https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70251
--- Comment #9 from Marc Glisse <glisse at gcc dot gnu.org> ---
I don't find the current situation very consistent. We seem to consider that in
gimple, the output of a vector comparison can only be used directly in a
vec_cond_expr (possibly also {and,ior,xor,bit_not}_expr?), except in the very
specific case where we apply this transformation a+(b?1:0) -> a-VCE(b), where
we only have a view_convert_expr. Unless we start transforming x?-1:0 into
view_convert_expr(x) under the same condition (VECTOR_MODE_P(...)), it doesn't
make sense to me. I am also not convinced that VECTOR_MODE_P is the right test.
If you are using a long vector (say of 2048 bits), the mode will never be a
vector mode, so at best the transformation is delayed until after vector
lowering. By the way, nothing seems to combine the following into a single
statement:
c_3 = VEC_COND_EXPR <x_1(D) < y_2(D), { -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1 }, { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }>;
_4 = VEC_COND_EXPR <c_3 != { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }, { 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 }, { 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0 }>;
(it does get combined in forwprop4 after the vector operation is lowered to
scalar operations (yes, we are missing the code to use sub-vectors here))
From a code generation point of view:
typedef int vec __attribute__((vector_size(32)));
vec f(vec x,vec y,vec z){
vec zero={};
vec one=zero+1;
vec c=x<y;
return z+(c?one:zero);
}
-mavx2 generates
vpcmpgtd %ymm0, %ymm1, %ymm1
vpsubd %ymm1, %ymm2, %ymm0
while -march=skylake-avx512 gives the inferior
vpcmpgtd %ymm0, %ymm1, %ymm1
vpandd .LC0(%rip), %ymm1, %ymm0
vpaddd %ymm2, %ymm0, %ymm0
(In reply to rguent...@suse.de from comment #5)
> On Wed, 16 Mar 2016, glisse at gcc dot gnu.org wrote:
> > A + (B vcmp C ? 1 : 0) -> A - (B vcmp C ? -1 : 0)
> I think that would be an odd transform.
As far as I understand, since the bool vector changes, this is currently the
proper gimple syntax for this transformation, with the added bonus that it
doesn't need to be disabled for avx512.
(I got "incorrect sharing of tree nodes" on the first argument of vec_cond_expr
when I wanted to try it out, the match-simplify machinery should probably do
the unshare_expr automatically, especially since gimplify now wants to keep the
comparison inside the vec_cond_expr)
