{% if msvc %}

_text segment
{{arch}}_mmm_f32_{{size}}_{{suffix}} proc

{% else %}

.intel_syntax noprefix
.text
.p2align 5
.globl {{G}}{{arch}}_mmm_f32_{{size}}_{{suffix}}
{{G}}{{arch}}_mmm_f32_{{size}}_{{suffix}}:
.cfi_startproc

{% endif %}

    push        rbp
    mov         rbp, rsp

{% if family == "windows" %}
// https://www.agner.org/optimize/calling_conventions.pdf xmm6-15 are not scratch
// https://stackoverflow.com/questions/43358429/save-value-of-xmm-registers
    and rsp,-16
    lea rsp,[rsp-160]
    vmovaps [rsp], xmm6
    vmovaps [rsp+16*1],xmm7
    vmovaps [rsp+16*2],xmm8
    vmovaps [rsp+16*3],xmm9
    vmovaps [rsp+16*4],xmm10
    vmovaps [rsp+16*5],xmm11
    vmovaps [rsp+16*6],xmm12
    vmovaps [rsp+16*7],xmm13
    vmovaps [rsp+16*8],xmm14
    vmovaps [rsp+16*9],xmm15

    push        rdi
    push        rsi

    mov         rdi, rcx

{% endif %}

    push        rbx
    push        r12
    push        r13
    push        r14
    push        r15

    sub         rsp, 8

{% if family == "unix" %}
.cfi_def_cfa_offset 64
{% endif %}
    stmxcsr     [rsp + 4]
{% if msvc %}
    mov         rax, 1FC0h
{% else %}
    mov         rax, 0x1FC0
{% endif %}
    mov         [rsp], eax
    ldmxcsr     [rsp]

{% include "dispatcher.tmpliq" %}