generate_normal_4wide.c

#include <x86intrin.h>

typedef struct TriangleX4 {
	__m128 ax, ay, az;
	__m128 bx, by, bz;
	__m128 cx, cy, cz;
} TriangleX4;

typedef struct TriangleNormalX4 {
	__m128 nx, ny, nz;
} TriangleNormalX4;

// Computes normals 4-wide
void gen_normals(const TriangleX4* tri, TriangleNormalX4* out) {
    // a.y * b.z - a.z * b.y
    __m128 nx = _mm_sub_ps(_mm_mul_ps(tri->ay, tri->bz), _mm_mul_ps(tri->az, tri->by));
            
    // a.z * b.x - a.x * b.z
    __m128 ny = _mm_sub_ps(_mm_mul_ps(tri->az, tri->bx), _mm_mul_ps(tri->ax, tri->bz));
            
    // a.x * b.y - a.y * b.x
    __m128 nz = _mm_sub_ps(_mm_mul_ps(tri->ax, tri->by), _mm_mul_ps(tri->ay, tri->bx));
            
    // sqrt((nx * nx) + (ny * ny) + (nz * nz))
    __m128 inv_length = _mm_rsqrt_ps(_mm_add_ps(_mm_add_ps(_mm_mul_ps(nx, nx), _mm_mul_ps(ny, ny)), _mm_mul_ps(nz, nz)));

    // NOTE(NeGate): This will shit out NaN if the 
    // triangle is a degenerate fuck
    out->nx = _mm_mul_ps(nx, inv_length);
    out->ny = _mm_mul_ps(ny, inv_length);
    out->nz = _mm_mul_ps(nz, inv_length);
}

output_clang.s

gen_normals:
        vmovaps xmm2, XMMWORD PTR [rdi+80]
        vmovaps xmm1, XMMWORD PTR [rdi+16]
        vmovaps xmm0, XMMWORD PTR [rdi+64]
        vmulps  xmm3, xmm2, xmm1
        vmovaps xmm4, XMMWORD PTR [rdi+32]
        vmovaps xmm6, XMMWORD PTR [rdi]
        vmovaps xmm5, XMMWORD PTR [rdi+48]
        vfnmadd231ps    xmm3, xmm0, xmm4
        vmulps  xmm0, xmm0, xmm6
        vmulps  xmm4, xmm4, xmm5
        vfnmadd132ps    xmm1, xmm0, xmm5
        vfnmadd132ps    xmm2, xmm4, xmm6
        vmulps  xmm0, xmm3, xmm3
        vfmadd231ps     xmm0, xmm2, xmm2
        vfmadd231ps     xmm0, xmm1, xmm1
        vrsqrtps        xmm0, xmm0
        vmulps  xmm3, xmm3, xmm0
        vmulps  xmm2, xmm2, xmm0
        vmulps  xmm1, xmm1, xmm0
        vmovaps XMMWORD PTR [rsi], xmm3
        vmovaps XMMWORD PTR [rsi+16], xmm2
        vmovaps XMMWORD PTR [rsi+32], xmm1
        ret