Fix CmpIX86/Jmp swap args bug

Some x86 cmp need to swap args, then Set/Jmp both need to swap `bop`.
Fix missing padding in ASM printer.
MergeSort test like brainfuck - runs on arm.

4 files changed, 34 insertions, 32 deletions

diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpIX86.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpIX86.java
index d274875a..91b6d43c 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpIX86.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpIX86.java
@@ -7,9 +7,9 @@ import com.seaofnodes.simple.node.*;
 // Compare immediate.  Sets flags.
 public class CmpIX86 extends MachConcreteNode implements MachNode {
     final int _imm;
-    CmpIX86( Node cmp, int imm ) {
+    CmpIX86( Node cmp, int imm, boolean swap ) {
         super(cmp);
-        _inputs.pop();
+        _inputs.del(swap ? 1 : 2);
         _imm = imm;
     }
     CmpIX86( CmpIX86 cmp ) {
@@ -18,11 +18,6 @@ public class CmpIX86 extends MachConcreteNode implements MachNode {
         addDef(cmp.in(1));
         _imm = cmp._imm;
     }
-    CmpIX86( Node lhs, int imm, double ignore ) {
-        super(lhs);
-        _inputs.del(1);
-        _imm = imm;
-    }
 
     @Override public String op() { return _imm==0 ? "test" : "cmp"; }
     @Override public RegMask regmap(int i) { return x86_64_v2.RMASK; }
diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpMemX86.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpMemX86.java
index f9cd597b..489853eb 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpMemX86.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpMemX86.java
@@ -7,10 +7,11 @@ import com.seaofnodes.simple.type.Type;
 import com.seaofnodes.simple.type.TypeFloat;
 
 public class CmpMemX86 extends MemOpX86 {
-    final boolean _invert;      // Op switched LHS, RHS
-    CmpMemX86( BoolNode bool, LoadNode ld, Node base, Node idx, int off, int scale, int imm, Node val, boolean invert ) {
+    final boolean _swap;      // Op switched LHS, RHS
+    CmpMemX86( BoolNode bool, LoadNode ld, Node base, Node idx, int off, int scale, int imm, Node val, boolean swap ) {
         super(bool,ld, base, idx, off, scale, imm, val );
-        _invert = invert;
+        _swap = swap;
+        assert !_swap || val() == null || bool.op()=="==" || bool.op()=="!="; // Cannot swap with immediate
     }
     @Override public String op() { return ((val()==null && _imm==0) ? "test" : "cmp") + _sz; }
     @Override public RegMask outregmap() { return x86_64_v2.FLAGS_MASK; }
@@ -23,52 +24,55 @@ public class CmpMemX86 extends MemOpX86 {
         short src = enc.reg(val());
 
         if(src == -1) {
+            assert !_swap;
             int log = _declaredType.log_size();
             x86_64_v2.rexF(-1, ptr, idx, log==3, enc);
             switch(log) {
-                case 0: enc.add1(0x80); break;
-                case 1: enc.add1(0x66).add1(0x81); break;
-                case 2: enc.add1(0x81); break;
-                case 3: enc.add1(0x81); break;
+            case 0: enc.add1(0x80); break;
+            case 1: enc.add1(0x66).add1(0x81); break;
+            case 2: enc.add1(0x81); break;
+            case 3: enc.add1(0x81); break;
             }
             x86_64_v2.indirectAdr(_scale, idx, ptr, _off,  7, enc);
             switch(log) {
-                case 0: enc.add1(_imm); break;
-                case 1: enc.add2(_imm); break;
-                case 2: enc.add4(_imm); break;
-                case 3: enc.add8(_imm); break;
+            case 0: enc.add1(_imm); break;
+            case 1: enc.add2(_imm); break;
+            case 2: enc.add4(_imm); break;
+            case 3: enc.add8(_imm); break;
             }
         } else {
-            encVal(enc, _declaredType, ptr, idx, src, _off, _scale);
+            encVal(enc, _declaredType, ptr, idx, src, _off, _scale, _swap);
         }
     }
 
-    static void encVal(Encoding enc, Type decl, short ptr, short idx, short src, int off, int scale) {
+    static void encVal(Encoding enc, Type decl, short ptr, short idx, short src, int off, int scale, boolean _swap) {
         if(decl instanceof TypeFloat) {
             src -= (short)x86_64_v2.XMM_OFFSET;
             enc.add1(decl==TypeFloat.F32 ? 0xF3 : 0xF2).add1(0x0F).add1(0xC2);
         } else {
             int log = decl.log_size();
             x86_64_v2.rexF(src, ptr, idx, log == 3, enc);
+            int swap = _swap ? 2 : 0;
             switch(log) {
-                case 0: enc.add1(0x38); break;
-                case 1: enc.add1(0x66).add1(0x3B); break;
-                case 2: enc.add1(0x3B); break;
-                case 3: enc.add1(0x3B); break;
+            case 0: enc.add1(0x38+swap); break;
+            case 1: enc.add1(0x66).add1(0x39+swap); break;
+            case 2: enc.add1(0x39+swap); break;
+            case 3: enc.add1(0x39+swap); break;
             }
         }
         x86_64_v2.indirectAdr(scale, idx, ptr, off, src, enc);
     }
+
     // General form: "cmp  dst = src, [base + idx<<2 + 12]"
     @Override public void asm(CodeGen code, SB sb) {
         String dst = code.reg(this);
         if( dst!="flags" )  sb.p(dst).p(" = ");
-        if( _invert ) asm_address(code,sb).p(",");
+        if( _swap ) asm_address(code,sb).p(",");
         if( val()==null ) {
             if( _imm!=0 )  sb.p("#").p(_imm);
         } else {
             sb.p(code.reg(val()));
         }
-        if( !_invert ) asm_address(code,sb.p(", "));
+        if( !_swap ) asm_address(code,sb.p(", "));
     }
 }
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 e6390386..23c12ae7 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
@@ -476,21 +476,23 @@ public class x86_64_v2 extends Machine {
         Node lhs = bool.in(1);
         Node rhs = bool.in(2);
 
+        // Since cmp does not record a BOP, SetX/Jmp need to know if the CMP is swapped.
+
         // Vs memory
         if( lhs instanceof LoadNode ld && ld.nOuts() == 1 && rhs._type.isa(ld._declaredType) )
             return new CmpMemX86(bool, address(ld), ld.ptr(), idx, off, scale, imm(rhs), val, false);
 
+        // Operands swap in the encoding directly, no need for Set/Jmp to swap `bop`
         if( rhs instanceof LoadNode ld && ld.nOuts() == 1 && lhs._type.isa(ld._declaredType))
             return new CmpMemX86(bool, address(ld), ld.ptr(), idx, off, scale, imm(lhs), val, true );
 
         // Vs immediate
         if( rhs instanceof ConstantNode con && con._con instanceof TypeInteger ti && imm32(ti.value()) )
-            return new CmpIX86(bool, (int)ti.value());
+            return new CmpIX86(bool, (int)ti.value(), false);
 
-        if( lhs instanceof ConstantNode con && con._con instanceof TypeInteger ti && imm32(ti.value()) ) {
-            swap = true;
-            return new CmpIX86(bool, (int)ti.value(), 0.5);
-        }
+        // Operand swap compare; Set and Jmp need to swap `bop`
+        if( lhs instanceof ConstantNode con && con._con instanceof TypeInteger ti && imm32(ti.value()) )
+            return new CmpIX86(bool, (int)ti.value(), swap=true);
 
         // x vs y
         return new CmpX86(bool);
@@ -519,7 +521,7 @@ public class x86_64_v2 extends Machine {
         // Most general arith ops will also set flags, which the Jmp needs directly.
         // Loads do not set the flags, and will need an explicit TEST
         String op = "!=";
-        if( iff.in(1) instanceof BoolNode bool ) op = bool.op();
+        if( iff.in(1) instanceof BoolNode bool ) op = swap ? IfNode.swap(bool.op()) : bool.op();
         else if( iff.in(1)==null ) op = "=="; // Never-node cutout
         else iff.setDef(1, new BoolNode.NE(iff.in(1), new ConstantNode(TypeInteger.ZERO)));
         return new JmpX86(iff, op);
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 8f3c9bb2..cf3bdde3 100644
--- a/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java
+++ b/chapter21/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java
@@ -79,6 +79,7 @@ public abstract class ASMPrinter {
         if( fun._name != null ) sb.p(fun._name).p(" ");
         fun.sig().print(sb);
         sb.p("---------------------------").nl();
+        iadr = (iadr + 15)&-16; // All function entries padded to 16 align
 
         if( fun._frameAdjust != 0 )
             iadr = doInst(iadr,sb,code,fun,cfgidx,fun,true,true);