Improved phi reachability computation a bit.

The use-def relation between phis is mainly "forwards" (i.e. from phis with
smaller IDs to ones with larger IDs), so the fixed point computation terminates
faster when iterate through the phis in a "backwards" manner. This is quite
visible in complex Mandreel-generated code, where a few hundred phis with
non-trivial use-def chains are generated.
Review URL: http://codereview.chromium.org/7848012

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@9185 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

diff --git a/src/hydrogen.cc b/src/hydrogen.cc
index 01046bf9bfe52438cfe2f7c65dbc6469a0d92b29..237352cd1ec339e3aa1d618fd1207b5c64d1a78f 100644 (file)
--- a/src/hydrogen.cc
+++ b/src/hydrogen.cc
@@ -1675,7 +1675,9 @@ void HInferRepresentation::Analyze() {
   bool change = true;
   while (change) {
     change = false;
-    for (int i = 0; i < phi_count; ++i) {
+    // We normally have far more "forward edges" than "backward edges",
+    // so we terminate faster when we walk backwards.
+    for (int i = phi_count - 1; i >= 0; --i) {
       HPhi* phi = phi_list->at(i);
       for (HUseIterator it(phi->uses()); !it.Done(); it.Advance()) {
         HValue* use = it.value();