Cleanup & backportclinit

The phase rework, made new phases explicit and add them to the CodeGen driver.  Fix the flag-spilling on X86 and ARM.  Fix the GCM conservative anti-deps and backport.

2 files changed, 28 insertions, 4 deletions

diff --git a/chapter15/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter15/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
index d17ec3b6..4c9fc779 100644
--- a/chapter15/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
+++ b/chapter15/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java
@@ -246,10 +246,11 @@ public abstract class GlobalCodeMotion {
                 lca = anti_dep(load,late[st._nid],st.cfg0(),lca,st,anti);
                 break;
             case PhiNode phi:
-                // Repeat anti-dep for matching Phi inputs.
-                // No anti-dep edges but may raise the LCA.
+                // Inputs carrying different memory might overwrite this alias.
+                // Raise the LCA for those paths; unchanged memory does not
+                // clobber the load. No scheduling edge to the Phi.
                 for( int i=1; i<phi.nIns(); i++ )
-                    if( phi.in(i)==load.mem() )
+                    if( phi.in(i)!=load.mem() )
                         lca = anti_dep(load,phi.region().cfg(i),load.mem().cfg0(),lca,null,anti);
                 break;
             default: throw Utils.TODO();
@@ -262,7 +263,9 @@ public abstract class GlobalCodeMotion {
     private static CFGNode anti_dep( LoadNode load, CFGNode stblk, CFGNode defblk, CFGNode lca, Node st, int[] anti ) {
         // Preserve the full store range for the earlier evaluator scheduler.
         // It places nodes independently of GCM and may hoist this store.
-        for( ; stblk != defblk.idom(); stblk = stblk.idom() ) {
+        // Another Phi input need not be dominated by this memory definition.
+        // Include its depth, and retain the sentinel for an entry's null idom.
+        for( ; stblk != defblk.idom() && stblk.idepth() >= defblk.idepth(); stblk = stblk.idom() ) {
             // Store and Load overlap, need anti-dependence
             if( anti[stblk._nid]==load._nid ) {
                 lca = stblk.domLCA(lca,null); // Raise Loads LCA
diff --git a/chapter15/src/test/java/com/seaofnodes/simple/SchedulerTest.java b/chapter15/src/test/java/com/seaofnodes/simple/SchedulerTest.java
index ad595822..55aa5d2b 100644
--- a/chapter15/src/test/java/com/seaofnodes/simple/SchedulerTest.java
+++ b/chapter15/src/test/java/com/seaofnodes/simple/SchedulerTest.java
@@ -10,6 +10,27 @@ import static org.junit.Assert.assertEquals;
 
 public class SchedulerTest {
 
+    @Test public void testMemoryPhiAcrossSiblingBranch() {
+        // The inner memory Phi defines the load's memory. The outer Phi's
+        // other input comes from a branch not dominated by that definition.
+        String src = """
+            struct S { int x; }; S a=new S; S b=new S;
+            a.x=10; b.x=20; S p=a; if(arg>3) p=b;
+            int result=0;
+            if(arg>0) {
+                if(arg>2) p.x=30; else p.x=40;
+                result=a.x;
+            }
+            return result;
+            """;
+        StopNode stop = new Parser(src).parse().iterate();
+        var eval = new Evaluator(stop);
+        long[] expected = {0,40,40,30,10,10};
+        for( int arg=0; arg<expected.length; arg++ )
+            assertEquals(expected[arg],eval.evaluate(arg,100));
+    }
+
+
     private static void assertObj(Object obj, String name, Object... fields) {
         assertEquals(Evaluator.Obj.class, obj.getClass());
         var o = (Evaluator.Obj)obj;