[LLVMdev] Number of spills/stores

Chris Lattner sabre at nondot.org
Wed Oct 4 20:52:58 PDT 2006

On Wed, 4 Oct 2006, Fernando Magno Quintao Pereira wrote:
>     After compiling 188.ammp, one of spec floating point benchmarks,
> using the linear scan register allocator, I got 34 spills, but only
> 28 store instructions. These data were given by -stats. Is this number
> correct? I mean, the number of stores should be equal or greater than
> the number of spills, shouldn't it?

Some spill code gets folded into instructions.  For example, if you have:

   vreg1234 += 7

they can be spilled to:

   [esp + 123] += 7

on x86.



More information about the llvm-dev mailing list