From 3df967056bf48bd9f87900f3a1ad9af2aedce948 Mon Sep 17 00:00:00 2001 From: Cliff Click Date: Wed, 28 Oct 2026 16:14:30 -0700 Subject: Backport some minor speedups --- .../com/seaofnodes/simple/GlobalCodeMotion.java | 55 +++++++++++++++------- .../main/java/com/seaofnodes/simple/IterPeeps.java | 4 +- .../java/com/seaofnodes/simple/node/CFGNode.java | 4 +- 3 files changed, 45 insertions(+), 18 deletions(-) (limited to 'chapter19/src/main/java/com/seaofnodes') diff --git a/chapter19/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java b/chapter19/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java index 9454e968..0f53ce36 100644 --- a/chapter19/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java +++ b/chapter19/src/main/java/com/seaofnodes/simple/GlobalCodeMotion.java @@ -134,24 +134,40 @@ public abstract class GlobalCodeMotion { } // ------------------------------------------------------------------------ + // Scheduling is not recursive. Retain scratch capacity between schedules, + // but release graph references and reset only the entries we used. + private static CFGNode[] LATE = new CFGNode[1]; + // Load NIDs at their CFG block choices, checked against Store blocks. + private static int[] ANTI = new int[1]; + private static final ArrayList TOUCHED = new ArrayList<>(); + private static final ArrayList PLACED = new ArrayList<>(); + private static final WorkList WORK = new WorkList<>(); + private static void schedLate( CodeGen code ) { - CFGNode[] late = new CFGNode[code.UID()]; - Node[] ns = new Node[code.UID()]; - // Record Load NIDs at all their CFG block choices, then check against - // Store block choices to force a Load above an anti-dependent Store. - int[] anti = new int[code.UID()]; + assert TOUCHED.isEmpty() && PLACED.isEmpty() && WORK.isEmpty(); + if( code.UID() > LATE.length ) { + int len = Math.max(code.UID(),LATE.length<<1); + LATE = new CFGNode[len]; + ANTI = new int[len]; + } + WORK.clear(); // Breadth-first scheduling - breadth(code._stop,ns,late,anti); + // Early scheduling has released the shared visit bits. + breadth(code._stop,PLACED,LATE,ANTI,code.visit()); // Copy the best placement choice into the control slot - for( int i=0; i placed, CFGNode[] late, int[] anti, BitSet visit) { // Things on the worklist have some (but perhaps not all) uses done. - WorkList work = new WorkList<>(); + WorkList work = WORK; work.push(stop); Node n; outer: @@ -175,16 +191,19 @@ public abstract class GlobalCodeMotion { continue outer; // All uses done, schedule - _doSchedLate(n,ns,late,anti); + _doSchedLate(n,placed,late,anti); } + TOUCHED.add(n); + // A use just finished; reconsider its inputs and waiting loads, // even when the shared memory input was already scheduled. for( Node def : n._inputs ) { if( def==null ) continue; if( late[def._nid]==null ) work.push(def); - wakeLoads(def,late,work,new BitSet()); + wakeLoads(def,late,work,visit); } + visit.clear(); if( n instanceof LoopNode loop ) for( Node phi : loop._outputs ) if( phi instanceof PhiNode && late[phi._nid]==null ) @@ -192,7 +211,7 @@ public abstract class GlobalCodeMotion { } } - private static void _doSchedLate(Node n, Node[] ns, CFGNode[] late, int[] anti) { + private static void _doSchedLate(Node n, ArrayList placed, CFGNode[] late, int[] anti) { // Walk uses, gathering the LCA (Least Common Ancestor) of uses CFGNode early = n.in(0) instanceof CFGNode cfg ? cfg : n.in(0).cfg0(); assert early != null; @@ -214,7 +233,7 @@ public abstract class GlobalCodeMotion { if( better(lca,best) ) best = lca; assert !(best instanceof IfNode); - ns [n._nid] = n; + placed.add(n); late[n._nid] = best; } @@ -255,7 +274,8 @@ public abstract class GlobalCodeMotion { private static CFGNode find_anti_dep(CFGNode lca, MemOpNode load, CFGNode early, CFGNode[] late, int[] anti) { // We could skip final-field loads here. // Walk LCA->early, flagging Load's block location choices - for( CFGNode cfg=lca; early!=null && cfg!=early.idom(); cfg = cfg.idom() ) + CFGNode marked = lca; + for( CFGNode cfg = marked; early!=null && cfg!=early.idom(); cfg = cfg.idom() ) anti[cfg._nid] = load._nid; // Walk load->mem uses, looking for Stores causing an anti-dep for( Node mem : load.antiDeps() ) { @@ -283,6 +303,9 @@ public abstract class GlobalCodeMotion { default: throw Utils.TODO(); } } + // Release this load's marks before another compilation reuses IDs. + for( CFGNode cfg = marked; early!=null && cfg!=early.idom(); cfg = cfg.idom() ) + anti[cfg._nid] = 0; return lca; } diff --git a/chapter19/src/main/java/com/seaofnodes/simple/IterPeeps.java b/chapter19/src/main/java/com/seaofnodes/simple/IterPeeps.java index 087a990b..14a7cb67 100644 --- a/chapter19/src/main/java/com/seaofnodes/simple/IterPeeps.java +++ b/chapter19/src/main/java/com/seaofnodes/simple/IterPeeps.java @@ -177,7 +177,7 @@ public class IterPeeps { * True if Node is on the WorkList */ boolean on( E x ) { return _on.get(x._nid); } - boolean isEmpty() { return _len==0; } + public boolean isEmpty() { return _len==0; } /** * Removes a random Node from the WorkList; null if WorkList is empty @@ -187,11 +187,13 @@ public class IterPeeps { int idx = _R.nextInt(_len); E x = (E)_es[idx]; _es[idx] = _es[--_len]; // Compress array + _es[_len] = null; _on.clear(x._nid); return x; } public void clear() { + Arrays.fill(_es,0,_len,null); _len = 0; _on.clear(); _R.setSeed(_seed); diff --git a/chapter19/src/main/java/com/seaofnodes/simple/node/CFGNode.java b/chapter19/src/main/java/com/seaofnodes/simple/node/CFGNode.java index 383f18d0..d7b2f362 100644 --- a/chapter19/src/main/java/com/seaofnodes/simple/node/CFGNode.java +++ b/chapter19/src/main/java/com/seaofnodes/simple/node/CFGNode.java @@ -109,7 +109,9 @@ public abstract class CFGNode extends Node { // Start contains the function bodies at depth 0; real loops add depth. public void buildLoopTree(StopNode stop) { _ltree = stop._ltree = Parser.XCTRL._ltree = new LoopTree((StartNode)this); - _bltWalk(2,null,stop, new BitSet()); + BitSet post = CodeGen.CODE.visit(); + _bltWalk(2,null,stop, post); + post.clear(); } int _bltWalk( int pre, FunNode fun, StopNode stop, BitSet post ) { // Pre-walked? -- cgit v1.2.3