<html>
    <head>
      <base href="https://bugs.freedesktop.org/" />
    </head>
    <body>
      <p>
        <div>
            <b><a class="bz_bug_link 
          bz_status_NEW "
   title="NEW - PhiMovesPass in register allocator broken"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=90887#c14">Comment # 14</a>
              on <a class="bz_bug_link 
          bz_status_NEW "
   title="NEW - PhiMovesPass in register allocator broken"
   href="https://bugs.freedesktop.org/show_bug.cgi?id=90887">bug 90887</a>
              from <span class="vcard"><a class="email" href="mailto:imirkin@alum.mit.edu" title="Ilia Mirkin <imirkin@alum.mit.edu>"> <span class="fn">Ilia Mirkin</span></a>
</span></b>
        <pre>OK, so among other changes, commit e43a3a66 had the following diff:

@@ -312,23 +337,26 @@ RegAlloc::PhiMovesPass::visit(BasicBlock *bb)
    Instruction *phi, *mov;
    BasicBlock *pb, *pn;

+   std::stack<BasicBlock *> stack;
+
    for (Graph::EdgeIterator ei = bb->cfg.incident(); !ei.end(); ei.next()) {
-      pb = pn = BasicBlock::get(ei.getNode());
+      pb = BasicBlock::get(ei.getNode());
       assert(pb);
-
-      if (needNewElseBlock(bb, pb)) {
-         pn = new BasicBlock(func);
-
-         // deletes an edge, iterator is invalid after this:
-         pb->cfg.detach(&bb->cfg);
-         pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
-         pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD); // XXX: check order !
-
-         assert(pb->getExit()->op != OP_CALL);
-         if (pb->getExit()->asFlow()->target.bb == bb)
-            pb->getExit()->asFlow()->target.bb = pn;
-         break;
-      }
+      if (needNewElseBlock(bb, pb))
+         stack.push(pb);
+   }
+   while (!stack.empty()) {
+      pb = stack.top();
+      pn = new BasicBlock(func);
+      stack.pop();
+
+      pb->cfg.detach(&bb->cfg);
+      pb->cfg.attach(&pn->cfg, Graph::Edge::TREE);
+      pn->cfg.attach(&bb->cfg, Graph::Edge::FORWARD);
+
+      assert(pb->getExit()->op != OP_CALL);
+      if (pb->getExit()->asFlow()->target.bb == bb)
+         pb->getExit()->asFlow()->target.bb = pn;
    }

    // insert MOVs (phi->src(j) should stem from j-th in-BB)

which the patch in <span class=""><a href="attachment.cgi?id=116339" name="attach_116339" title="Candidate Fix for nv50 PhiMovesPass">attachment 116339</a> <a href="attachment.cgi?id=116339&action=edit" title="Candidate Fix for nv50 PhiMovesPass">[details]</a></span> effectively undoes. [Why was that in
there? No idea. But presumably futzing with bb's affects needNewElseBlock's
decisions.] So I'm still tending towards my solution. But I'm not particularly
happy that it has to build up and use that hash table every time even if
nothing's changed.</pre>
        </div>
      </p>
      <hr>
      <span>You are receiving this mail because:</span>
      
      <ul>
          <li>You are the QA Contact for the bug.</li>
          <li>You are the assignee for the bug.</li>
      </ul>
    </body>
</html>