[PATCH] D28117: [NewGVN] Merge conditional branches before value numbering

[Davide Italiano via Phabricator via llvm-commits]

davide created this revision.
davide added a reviewer: dberlin.
davide added a subscriber: llvm-commits.

We're discussing it in https://llvm.org/bugs/show_bug.cgi?id=31468
Patch on phab for easier reviews.


https://reviews.llvm.org/D28117

Files:
  lib/Transforms/Scalar/NewGVN.cpp
  test/Transforms/NewGVN/basic.ll


Index: test/Transforms/NewGVN/basic.ll
===================================================================
--- test/Transforms/NewGVN/basic.ll
+++ test/Transforms/NewGVN/basic.ll
@@ -1,4 +1,3 @@
-; XFAIL: *
 ; RUN: opt < %s -newgvn -S | FileCheck %s

 define i32 @main() {
Index: lib/Transforms/Scalar/NewGVN.cpp
===================================================================
--- lib/Transforms/Scalar/NewGVN.cpp
+++ lib/Transforms/Scalar/NewGVN.cpp
@@ -1378,6 +1378,20 @@
   unsigned ICount = 0;
   SmallPtrSet<BasicBlock *, 16> VisitedBlocks;

+  // Do a sweep over all the basic blocks to merge unconditional branches.
+  for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE;) {
+    BasicBlock *BB = &*FI++;
+
+    // In theory here we could pass MemSSA if/when MergeBlockIntoPredecessor
+    // will grow version which accepts the analysis.
+    bool removedBlock = MergeBlockIntoPredecessor(
+        BB, DT, nullptr /* LoopInfo */, nullptr /* MemDep */);
+
+    if (removedBlock)
+      NumGVNBlocksDeleted++;
+    Changed |= removedBlock;
+  }
+
   // Note: We want RPO traversal of the blocks, which is not quite the same as
   // dominator tree order, particularly with regard whether backedges get
   // visited first or second, given a block with multiple successors.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D28117.82498.patch
Type: text/x-patch
Size: 1295 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161226/6a9ff93f/attachment.bin>