diff options
Diffstat (limited to 'chapter22/src/main/java/com')
9 files changed, 40 insertions, 14 deletions
diff --git a/chapter22/src/main/java/com/seaofnodes/simple/codegen/Encoding.java b/chapter22/src/main/java/com/seaofnodes/simple/codegen/Encoding.java index 7a43f099..db799b98 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/codegen/Encoding.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/codegen/Encoding.java @@ -359,11 +359,6 @@ public class Encoding { slide = 0; for( int i=0; i<len; i++ ) { CFGNode bb = _code._cfg.at(i); - // Functions pad to align 16 - if( bb instanceof FunNode ) { - int newStart = _opStart[bb._nid]+slide; - slide += (newStart+15 & -16)-newStart; - } _opStart[bb._nid] += slide; // Slide down all other (non-CFG) ops in the block for( Node n : bb._outputs ) @@ -387,6 +382,23 @@ public class Encoding { } + // Go again, inserting padding on function headers. Since no + // short-jumps span function headers, the padding will not make any + // short jumps fail. + for( int i=0; i<len; i++ ) { + CFGNode bb = _code._cfg.at(i); + // Functions pad to align 16 + if( bb instanceof FunNode ) { + int newStart = _opStart[bb._nid]+slide; + slide += (newStart+15 & -16)-newStart; + } + _opStart[bb._nid] += slide; + for( Node n : bb._outputs ) + if( n instanceof MachNode && !(n instanceof CFGNode) ) + _opStart[n._nid] += slide; + } + + // Copy/slide the bits to make space for all the longer branches int grow = _opStart[_code._cfg.at(len-1)._nid] - oldStarts[len-1]; if( grow > 0 ) { // If no short-form ops, nothing to do here diff --git a/chapter22/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java b/chapter22/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java index 7e0a58ac..04a119c2 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/codegen/GlobalCodeMotion.java @@ -177,14 +177,22 @@ public abstract class GlobalCodeMotion { } // Walk all inputs and put on worklist, as their last-use might now be done - for( Node def : n._inputs ) - if( def!=null && late[def._nid]==null ) { + for( Node def : n._inputs ) { + if( def!=null && late[def._nid]==null ) work.push(def); - // if the def has a load use, maybe the load can fire + // if the def has a load use, maybe the load can fire + if( def!=null ) for( Node out : def._outputs ) if( out instanceof MemOpNode ld && ld._isLoad && late[ld._nid]==null ) work.push(ld); + } + + if( n instanceof LoopNode loop ) { + for( Node phi : loop._outputs ) { + if( phi instanceof PhiNode && late[phi._nid]==null ) + work.push(phi); } + } } } diff --git a/chapter22/src/main/java/com/seaofnodes/simple/codegen/RegAlloc.java b/chapter22/src/main/java/com/seaofnodes/simple/codegen/RegAlloc.java index 68b55f6f..57d0d90c 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/codegen/RegAlloc.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/codegen/RegAlloc.java @@ -559,7 +559,7 @@ public class RegAlloc { private void postColor() { int maxReg = -1; for( CFGNode bb : _code._cfg ) { // For all ops - if( bb instanceof FunNode fun ) + if( bb instanceof FunNode ) maxReg = -1; // Reset for new function // Compute frame size, based on arguments and largest reg seen if( bb instanceof ReturnNode ret ) diff --git a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpX86.java b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpX86.java index 06a36d99..3e4200c0 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpX86.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/CmpX86.java @@ -9,6 +9,12 @@ public class CmpX86 extends MachConcreteNode implements MachNode { @Override public String op() { return "cmp"; } @Override public RegMask regmap(int i) { return x86_64_v2.RMASK; } @Override public RegMask outregmap() { return x86_64_v2.FLAGS_MASK; } + // This one is marginal: cloning a Cmp[reg,reg] means stretching the + // lifetimes of 2 normal registers in exchange for shortening a flags + // register lifetime. Spilling flags is (probably) relatively expensive + // compared to some normal registers - and those normal registers might not + // be spilling! + @Override public boolean isClone() { return true; } @Override public void encoding( Encoding enc ) { short dst = enc.reg(in(1)); diff --git a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/DivX86.java b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/DivX86.java index 87b8b0c3..06f001b0 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/DivX86.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/DivX86.java @@ -8,7 +8,7 @@ public class DivX86 extends MachConcreteNode implements MachNode { DivX86( Node div ) { super(div); } @Override public String op() { return "div"; } @Override public RegMask regmap(int i) { - return (i==1) ? x86_64_v2.RAX_MASK : x86_64_v2.RMASK; + return (i==1) ? x86_64_v2.RAX_MASK : x86_64_v2.DIV_MASK; } @Override public RegMask outregmap() { return x86_64_v2.RAX_MASK; } @Override public RegMask killmap() { return x86_64_v2.RDX_MASK; } diff --git a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/ImmX86.java b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/ImmX86.java index 760d6805..9e602b47 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/ImmX86.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/ImmX86.java @@ -15,6 +15,7 @@ public abstract class ImmX86 extends MachConcreteNode implements MachNode { @Override public RegMask regmap(int i) { return x86_64_v2.RMASK; } @Override public RegMask outregmap() { return x86_64_v2.WMASK; } @Override public int twoAddress() { return 1; } + @Override public RegMask killmap() { return x86_64_v2.FLAGS_MASK; } abstract int opcode(); abstract int mod(); diff --git a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/RegX86.java b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/RegX86.java index 0997644d..a4a6a814 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/RegX86.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/RegX86.java @@ -8,6 +8,7 @@ public abstract class RegX86 extends MachConcreteNode { RegX86( Node add ) { super(add); } @Override public RegMask regmap(int i) { return x86_64_v2.RMASK; } @Override public RegMask outregmap() { return x86_64_v2.WMASK; } + @Override public RegMask killmap() { return x86_64_v2.FLAGS_MASK; } @Override public int twoAddress() { return 1; } abstract int opcode(); @Override public void encoding( Encoding enc ) { diff --git a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java index 60e10d79..94c1511a 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/node/cpus/x86_64_v2/x86_64_v2.java @@ -43,6 +43,8 @@ public class x86_64_v2 extends Machine { static RegMask XMASK = new RegMask(FP_BITS); static RegMask FLAGS_MASK = new RegMask(FLAGS); static RegMask RPC_MASK = new RegMask(RPC); + // Divide mask: exclude RDX:RAX + static RegMask DIV_MASK = new RegMask(RD_BITS & ~(1L<<RAX) & ~(1L<<RDX)); static final long SPILLS = -(1L << MAX_REG); static final RegMask SPLIT_MASK = new RegMask(WR_BITS | FP_BITS /* | (1L<<FLAGS)*/ | SPILLS, -1L ); diff --git a/chapter22/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java b/chapter22/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java index 45f71e40..366ae376 100644 --- a/chapter22/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java +++ b/chapter22/src/main/java/com/seaofnodes/simple/print/ASMPrinter.java @@ -22,10 +22,6 @@ public abstract class ASMPrinter { if( code._cfg.at(i) instanceof FunNode fun ) iadr = print(iadr,sb,code,fun,i); - // Skip padding - while( ((iadr+7) & -8) > iadr ) - iadr++; - // constant pool iadr = (iadr+15)&-16; // pad to 16 Encoding enc = code._encoding; |