On Wed, Jan 18, 2017 at 06:34:05PM +0300, Alexander Monakov wrote:
> My main concern is that nothing indicates to optimization passes that after
> SIMT_EXIT(), underlying storage is unavailable. What would prevent the
> compiler
> from transforming, say
>
> SIMT_ENTER();
> // originally omp simd private(inner)
> for (...)
> { ... }
> outer = inner;
> SIMT_EXIT();
>
> into
>
> SIMT_ENTER();
> // omp simd private(inner)
> for (...)
> { ... }
> SIMT_EXIT();
> outer = inner; // too late
>
>
> Also, we'd need to ensure IPA-SRA propagates the magic flag when decomposing
> structs.
We are talking here about addressable vars, right (so if we turn it into
non-addressable, in the SIMT region we just use the normal PTX pseudos),
right? We could emit inner ={v} {CLOBBER}; before SIMT_EXIT() to make it
clear it shouldn't be moved afterwards. For the private vars used directly
in SIMD region, for the vars from inlined functions I assume if they are
addressable we emit clobbers for them too. Or perhaps the alias oracle can
say that SIMT_EXIT ifn can clobber any addressable var with that
flag/attribute. And yes, SRA would need to propagate it.
But I believe it is worth it, because inlining is essential for good
performance of the simd regions.
Jakub
