RE: [Algorithms] non-branching zbuf code tricks?

[Alex C. <al...@ar...>]

The intel compiler produces some interesting code for:

void foobar(int z, int *restrict zBuf, int len)
{
	for (int i = 0; i < len; i++)
		zBuf[i] = z < zBuf[i] ? z : zBuf[i];
}

1284: void foobar(int z, int *restrict zBuf, int len)
1285: {
004CED10   push        esi
004CED11   push        ebx
004CED12   sub         esp,10h
004CED15   mov         esi,dword ptr [esp+1Ch]
1286:     for (int i = 0; i < len; i++)
004CED19   xor         ebx,ebx
004CED1B   test        esi,esi
004CED1D   jle         foobar+80h (004ced90)
004CED23   cmp         esi,6
004CED26   jl          foobar+70h (004ced80)
004CED28   lea         eax,[esi-6]
004CED2B   mov         dword ptr [esp+0Ch],ebp
004CED2F   nop
1287:         zBuf[i] = z < zBuf[i] ? z : zBuf[i];
004CED30   mov         ebp,dword ptr [edx+ebx*4]
004CED33   cmp         ecx,ebp
004CED35   cmovl       ebp,ecx
004CED38   mov         dword ptr [edx+ebx*4],ebp
004CED3B   mov         ebp,dword ptr [edx+ebx*4+4]
004CED3F   cmp         ecx,ebp
004CED41   cmovl       ebp,ecx
004CED44   mov         dword ptr [edx+ebx*4+4],ebp
004CED48   mov         ebp,dword ptr [edx+ebx*4+8]
004CED4C   cmp         ecx,ebp
004CED4E   cmovl       ebp,ecx
004CED51   mov         dword ptr [edx+ebx*4+8],ebp
004CED55   mov         ebp,dword ptr [edx+ebx*4+0Ch]
004CED59   cmp         ecx,ebp
004CED5B   cmovl       ebp,ecx
004CED5E   mov         dword ptr [edx+ebx*4+0Ch],ebp
004CED62   mov         ebp,dword ptr [edx+ebx*4+10h]
004CED66   cmp         ecx,ebp
004CED68   cmovl       ebp,ecx
004CED6B   mov         dword ptr [edx+ebx*4+10h],ebp
1286:     for (int i = 0; i < len; i++)
004CED6F   add         ebx,5
004CED72   cmp         ebx,eax
004CED74   jle         foobar+20h (004ced30)
004CED76   mov         ebp,dword ptr [esp+0Ch]
004CED7A   lea         esi,[esi]
1287:         zBuf[i] = z < zBuf[i] ? z : zBuf[i];
004CED80   mov         eax,dword ptr [edx+ebx*4]
004CED83   cmp         ecx,eax
004CED85   cmovl       eax,ecx
004CED88   mov         dword ptr [edx+ebx*4],eax
1286:     for (int i = 0; i < len; i++)
004CED8B   inc         ebx
004CED8C   cmp         ebx,esi
004CED8E   jl          foobar+70h (004ced80)
1288: }
004CED90   add         esp,10h
004CED93   pop         ebx
004CED94   pop         esi
004CED95   ret         4

If you have the time, I'd reccomend evaluating the Intel compiler.  
I tend to find it produces good code and it often has better error
messages than VC++, it also integrates into the VC IDE which is great.

Alex Clarke, Programmer
Argonaut Games PLC