On 12/21/2015 06:13 AM, Alan Lawrence wrote:
This is a respin of patches
https://gcc.gnu.org/ml/gcc-patches/2015-10/msg03266.html and
https://gcc.gnu.org/ml/gcc-patches/2015-10/msg03267.html, which were
"too quickly" approved before concerns with efficiency were pointed out.
I tried to change the hashing just in tree-ssa-dom.c using C++ subclassing, but
couldn't cleanly separate this out from tree-ssa-scopedtables and
tree-ssa-threadedge.c due to use of avail_exprs_stack. So I figured it was
probably appropriate to use the equivalences in jump threading too. Also,
using get_ref_base_and_extent unifies handling of MEM_REFs and ARRAY_REFs
(hence only one patch rather than two).
It is appropriate.
I've added a couple of testcases that show the improvement in DOM, but in all
cases I had to disable FRE, even PRE, to get any improvement, apart from on
ssa-dom-cse-2.c itself (where on the affected platforms FRE still does not do
the optimization). This makes me wonder if this is the right approach or whether
changing the references output by SRA (as per
https://gcc.gnu.org/ml/gcc-patches/2015-08/msg01490.html , judged as a hack to
SRA to work around limitations in DOM - or is it?) would be better.
I just doubt it happens all that much.
Jeff
