[llvm-commits] CVS: llvm/lib/Target/ARM/ARMInstrVFP.td

[Chris Lattner]

> Evan's patch to avoid FPreg->intreg copy for cvt; store to mem
>
> +// avoid copying result of FTOUIZ to int reg when we're just  
> storing it
> +let AddedComplexity = 1 in
> +def : ARMPat<(store (i32 (bitconvert SPR:$src)), addrmode5:$addr),
> +             (FSTS SPR:$src, addrmode5:$addr)>;
> +
>  // 
> ===------------------------------------------------------------------- 
> ---===//
>  //  Load / store multiple Instructions.
>  //

Interesting approach.  Instead of adding this to one target's .td  
file, can you please investigate whether we can enable the dag  
combine xform at DAGCombiner.cpp:3412 now?

The primary thing blocking that xform was that we need to know (in  
this case) that an f32 store doesn't require greater alignment than  
an i32 store has.

-Chris