20190709-CircleAssemblyOriginal

.NET Framework 4.7.2 (CLR 4.0.30319.42000), 64bit RyuJIT-v4.7.3062.0
Raytracer2.Benchmarking.CircleTest.Intersection()
            for (int j = 0, l = n; j < l; j++)
                 ^^^^^^^^^
xor     edi,edi
            for (int j = 0, l = n; j < l; j++)
                            ^^^^^
mov     ebx,dword ptr [rsi+20h]
            for (int j = 0, l = n; j < l; j++)
                                   ^^^^^
test    ebx,ebx
jle     M00_L01
                rs[j].Intersect(c);
                ^^^^^^^^^^^^^^^^^^^
M00_L00

mov     rcx,qword ptr [rsi+8]
cmp     edi,dword ptr [rcx+8]
jae     00007ffe`3aa65f8f
movsxd  rdx,edi
lea     rdx,[rdx+rdx*4]
lea     rcx,[rcx+rdx*4+10h]
mov     rdx,qword ptr [rsi+10h]
call    Raytracer2.Primitives.Infinite.Ray.Intersect(Raytracer2.Primitives.Circle)
            for (int j = 0, l = n; j < l; j++)
                                          ^^^
inc     edi
cmp     edi,ebx
jl      M00_L00
        }
        ^
M00_L01

add     rsp,20h

Raytracer2.Primitives.Infinite.Ray.Intersect(Raytracer2.Primitives.Circle)
            float dx = this.direction.X;
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
lea     rax,[rsi+8]
mov     rcx,rax
vmovss  xmm1,dword ptr [rcx]
vmovss  xmm2,dword ptr [rsi+4]
cmp     dword ptr [rsi],esi
vmovss  xmm3,dword ptr [rax+4]
lea     rax,[rdx+10h]
mov     rcx,rax
vmovss  xmm4,dword ptr [rcx]
vmovss  xmm5,dword ptr [rax+4]
vmovss  xmm6,dword ptr [rdx+8]
vsubss  xmm0,xmm0,xmm4
vsubss  xmm2,xmm2,xmm5
            float r = 2 * p * dx;
            ^^^^^^^^^^^^^^^^^^^^^
vmovaps xmm4,xmm2
vmulss  xmm4,xmm4,dword ptr [00007ffe`3aa66138]
vmulss  xmm4,xmm4,xmm3
vmovaps xmm5,xmm1
vmulss  xmm5,xmm5,xmm1
vmulss  xmm3,xmm3,xmm3
vaddss  xmm3,xmm3,xmm5
vmovaps xmm5,xmm0
vmulss  xmm5,xmm5,dword ptr [00007ffe`3aa6613c]
vmulss  xmm5,xmm5,xmm1
vmovaps xmm1,xmm5
vaddss  xmm1,xmm1,xmm4
            float c = p * p + q * q - cr * cr;
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmulss  xmm0,xmm0,xmm0
vmulss  xmm2,xmm2,xmm2
vaddss  xmm0,xmm0,xmm2
vmulss  xmm6,xmm6,xmm6
vsubss  xmm0,xmm0,xmm6
            float DSqrt = b * b - 4 * a * c;
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovaps xmm2,xmm1
vmulss  xmm2,xmm2,xmm1
vmovaps xmm4,xmm3
vmulss  xmm4,xmm4,dword ptr [00007ffe`3aa66140]
vmulss  xmm4,xmm4,xmm0
vmovaps xmm0,xmm2
vsubss  xmm0,xmm0,xmm4
            if (DSqrt < 0)
            ^^^^^^^^^^^^^^
vxorps  xmm2,xmm2,xmm2
vucomiss xmm2,xmm0
jbe     M01_L00
            { return false; }
              ^^^^^^^^^^^^^
xor     eax,eax
vmovaps xmm6,xmmword ptr [rsp+20h]
add     rsp,30h
pop     rsi
ret
            float D = (float)Math.Sqrt(DSqrt);
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
M01_L00

vcvtss2sd xmm0,xmm0,xmm0
vsqrtsd xmm0,xmm0,xmm0
vcvtsd2ss xmm0,xmm0,xmm0
            float t0 = (-b + D) / (2 * a);
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovss  xmm2,dword ptr [00007ffe`3aa66144]
vxorps  xmm2,xmm2,xmm1
vaddss  xmm2,xmm2,xmm0
vmulss  xmm3,xmm3,dword ptr [00007ffe`3aa66148]
vdivss  xmm2,xmm2,xmm3
            float t1 = (-b - D) / (2 * a);
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovss  xmm4,dword ptr [00007ffe`3aa66144]
vxorps  xmm1,xmm1,xmm4
vsubss  xmm1,xmm1,xmm0
vdivss  xmm1,xmm1,xmm3
vmovaps xmm0,xmm2
call    System.Math.Min(Single, Single)
            if(ti > 0 && ti < t)
            ^^^^^^^^^^^^^^^^^^^^
vxorps  xmm1,xmm1,xmm1
vucomiss xmm0,xmm1
jbe     M01_L01
vmovss  xmm1,dword ptr [rsi+10h]
vucomiss xmm1,xmm0
jbe     M01_L01
                t = ti;
                ^^^^^^^
vmovss  dword ptr [rsi+10h],xmm0
                return true;
                ^^^^^^^^^^^^
mov     eax,1
vmovaps xmm6,xmmword ptr [rsp+20h]
add     rsp,30h
pop     rsi
ret
            return false;
            ^^^^^^^^^^^^^
M01_L01

xor     eax,eax

.NET Framework 4.7.2 (CLR 4.0.30319.42000), 64bit RyuJIT-v4.7.3062.0
Raytracer2.Benchmarking.CircleTest.IntersectionPacked()
            for (int j = 0, l = n / 8; j < l; j++)
                 ^^^^^^^^^
xor     edi,edi
            for (int j = 0, l = n / 8; j < l; j++)
                            ^^^^^^^^^
mov     edx,dword ptr [rsi+20h]
mov     ecx,edx
sar     ecx,1Fh
and     ecx,7
add     ecx,edx
mov     ebx,ecx
sar     ebx,3
            for (int j = 0, l = n / 8; j < l; j++)
                                       ^^^^^
test    ebx,ebx
jle     M00_L01
lea     rbp,[rsi+28h]
                rps[j].Intersect(cp);
                ^^^^^^^^^^^^^^^^^^^^^
M00_L00

mov     rdx,qword ptr [rsi+18h]
cmp     edi,dword ptr [rdx+8]
jae     00007ffe`3aa75fd7
movsxd  rcx,edi
imul    rcx,rcx,0A0h
lea     r14,[rdx+rcx+10h]
lea     r15,[rsp+80h]
lea     rcx,[rsp+20h]
mov     rdx,rbp
mov     r8d,60h
call    clr+0x1350
mov     rcx,r14
mov     rdx,r15
lea     r8,[rsp+20h]
call    Raytracer2.Primitives.Infinite.RayPack.Intersect(Raytracer2.Primitives.CirclePack)
            for (int j = 0, l = n / 8; j < l; j++)
                                              ^^^
inc     edi
cmp     edi,ebx
jl      M00_L00
        }
        ^
M00_L01

add     rsp,0A8h

Raytracer2.Primitives.Infinite.RayPack.Intersect(Raytracer2.Primitives.CirclePack)
    Vector ox = this.origin.x;
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovupd ymm1,ymmword ptr [rsi]
    Vector dx = this.direction.x;
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovupd ymm6,ymmword ptr [rsi+40h]
    Vector oy = this.origin.y;
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovupd ymm0,ymmword ptr [rsi+20h]
vmovupd ymm7,ymmword ptr [rsi+60h]
vmovupd ymm2,ymmword ptr [r8]
vmovupd ymm3,ymmword ptr [r8+20h]
vmovupd ymm8,ymmword ptr [r8+40h]
vsubps  ymm9,ymm1,ymm2
vsubps  ymm10,ymm0,ymm3
    Vector r = 2 * p * dx;
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
lea     rcx,[rsp+0E0h]
vmovupd ymmword ptr [rsp+40h],ymm9
vmovss  xmm1,dword ptr [00007ffe`3aa76ab8]
lea     r8,[rsp+40h]
vextractf128 xmm11,ymm9,1
vextractf128 xmm12,ymm10,1
vextractf128 xmm13,ymm6,1
vextractf128 xmm14,ymm7,1
vextractf128 xmm15,ymm8,1
call    System.Numerics.Vector`1[[System.Single, mscorlib]].op_Multiply(Single, System.Numerics.Vector`1)
vinsertf128 ymm8,ymm8,xmm15,1
vinsertf128 ymm7,ymm7,xmm14,1
vinsertf128 ymm6,ymm6,xmm13,1
vinsertf128 ymm10,ymm10,xmm12,1
vinsertf128 ymm9,ymm9,xmm11,1
lea     rcx,[rsp+0C0h]
vmovupd ymmword ptr [rsp+40h],ymm10
vmovss  xmm1,dword ptr [00007ffe`3aa76abc]
lea     r8,[rsp+40h]
vextractf128 xmm11,ymm9,1
vextractf128 xmm12,ymm10,1
vextractf128 xmm13,ymm6,1
vextractf128 xmm14,ymm7,1
vextractf128 xmm15,ymm8,1
call    System.Numerics.Vector`1[[System.Single, mscorlib]].op_Multiply(Single, System.Numerics.Vector`1)
vinsertf128 ymm8,ymm8,xmm15,1
vinsertf128 ymm7,ymm7,xmm14,1
vinsertf128 ymm6,ymm6,xmm13,1
vinsertf128 ymm10,ymm10,xmm12,1
vinsertf128 ymm9,ymm9,xmm11,1
vmovupd ymm1,ymmword ptr [rsp+0C0h]
vmulps  ymm1,ymm1,ymm7
vmulps  ymm0,ymm6,ymm6
vmulps  ymm7,ymm7,ymm7
vaddps  ymm7,ymm0,ymm7
vmulps  ymm9,ymm9,ymm9
vmulps  ymm10,ymm10,ymm10
vaddps  ymm9,ymm9,ymm10
vmulps  ymm8,ymm8,ymm8
vsubps  ymm9,ymm9,ymm8
vmovupd ymm0,ymmword ptr [rsp+0E0h]
vmulps  ymm0,ymm0,ymm6
vaddps  ymm6,ymm0,ymm1
lea     rcx,[rsp+0A0h]
vmovupd ymmword ptr [rsp+40h],ymm7
vmovss  xmm1,dword ptr [00007ffe`3aa76ac0]
lea     r8,[rsp+40h]
vextractf128 xmm8,ymm6,1
vextractf128 xmm10,ymm7,1
vextractf128 xmm11,ymm9,1
call    System.Numerics.Vector`1[[System.Single, mscorlib]].op_Multiply(Single, System.Numerics.Vector`1)
vinsertf128 ymm9,ymm9,xmm11,1
vinsertf128 ymm7,ymm7,xmm10,1
vinsertf128 ymm6,ymm6,xmm8,1
vmulps  ymm1,ymm6,ymm6
vmovupd ymm0,ymmword ptr [rsp+0A0h]
vmulps  ymm0,ymm0,ymm9
vsubps  ymm1,ymm1,ymm0
mov     rcx,qword ptr [13465D10h]
vmovupd ymm0,ymmword ptr [rcx+8]
vmovaps ymm8,ymm0
vcmpltps ymm8,ymm8,ymm1
vsqrtps ymm9,ymm1
vxorps  ymm1,ymm1,ymm1
vsubps  ymm10,ymm1,ymm6
lea     rcx,[rsp+80h]
vmovupd ymmword ptr [rsp+40h],ymm7
vmovss  xmm1,dword ptr [00007ffe`3aa76ac4]
lea     r8,[rsp+40h]
vextractf128 xmm11,ymm6,1
vextractf128 xmm12,ymm7,1
vextractf128 xmm13,ymm10,1
vextractf128 xmm14,ymm9,1
vextractf128 xmm15,ymm8,1
call    System.Numerics.Vector`1[[System.Single, mscorlib]].op_Multiply(Single, System.Numerics.Vector`1)
vinsertf128 ymm8,ymm8,xmm15,1
vinsertf128 ymm9,ymm9,xmm14,1
vinsertf128 ymm10,ymm10,xmm13,1
vinsertf128 ymm7,ymm7,xmm12,1
vinsertf128 ymm6,ymm6,xmm11,1
vaddps  ymm10,ymm10,ymm9
vmovupd ymm1,ymmword ptr [rsp+80h]
vdivps  ymm10,ymm10,ymm1
vxorps  ymm1,ymm1,ymm1
vsubps  ymm1,ymm1,ymm6
vsubps  ymm6,ymm1,ymm9
lea     rcx,[rsp+60h]
vmovupd ymmword ptr [rsp+40h],ymm7
vmovss  xmm1,dword ptr [00007ffe`3aa76ac8]
lea     r8,[rsp+40h]
vextractf128 xmm7,ymm10,1
vextractf128 xmm9,ymm6,1
vextractf128 xmm11,ymm8,1
call    System.Numerics.Vector`1[[System.Single, mscorlib]].op_Multiply(Single, System.Numerics.Vector`1)
vinsertf128 ymm8,ymm8,xmm11,1
vinsertf128 ymm6,ymm6,xmm9,1
vinsertf128 ymm10,ymm10,xmm7,1
vmovupd ymm0,ymmword ptr [rsp+60h]
vdivps  ymm1,ymm6,ymm0
    Vector tm = Vector.ConditionalSelect(Vector.LessThan(t1, t0), t1, t0);
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovaps ymm0,ymm1
vcmpltps ymm0,ymm0,ymm10
vpand   ymm0,ymm1,ymm0
vcmpltps ymm1,ymm1,ymm10
vpandn  ymm1,ymm1,ymm10
vpor    ymm0,ymm0,ymm1
    Vector maskBiggerThanZero = Vector.GreaterThan(tm, Constants.zeros);
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
mov     rax,qword ptr [13465D10h]
vmovupd ymm1,ymmword ptr [rax+8]
vcmpltps ymm1,ymm1,ymm0
    Vector maskSmallerThanT = Vector.LessThan(tm, this.t);
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
vmovupd ymm2,ymmword ptr [rsi+80h]
vmovaps ymm3,ymm0
vcmpltps ymm3,ymm3,ymm2
vpand   ymm1,ymm1,ymm3
vpand   ymm1,ymm8,ymm1
    this.t = Vector.ConditionalSelect(
    ^^^^^^^
        mask,                                                           // the bit mask that allows us to choose.
    ^^^^^^^
        tm,                                                             // the smallest of the t's.
    ^^^^^^^
        t);                                                             // if the bit mask is false (0), then we get our original t.
    ^^^^^^^
vpand   ymm0,ymm0,ymm1
vmovupd ymm2,ymmword ptr [rsi+80h]
vpandn  ymm3,ymm1,ymm2
vpor    ymm0,ymm0,ymm3
vmovupd ymmword ptr [rsi+80h],ymm0
    return mask;
    ^^^^^^^^^^^^
vmovupd ymmword ptr [rdi],ymm1
mov     rax,rdi