Small load hoist bugfix

2 files changed, 10 insertions, 11 deletions

diff --git a/chapter31/src/main/java/com/seaofnodes/simple/node/LoadNode.java b/chapter31/src/main/java/com/seaofnodes/simple/node/LoadNode.java
index eb78ed1f..35db4c2f 100644
--- a/chapter31/src/main/java/com/seaofnodes/simple/node/LoadNode.java
+++ b/chapter31/src/main/java/com/seaofnodes/simple/node/LoadNode.java
@@ -364,15 +364,13 @@ public class LoadNode extends MemOpNode {
         if( ptr instanceof PhiNode pphi && pphi.region() == r )
             return true;
 
-        // No, so can we lift this ptr?
-        CFGNode cptr = ptr.cfg0();
-        if( cptr != null )
-            // Pointer is controlled high
-            // TODO: Really needs to be the LCA of all inputs is high
-            return cptr.idepth() <= r.idepth();
-
-        // Dunno without a longer walk
-        return false;
+        // Each copied Load needs its pointer before the Region, including
+        // a loop's entry arm. input 0 alone misses floating Loads whose own
+        // pointer was produced inside the loop.
+        CFGNode pinned = ptr.cfg0();
+        if( pinned==null || pinned.idepth()>r.idepth() ) return false;
+        CFGNode cptr = ptr instanceof PhiNode phi ? phi.region() : CFGNode.earlyCFG(ptr);
+        return cptr==null || cptr.dominates(r.idom(memphi),memphi);
     }
 
     @Override public Parser.ParseException err() {
diff --git a/chapter31/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java b/chapter31/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
index 3d71b4ea..673339a9 100644
--- a/chapter31/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
+++ b/chapter31/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
@@ -370,8 +370,9 @@ public class x86_64_v2 extends Machine {
 
     private Node _lea(Node add, Node base, Node idx, int off) {
         int scale = 0;
-        if( base instanceof ShlNode && !(idx instanceof ShlNode) )
-            throw Utils.TODO(); // Bug in canonicalization, should on RHS
+        if( base instanceof ShlNode && !(idx instanceof ShlNode) ) {
+            Node tmp = base; base = idx; idx = tmp;
+        }
         if( idx instanceof ShlNode shift && shift.in(2) instanceof ConstantNode shfcon &&
             shfcon._con instanceof TypeInteger tscale && 0 <= tscale.value() && tscale.value() <= 3 ) {
             idx = shift.in(1);