[llvm-dev] How to avoid register spills at wide integer addition?

[MITSUNARI Shigeo via llvm-dev]

Hi, all
I want to make a function such as

void add512(uint512_t *pz, const uint512_t *px, const uint512_t *py)
{
  *pz = *px + *py;
}

# uint512_t means 512-bit unsigned integer register.

Then, I make a sample code:
>cat t.ll
define void @add512(i512*noalias %pz, i512*noalias %px, i512*noalias %py)
{
  %x = load i512* %px
  %y = load i512* %py
  %z = add i512 %x, %y
  store i512 %z, i512* %pz
  ret void
}

>llc-3.6 -O3 -march=x86 t.ll -o -

add512:
        pushl   %ebp
        pushl   %ebx
        pushl   %edi
        pushl   %esi
        subl    $56, %esp
        movl    84(%esp), %eax
        movl    80(%esp), %edi
        movl    8(%edi), %esi
        movl    (%edi), %edx
        movl    4(%edi), %ebx
        movl    60(%eax), %ecx
        movl    %ecx, 52(%esp)          # 4-byte Spill
        movl    56(%eax), %ecx
        movl    %ecx, 48(%esp)          # 4-byte Spill
        addl    (%eax), %edx
        movl    %edx, 24(%esp)          # 4-byte Spill
        adcl    4(%eax), %ebx
        movl    %ebx, 20(%esp)          # 4-byte Spill
        adcl    8(%eax), %esi
        movl    %esi, 16(%esp)          # 4-byte Spill
        movl    52(%eax), %ecx
        movl    %ecx, 40(%esp)          # 4-byte Spill
        movl    48(%eax), %ecx
        movl    %ecx, 36(%esp)          # 4-byte Spill
        ....

I expect a code such as

add512:
    pushl %ebx
    movl 8(%esp), %ebx  ; pz
    movl 12(%esp), %ecx ; px
    movl 16(%esp), %edx ; py
    movl (%ecx), %eax
    addl (%edx), %eax
    movl %eax, (%ebx)
    movl 4(%ecx), %eax
    adcl 4(%edx), %eax
    movl %eax, 4(%ebx)
    movl 8(%ecx), %eax
    adcl 8(%edx), %eax
    movl %eax, 8(%ebx)
    ...

How can I let llvm to assign *%pz for work area of %z?

For example,

void twice512(uint512_t *pz, const uint512_t *px)
{
  *pz = *px + *px;
}
generates good code without spills.

Yours,
 Shigeo