On 05/08/2017 12:44 PM, Richard Biener wrote:
On Wed, 26 Apr 2017, Peter Bergner wrote:
One difference from the last patch is that I am no longer setting
default_label to NULL when we emit a decision tree. I noticed that
the decision tree code seemed to generate slightly better code for
some of my unit tests if I left it alone. This simplified the
patch somewhat by removing the changes to emit_case_nodes().
[snip]
Can you do the gimple_unreachable_bb_p check earlier in
expand_case so it covers the emit_case_decision_tree path as well
(and verify that works, of course)? So basically right at
/* Find the default case target label. */
default_label = jump_target_rtx
(CASE_LABEL (gimple_switch_default_label (stmt)));
edge default_edge = EDGE_SUCC (bb, 0);
int default_prob = default_edge->probability;
handle this case.
That is what the previous patch did, but as I mention above,
we generate slightly better code for some test cases (other
tests seemed to generate the same code) if we don't attempt
to handle the decision tree case. I'll note that the current
unpatched compiler already knows how to remove unreachable
case statement blocks when we expand to a decision tree.
I can add that code back if you think that it will have a
positive benefit for some test case I haven't tried yet.
Any comment on the above?
Peter
