Request for review (L): 6378821: where available, bitCount() should use POPC on SPARC processors and AMD+10h

[Christian Thalinger]

On Thu, 2009-03-12 at 16:16 -0700, Tom Rodriguez wrote:
> The PopCount nodes should only take one argument since they are never  
> required to have control.
> 
> I think there are problems with register usage with PopCountL in  
> x86_32.ad.  USEs (inputs) and TEMPs are guaranteed to be different  
> registers but DEFs are not.  KILLs and DEFs are guaranteed to be  
> different.  The model of register interference is like a single  
> hardware instruction where all the effects occur simultaneously so the  
> DEFs and USEs can't interfere.  So for this:
> 
> +instruct popCountL(eRegI dst, eRegL src, eRegI tmp, eFlagsReg cr) %{
> +  match(Set dst (PopCountL src));
> +  effect(KILL cr, TEMP tmp);
> 
> dst isn't required to be different than src or tmp which will break  
> you code generation.  If you add TEMP dst to the effects then the  
> register allocator will make the output interfere with the inputs so  
> you can use it as a temp in the code generation.  The memory one has  
> the same problem.

I see.  Thanks for the detailed explanation, I will fix that.

-- Christian