[Turbofan] Fix perf regression introduced by per-range change.

When the range ends just at the gap of a non-deferred block, the last
instruction the range covers is in the predecessor. If that predecessor is
a deferred block, before this CL, we would splinter the remainder of the
range all the way to the end. That leads to inefficient codegen, because
we still want a split inside the deferred block.

Also, opportunistically added a trace before we splinter, for better
diagnostics.

BUG= chromium:546416
LOG=N

Review URL: https://codereview.chromium.org/1412123009

Cr-Commit-Position: refs/heads/master@{#31529}

src/compiler/live-range-separator.cc

@@ -51,6 +51,8 @@ void CreateSplinter(TopLevelLiveRange *range, RegisterAllocationData *data,
       range->SetSplinter(splinter);
     }
     Zone *zone = data->allocation_zone();
+    TRACE("creating splinter for range %d between %d and %d\n", range->vreg(),
+          start.ToInstructionIndex(), end.ToInstructionIndex());
     range->Splinter(start, end, zone);
   }
 }
@@ -112,8 +114,9 @@ void SplinterLiveRange(TopLevelLiveRange *range, RegisterAllocationData *data) {
     interval = next_interval;
   }
   // When the range ends in deferred blocks, first_cut will be valid here.
+  // Splinter from there to the last instruction that was in a deferred block.
   if (first_cut.IsValid()) {
-    CreateSplinter(range, data, first_cut, range->End());
+    CreateSplinter(range, data, first_cut, last_cut);
   }
 }
 }  // namespace