Request for reviews (L): 7008866: Missing loop predicate for loop with multiple entries
vladimir.kozlov at oracle.com
Mon Mar 14 14:54:12 PDT 2011
It is the first of coming loop opts changes. No benchmarks performance
improvement is expected since main loops in our benchmarks have predicates already.
Fixed 7008866: Missing loop predicate for loop with multiple entries
Synopsis is a little misleading since I don't mean irreducible loop entries
(new code has a check to not add predicates to irreducible loops). I mean there
could be several entry and backbranch paths to loop head.
Currently loop predicates are generated when each backbranch path is parsed.
Later in IdealLoopTree::beautify_loops() a merge region will be created if there
are several entry paths separating predicates from the loop. As result those
predicates will be removed.
Add predicates when loop head bytecode is parsed instead of when back branch
bytecode is parsed. All fall in paths will be merged already at this point and
for back branches a new merge region is created.
Add missing loop predicate when there are back branches to the method entry
Add missing predicates for loops generated in Ideal graph, for example, for
Rearrange some code to help next loop optimization changes. For example, in
PhaseIdealLoop::is_counted_loop() move bailout checks before new ideal nodes are
generated. Add new counted loop verification code.
Perform local Ideal transformation for parallel induction variables replacement
since in most cases they use the same stride (ratio == 1).
In PhaseIdealLoop::split_thru_phi() don't move split node (or existing node
returned by Identity()) inside a loop if all its uses are outside of this loop.
Otherwise the node will be cloned for each outside use and this will mess up an
external canonical counted loop if the node is loop's increment, for example.
Preserve unswitch count when partial peel is executed and when a loop is
converted to counted.
Add debug flag TraceLoopOpts to show executed loop optimizations and also print
whole loop tree at the beginning of each iteration of loop optimizations:
PartialPeel Loop: N22966/N22965
Tested: ctw, nsk, jprt
Ran refworkload on x86 - nor regression (or improvement).
More information about the hotspot-compiler-dev