[PATCH] [SDAG] Introduce a combined set to the DAG combiner which tracks nodes which have successfully round-tripped through the combine phase, and use this to ensure all operands to DAG nodes are visited by the combiner, even if they are only added...

Chandler Carruth chandlerc at gmail.com
Wed Jul 23 04:09:39 PDT 2014


A couple of random other notes about this patch....

- I did benchmark the compile time impact. For a "worst case" scenario in
my mind of taking a single bitcode for all of clang and running it through
just llc I observed a 3% slowdown of llc with this patch. I expect that to
translate into more like 1% or 2% of slowdown for LTO, and well under that
for normal compiles (which have the frontend in the critical path). So the
runtime cost isn't great, but isn't terribly scary to me.

- An obvious question is, why still populate the worklist with all the
nodes? The answer is because just starting from the root and letting the
de-dup'ed operand walk handle the rest causes more tests to fail, and I'm
already struggling to drag the test suite kicking and screaming through
these changes. It still might be nice eventually so that we move to a more
principled graph-traversal strategy. It may recover a (very minor) fraction
of the runtime cost, but nothing huge.

- With this functionality added, I can add legalization support to the last
DAG combine phase and successfully legalize newly created illegal nodes as
we effectively traverse the new nodes. I've successfully implemented the
PSHUFB matching logic on top of this. So this isn't just idle hardening of
the DAG combiner.

http://reviews.llvm.org/D4638






More information about the llvm-commits mailing list