PowerPC SVR4 64bit parameters read from wrong registers

Version and Platform (required):

• Binary Ninja Version: 5.3.9369-dev (f52453d9)

Bug Description:
64bit register parameters in PPC SVR4 are "aligned" in terms of their offset from arg1/r3. Binja doesn't know that. That is, a function like:

int math2(int a, unsigned long long b)

Will take a in r3, skip r4, and take b in r5:r6.

Steps To Reproduce:
Dumb repro source:

int
blackhole(char c)
{
    *(volatile char *)0x41414140 = c;
    return (*(volatile int *)0x42424240);
}

int
math2(int a, unsigned long long b)
{
    blackhole(a >> 24);
    blackhole(a >> 16);
    blackhole(a >> 8); 
    blackhole(a);
    for (unsigned i=0;i < 8;i++) {
        blackhole(b >> (8 * (i % 8)));
    }   
    return (0);
}

unsigned long long
math(unsigned long long a)
{
    return ((a << 3) | (a >> 61));
}

void _start(void)
{
    unsigned long long foo;

    foo = math(0x0102030405060708ULL);

    blackhole(math(0x1121314151617181ULL));

    math2(0xcafebeef, foo);

    return;
}

Build: powerpc-none-eabi-gcc test.c -o test

Binary: polygonal entropy splits intuitively

HLIL in _start:

01800424        r3, r4 = math(0x1020304, 0x5060708)
01800440        int32_t r4_1 = math(0x11213141, 0x51617181)
01800454        blackhole(r4_1.b & 0xff)
01800468        math2(0xcafebeef, r4_1, r3, r4)

The extra r4_1 argument to math2 is incorrect. Not catastrophic here in this tiny case, but in my real binaries, printf calls with %llus sprinkled in get super messed up and confusing.

Of course this should also be concatenating the upper/lower halves back together in theory, but I imagine that's a separate, larger issue.

Start by building the supplied test with powerpc-none-eabi-gcc and inspect the _start HLIL, especially the calls to math and math2. Trace how the PPC SVR4 calling convention assigns 64-bit parameters from r3 onward; done means the math2 call no longer includes the spurious r4_1 argument and the affected 64-bit values are represented correctly.