movq %rcx, 688(%rsp) movq 216(%rdx), %rcx movq %rcx, 696(%rsp) movq 224(%rdx), %rcx movq %rcx, 640(%rsp) movq 232(%rdx), %rcx movq %rcx, 648(%rsp) movq 240(%rdx), %rcx movq %rcx, 656(%rsp) movq 248(%rdx), %rcx movq %rcx, 664(%rsp) movaps 608(%rsp), %xmm1 mulps 864(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rsi movaps %xmm1, 64(%rsp) movaps 624(%rsp), %xmm1 mulps 880(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 576(%rsp) movq 72(%rsp), %rsi movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq 88(%rsp), %rcx movq %rcx, 600(%rsp) movaps 832(%rsp), %xmm1 mulps 544(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 848(%rsp), %xmm1 mulps 560(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 544(%rsp) movq 72(%rsp), %rsi movq %rsi, 552(%rsp) movq %rcx, 560(%rsp) movq 88(%rsp), %rcx movq %rcx, 568(%rsp) movaps 800(%rsp), %xmm1 mulps 512(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 816(%rsp), %xmm1 mulps 528(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rsi movaps 32(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 48(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 512(%rsp) movq 72(%rsp), %rsi movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq 88(%rsp), %rcx movq %rcx, 536(%rsp) movaps 480(%rsp), %xmm1 mulps 768(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 496(%rsp), %xmm1 mulps 784(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 480(%rsp) movq 72(%rsp), %rsi movq %rsi, 488(%rsp) movq %rcx, 496(%rsp) movq 88(%rsp), %rcx movq %rcx, 504(%rsp) movaps 448(%rsp), %xmm1 mulps 736(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 464(%rsp), %xmm1 mulps 752(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 448(%rsp) movq 72(%rsp), %rsi movq %rsi, 456(%rsp) movq %rcx, 464(%rsp) movq 88(%rsp), %rcx movq %rcx, 472(%rsp) movaps 416(%rsp), %xmm1 mulps 704(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 432(%rsp), %xmm1 mulps 720(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 416(%rsp) movq 72(%rsp), %rsi movq %rsi, 424(%rsp) movq %rcx, 432(%rsp) movq 88(%rsp), %rcx movq %rcx, 440(%rsp) movaps 384(%rsp), %xmm1 mulps 672(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 400(%rsp), %xmm1 mulps 688(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 384(%rsp) movq 72(%rsp), %rsi movq %rsi, 392(%rsp) movq %rcx, 400(%rsp) movq 88(%rsp), %rcx movq %rcx, 408(%rsp) movaps 352(%rsp), %xmm1 mulps 640(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 368(%rsp), %xmm1 mulps 656(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 352(%rsp) movq 72(%rsp), %rsi movq %rsi, 360(%rsp) movq %rcx, 368(%rsp) movq 88(%rsp), %rcx movq %rcx, 376(%rsp) addq $256, %rdx subl $64, %eax cmpl $63, %eax jbe .L443 movq 576(%rsp), %rcx movq %rcx, 608(%rsp) movq 584(%rsp), %rcx movq %rcx, 616(%rsp) movq 592(%rsp), %rcx movq %rcx, 624(%rsp) movq 600(%rsp), %rcx movq %rcx, 632(%rsp) jmp .L444 .L443: movq 576(%rsp), %r9 movq %r9, 320(%rsp) movq 584(%rsp), %r8 movq %r8, 328(%rsp) movq 592(%rsp), %rsi movq %rsi, 336(%rsp) movq 600(%rsp), %rcx movq %rcx, 344(%rsp) leal -64(%rdi), %edx movl %edx, %eax shrl $6, %eax movl %eax, %edi addq $1, %rdi salq $8, %rdi addq %rdi, %rbx movq %r9, 96(%rsp) movq %r8, 104(%rsp) movq %rsi, 112(%rsp) movq %rcx, 120(%rsp) sall $6, %eax subl %eax, %edx .L441: testl %edx, %edx je .L445 .L446: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L446 jmp .L445 .L448: mulss 96(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L448 movss %xmm0, (%r12) jmp .L451 .L450: movss .LC0(%rip), %xmm0 .L438: movq 544(%rsp), %rdi movq %rdi, 352(%rsp) movq 552(%rsp), %rsi movq %rsi, 360(%rsp) movq 560(%rsp), %rcx movq %rcx, 368(%rsp) movq 568(%rsp), %rax movq %rax, 376(%rsp) movq %rdi, 384(%rsp) movq %rsi, 392(%rsp) movq %rcx, 400(%rsp) movq %rax, 408(%rsp) movq %rdi, 416(%rsp) movq %rsi, 424(%rsp) movq %rcx, 432(%rsp) movq %rax, 440(%rsp) movq %rdi, 448(%rsp) movq %rsi, 456(%rsp) movq %rcx, 464(%rsp) movq %rax, 472(%rsp) movq %rdi, 480(%rsp) movq %rsi, 488(%rsp) movq %rcx, 496(%rsp) movq %rax, 504(%rsp) movq %rdi, 512(%rsp) movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq %rax, 536(%rsp) jmp .L441 .L445: movq 96(%rsp), %rax movq %rax, 128(%rsp) movq 104(%rsp), %rax movq %rax, 136(%rsp) movq 112(%rsp), %rax movq %rax, 144(%rsp) movq 120(%rsp), %rax movq %rax, 152(%rsp) movaps 128(%rsp), %xmm1 mulps 544(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 144(%rsp), %xmm1 mulps 560(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 160(%rsp) movq 72(%rsp), %rdx movq %rdx, 168(%rsp) movq %rax, 176(%rsp) movq 88(%rsp), %rax movq %rax, 184(%rsp) movaps 512(%rsp), %xmm1 mulps 480(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 528(%rsp), %xmm1 mulps 496(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 192(%rsp) movq 72(%rsp), %rdx movq %rdx, 200(%rsp) movq %rax, 208(%rsp) movq 88(%rsp), %rax movq %rax, 216(%rsp) movaps 48(%rsp), %xmm6 mulps 16(%rsp), %xmm6 movaps %xmm6, 48(%rsp) movaps 32(%rsp), %xmm7 mulps %xmm1, %xmm7 movaps %xmm7, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm7, 80(%rsp) movq %rdx, 224(%rsp) movq 72(%rsp), %rdx movq %rdx, 232(%rsp) movq %rax, 240(%rsp) movq 88(%rsp), %rax movq %rax, 248(%rsp) movaps 448(%rsp), %xmm1 mulps 416(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 464(%rsp), %xmm1 mulps 432(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 256(%rsp) movq 72(%rsp), %rdx movq %rdx, 264(%rsp) movq %rax, 272(%rsp) movq 88(%rsp), %rax movq %rax, 280(%rsp) movaps 48(%rsp), %xmm6 mulps 16(%rsp), %xmm6 movaps %xmm6, 48(%rsp) movaps 32(%rsp), %xmm7 mulps %xmm1, %xmm7 movaps %xmm7, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm7, 80(%rsp) movq %rdx, 288(%rsp) movq 72(%rsp), %rdx movq %rdx, 296(%rsp) movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rax, 312(%rsp) movaps 384(%rsp), %xmm1 mulps 352(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 400(%rsp), %xmm1 mulps 368(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 896(%rsp) movq 72(%rsp), %rdx movq %rdx, 904(%rsp) movq %rax, 912(%rsp) movq 88(%rsp), %rax movq %rax, 920(%rsp) movaps 48(%rsp), %xmm6 mulps 16(%rsp), %xmm6 movaps %xmm6, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 64(%rsp) movaps 32(%rsp), %xmm7 mulps %xmm1, %xmm7 movaps %xmm7, 48(%rsp) movq 48(%rsp), %rax movaps %xmm7, 80(%rsp) movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rdx, 104(%rsp) movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rax, 120(%rsp) movl $0, %edx movl $0, %eax jmp .L448 .L451: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v10_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $1152, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L455: movss %xmm0, 96(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L455 movq 96(%rsp), %rcx movq %rcx, 736(%rsp) movq 104(%rsp), %rcx movq %rcx, 744(%rsp) movq 112(%rsp), %rcx movq %rcx, 752(%rsp) movq 120(%rsp), %rcx movq %rcx, 760(%rsp) testb $31, %bl je .L468 testl %eax, %eax je .L469 movss .LC0(%rip), %xmm0 .L461: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L458 jmp .L456 .L468: movss .LC0(%rip), %xmm0 .L456: movl %edx, %esi cmpl $79, %edx ja .L459 movq 736(%rsp), %rdi movq %rdi, 480(%rsp) movq 744(%rsp), %rsi movq %rsi, 488(%rsp) movq 752(%rsp), %rcx movq %rcx, 496(%rsp) movq 760(%rsp), %rax movq %rax, 504(%rsp) movq %rdi, 512(%rsp) movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq %rax, 536(%rsp) movq %rdi, 544(%rsp) movq %rsi, 552(%rsp) movq %rcx, 560(%rsp) movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) jmp .L460 .L458: testl %edx, %edx jne .L461 jmp .L457 .L459: movq 736(%rsp), %rdi movq %rdi, 768(%rsp) movq 744(%rsp), %rcx movq %rcx, 776(%rsp) movq 752(%rsp), %rdx movq %rdx, 784(%rsp) movq 760(%rsp), %rax movq %rax, 792(%rsp) movq %rdi, 480(%rsp) movq %rcx, 488(%rsp) movq %rdx, 496(%rsp) movq %rax, 504(%rsp) movq %rdi, 512(%rsp) movq %rcx, 520(%rsp) movq %rdx, 528(%rsp) movq %rax, 536(%rsp) movq %rdi, 544(%rsp) movq %rcx, 552(%rsp) movq %rdx, 560(%rsp) movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rcx, 584(%rsp) movq %rdx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rcx, 616(%rsp) movq %rdx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rcx, 648(%rsp) movq %rdx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rcx, 680(%rsp) movq %rdx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rcx, 712(%rsp) movq %rdx, 720(%rsp) movq %rax, 728(%rsp) .L463: movq (%rbx), %rax movq %rax, 1120(%rsp) movq 8(%rbx), %rax movq %rax, 1128(%rsp) movq 16(%rbx), %rax movq %rax, 1136(%rsp) movq 24(%rbx), %rax movq %rax, 1144(%rsp) movq 32(%rbx), %rax movq %rax, 1088(%rsp) movq 40(%rbx), %rax movq %rax, 1096(%rsp) movq 48(%rbx), %rax movq %rax, 1104(%rsp) movq 56(%rbx), %rax movq %rax, 1112(%rsp) movq 64(%rbx), %rax movq %rax, 1056(%rsp) movq 72(%rbx), %rax movq %rax, 1064(%rsp) movq 80(%rbx), %rax movq %rax, 1072(%rsp) movq 88(%rbx), %rax movq %rax, 1080(%rsp) movq 96(%rbx), %rax movq %rax, 1024(%rsp) movq 104(%rbx), %rax movq %rax, 1032(%rsp) movq 112(%rbx), %rax movq %rax, 1040(%rsp) movq 120(%rbx), %rax movq %rax, 1048(%rsp) movq 128(%rbx), %rax movq %rax, 992(%rsp) movq 136(%rbx), %rax movq %rax, 1000(%rsp) movq 144(%rbx), %rax movq %rax, 1008(%rsp) movq 152(%rbx), %rax movq %rax, 1016(%rsp) movq 160(%rbx), %rax movq %rax, 960(%rsp) movq 168(%rbx), %rax movq %rax, 968(%rsp) movq 176(%rbx), %rax movq %rax, 976(%rsp) movq 184(%rbx), %rax movq %rax, 984(%rsp) movq 192(%rbx), %rax movq %rax, 928(%rsp) movq 200(%rbx), %rax movq %rax, 936(%rsp) movq 208(%rbx), %rax movq %rax, 944(%rsp) movq 216(%rbx), %rax movq %rax, 952(%rsp) movq 224(%rbx), %rax movq %rax, 896(%rsp) movq 232(%rbx), %rax movq %rax, 904(%rsp) movq 240(%rbx), %rax movq %rax, 912(%rsp) movq 248(%rbx), %rax movq %rax, 920(%rsp) movq 256(%rbx), %rax movq %rax, 864(%rsp) movq 264(%rbx), %rax movq %rax, 872(%rsp) movq 272(%rbx), %rax movq %rax, 880(%rsp) movq 280(%rbx), %rax movq %rax, 888(%rsp) movq 288(%rbx), %rax movq %rax, 832(%rsp) movq 296(%rbx), %rax movq %rax, 840(%rsp) movq 304(%rbx), %rax movq %rax, 848(%rsp) movq 312(%rbx), %rax movq %rax, 856(%rsp) movaps 768(%rsp), %xmm1 mulps 1120(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm1, 64(%rsp) movaps 1136(%rsp), %xmm1 mulps 784(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 800(%rsp) movq 72(%rsp), %rdx movq %rdx, 808(%rsp) movq %rax, 816(%rsp) movq 88(%rsp), %rax movq %rax, 824(%rsp) movaps 1088(%rsp), %xmm1 mulps 736(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1104(%rsp), %xmm1 mulps 752(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 736(%rsp) movq 72(%rsp), %rdx movq %rdx, 744(%rsp) movq %rax, 752(%rsp) movq 88(%rsp), %rax movq %rax, 760(%rsp) movaps 704(%rsp), %xmm1 mulps 1056(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1072(%rsp), %xmm1 mulps 720(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 704(%rsp) movq 72(%rsp), %rdx movq %rdx, 712(%rsp) movq %rax, 720(%rsp) movq 88(%rsp), %rax movq %rax, 728(%rsp) movaps 1024(%rsp), %xmm1 mulps 672(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1040(%rsp), %xmm1 mulps 688(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 672(%rsp) movq 72(%rsp), %rdx movq %rdx, 680(%rsp) movq %rax, 688(%rsp) movq 88(%rsp), %rax movq %rax, 696(%rsp) movaps 992(%rsp), %xmm1 mulps 640(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 1008(%rsp), %xmm1 mulps 656(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 640(%rsp) movq 72(%rsp), %rdx movq %rdx, 648(%rsp) movq %rax, 656(%rsp) movq 88(%rsp), %rax movq %rax, 664(%rsp) movaps 960(%rsp), %xmm1 mulps 608(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 976(%rsp), %xmm1 mulps 624(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 608(%rsp) movq 72(%rsp), %rdx movq %rdx, 616(%rsp) movq %rax, 624(%rsp) movq 88(%rsp), %rax movq %rax, 632(%rsp) movaps 576(%rsp), %xmm1 mulps 928(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 592(%rsp), %xmm1 mulps 944(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 576(%rsp) movq 72(%rsp), %rdx movq %rdx, 584(%rsp) movq %rax, 592(%rsp) movq 88(%rsp), %rax movq %rax, 600(%rsp) movaps 544(%rsp), %xmm1 mulps 896(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 560(%rsp), %xmm1 mulps 912(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 544(%rsp) movq 72(%rsp), %rdx movq %rdx, 552(%rsp) movq %rax, 560(%rsp) movq 88(%rsp), %rax movq %rax, 568(%rsp) movaps 512(%rsp), %xmm1 mulps 864(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 528(%rsp), %xmm1 mulps 880(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rcx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rcx, 512(%rsp) movq 72(%rsp), %rcx movq %rcx, 520(%rsp) movq %rax, 528(%rsp) movq 88(%rsp), %rax movq %rax, 536(%rsp) movaps 480(%rsp), %xmm1 mulps 832(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 496(%rsp), %xmm1 mulps 848(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rcx movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rcx, 480(%rsp) movq 72(%rsp), %rcx movq %rcx, 488(%rsp) movq %rax, 496(%rsp) movq 88(%rsp), %rax movq %rax, 504(%rsp) addq $320, %rbx leal -80(%rsi), %edx cmpl $79, %edx jbe .L462 movl %edx, %esi movq 800(%rsp), %rax movq %rax, 768(%rsp) movq 808(%rsp), %rax movq %rax, 776(%rsp) movq 816(%rsp), %rax movq %rax, 784(%rsp) movq 824(%rsp), %rax movq %rax, 792(%rsp) jmp .L463 .L462: movq 800(%rsp), %rdi movq %rdi, 128(%rsp) movq 808(%rsp), %rsi movq %rsi, 136(%rsp) movq 816(%rsp), %rcx movq %rcx, 144(%rsp) movq 824(%rsp), %rax movq %rax, 152(%rsp) movq %rdi, 96(%rsp) movq %rsi, 104(%rsp) movq %rcx, 112(%rsp) movq %rax, 120(%rsp) .L460: testl %edx, %edx je .L464 .L465: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L465 .L464: movq 96(%rsp), %rax movq %rax, 448(%rsp) movq 104(%rsp), %rax movq %rax, 456(%rsp) movq 112(%rsp), %rax movq %rax, 464(%rsp) movq 120(%rsp), %rax movq %rax, 472(%rsp) movaps 448(%rsp), %xmm1 mulps 736(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 464(%rsp), %xmm1 mulps 752(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, %xmm5 movaps %xmm1, 80(%rsp) movq %rdx, 416(%rsp) movq 72(%rsp), %rdx movq %rdx, 424(%rsp) movq %rax, 432(%rsp) movq 88(%rsp), %rax movq %rax, 440(%rsp) movaps 704(%rsp), %xmm1 mulps 672(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 720(%rsp), %xmm1 mulps 688(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 384(%rsp) movq 72(%rsp), %rdx movq %rdx, 392(%rsp) movq %rax, 400(%rsp) movq 88(%rsp), %rax movq %rax, 408(%rsp) mulps %xmm6, %xmm4 movaps %xmm4, 48(%rsp) mulps %xmm1, %xmm5 movaps %xmm5, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm5, 80(%rsp) movq %rdx, 352(%rsp) movq 72(%rsp), %rdx movq %rdx, 360(%rsp) movq %rax, 368(%rsp) movq 88(%rsp), %rax movq %rax, 376(%rsp) movaps 640(%rsp), %xmm1 mulps 608(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 656(%rsp), %xmm1 mulps 624(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 320(%rsp) movq 72(%rsp), %rdx movq %rdx, 328(%rsp) movq %rax, 336(%rsp) movq 88(%rsp), %rax movq %rax, 344(%rsp) movaps %xmm4, %xmm6 mulps 16(%rsp), %xmm6 movaps %xmm6, 48(%rsp) movaps %xmm5, %xmm7 mulps %xmm1, %xmm7 movaps %xmm7, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm6, %xmm4 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm7, %xmm5 movaps %xmm7, 80(%rsp) movq %rdx, 288(%rsp) movq 72(%rsp), %rdx movq %rdx, 296(%rsp) movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rax, 312(%rsp) movaps 576(%rsp), %xmm1 mulps 544(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 592(%rsp), %xmm1 mulps 560(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 256(%rsp) movq 72(%rsp), %rdx movq %rdx, 264(%rsp) movq %rax, 272(%rsp) movq 88(%rsp), %rax movq %rax, 280(%rsp) mulps %xmm6, %xmm4 movaps %xmm4, 48(%rsp) mulps %xmm1, %xmm5 movaps %xmm5, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm5, 80(%rsp) movq %rdx, 224(%rsp) movq 72(%rsp), %rdx movq %rdx, 232(%rsp) movq %rax, 240(%rsp) movq 88(%rsp), %rax movq %rax, 248(%rsp) movaps 512(%rsp), %xmm1 mulps 480(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 528(%rsp), %xmm1 mulps 496(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 192(%rsp) movq 72(%rsp), %rdx movq %rdx, 200(%rsp) movq %rax, 208(%rsp) movq 88(%rsp), %rax movq %rax, 216(%rsp) movaps %xmm4, %xmm6 mulps 16(%rsp), %xmm6 movaps %xmm6, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm6, 64(%rsp) movaps %xmm5, %xmm7 mulps %xmm1, %xmm7 movaps %xmm7, 48(%rsp) movq 48(%rsp), %rax movaps %xmm7, 80(%rsp) movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rdx, 104(%rsp) movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rax, 120(%rsp) movl $0, %edx movl $0, %eax .L467: mulss 96(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L467 movss %xmm0, (%r12) jmp .L470 .L469: movss .LC0(%rip), %xmm0 .L457: movq 736(%rsp), %rdi movq %rdi, 480(%rsp) movq 744(%rsp), %rsi movq %rsi, 488(%rsp) movq 752(%rsp), %rcx movq %rcx, 496(%rsp) movq 760(%rsp), %rax movq %rax, 504(%rsp) movq %rdi, 512(%rsp) movq %rsi, 520(%rsp) movq %rcx, 528(%rsp) movq %rax, 536(%rsp) movq %rdi, 544(%rsp) movq %rsi, 552(%rsp) movq %rcx, 560(%rsp) movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) jmp .L460 .L470: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v12_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $1344, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L474: movss %xmm0, 96(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L474 movq 96(%rsp), %rcx movq %rcx, 864(%rsp) movq 104(%rsp), %rcx movq %rcx, 872(%rsp) movq 112(%rsp), %rcx movq %rcx, 880(%rsp) movq 120(%rsp), %rcx movq %rcx, 888(%rsp) testb $31, %bl je .L487 testl %eax, %eax je .L488 movss .LC0(%rip), %xmm0 .L480: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L477 jmp .L475 .L487: movss .LC0(%rip), %xmm0 .L475: movl %edx, %esi cmpl $95, %edx ja .L478 movq 864(%rsp), %rdi movq %rdi, 544(%rsp) movq 872(%rsp), %rsi movq %rsi, 552(%rsp) movq 880(%rsp), %rcx movq %rcx, 560(%rsp) movq 888(%rsp), %rax movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) movq %rdi, 736(%rsp) movq %rsi, 744(%rsp) movq %rcx, 752(%rsp) movq %rax, 760(%rsp) movq %rdi, 768(%rsp) movq %rsi, 776(%rsp) movq %rcx, 784(%rsp) movq %rax, 792(%rsp) movq %rdi, 800(%rsp) movq %rsi, 808(%rsp) movq %rcx, 816(%rsp) movq %rax, 824(%rsp) movq %rdi, 832(%rsp) movq %rsi, 840(%rsp) movq %rcx, 848(%rsp) movq %rax, 856(%rsp) jmp .L479 .L477: testl %edx, %edx jne .L480 jmp .L476 .L478: movq 864(%rsp), %rdi movq %rdi, 896(%rsp) movq 872(%rsp), %rcx movq %rcx, 904(%rsp) movq 880(%rsp), %rdx movq %rdx, 912(%rsp) movq 888(%rsp), %rax movq %rax, 920(%rsp) movq %rdi, 544(%rsp) movq %rcx, 552(%rsp) movq %rdx, 560(%rsp) movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rcx, 584(%rsp) movq %rdx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rcx, 616(%rsp) movq %rdx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rcx, 648(%rsp) movq %rdx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rcx, 680(%rsp) movq %rdx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rcx, 712(%rsp) movq %rdx, 720(%rsp) movq %rax, 728(%rsp) movq %rdi, 736(%rsp) movq %rcx, 744(%rsp) movq %rdx, 752(%rsp) movq %rax, 760(%rsp) movq %rdi, 768(%rsp) movq %rcx, 776(%rsp) movq %rdx, 784(%rsp) movq %rax, 792(%rsp) movq %rdi, 800(%rsp) movq %rcx, 808(%rsp) movq %rdx, 816(%rsp) movq %rax, 824(%rsp) movq %rdi, 832(%rsp) movq %rcx, 840(%rsp) movq %rdx, 848(%rsp) movq %rax, 856(%rsp) .L482: movq (%rbx), %rax movq %rax, 1312(%rsp) movq 8(%rbx), %rax movq %rax, 1320(%rsp) movq 16(%rbx), %rax movq %rax, 1328(%rsp) movq 24(%rbx), %rax movq %rax, 1336(%rsp) movq 32(%rbx), %rax movq %rax, 1280(%rsp) movq 40(%rbx), %rax movq %rax, 1288(%rsp) movq 48(%rbx), %rax movq %rax, 1296(%rsp) movq 56(%rbx), %rax movq %rax, 1304(%rsp) movq 64(%rbx), %rax movq %rax, 1248(%rsp) movq 72(%rbx), %rax movq %rax, 1256(%rsp) movq 80(%rbx), %rax movq %rax, 1264(%rsp) movq 88(%rbx), %rax movq %rax, 1272(%rsp) movq 96(%rbx), %rax movq %rax, 1216(%rsp) movq 104(%rbx), %rax movq %rax, 1224(%rsp) movq 112(%rbx), %rax movq %rax, 1232(%rsp) movq 120(%rbx), %rax movq %rax, 1240(%rsp) movq 128(%rbx), %rax movq %rax, 1184(%rsp) movq 136(%rbx), %rax movq %rax, 1192(%rsp) movq 144(%rbx), %rax movq %rax, 1200(%rsp) movq 152(%rbx), %rax movq %rax, 1208(%rsp) movq 160(%rbx), %rax movq %rax, 1152(%rsp) movq 168(%rbx), %rax movq %rax, 1160(%rsp) movq 176(%rbx), %rax movq %rax, 1168(%rsp) movq 184(%rbx), %rax movq %rax, 1176(%rsp) movq 192(%rbx), %rax movq %rax, 1120(%rsp) movq 200(%rbx), %rax movq %rax, 1128(%rsp) movq 208(%rbx), %rax movq %rax, 1136(%rsp) movq 216(%rbx), %rax movq %rax, 1144(%rsp) movq 224(%rbx), %rax movq %rax, 1088(%rsp) movq 232(%rbx), %rax movq %rax, 1096(%rsp) movq 240(%rbx), %rax movq %rax, 1104(%rsp) movq 248(%rbx), %rax movq %rax, 1112(%rsp) movq 256(%rbx), %rax movq %rax, 1056(%rsp) movq 264(%rbx), %rax movq %rax, 1064(%rsp) movq 272(%rbx), %rax movq %rax, 1072(%rsp) movq 280(%rbx), %rax movq %rax, 1080(%rsp) movq 288(%rbx), %rax movq %rax, 1024(%rsp) movq 296(%rbx), %rax movq %rax, 1032(%rsp) movq 304(%rbx), %rax movq %rax, 1040(%rsp) movq 312(%rbx), %rax movq %rax, 1048(%rsp) movq 320(%rbx), %rax movq %rax, 992(%rsp) movq 328(%rbx), %rax movq %rax, 1000(%rsp) movq 336(%rbx), %rax movq %rax, 1008(%rsp) movq 344(%rbx), %rax movq %rax, 1016(%rsp) movq 352(%rbx), %rax movq %rax, 960(%rsp) movq 360(%rbx), %rax movq %rax, 968(%rsp) movq 368(%rbx), %rax movq %rax, 976(%rsp) movq 376(%rbx), %rax movq %rax, 984(%rsp) movaps 896(%rsp), %xmm1 mulps 1312(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm1, 64(%rsp) movaps 912(%rsp), %xmm1 mulps 1328(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 928(%rsp) movq 72(%rsp), %rdx movq %rdx, 936(%rsp) movq %rax, 944(%rsp) movq 88(%rsp), %rax movq %rax, 952(%rsp) movaps 1280(%rsp), %xmm1 mulps 864(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1296(%rsp), %xmm1 mulps 880(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 864(%rsp) movq 72(%rsp), %rdx movq %rdx, 872(%rsp) movq %rax, 880(%rsp) movq 88(%rsp), %rax movq %rax, 888(%rsp) movaps 1248(%rsp), %xmm1 mulps 832(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 1264(%rsp), %xmm1 mulps 848(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 832(%rsp) movq 72(%rsp), %rdx movq %rdx, 840(%rsp) movq %rax, 848(%rsp) movq 88(%rsp), %rax movq %rax, 856(%rsp) movaps 1216(%rsp), %xmm1 mulps 800(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1232(%rsp), %xmm1 mulps 816(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 800(%rsp) movq 72(%rsp), %rdx movq %rdx, 808(%rsp) movq %rax, 816(%rsp) movq 88(%rsp), %rax movq %rax, 824(%rsp) movaps 1184(%rsp), %xmm1 mulps 768(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1200(%rsp), %xmm1 mulps 784(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 768(%rsp) movq 72(%rsp), %rdx movq %rdx, 776(%rsp) movq %rax, 784(%rsp) movq 88(%rsp), %rax movq %rax, 792(%rsp) movaps 1152(%rsp), %xmm1 mulps 736(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 1168(%rsp), %xmm1 mulps 752(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 736(%rsp) movq 72(%rsp), %rdx movq %rdx, 744(%rsp) movq %rax, 752(%rsp) movq 88(%rsp), %rax movq %rax, 760(%rsp) movaps 1120(%rsp), %xmm1 mulps 704(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 1136(%rsp), %xmm1 mulps 720(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rdx movaps 32(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 48(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 704(%rsp) movq 72(%rsp), %rdx movq %rdx, 712(%rsp) movq %rax, 720(%rsp) movq 88(%rsp), %rax movq %rax, 728(%rsp) movaps 672(%rsp), %xmm1 mulps 1088(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 688(%rsp), %xmm1 mulps 1104(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 672(%rsp) movq 72(%rsp), %rdx movq %rdx, 680(%rsp) movq %rax, 688(%rsp) movq 88(%rsp), %rax movq %rax, 696(%rsp) movaps 640(%rsp), %xmm1 mulps 1056(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 656(%rsp), %xmm1 mulps 1072(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 640(%rsp) movq 72(%rsp), %rdx movq %rdx, 648(%rsp) movq %rax, 656(%rsp) movq 88(%rsp), %rax movq %rax, 664(%rsp) movaps 608(%rsp), %xmm1 mulps 1024(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 624(%rsp), %xmm1 mulps 1040(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 608(%rsp) movq 72(%rsp), %rdx movq %rdx, 616(%rsp) movq %rax, 624(%rsp) movq 88(%rsp), %rax movq %rax, 632(%rsp) movaps 576(%rsp), %xmm1 mulps 992(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 592(%rsp), %xmm1 mulps 1008(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rcx movaps 48(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rcx, 576(%rsp) movq 72(%rsp), %rcx movq %rcx, 584(%rsp) movq %rax, 592(%rsp) movq 88(%rsp), %rax movq %rax, 600(%rsp) movaps 544(%rsp), %xmm1 mulps 960(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 560(%rsp), %xmm1 mulps 976(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rcx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rcx, 544(%rsp) movq 72(%rsp), %rcx movq %rcx, 552(%rsp) movq %rax, 560(%rsp) movq 88(%rsp), %rax movq %rax, 568(%rsp) addq $384, %rbx leal -96(%rsi), %edx cmpl $95, %edx jbe .L481 movl %edx, %esi movq 928(%rsp), %rax movq %rax, 896(%rsp) movq 936(%rsp), %rax movq %rax, 904(%rsp) movq 944(%rsp), %rax movq %rax, 912(%rsp) movq 952(%rsp), %rax movq %rax, 920(%rsp) jmp .L482 .L481: movq 928(%rsp), %rdi movq %rdi, 128(%rsp) movq 936(%rsp), %rsi movq %rsi, 136(%rsp) movq 944(%rsp), %rcx movq %rcx, 144(%rsp) movq 952(%rsp), %rax movq %rax, 152(%rsp) movq %rdi, 96(%rsp) movq %rsi, 104(%rsp) movq %rcx, 112(%rsp) movq %rax, 120(%rsp) .L479: testl %edx, %edx je .L483 .L484: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L484 .L483: movq 96(%rsp), %rax movq %rax, 512(%rsp) movq 104(%rsp), %rax movq %rax, 520(%rsp) movq 112(%rsp), %rax movq %rax, 528(%rsp) movq 120(%rsp), %rax movq %rax, 536(%rsp) movaps 512(%rsp), %xmm1 mulps 864(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 528(%rsp), %xmm1 mulps 880(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rdx movaps 48(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 480(%rsp) movq 72(%rsp), %rdx movq %rdx, 488(%rsp) movq %rax, 496(%rsp) movq 88(%rsp), %rax movq %rax, 504(%rsp) movaps 832(%rsp), %xmm1 mulps 800(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 848(%rsp), %xmm1 mulps 816(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 448(%rsp) movq 72(%rsp), %rdx movq %rdx, 456(%rsp) movq %rax, 464(%rsp) movq 88(%rsp), %rax movq %rax, 472(%rsp) movaps 48(%rsp), %xmm2 mulps %xmm6, %xmm2 movaps %xmm2, 48(%rsp) movaps 32(%rsp), %xmm3 mulps %xmm1, %xmm3 movaps %xmm3, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm2, 64(%rsp) movq 32(%rsp), %rax movaps %xmm3, 80(%rsp) movq %rdx, 416(%rsp) movq 72(%rsp), %rdx movq %rdx, 424(%rsp) movq %rax, 432(%rsp) movq 88(%rsp), %rax movq %rax, 440(%rsp) movaps 768(%rsp), %xmm1 mulps 736(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 784(%rsp), %xmm1 mulps 752(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 384(%rsp) movq 72(%rsp), %rdx movq %rdx, 392(%rsp) movq %rax, 400(%rsp) movq 88(%rsp), %rax movq %rax, 408(%rsp) movaps %xmm2, %xmm4 mulps 16(%rsp), %xmm4 movaps %xmm4, 48(%rsp) movaps %xmm3, %xmm5 mulps %xmm1, %xmm5 movaps %xmm5, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm4, %xmm6 movaps %xmm4, 64(%rsp) movq 32(%rsp), %rax movaps %xmm5, %xmm7 movaps %xmm5, 80(%rsp) movq %rdx, 352(%rsp) movq 72(%rsp), %rdx movq %rdx, 360(%rsp) movq %rax, 368(%rsp) movq 88(%rsp), %rax movq %rax, 376(%rsp) movaps 704(%rsp), %xmm1 mulps 672(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 720(%rsp), %xmm1 mulps 688(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 320(%rsp) movq 72(%rsp), %rdx movq %rdx, 328(%rsp) movq %rax, 336(%rsp) movq 88(%rsp), %rax movq %rax, 344(%rsp) mulps %xmm2, %xmm6 movaps %xmm6, 48(%rsp) mulps %xmm1, %xmm7 movaps %xmm7, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm6, %xmm4 movaps %xmm6, 64(%rsp) movq 32(%rsp), %rax movaps %xmm7, %xmm5 movaps %xmm7, 80(%rsp) movq %rdx, 288(%rsp) movq 72(%rsp), %rdx movq %rdx, 296(%rsp) movq %rax, 304(%rsp) movq 88(%rsp), %rax movq %rax, 312(%rsp) movaps 640(%rsp), %xmm1 mulps 608(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 656(%rsp), %xmm1 mulps 624(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 256(%rsp) movq 72(%rsp), %rdx movq %rdx, 264(%rsp) movq %rax, 272(%rsp) movq 88(%rsp), %rax movq %rax, 280(%rsp) movaps %xmm4, %xmm2 mulps %xmm6, %xmm2 movaps %xmm2, 48(%rsp) movaps %xmm5, %xmm3 mulps %xmm1, %xmm3 movaps %xmm3, 32(%rsp) movq 48(%rsp), %rdx movaps %xmm2, 64(%rsp) movq 32(%rsp), %rax movaps %xmm3, 80(%rsp) movq %rdx, 224(%rsp) movq 72(%rsp), %rdx movq %rdx, 232(%rsp) movq %rax, 240(%rsp) movq 88(%rsp), %rax movq %rax, 248(%rsp) movaps 576(%rsp), %xmm1 mulps 544(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 592(%rsp), %xmm1 mulps 560(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rdx movaps 16(%rsp), %xmm4 movaps %xmm4, 64(%rsp) movq (%rsp), %rax movaps %xmm1, 80(%rsp) movq %rdx, 192(%rsp) movq 72(%rsp), %rdx movq %rdx, 200(%rsp) movq %rax, 208(%rsp) movq 88(%rsp), %rax movq %rax, 216(%rsp) movaps %xmm2, %xmm4 mulps 16(%rsp), %xmm4 movaps %xmm4, 48(%rsp) movq 48(%rsp), %rdx movaps %xmm4, 64(%rsp) movaps %xmm3, %xmm5 mulps %xmm1, %xmm5 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rax movaps %xmm5, 80(%rsp) movq %rdx, 96(%rsp) movq 72(%rsp), %rdx movq %rdx, 104(%rsp) movq %rax, 112(%rsp) movq 88(%rsp), %rax movq %rax, 120(%rsp) movl $0, %edx movl $0, %eax .L486: mulss 96(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L486 movss %xmm0, (%r12) jmp .L489 .L488: movss .LC0(%rip), %xmm0 .L476: movq 864(%rsp), %rdi movq %rdi, 544(%rsp) movq 872(%rsp), %rsi movq %rsi, 552(%rsp) movq 880(%rsp), %rcx movq %rcx, 560(%rsp) movq 888(%rsp), %rax movq %rax, 568(%rsp) movq %rdi, 576(%rsp) movq %rsi, 584(%rsp) movq %rcx, 592(%rsp) movq %rax, 600(%rsp) movq %rdi, 608(%rsp) movq %rsi, 616(%rsp) movq %rcx, 624(%rsp) movq %rax, 632(%rsp) movq %rdi, 640(%rsp) movq %rsi, 648(%rsp) movq %rcx, 656(%rsp) movq %rax, 664(%rsp) movq %rdi, 672(%rsp) movq %rsi, 680(%rsp) movq %rcx, 688(%rsp) movq %rax, 696(%rsp) movq %rdi, 704(%rsp) movq %rsi, 712(%rsp) movq %rcx, 720(%rsp) movq %rax, 728(%rsp) movq %rdi, 736(%rsp) movq %rsi, 744(%rsp) movq %rcx, 752(%rsp) movq %rax, 760(%rsp) movq %rdi, 768(%rsp) movq %rsi, 776(%rsp) movq %rcx, 784(%rsp) movq %rax, 792(%rsp) movq %rdi, 800(%rsp) movq %rsi, 808(%rsp) movq %rcx, 816(%rsp) movq %rax, 824(%rsp) movq %rdi, 832(%rsp) movq %rsi, 840(%rsp) movq %rcx, 848(%rsp) movq %rax, 856(%rsp) jmp .L479 .L489: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v2a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $288, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L493: movss %xmm0, 64(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L493 testb $31, %bl je .L506 testl %eax, %eax je .L507 movss .LC0(%rip), %xmm0 .L499: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L496 jmp .L494 .L506: movss .LC0(%rip), %xmm0 .L494: movl %edx, %edi cmpl $15, %edx ja .L497 jmp .L498 .L496: testl %edx, %edx jne .L499 jmp .L498 .L497: movq 64(%rsp), %rax movq %rax, 128(%rsp) movq 72(%rsp), %rax movq %rax, 136(%rsp) movq 80(%rsp), %rax movq %rax, 144(%rsp) movq 88(%rsp), %rax movq %rax, 152(%rsp) movl %edx, %eax movq %rbx, %rdx .L501: movq (%rdx), %rcx movq %rcx, 256(%rsp) movq 8(%rdx), %rcx movq %rcx, 264(%rsp) movq 16(%rdx), %rcx movq %rcx, 272(%rsp) movq 24(%rdx), %rcx movq %rcx, 280(%rsp) movq 32(%rdx), %rcx movq %rcx, 224(%rsp) movq 40(%rdx), %rcx movq %rcx, 232(%rsp) movq 48(%rdx), %rcx movq %rcx, 240(%rsp) movq 56(%rdx), %rcx movq %rcx, 248(%rsp) movaps 224(%rsp), %xmm1 mulps 256(%rsp), %xmm1 movaps %xmm1, (%rsp) movaps 240(%rsp), %xmm1 mulps 272(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movq (%rsp), %rsi movaps (%rsp), %xmm4 movaps %xmm4, 32(%rsp) movq 16(%rsp), %rcx movaps %xmm1, 48(%rsp) movq %rsi, 192(%rsp) movq 40(%rsp), %rsi movq %rsi, 200(%rsp) movq %rcx, 208(%rsp) movq 56(%rsp), %rcx movq %rcx, 216(%rsp) movaps %xmm4, %xmm2 mulps 128(%rsp), %xmm2 movaps %xmm2, (%rsp) movq (%rsp), %rsi movaps %xmm2, 32(%rsp) movaps %xmm1, %xmm3 mulps 144(%rsp), %xmm3 movaps %xmm3, 16(%rsp) movq 16(%rsp), %rcx movaps %xmm3, 48(%rsp) movq %rsi, 160(%rsp) movq 40(%rsp), %rsi movq %rsi, 168(%rsp) movq %rcx, 176(%rsp) movq 56(%rsp), %rcx movq %rcx, 184(%rsp) addq $64, %rdx subl $16, %eax cmpl $15, %eax jbe .L500 movq 160(%rsp), %rcx movq %rcx, 128(%rsp) movq %rsi, 136(%rsp) movq 176(%rsp), %rcx movq %rcx, 144(%rsp) movq 184(%rsp), %rcx movq %rcx, 152(%rsp) jmp .L501 .L500: movq 160(%rsp), %r9 movq %r9, 96(%rsp) movq 168(%rsp), %r8 movq %r8, 104(%rsp) movq 176(%rsp), %rsi movq %rsi, 112(%rsp) movq 184(%rsp), %rcx movq %rcx, 120(%rsp) leal -16(%rdi), %edx movl %edx, %eax shrl $4, %eax movl %eax, %edi addq $1, %rdi salq $6, %rdi addq %rdi, %rbx movq %r9, 64(%rsp) movq %r8, 72(%rsp) movq %rsi, 80(%rsp) movq %rcx, 88(%rsp) sall $4, %eax subl %eax, %edx .L498: testl %edx, %edx je .L502 .L503: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L503 jmp .L502 .L505: mulss 64(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L505 movss %xmm0, (%r12) jmp .L508 .L507: movss .LC0(%rip), %xmm0 jmp .L498 .L502: movl $0, %edx movl $0, %eax jmp .L505 .L508: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v4a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $448, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L512: movss %xmm0, 96(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L512 testb $31, %bl je .L525 testl %eax, %eax je .L526 movss .LC0(%rip), %xmm0 .L518: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L515 jmp .L513 .L525: movss .LC0(%rip), %xmm0 .L513: movl %edx, %edi cmpl $31, %edx ja .L516 jmp .L517 .L515: testl %edx, %edx jne .L518 jmp .L517 .L516: movq 96(%rsp), %rax movq %rax, 160(%rsp) movq 104(%rsp), %rax movq %rax, 168(%rsp) movq 112(%rsp), %rax movq %rax, 176(%rsp) movq 120(%rsp), %rax movq %rax, 184(%rsp) movl %edx, %eax movq %rbx, %rdx .L520: movq (%rdx), %rcx movq %rcx, 416(%rsp) movq 8(%rdx), %rcx movq %rcx, 424(%rsp) movq 16(%rdx), %rcx movq %rcx, 432(%rsp) movq 24(%rdx), %rcx movq %rcx, 440(%rsp) movq 32(%rdx), %rcx movq %rcx, 384(%rsp) movq 40(%rdx), %rcx movq %rcx, 392(%rsp) movq 48(%rdx), %rcx movq %rcx, 400(%rsp) movq 56(%rdx), %rcx movq %rcx, 408(%rsp) movq 64(%rdx), %rcx movq %rcx, 352(%rsp) movq 72(%rdx), %rcx movq %rcx, 360(%rsp) movq 80(%rdx), %rcx movq %rcx, 368(%rsp) movq 88(%rdx), %rcx movq %rcx, 376(%rsp) movq 96(%rdx), %rcx movq %rcx, 320(%rsp) movq 104(%rdx), %rcx movq %rcx, 328(%rsp) movq 112(%rdx), %rcx movq %rcx, 336(%rsp) movq 120(%rdx), %rcx movq %rcx, 344(%rsp) movaps 384(%rsp), %xmm1 mulps 416(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movaps 400(%rsp), %xmm1 mulps 432(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movq 32(%rsp), %rsi movaps 32(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq 48(%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 288(%rsp) movq 72(%rsp), %rsi movq %rsi, 296(%rsp) movq %rcx, 304(%rsp) movq 88(%rsp), %rcx movq %rcx, 312(%rsp) movaps 320(%rsp), %xmm1 mulps 352(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 368(%rsp), %xmm1 mulps 336(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rsi movaps 16(%rsp), %xmm6 movaps %xmm6, 64(%rsp) movq (%rsp), %rcx movaps %xmm1, 80(%rsp) movq %rsi, 256(%rsp) movq 72(%rsp), %rsi movq %rsi, 264(%rsp) movq %rcx, 272(%rsp) movq 88(%rsp), %rcx movq %rcx, 280(%rsp) movaps 32(%rsp), %xmm2 mulps %xmm6, %xmm2 movaps %xmm2, 32(%rsp) movaps 48(%rsp), %xmm3 mulps %xmm1, %xmm3 movaps %xmm3, 48(%rsp) movq 32(%rsp), %rsi movaps %xmm2, 64(%rsp) movq 48(%rsp), %rcx movaps %xmm3, 80(%rsp) movq %rsi, 224(%rsp) movq 72(%rsp), %rsi movq %rsi, 232(%rsp) movq %rcx, 240(%rsp) movq 88(%rsp), %rcx movq %rcx, 248(%rsp) movaps %xmm2, %xmm4 mulps 160(%rsp), %xmm4 movaps %xmm4, 32(%rsp) movq 32(%rsp), %rsi movaps %xmm4, 64(%rsp) movaps %xmm3, %xmm5 mulps 176(%rsp), %xmm5 movaps %xmm5, 48(%rsp) movq 48(%rsp), %rcx movaps %xmm5, 80(%rsp) movq %rsi, 192(%rsp) movq 72(%rsp), %rsi movq %rsi, 200(%rsp) movq %rcx, 208(%rsp) movq 88(%rsp), %rcx movq %rcx, 216(%rsp) subq $-128, %rdx subl $32, %eax cmpl $31, %eax jbe .L519 movq 192(%rsp), %rcx movq %rcx, 160(%rsp) movq %rsi, 168(%rsp) movq 208(%rsp), %rcx movq %rcx, 176(%rsp) movq 216(%rsp), %rcx movq %rcx, 184(%rsp) jmp .L520 .L519: movq 192(%rsp), %r9 movq %r9, 128(%rsp) movq 200(%rsp), %r8 movq %r8, 136(%rsp) movq 208(%rsp), %rsi movq %rsi, 144(%rsp) movq 216(%rsp), %rcx movq %rcx, 152(%rsp) leal -32(%rdi), %edx movl %edx, %eax shrl $5, %eax movl %eax, %edi addq $1, %rdi salq $7, %rdi addq %rdi, %rbx movq %r9, 96(%rsp) movq %r8, 104(%rsp) movq %rsi, 112(%rsp) movq %rcx, 120(%rsp) sall $5, %eax subl %eax, %edx .L517: testl %edx, %edx je .L521 .L522: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L522 jmp .L521 .L524: mulss 96(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L524 movss %xmm0, (%r12) jmp .L527 .L526: movss .LC0(%rip), %xmm0 jmp .L517 .L521: movl $0, %edx movl $0, %eax jmp .L524 .L527: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret simd_v8a_combine: pushq %rbp movq %rsp, %rbp pushq %r13 pushq %r12 pushq %rbx andq $-32, %rsp subq $736, %rsp movq %rdi, %r13 movq %rsi, %r12 call get_vec_start movq %rax, %rbx movq %r13, %rdi call vec_length movl %eax, %edx movl $0, %esi movl $0, %ecx movss .LC0(%rip), %xmm0 .L531: movss %xmm0, 128(%rsp,%rsi,4) addq $1, %rcx movq %rcx, %rsi cmpq $8, %rcx jne .L531 testb $31, %bl je .L544 testl %eax, %eax je .L545 movss .LC0(%rip), %xmm0 .L537: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx testb $31, %bl jne .L534 jmp .L532 .L544: movss .LC0(%rip), %xmm0 .L532: movl %edx, %edi cmpl $63, %edx ja .L535 jmp .L536 .L534: testl %edx, %edx jne .L537 jmp .L536 .L535: movq 128(%rsp), %rax movq %rax, 192(%rsp) movq 136(%rsp), %rax movq %rax, 200(%rsp) movq 144(%rsp), %rax movq %rax, 208(%rsp) movq 152(%rsp), %rax movq %rax, 216(%rsp) movl %edx, %eax movq %rbx, %rdx .L539: movq (%rdx), %rcx movq %rcx, 704(%rsp) movq 8(%rdx), %rcx movq %rcx, 712(%rsp) movq 16(%rdx), %rcx movq %rcx, 720(%rsp) movq 24(%rdx), %rcx movq %rcx, 728(%rsp) movq 32(%rdx), %rcx movq %rcx, 672(%rsp) movq 40(%rdx), %rcx movq %rcx, 680(%rsp) movq 48(%rdx), %rcx movq %rcx, 688(%rsp) movq 56(%rdx), %rcx movq %rcx, 696(%rsp) movq 64(%rdx), %rcx movq %rcx, 640(%rsp) movq 72(%rdx), %rcx movq %rcx, 648(%rsp) movq 80(%rdx), %rcx movq %rcx, 656(%rsp) movq 88(%rdx), %rcx movq %rcx, 664(%rsp) movq 96(%rdx), %rcx movq %rcx, 608(%rsp) movq 104(%rdx), %rcx movq %rcx, 616(%rsp) movq 112(%rdx), %rcx movq %rcx, 624(%rsp) movq 120(%rdx), %rcx movq %rcx, 632(%rsp) movq 128(%rdx), %rcx movq %rcx, 576(%rsp) movq 136(%rdx), %rcx movq %rcx, 584(%rsp) movq 144(%rdx), %rcx movq %rcx, 592(%rsp) movq 152(%rdx), %rcx movq %rcx, 600(%rsp) movq 160(%rdx), %rcx movq %rcx, 544(%rsp) movq 168(%rdx), %rcx movq %rcx, 552(%rsp) movq 176(%rdx), %rcx movq %rcx, 560(%rsp) movq 184(%rdx), %rcx movq %rcx, 568(%rsp) movq 192(%rdx), %rcx movq %rcx, 512(%rsp) movq 200(%rdx), %rcx movq %rcx, 520(%rsp) movq 208(%rdx), %rcx movq %rcx, 528(%rsp) movq 216(%rdx), %rcx movq %rcx, 536(%rsp) movq 224(%rdx), %rcx movq %rcx, 480(%rsp) movq 232(%rdx), %rcx movq %rcx, 488(%rsp) movq 240(%rdx), %rcx movq %rcx, 496(%rsp) movq 248(%rdx), %rcx movq %rcx, 504(%rsp) movaps 672(%rsp), %xmm1 mulps 704(%rsp), %xmm1 movaps %xmm1, 64(%rsp) movaps 688(%rsp), %xmm1 mulps 720(%rsp), %xmm1 movaps %xmm1, 80(%rsp) movq 64(%rsp), %rsi movaps 64(%rsp), %xmm4 movaps %xmm4, 96(%rsp) movq 80(%rsp), %rcx movaps %xmm1, %xmm5 movaps %xmm1, 112(%rsp) movq %rsi, 448(%rsp) movq 104(%rsp), %rsi movq %rsi, 456(%rsp) movq %rcx, 464(%rsp) movq 120(%rsp), %rcx movq %rcx, 472(%rsp) movaps 608(%rsp), %xmm1 mulps 640(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 656(%rsp), %xmm1 mulps 624(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm6 movaps %xmm6, 96(%rsp) movq 32(%rsp), %rcx movaps %xmm1, 112(%rsp) movq %rsi, 416(%rsp) movq 104(%rsp), %rsi movq %rsi, 424(%rsp) movq %rcx, 432(%rsp) movq 120(%rsp), %rcx movq %rcx, 440(%rsp) movaps %xmm4, %xmm2 mulps %xmm6, %xmm2 movaps %xmm2, 64(%rsp) movaps %xmm5, %xmm3 mulps %xmm1, %xmm3 movaps %xmm3, 80(%rsp) movq 64(%rsp), %rsi movaps %xmm2, 96(%rsp) movq 80(%rsp), %rcx movaps %xmm3, 112(%rsp) movq %rsi, 384(%rsp) movq 104(%rsp), %rsi movq %rsi, 392(%rsp) movq %rcx, 400(%rsp) movq 120(%rsp), %rcx movq %rcx, 408(%rsp) movaps 576(%rsp), %xmm1 mulps 544(%rsp), %xmm1 movaps %xmm1, 48(%rsp) movaps 592(%rsp), %xmm1 mulps 560(%rsp), %xmm1 movaps %xmm1, 32(%rsp) movq 48(%rsp), %rsi movaps 48(%rsp), %xmm6 movaps %xmm6, 96(%rsp) movq 32(%rsp), %rcx movaps %xmm1, %xmm7 movaps %xmm1, 112(%rsp) movq %rsi, 352(%rsp) movq 104(%rsp), %rsi movq %rsi, 360(%rsp) movq %rcx, 368(%rsp) movq 120(%rsp), %rcx movq %rcx, 376(%rsp) movaps 512(%rsp), %xmm1 mulps 480(%rsp), %xmm1 movaps %xmm1, 16(%rsp) movaps 528(%rsp), %xmm1 mulps 496(%rsp), %xmm1 movaps %xmm1, (%rsp) movq 16(%rsp), %rsi movaps 16(%rsp), %xmm2 movaps %xmm2, 96(%rsp) movq (%rsp), %rcx movaps %xmm1, 112(%rsp) movq %rsi, 320(%rsp) movq 104(%rsp), %rsi movq %rsi, 328(%rsp) movq %rcx, 336(%rsp) movq 120(%rsp), %rcx movq %rcx, 344(%rsp) movaps %xmm6, %xmm4 mulps %xmm2, %xmm4 movaps %xmm4, 48(%rsp) movaps %xmm7, %xmm5 mulps %xmm1, %xmm5 movaps %xmm5, 32(%rsp) movq 48(%rsp), %rsi movaps %xmm4, 96(%rsp) movq 32(%rsp), %rcx movaps %xmm5, 112(%rsp) movq %rsi, 288(%rsp) movq 104(%rsp), %rsi movq %rsi, 296(%rsp) movq %rcx, 304(%rsp) movq 120(%rsp), %rcx movq %rcx, 312(%rsp) movaps 64(%rsp), %xmm6 mulps %xmm4, %xmm6 movaps %xmm6, 64(%rsp) movaps 80(%rsp), %xmm7 mulps %xmm5, %xmm7 movaps %xmm7, 80(%rsp) movq 64(%rsp), %rsi movaps %xmm6, 96(%rsp) movq 80(%rsp), %rcx movaps %xmm7, 112(%rsp) movq %rsi, 256(%rsp) movq 104(%rsp), %rsi movq %rsi, 264(%rsp) movq %rcx, 272(%rsp) movq 120(%rsp), %rcx movq %rcx, 280(%rsp) movaps %xmm6, %xmm2 mulps 192(%rsp), %xmm2 movaps %xmm2, 64(%rsp) movq 64(%rsp), %rsi movaps %xmm2, 96(%rsp) movaps %xmm7, %xmm3 mulps 208(%rsp), %xmm3 movaps %xmm3, 80(%rsp) movq 80(%rsp), %rcx movaps %xmm3, 112(%rsp) movq %rsi, 224(%rsp) movq 104(%rsp), %rsi movq %rsi, 232(%rsp) movq %rcx, 240(%rsp) movq 120(%rsp), %rcx movq %rcx, 248(%rsp) addq $256, %rdx subl $64, %eax cmpl $63, %eax jbe .L538 movq 224(%rsp), %rcx movq %rcx, 192(%rsp) movq %rsi, 200(%rsp) movq 240(%rsp), %rcx movq %rcx, 208(%rsp) movq 248(%rsp), %rcx movq %rcx, 216(%rsp) jmp .L539 .L538: movq 224(%rsp), %r9 movq %r9, 160(%rsp) movq 232(%rsp), %r8 movq %r8, 168(%rsp) movq 240(%rsp), %rsi movq %rsi, 176(%rsp) movq 248(%rsp), %rcx movq %rcx, 184(%rsp) leal -64(%rdi), %edx movl %edx, %eax shrl $6, %eax movl %eax, %edi addq $1, %rdi salq $8, %rdi addq %rdi, %rbx movq %r9, 128(%rsp) movq %r8, 136(%rsp) movq %rsi, 144(%rsp) movq %rcx, 152(%rsp) sall $6, %eax subl %eax, %edx .L536: testl %edx, %edx je .L540 .L541: addq $4, %rbx mulss -4(%rbx), %xmm0 subl $1, %edx jne .L541 jmp .L540 .L543: mulss 128(%rsp,%rdx,4), %xmm0 addq $1, %rax movq %rax, %rdx cmpq $8, %rax jne .L543 movss %xmm0, (%r12) jmp .L546 .L545: movss .LC0(%rip), %xmm0 jmp .L536 .L540: movl $0, %edx movl $0, %eax jmp .L543 .L546: leaq -24(%rbp), %rsp popq %rbx popq %r12 popq %r13 popq %rbp ret unroll4x2as_combine: pushq %r14 pushq %r13 pushq %r12 pushq %rbp pushq %rbx movq %rdi, %r14 movq %rsi, %r12 call vec_length movq %rax, %rbx movq %rax, %r13 shrq $63, %r13 addq %rax, %r13 sarq %r13 movq %r14, %rdi call get_vec_start movq %rax, %rcx leaq (%rax,%r13,4), %rax testq %r13, %r13 jle .L553 movq %r13, %rbp movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 movl $0, %edx .L550: mulss (%rcx,%rdx,4), %xmm1 mulss (%rax,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbp, %rdx jne .L550 jmp .L549 .L553: movss .LC0(%rip), %xmm0 movaps %xmm0, %xmm1 .L549: leaq (%r13,%r13), %rdx cmpq %rdx, %rbx jle .L551 .L552: mulss (%rcx,%rdx,4), %xmm0 addq $1, %rdx cmpq %rbx, %rdx jne .L552 .L551: mulss %xmm1, %xmm0 movss %xmm0, (%r12) popq %rbx popq %rbp popq %r12 popq %r13 popq %r14 ret register_combiners: subq $8, %rsp movl $combine1_descr, %edx movl $combine1, %esi 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_descr, %edx movl $combine1, %esi movl $simd_v8a_combine, %edi call add_combiner movsd .LC1(%rip), %xmm1 movsd .LC2(%rip), %xmm0 movl $simd_v8a_combine, %edi call log_combiner addq $8, %rsp ret 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: