llvm dev - Nov 2009 - [LLVMdev] Overflow intrinsics

Hello-

Over the past couple of days I've been writing some very basic  
optimisations in -instcombine for the arithmetic with overflow  
intrinsics. As Duncan Sands pointed out on IRC however, this is  
duplicated effort when one can get equivalent results using weird  
sizes of ints in the IR.

At present (at least, on x86/x86-64), the overflow intrinsics generate  
better code. So what would be the better route: keeping the overflow  
intrinsics, or getting rid of them and improving codegen to handle  
flags better?

Alastair

Hi,
> At present (at least, on x86/x86-64), the overflow intrinsics generate  
> better code. So what would be the better route: keeping the overflow  
> intrinsics, or getting rid of them and improving codegen to handle  
> flags better?
http://llvm.org/bugs/show_bug.cgi?id=5443

Ciao,

Duncan.

On Nov 10, 2009, at 9:56 AM, Alastair Lynn wrote:
> Hello-
>
> Over the past couple of days I've been writing some very basic
> optimisations in -instcombine for the arithmetic with overflow
> intrinsics. As Duncan Sands pointed out on IRC however, this is
> duplicated effort when one can get equivalent results using weird
> sizes of ints in the IR.
>
> At present (at least, on x86/x86-64), the overflow intrinsics generate
> better code. So what would be the better route: keeping the overflow
> intrinsics, or getting rid of them and improving codegen to handle
> flags better?
Both.  Until codegen improves to the point where it is always better  
or equal to the intrinsics, we should keep the intrinsics and keep  
improving codegen.  If they ever become "always better than" the  
intrinsics, we can remove the intrinsics.  We're a long way away from  
being able to remove them, but I agree in principle that it would be  
nice to eliminate them.

-Chris