[PATCH] Add basic support for removal of load that are fed by a store of an aggregate

[Björn Steinbrink]

[Added llvm-commit back to Cc, sorry, removing it wasn't intentional]

On 2015.01.02 14:05:28 -0800, Chandler Carruth wrote:
> On Fri, Jan 2, 2015 at 2:02 PM, Björn Steinbrink <bsteinbr at gmail.com> wrote:
> 
> > 2015-01-02 22:36 GMT+01:00 Chandler Carruth <chandlerc at google.com>:
> > >
> > > On Fri, Jan 2, 2015 at 1:27 PM, Björn Steinbrink <bsteinbr at gmail.com> wrote:
> > >>
> > >> If we have a simple load through a GEP that is fed by a store of an
> > >> aggregate, we can use the GEP indices to walk the stored aggregate and
> > >> extract the appropriate value to replace the load.
> > >
> > >
> > > What's the motivation for this change?
> >
> > The rust compiler hit a case where not having this optimization caused
> > a branch not to be removed. The corresponding issue for rustc is
> > https://github.com/rust-lang/rust/issues/20149
> >
> > > Note that SROA replaces all stores of aggregates with scalar stores of
> > > the components specifically so that neither it nor GVN needs to
> > > cope with aggregate loads or stores.
> >
> > SROA did the opposite thing in this case, replacing individual stores
> > insertvalues and a store of an FCA. I'm attaching the full output of
> > opt -print-after-all -O2 -S for the failing test case given in the
> > rust issue. The first SROA creates the FCA store, and later when the
> > call to `unwrap` is inlined the existing optimizations can't eliminate
> > the branch that comes with it, because of the FCA store.
> 
> 
> SROA doesn't create FCA stores in any cases I'm aware of... Can you provide
> a somewhat reduced test case rather than a massive log?

I got myself confused because it introduced the insertvalue
instructions. The FCA store was already present and not constructed by
SROA. I've attach a small test case this time.

Björn