[Intel-gfx] [PATCH i-g-t] benchmarks/gem_wsim: Heap allocate VLA structs

Chris Wilson chris at chris-wilson.co.uk
Fri May 24 08:33:52 UTC 2019


Quoting Tvrtko Ursulin (2019-05-24 09:20:47)
> 
> On 24/05/2019 08:25, Chris Wilson wrote:
> > Apparently VLA structs (e.g. struct { int array[count] }) is a gcc
> > extension that clang refuses to support as handling memory layout is too
> > difficult for it.
> > 
> > Move the on-stack VLA to the heap.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > ---
> >   benchmarks/gem_wsim.c | 146 +++++++++++++++++++++++++++---------------
> >   1 file changed, 95 insertions(+), 51 deletions(-)
> > 
> > diff --git a/benchmarks/gem_wsim.c b/benchmarks/gem_wsim.c
> > index e2ffb93a9..0a0032bff 100644
> > --- a/benchmarks/gem_wsim.c
> > +++ b/benchmarks/gem_wsim.c
> > @@ -1441,6 +1441,48 @@ set_ctx_sseu(struct ctx *ctx, uint64_t slice_mask)
> >       return slice_mask;
> >   }
> >   
> > +static size_t sizeof_load_balance(int count)
> > +{
> > +     struct i915_context_engines_load_balance *ptr;
> > +
> > +     assert(sizeof(ptr->engines[count]) == count * sizeof(ptr->engines[0]));
> 
> This seems wrong - is bound to trigger.

Why does it seem wrong? That's the calculation used previously, and the
ptr->engines[] was meant to be packed in order for
sizeof(ptr->engines[count]) == count * sizeof(ptr->engines[0]). Anyway,
I threw it in there to check if the calculation was sane.

> > +     return sizeof(*ptr) + sizeof(ptr->engines[count]);
> 
> So size of of engine needs to be multiplied by count.

(Just note this is the what the current VLA evaluates to :)

> > +}
> > +
> > +static struct i915_context_engines_load_balance *
> > +alloc_load_balance(int count)
> > +{
> > +     return calloc(1, sizeof_load_balance(count));
> 
> How about alloca so cleanup is simpler? Or is alloca also on the 
> unpopular list?

I don't mind. Would shave a few lines indeed, but we need the memsets
back. #define alloca0()?

> Or possibly what Simon suggested, just a large temporary stack arrays 
> would be enough and easiest diff. Just with an assert that it fits.

I don't think that is as clean for the long term.
-Chris


More information about the Intel-gfx mailing list