[llvm-commits] [llvm] r48909 - /llvm/trunk/lib/Transforms/Scalar/GVN.cpp

[Chris Lattner]

On Mar 28, 2008, at 2:19 AM, Duncan Sands wrote:

> Hi Chris,
>
>> +    // If we found less than 4 stores to merge, ignore the  
>> subrange: it isn't
>> +    // worth losing type information in llvm IR to do the  
>> transformation.
>> +    if (Range.TheStores.size() < 4)
>> +      continue;
>
> is this counting bytes stored, or store instructions?

This is checking # stores.  I'm open to suggestions, this is certainly  
work in progress.  The idea is that merging just a few stores could be  
done by the code generator, turning them into a larger scalar store.   
What I want the optimizer to get is to turn huge sequences of stores  
into memset.  For example, the testcase I'm working from looks like:

typedef struct {
   short mvx;
   short mvy;
} MV;

extern void foo(MV *, MV *, char *);

int t()
{
   MV up_mvd[8], left_mvd[8], mv_zero = {0,0};
   char ref_idx[8];

   ref_idx[0] = ref_idx[1] = ref_idx[2] = ref_idx[3] = ref_idx[4] =  
ref_idx[5] = ref_idx[6] = ref_idx[7] = -1;
   up_mvd[0] = up_mvd[1] = up_mvd[2] = up_mvd[3] = up_mvd[4] =  
up_mvd[5] = up_mvd[6] = up_mvd[7] = mv_zero;
   left_mvd[0] = left_mvd[1] = left_mvd[2] = left_mvd[3] = left_mvd[4]  
= left_mvd[5] = left_mvd[6] = left_mvd[7] = mv_zero;  	
   foo(up_mvd, left_mvd, ref_idx);
}

The assignments into up_mvd/left_mvd should be memsets.  The ref_idx  
should be one 8-byte store or 2 32-bit ones.

Suggestions for heuristics are welcome!

> PS: Feel free to ignore my previous email: I thought you were trying  
> to
> maintain one range rather than a bunch of them.

Ok :)

-Chris