llvm dev - Jan 2012 - [LLVMdev] Integer factorizations in clang

Hello,

I have a small test case :

a = b * c + b * d

and clang is fatorizing b to get only one multiplication (-O3) which is 
normally good for most architectures. But in my case, that pattern takes 
just one instruction and I would like to recover it as is.
Where can I find this particular optimization in clang ?
Is there a way of setting multiplications as being cheaper than 
additions/subtracions in clang ?

Thanks in advance,

Ivan

Hi Ivan,
> I have a small test case :
>
> a = b * c + b * d
>
> and clang is fatorizing b to get only one multiplication (-O3) which is
> normally good for most architectures. But in my case, that pattern takes
> just one instruction and I would like to recover it as is.
> Where can I find this particular optimization in clang ?
> Is there a way of setting multiplications as being cheaper than
> additions/subtracions in clang ?
if you see the b*(c+d) pattern can't you just generate your b*c+b*d
instruction
for it?

Ciao, Duncan.

Le 5 janv. 2012 à 15:59, Ivan Llopard a écrit :
> Hello,
> 
> I have a small test case :
> 
> a = b * c + b * d
> 
> and clang is fatorizing b to get only one multiplication (-O3) which is 
> normally good for most architectures. But in my case, that pattern takes 
> just one instruction and I would like to recover it as is.
> Where can I find this particular optimization in clang ?
> Is there a way of setting multiplications as being cheaper than 
> additions/subtracions in clang ?
Wouldn't it be simpler to detect factorize pattern in you code and generate
the appropriate instruction as it if was a decomposed form.
It would have the benefit to properly optimize the code if the developer choose
to use the factorized form directly.


-- Jean-Daniel

Le 05/01/2012 16:35, Jean-Daniel Dupas a écrit :
> Le 5 janv. 2012 à 15:59, Ivan Llopard a écrit :
>
>> Hello,
>>
>> I have a small test case :
>>
>> a = b * c + b * d
>>
>> and clang is fatorizing b to get only one multiplication (-O3) which is
>> normally good for most architectures. But in my case, that pattern
takes
>> just one instruction and I would like to recover it as is.
>> Where can I find this particular optimization in clang ?
>> Is there a way of setting multiplications as being cheaper than
>> additions/subtracions in clang ?
> Wouldn't it be simpler to detect factorize pattern in you code and
generate the appropriate instruction as it if was a decomposed form.
> It would have the benefit to properly optimize the code if the developer
choose to use the factorized form directly.
Ok, I will try to match the factorized form from my back-end. Thanks you 
all.
>
> -- Jean-Daniel
>
>
>
>