U++ forum: Welcome to the forum

Re: "better" version of Iscale functions [message #15167 is a reply to message #15166]

Sun, 06 April 2008 20:05

mdelfede
Messages: 1308
Registered: September 2007

Ultimate Contributor

Btw, I guess GCC does some better optimizazions in this case :

int iscale(int x, int y, int z)
{
  int64_t res = x;
  res *= y;
  res /= z;
  return (int)res;
}

gets translated as :

.globl _ZN3Upp6iscaleEiii
	.type	_ZN3Upp6iscaleEiii, @function
_ZN3Upp6iscaleEiii:
.LFB4039:
	.file 4 "/home/massimo/sources/upp-svn/uppsrc/Core/mathutil.cpp"
	.loc 4 11 0
	pushq	%rbp
.LCFI15:
	movq	%rsp, %rbp
.LCFI16:
	movl	%edi, -20(%rbp)
	movl	%esi, -24(%rbp)
	movl	%edx, -28(%rbp)
.LBB2:
	.loc 4 17 0
	movl	-20(%rbp), %eax
	cltq
	movq	%rax, -8(%rbp)
	.loc 4 18 0
	movl	-24(%rbp), %eax
	movslq	%eax,%rdx
	movq	-8(%rbp), %rax
	imulq	%rdx, %rax
	movq	%rax, -8(%rbp)
	.loc 4 19 0
	movl	-28(%rbp), %eax
	cltq
	movq	-8(%rbp), %rdx
	movq	%rax, %rcx
	movq	%rdx, %rax
	sarq	$63, %rdx
	idivq	%rcx
	movq	%rax, -8(%rbp)
	.loc 4 20 0
	movq	-8(%rbp), %rax
.LBE2:
	.loc 4 30 0
	leave
	ret

If you drop assembly directive and/or line references and other stuffs, you can see that work is done with just one imul and one idiv... like your assembly version. The rest is just registry moving stuffs. I don't know what happens if you compile it with full optimization, but I guess mostly of them just disappears.
And I think that's still much quicker than floating point version.

Max

[Updated on: Sun, 06 April 2008 20:06]

Report message to a moderator