On Mon, 8 Oct 2012, Richard Guenther wrote:
On Mon, Oct 8, 2012 at 11:34 AM, Marc Glisse <marc.gli...@inria.fr> wrote:
On Mon, 8 Oct 2012, Richard Guenther wrote:
VEC_COND_EXPR is more complicated. We could for instance require that it
takes as first argument a vector of -1 and 0 (thus <0, !=0 and the neon
thing are equivalent). Which would leave to decide what the expansion of
vec_cond_expr passes to the targets when the first argument is not a
comparison, between !=0, <0, ==-1 or others (I vote for <0 because of
opencl). One issue is that targets wouldn't know if it was a dummy
comparison that can safely be ignored because the other part is the
result
of logical operations on comparisons (thus composed of -1 and 0) or a
genuine comparison with an arbitrary vector, so a new optimization would
be
needed (in the back-end I guess or we would need an alternate instruction
to
vcond) to detect if a vector is a "signed boolean" vector.
We could instead say that vec_cond_expr really follows OpenCL's semantics
and looks at the MSB of each element. I am not sure that would change
much,
it would mostly delay the apparition of <0 to RTL expansion time (and
thus
make gimple slightly lighter).
I think we should delay the decision on how to optimize this. It's indeed
not trivial and the GIMPLE middle-end aggressively forwards feeding
comparisons into the VEC_COND_EXPR expressions already (somewhat
defeating any CSE that might be possible here) in forwprop.
Thanks for going through the long email :-)
What does that imply for the first argument of VEC_COND_EXPR? Currently, the
expander asserts that it is a comparison, but that is not reflected in the
gimple checkers.
And I don't think we should reflect that in the gimple checkers rather fixup the
expander (transparently use p != 0 or p < 0).
I guess I'll pick p < 0 then (just because I am more interested in x86 and
it makes the optimization easier on x86). Having another expander than
vcond (one that takes the mask directly instead of a comparison, and for
which we promise that the argument will be a vector of -1/0) would be
convenient...
So is the best choice to document that VEC_COND_EXPR takes as
first argument a comparison and make gimple checking reflect that? (seems
sad, but at least that would tell me what I can/can't do)
No, that would just mean that in GIMPLE you'd add this p != 0 or p < 0.
And at some point in the future I really really want to push this embedded
expression to a separate statement so I have a SSA definition for it.
Once the expander is ready to accept it, ok. It seems to me that the
scalar COND_EXPR may also have an embedded expression, so I assume
COND_EXPR and VEC_COND_EXPR are meant to diverge (or maybe you also want
to do the same for COND_EXPR?).
By the way, since we are documenting comparisons as returning 0 and -1, does
that bring back the integer_truep predicate?
Not sure, true would still be != 0 or all_onesp (all bits of the
precision are 1), no?
I was going to make truep equivalent to onep for scalars and all_onesp for
vectors (since -1 will be the only value documented as "true" for
vectors). I guess it can wait, I can manually inline it for now.
Since we are documenting that comparisons of vectors return -1 and 0 in
the middle-end, I was wondering whether the comparison expanders would
need updating so they forward to vcond(...,-1,0), at least on platforms
that don't define VECTOR_STORE_FLAG_VALUE to constm1_rtx for this mode.
But a simple test on sparc shows it is already fine :-)
--
Marc Glisse