changed relo ABI

Missing patch/merge from ARM ldst-encoding
Explicit external calls.
Better printing  for multi-op template encodings

20 files changed, 115 insertions, 78 deletions

diff --git a/chapter21/src/main/java/com/seaofnodes/simple/SB.java b/chapter21/src/main/java/com/seaofnodes/simple/SB.java
index 31fbb7d4..af49e52b 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/SB.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/SB.java
@@ -54,6 +54,7 @@ public final class SB {
       _sb.append( i < s.length() ? s.charAt(i) : ' ');
     return this;
   }
+  public char at(int idx ) { return _sb.charAt(idx); }
 
   // Not spelled "p" on purpose: too easy to accidentally say "p(1.0)" and
   // suddenly call the autoboxed version.
@@ -93,6 +94,8 @@ public final class SB {
   //
   public SB unchar() { return unchar(1); }
   public SB unchar(int x) { _sb.setLength(_sb.length()-x); return this; }
+  public SB setLen(int len) { _sb.setLength(len); return this; }
+  public String subString(int start, int end ) { return _sb.substring(start,end); }
 
   public SB clear() { _sb.setLength(0); return this; }
   public int len() { return _sb.length(); }
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/codegen/ElfFile.java b/chapter21/src/main/java/com/seaofnodes/simple/codegen/ElfFile.java
index c01eadd1..c39c5dc2 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/codegen/ElfFile.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/codegen/ElfFile.java
@@ -284,10 +284,10 @@ public class ElfFile {
 
         // create .text relocations
         DataSection text_rela = new DataSection(".rela.text", 4 /* SHT_RELA */);
-        for( Node n : _code._encoding._externals) {
+        for( Node n : _code._encoding._externals.keySet()) {
             int nid    = n._nid;
-            TypeFunPtr tfp = n instanceof CallNode call ? call.tfp() : (TypeFunPtr)n._type;
-            int sym_id = _funcs.get(tfp)._index;
+            String extern = _code._encoding._externals.get(n);
+            int sym_id = _funcs.get(extern)._index;
             int offset = _code._encoding._opStart[nid] + _code._encoding._opLen[nid] - 4;
 
             // u64 offset
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/codegen/Encoding.java b/chapter21/src/main/java/com/seaofnodes/simple/codegen/Encoding.java
index 7cdbf1c8..56e5960a 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/codegen/Encoding.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/codegen/Encoding.java
@@ -82,17 +82,31 @@ public class Encoding {
 
 
     // Nodes need "relocation" patching; things done after code is placed.
-    private final Ary<Node> _internals = new Ary<>(Node.class);
-    public  final Ary<Node> _externals = new Ary<>(Node.class);
+    // Record src and dst Nodes.
+    private final HashMap<Node,CFGNode> _internals = new HashMap<>();
+    // Source is a Call, destination in the Fun.
     public Encoding relo( CallNode call ) {
-        (call.external() ? _externals : _internals).add(call);
+        _internals.put(call,_code.link(call.tfp()));
         return this;
     }
-    public Encoding reloTFP( ConstantNode con ) { _internals.add(con); return this; }
+    public Encoding relo( ConstantNode con ) {
+        TypeFunPtr tfp = (TypeFunPtr)con._con;
+        _internals.put(con,_code.link(tfp));
+        return this;
+    }
+    public void jump( CFGNode jmp, CFGNode dst ) {
+        while( dst.nOuts() == 1 ) // Skip empty blocks
+            dst = dst.uctrl();
+        _internals.put(jmp,dst);
+    }
 
-    public void relo( NewNode nnn ) {
-        // TODO: record alloc relocation info
+
+    final HashMap<Node,String> _externals = new HashMap<>();
+    public Encoding external( Node call, String extern ) {
+        _externals.put(call,extern);
+        return this;
     }
+
     // Store t as a 32/64 bit constant in the code space; generate RIP-relative
     // addressing to load it
 
@@ -103,10 +117,6 @@ public class Encoding {
         // TODO:
     }
 
-    // Local relocation info for patching local jumps, once targets are known.
-    private final HashMap<CFGNode,CFGNode> _jmps = new HashMap<>();
-    public void jump( CFGNode jmp, CFGNode target ) { _jmps.put(jmp,target); }
-
     void encode() {
         // Basic block layout: invert branches to keep blocks in-order; insert
         // unconditional jumps.  Attempt to keep backwards branches taken,
@@ -286,27 +296,11 @@ public class Encoding {
 
     // Patch local encodings now
     private void patchLocalRelocations() {
-        // Walk all the jumps.  Re-patch them all now with but with the Real Offset
-        for( CFGNode jmp : _jmps.keySet() ) {
-            CFGNode target = jmp instanceof IfNode iff ? iff.cproj(0) : jmp.uctrl();
-            while( target.nOuts() == 1 ) // Skip empty blocks
-                target = target.uctrl();
-            int start = _opStart[jmp._nid];
-            ((RIPRelSize)jmp).patch(this, start, _opLen[jmp._nid], _opStart[target._nid] - start);
-        }
-
-
         // Walk the local code-address relocations
-        for( Node n : _internals ) {
-            TypeFunPtr tfp = n instanceof CallNode call ? call.tfp() : (TypeFunPtr)n._type;
-            FunNode fun = _code.link(tfp);
-            int start = _opStart[n._nid];
-            ((RIPRelSize)n).patch(this, start, _opLen[n._nid], _opStart[fun._nid] - start);
-        }
-
-        // Walk the external call relocations
-        for( Node call : _externals ) {
-          throw Utils.TODO();
+        for( Node src : _internals.keySet() ) {
+            Node dst = _internals.get(src);
+            int start = _opStart[src._nid];
+            ((RIPRelSize)src).patch(this, start, _opLen[src._nid], _opStart[dst._nid] - start);
         }
     }
 
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/codegen/Machine.java b/chapter21/src/main/java/com/seaofnodes/simple/codegen/Machine.java
index 13d7abc1..fbc3f6c2 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/codegen/Machine.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/codegen/Machine.java
@@ -7,6 +7,8 @@ abstract public class Machine {
     public abstract String name();
     // Human-readable name for a register number, e.g. "RAX" or "R0"
     public abstract String reg( int reg );
+    // Default opcode size for printing; 4 bytes for most 32-bit risc chips
+    public int defaultOpSize() { return 4; }
     // Create a split op; any register to any register, including stack slots
     public abstract SplitNode split( String kind, byte round, LRG lrg);
     // List of caller-save registers
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/MachNode.java b/chapter21/src/main/java/com/seaofnodes/simple/node/MachNode.java
index 5486b1c1..3fc6e031 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/MachNode.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/MachNode.java
@@ -52,6 +52,9 @@ public interface MachNode {
     // Just something like "R17=[R18+12] // Load array base".
     // General form: "dst=src+src".
     default void asm(CodeGen code, SB sb) { throw Utils.TODO(); }
+    // Multi-op encodings, printing 1 op per line.  Template-style,
+    // no RA, no scheduling.
+    default void asm(CodeGen code, SB sb, int multiN) {  }
 
     // Human readable opcode; e.g. "ld4" or "call".
     default String op() { throw Utils.TODO(); }
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/BranchARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/BranchARM.java
index b3384f21..c9f47ba5 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/BranchARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/BranchARM.java
@@ -30,12 +30,11 @@ public class BranchARM extends IfNode implements MachNode, RIPRelSize {
 
     // Encoding is appended into the byte array; size is returned
     @Override public void encoding( Encoding enc ) {
-        enc.jump(this,cproj(0));
         // Assuming that condition flags are already set.  These flags are set
         // by comparison (or sub).  No need for regs because it uses flags
+        enc.jump(this,cproj(0));
         // B.cond
-        int body = arm.b_cond(0b01010100, 0, arm.make_condition(_bop));
-        enc.add4(body);
+        enc.add4( arm.b_cond(0b01010100, 0, arm.make_condition(_bop)) );
     }
 
     // Delta is from opcode start
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/LoadARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/LoadARM.java
index f09d208a..800ad34c 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/LoadARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/LoadARM.java
@@ -18,7 +18,7 @@ public class LoadARM extends MemOpARM{
     @Override public String op() { return "ld"+_sz; }
     @Override public RegMask outregmap() { return arm.MEM_MASK; }
     // ldr(immediate - unsigned offset) | ldr(register)
-    @Override public void encoding( Encoding enc ) { arm.ldst_encode(enc,0b11111000010, this,this); }
+    @Override public void encoding( Encoding enc ) { ldst_encode(enc,0b11111000010,this); }
     @Override public void asm(CodeGen code, SB sb) {
         sb.p(code.reg(this)).p(",");
         asm_address(code,sb);
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
index 922276d4..0987b7bf 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/MemOpARM.java
@@ -47,6 +47,17 @@ public abstract class MemOpARM extends MemOpNode implements MachNode {
         throw Utils.TODO();
     }
 
+    // Shared encoding for loads and stores
+    public void ldst_encode( Encoding enc, int opcode, Node xval ) {
+        short ptr = enc.reg(ptr());
+        short off = enc.reg(off());
+        short val = enc.reg(xval );
+        int body = off() == null
+            ? arm.load_str_imm(0b1111100101, off, ptr, val)
+            : arm.indr_adr(0b11111000011, off, arm.STORE_LOAD_OPTION.SXTX, 0, ptr, val);
+        enc.add4(body);
+    }
+
     SB asm_address(CodeGen code, SB sb) {
         sb.p("[").p(code.reg(ptr())).p("+");
         if( off() != null ) sb.p(code.reg(off()));
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
index 8ddf1bd6..8a465b66 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NewARM.java
@@ -14,9 +14,8 @@ public class NewARM extends NewNode implements MachNode {
     @Override public RegMask killmap() { return arm.armCallerSave(); }
 
     @Override public void encoding( Encoding enc ) {
-        enc.relo(this);
         // bl
-        enc.add4(arm.b(0b100101, 0));
+        enc.external(this,"calloc").add4(arm.b(0b100101, 0));
     }
 
     @Override public void asm(CodeGen code, SB sb) {
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
index 094348ce..e1fb592c 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/StoreARM.java
@@ -21,7 +21,7 @@ public class StoreARM extends MemOpARM {
     }
     @Override public String op() { return "st"+_sz; }
     @Override public RegMask outregmap() { return null; }
-    @Override public void encoding( Encoding enc ) {   arm.ldst_encode(enc,0b11111000000, this,val()); }
+    @Override public void encoding( Encoding enc ) {  ldst_encode(enc,0b11111000000, val()); }
     @Override public void asm(CodeGen code, SB sb) {
         asm_address(code,sb).p(",");
         if( val()==null ) sb.p("#").p(_imm);
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/TFPARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/TFPARM.java
index 429148d2..0ba47317 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/TFPARM.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/TFPARM.java
@@ -14,7 +14,7 @@ public class TFPARM extends ConstantNode implements MachNode {
     @Override public boolean isClone() { return true; }
     @Override public TFPARM copy() { return new TFPARM(this); }
     @Override public void encoding( Encoding enc ) {
-        enc.reloTFP(this);
+        enc.relo(this);
         short self = enc.reg(this);
         // adrp    x0, 0
         int adrp = arm.adrp(1,0, 0b10000, 0,self);
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
index a3f08730..d90ea15e 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/arm.java
@@ -298,16 +298,6 @@ public class arm extends Machine {
     public static int load_str_imm(int opcode, int imm9, int ptr, int rt) {
         return (opcode << 21) | (imm9 << 12) |(1 << 10) |(ptr << 5) | rt;
     }
-    // Todo: provide op as binary here
-    public static void ldst_encode( Encoding enc, int opcode, MemOpARM n, Node xval ) {
-        short ptr = enc.reg(n.ptr());
-        short off = enc.reg(n.off());
-        short val = enc.reg(xval)   ;
-        int body = n.off() == null
-            ? load_str_imm(0b1111100101, off, ptr, val)
-            : indr_adr(0b11111000011, off, STORE_LOAD_OPTION.SXTX, 0, ptr, val);
-        enc.add4(body);
-    }
 
     // encoding for vcvt, size is encoded in operand
     // <Qd>, <Qm>
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/AUIPC.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/AUIPC.java
index 244e0456..59075190 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/AUIPC.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/AUIPC.java
@@ -15,7 +15,7 @@ public class AUIPC extends ConstantNode implements MachNode, RIPRelSize {
     @Override public AUIPC copy() { return new AUIPC((TypeFunPtr)_con); }
     @Override public String op() { return "auipc"; }
     @Override public void encoding( Encoding enc ) {
-        enc.reloTFP(this);
+        enc.relo(this);
         short dst = enc.reg(this);
         enc.add4(riscv.u_type(0x17, dst, 0));
     }
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
index 2bdfc262..7d43e1ec 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NewRISC.java
@@ -3,20 +3,14 @@ package com.seaofnodes.simple.node.cpus.riscv;
 import com.seaofnodes.simple.*;
 import com.seaofnodes.simple.codegen.*;
 import com.seaofnodes.simple.node.*;
+import com.seaofnodes.simple.print.ASMPrinter;
 
 public class NewRISC extends NewNode implements MachNode {
     // A pre-zeroed chunk of memory.
-    NewRISC( NewNode nnn, AUIPC auipc ) {
-        super(nnn);
-        _inputs.add(auipc);     // Add high-half address
-    }
+    NewRISC( NewNode nnn ) { super(nnn); }
     @Override public String op() { return "alloc"; }
-    @Override public RegMask    regmap(int i) {
-        if( i == 1 ) return riscv.A0_MASK; // Size input
-        // Last call input is AUIPC
-        if( i == nIns()-1 ) return riscv.RMASK;
-        // In-between are mem aliases?
-        return null;
+    @Override public RegMask regmap(int i) {
+        return i==1 ? riscv.A0_MASK : null; // Size input or mem aliases
     }
     @Override public RegMask outregmap(int i) { return i == 1 ? riscv.A0_MASK : null; }
     @Override public RegMask outregmap() { return null; }
@@ -24,15 +18,16 @@ public class NewRISC extends NewNode implements MachNode {
 
     // Encoding is appended into the byte array; size is returned
     @Override public void encoding( Encoding enc ) {
-        enc.relo(this);
-        // High half is where the TFP constant used to be, the last input
-        short auipc = enc.reg(in(nIns()-1));
-        int body = riscv.i_type(0x67, riscv.RPC, 0, auipc, 0);
-        enc.add4(body);
+        // Generic external encoding; 2 ops.
+        enc.external(this,"calloc");
+        // A1 is a caller-save, allowed to crush building external address
+        enc.add4(riscv.u_type(0x17, riscv.A1, 0));
+        enc.add4(riscv.i_type(0x67, riscv.RPC, riscv.I_JALR, riscv.A1, 0));
     }
 
     // General form: "alloc #bytes  PC"
     @Override public void asm(CodeGen code, SB sb) {
-        sb.p(code.reg(size())).p(" ").p(code.reg(in(nIns()-1)));
+        sb.p("auipc a1=#calloc\n");
+        sb.p("call  a1+#calloc, a0");
     }
 }
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/TFPRISC.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/TFPRISC.java
index 4b096e7b..873a07ce 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/TFPRISC.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/TFPRISC.java
@@ -14,7 +14,8 @@ public class TFPRISC extends ConstantNode implements MachNode {
     @Override public boolean isClone() { return true; }
     @Override public TFPRISC copy() { return new TFPRISC(this); }
     @Override public void encoding( Encoding enc ) {
-        enc.reloTFP(this);
+        enc.relo(this);
+        // TODO: 1 op encoding, plus a TODO if it does not fit
         short dst = enc.reg(this);
         TypeFunPtr tfp = (TypeFunPtr)_con;
         // auipc  t0,0
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
index e3cc8724..285b4583 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/riscv.java
@@ -384,7 +384,7 @@ public class riscv extends Machine {
     }
     private Node nnn(NewNode nnn) {
         // TODO: pass in the TFP for alloc
-        return new NewRISC(nnn, new AUIPC(null));
+        return new NewRISC(nnn);
     }
 
     private Node cmp(BoolNode bool) {
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
index 19ebcbcf..c846f341 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/NewX86.java
@@ -17,7 +17,7 @@ public class NewX86 extends NewNode implements MachNode {
 
     // Encoding is appended into the byte array; size is returned
     @Override public void encoding( Encoding enc ) {
-        enc.relo(this);
+        enc.external(this,"calloc");
         // E8 cd    CALL rel32;
         enc.add1(0xE8);
         enc.add4(0);            // offset
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/TFPX86.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/TFPX86.java
index 6131fea6..13b9f8a3 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/TFPX86.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/TFPX86.java
@@ -35,7 +35,7 @@ public class TFPX86 extends ConstantNode implements MachNode {
 
         // lea to load function pointer address
         // lea rax, [rip+disp32]
-        enc.reloTFP(this);
+        enc.relo(this);
         enc.add1(x86_64_v2.rex(dst, 0, 0));
         enc.add1(0x8D); // opcode
         enc.add1(x86_64_v2.modrm(x86_64_v2.MOD.INDIRECT, dst, 0b101));
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
index ee0f89d6..20abb4e8 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java
@@ -10,6 +10,7 @@ import java.io.ByteArrayOutputStream;
 public class x86_64_v2 extends Machine {
     // X86-64 V2.  Includes e.g. SSE4.2 and POPCNT.
     @Override public String name() { return "x86_64_v2"; }
+    @Override public int defaultOpSize() { return 5; }
 
     public static int RAX = 0, RCX = 1, RDX = 2, RBX = 3, RSP = 4, RBP = 5, RSI = 6, RDI = 7;
     public static int R08 = 8, R09 = 9, R10 = 10, R11 = 11, R12 = 12, R13 = 13, R14 = 14, R15 = 15;
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java b/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java
index d511d8ad..9b0074c5 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java
@@ -1,7 +1,8 @@
 package com.seaofnodes.simple.print;
 
-import com.seaofnodes.simple.codegen.CodeGen;
 import com.seaofnodes.simple.SB;
+import com.seaofnodes.simple.Utils;
+import com.seaofnodes.simple.codegen.CodeGen;
 import com.seaofnodes.simple.node.*;
 import com.seaofnodes.simple.type.TypeMem;
 import com.seaofnodes.simple.type.TypeRPC;
@@ -38,10 +39,10 @@ public abstract class ASMPrinter {
         return iadr;
     }
 
-    static private final int encWidth = 10;
     static private final int opWidth = 5;
     static private final int argWidth = 30;
     static int doBlock(int iadr, SB sb, CodeGen code, FunNode fun, int cfgidx) {
+        final int encWidth = code._mach.defaultOpSize()*2;
         CFGNode bb = code._cfg.at(cfgidx);
         if( bb != fun && !(bb instanceof IfNode) && !(bb instanceof CallEndNode) && !(bb instanceof CallNode)  && !(bb instanceof CProjNode && bb.in(0) instanceof CallEndNode ))
             sb.p(label(bb)).p(":").nl();
@@ -84,6 +85,7 @@ public abstract class ASMPrinter {
         if( postAlloc && n instanceof CalleeSaveNode ) return iadr;
         if( postEncode && n instanceof ProjNode ) return iadr;
         if( n instanceof MemMergeNode ) return iadr;
+        final int encWidth = code._mach.defaultOpSize()*2;
 
         // All blocks ending in a Region will need to either fall into or jump
         // to this block.  Until the post-reg-alloc block layout cleanup, we
@@ -127,14 +129,16 @@ public abstract class ASMPrinter {
         sb.p("  ");
 
         // Op; generally "ld4" or "call"
-        sb.fix(opWidth, n instanceof MachNode mach ? mach.op() : n.label());
-        sb.p(" ");
+        sb.fix(opWidth, n instanceof MachNode mach ? mach.op() : n.label()).p(" ");
 
         // General asm args
+        String isMultiOp = null;
         if( n instanceof MachNode mach ) {
             int old = sb.len();
             mach.asm(code,sb);
-            sb.fix(argWidth-(sb.len()-old),""); // Pad out
+            int len = sb.len();
+            isMultiOp = isMultiOp(sb,old,len);
+            sb.fix(argWidth-len,""); // Pad out
 
         } else if( !(n._type instanceof TypeMem) ) {
             // Room for some inputs
@@ -155,8 +159,33 @@ public abstract class ASMPrinter {
 
         sb.nl();
 
+        // Printing more op bits than fit
+        if( isMultiOp != null ) {
+            // Multiple ops, template style, no RA, no scheduling.  Print out
+            // one-line-per-newline, with encoding bits up front.
+            int size = code._encoding._opLen[n._nid];
+            int off = code._mach.defaultOpSize();
+            while( isMultiOp!=null ) {
+                sb.hex2(iadr).p(" ");
+                int len = Math.min(off,size);
+                for( int i=0; i<len; i++ )
+                    sb.hex1(code._encoding._bits.buf()[iadr++]);
+                off += code._mach.defaultOpSize();
+                sb.p("  ");
+                int x = isMultiOp.indexOf('\n');
+                if( x== -1 ) {  // Last line
+                    sb.p(isMultiOp).nl();
+                    isMultiOp = null;
+                } else {
+                    sb.p(isMultiOp.substring(0,x+1)); // Includes the newline
+                    isMultiOp = isMultiOp.substring(x+1);
+                }
+
+            }
 
-        if( fatEncoding > 0 ) {
+        } else if( fatEncoding > 0 ) {
+            // Extra bytes past the default encoding width, all put on a line by
+            // themselves.  X86 special for super long encodings
             sb.hex2(iadr).p(" ");
             for( int i=0; i<fatEncoding; i++ )
                 sb.hex1(code._encoding._bits.buf()[iadr++]);
@@ -179,4 +208,14 @@ public abstract class ASMPrinter {
         return (bb instanceof LoopNode ? "LOOP" : "L")+bb._nid;
     }
 
+    private static String isMultiOp(SB sb, int old, int len) {
+        for( int i=old; i<len; i++ )
+            if( sb.at(i)=='\n' ) {
+                String s = sb.subString(i+1,len);
+                sb.setLen(i);
+                return s;
+            }
+        return null;
+    }
+
 }