Give a small negative bias to giant edge bundles.

This helps compile time when the greedy register allocator splits live
ranges in giant functions. Without the bias, we would try to grow
regions through the giant edge bundles, usually to find out that the
region became too big and expensive.

If a live range has many uses in blocks near the giant bundle, the small
negative bias doesn't make a big difference, and we still consider
regions including the giant edge bundle.

Giant edge bundles are usually connected to landing pads or indirect
branches.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@157174 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 11 insertions, 0 deletions

diff --git a/lib/CodeGen/SpillPlacement.cpp b/lib/CodeGen/SpillPlacement.cpp
index 6f33f5465c..320128a999 100644
--- a/lib/CodeGen/SpillPlacement.cpp
+++ b/lib/CodeGen/SpillPlacement.cpp
@@ -207,6 +207,17 @@ void SpillPlacement::activate(unsigned n) {
     return;
   ActiveNodes->set(n);
   nodes[n].clear();
+
+  // Very large bundles usually come from big switches, indirect branches,
+  // landing pads, or loops with many 'continue' statements. It is difficult to
+  // allocate registers when so many different blocks are involved.
+  //
+  // Give a small negative bias to large bundles such that 1/32 of the
+  // connected blocks need to be interested before we consider expanding the
+  // region through the bundle. This helps compile time by limiting the number
+  // of blocks visited and the number of links in the Hopfield network.
+  if (bundles->getBlocks(n).size() > 100)
+    nodes[n].Bias = -0.0625f;
 }