[PATCH] Switch lowering: extract jump tables and bit tests before building binary tree (PR22262)

[Daniel Jasper]

================
Comment at: lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp:7430
@@ +7429,3 @@
+
+void SelectionDAGBuilder::findBitTestClusters(CaseClusterVector &Clusters,
+                                              const SwitchInst *SI) {
----------------
djasper wrote:
> With the new structure, I'd probably approach this entirely differently. IIUC, bit tests are really good to handle many (potentially relatively spares) case labels leading to the same destination.
> 
> Now, to make use of this, I'd first group all case clusters (maybe even before rangeifying them) and group them by destination. Now, start with the biggest group (possibly biggest weight by taking branch probabilities into account) and look at whether they can be lowered as bit tests. Thereby, it is possible to peel away the most bit-test-worthy groups (grouped by common destination) one at a time. I haven't completely thought about how to do this and there are a few additional things to consider:
> 
> - The range of a group might be too large (greater than the bit width). However, it might be possible to split that one group into multiple sub groups.
> - It might be possible to lower multiple groups (Group1 leading to Dest1, Group2 leading to Dest2, ..) as a single bit test to conserve extra shifts (if the entire range fits into the bit width). This is currently (and in the previous implementation) sort of done by saying that there can be up to three unique destinations, but I don't that this is actually a good heuristic.
> - At some point there will be a cut-off, i.e. lowering a group as bit test is not worth it. Trivially, single-destination groups are never worth to be lowered as bit tests (a single cmp is faster than and+cmp). But as we might be able to lower the others as jump tables, the cut-off might be higher.
> 
> Lets chat some more about this.
"relatively sparse" was what I meant to say.

http://reviews.llvm.org/D8649

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/