On Tue, 27 Nov 2012, Richard Biener wrote:
On Mon, Nov 12, 2012 at 2:31 PM, Marc Glisse <marc.gli...@inria.fr> wrote:
Hello,
this patch lets us fold a bit_field_ref of a constructor even when the
elements are vectors. Writing a testcase is not that convenient because of
the lack of a dead code elimination pass after forwprop4 (RTL doesn't always
remove everything either), but I see a clear difference on this test on
x86_64 without AVX.
typedef double vec __attribute__((vector_size(4*sizeof(double))));
void f(vec*x){
*x+=*x+*x;
}
double g(vec*x){
return (*x+*x)[3];
}
void h(vec*x, double a, double b, double c, double d){
vec y={a,b,c,d};
*x+=y;
}
I don't know if the patch (assuming it is ok) is suitable for stage 3, it
may have to wait. It passes bootstrap+testsuite.
... on x86_64-linux-gnu, sorry for the lack of precision (default
languages, Xeon E5520, graphite enabled).
I am still not quite sure why we even have a valid_gimple_rhs_p function
(after which we usually give up if it says no) instead of gimplifying, so I
may have missed a reason why CONSTRUCTOR or BIT_FIELD_REF shouldn't be ok.
Boostrapped and tested on ... ?
Ok if bootstrap / testing passes.
Thanks, I'll retest it to make sure it still works.
2012-11-12 Marc Glisse <marc.gli...@inria.fr>
* fold-const.c (fold_ternary_loc) [BIT_FIELD_REF]: Handle
CONSTRUCTOR with vector elements.
* tree-ssa-propagate.c (valid_gimple_rhs_p): Handle CONSTRUCTOR
and BIT_FIELD_REF.
--
Marc Glisse
