[Bug tree-optimization/54855] Unnecessary duplication when performing scalar operation on vector element

[glisse at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54855



--- Comment #7 from Marc Glisse <glisse at gcc dot gnu.org> 2012-11-30 01:31:25 
UTC ---

Created attachment 28832

  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28832

simplify-rtx patch



With the patch, this code takes a single instruction. However, if I replace 'a'

with 1., it doesn't work (paradoxical subreg of mem...). I also see in the

combine dump some weird things like:

Failed to match this instruction:

(set (reg:V2DF 66 [ D.2211 ])

    (zero_extend:V2DF (mem/u/c:DF (symbol_ref/u:DI ("*.LC0") [flags 0x2]) [2 S8

A64])))







typedef double vec __attribute__((vector_size(16)));

vec f(vec x,double a)

{

  x[0] -= a;

  return x;

}