Re: [Caml-list] Re: optimize div to right shift (NOT!)

[Benedikt Meurer <benedikt.meurer@googlemail.com>]

On Dec 13, 2010, at 13:58 , Török Edwin wrote:

> It is still possible to avoid the division, gcc generates this:
> 	movq	%rdi, %rax
> 	shrq	$63, %rax
> 	addq	%rdi, %rax
> 	sarq	%rax
> 
> Or a better example with division by 8:
> 	leaq	7(%rdi), %rax
> 	testq	%rdi, %rdi
> 	cmovns	%rdi, %rax
> 	sarq	$3, %rax

ocamlopt does exactly this, atleast for x86-64.

> And division by non-power of two integers can optimized by replacing it
> with multiplication with its inverse (which gcc and llvm don't do for
> signed divisions, only for unsigned ones):
> http://www.hackersdelight.org/HDcode/magic.c.txt
> http://www.hackersdelight.org/HDcode/magicu.c.txt

The AMD64 optimization guide gives additional pointers here, i.e. it also lists replacement sequences for multiplication by constant.

> Best regards,
> --Edwin

greets,
Benedikt