[Bug tree-optimization/78687] inefficient code generated for eggs.variant

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78687

--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 6 Dec 2016, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78687
> 
> --- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> From looking at the testcase, I think it is solely SRA that can help here, if
> it manages to scalarize everything, we are good, if not, you get terrible code.
> The classes are huge and largely uninitialized, as can be seen on the llvm's
> optimized code (first 40 bytes uninitialized, then 12 bytes initialized).

I guess the main issue is that SRA is not flow-sensitive so it would
need to materialize all aggregate elements as replacement (later DCE
would clean up of course).  For this case it's easy though -- only
parts of the structures are ever written to (the aggregate assignment
might confuse SRA enough here though)