From 87a75c83b75480e65147a4d425fddd14a05df4d9 Mon Sep 17 00:00:00 2001 From: Cliff Click Date: Wed, 7 Oct 2026 21:31:18 -0700 Subject: More test case vs RA cleanup Missing a few risc/arm encodings. Make private many tests, so they don't escape and the escaped fcn inflate x86 spill numbers --- chapter21/README.md | 20 +++++++++++--------- .../com/seaofnodes/simple/node/cpus/arm/LoadARM.java | 2 +- .../com/seaofnodes/simple/node/cpus/arm/NotARM.java | 4 ++-- .../com/seaofnodes/simple/node/cpus/arm/arm.java | 4 ++-- .../seaofnodes/simple/node/cpus/riscv/NotRISC.java | 3 ++- .../seaofnodes/simple/node/cpus/riscv/SetIRISC.java | 6 +++++- .../com/seaofnodes/simple/node/cpus/riscv/riscv.java | 2 +- 7 files changed, 24 insertions(+), 17 deletions(-) (limited to 'chapter21') diff --git a/chapter21/README.md b/chapter21/README.md index 63b02ae9..80304c1a 100644 --- a/chapter21/README.md +++ b/chapter21/README.md @@ -147,22 +147,24 @@ its two-address operand. For example, `mov rpc=s8; st1 [rpc+4],s1` becomes Phi and control-flow users are excluded. This local cleanup leaves spill selection unchanged; it complements coalescing without another CFG analysis. -For looking at allocator progress, we will keep a table of weighted spill costs +For looking at allocator progress, we will keep a table of execution and allocation costs a chapters' tests - and use the later chapter compiler and register allocator on them. The later chapters also modify the graph (sometimes adding boiler plate graph) so its not really a 100% fair comparison of how well this *heuristic* does, but it does give us some idea that we're not losing ground. -Rows are fixed test cohorts; columns are the compiler/allocator chapters. Each -cell is the sum of `_spillScaled`: retained split moves, including register -copies, weighted by loopDepth. Lower is better. **# tests counts -compilations** (program/CPU/ABI cases), including zero-spill cases, rather than -JUnit methods. The count and membership of each row stay fixed across columns. +Rows are fixed test cohorts; columns are compiler chapters. Each cell gives +**Ops / RA / X**: executed ARM+RISC-V instructions, the subset emitted for +allocator copies/rematerializations, and x86's loop-weighted split-move estimate. +**Cases D / X** counts measured dynamic program/target cases and x86 compilations; +membership and inputs stay fixed across columns. Native-library work is excluded. +Chapter 20 has no encoder, hence no dynamic counts. Lower is better within each +metric; see the [measurement details](../docs/regalloc-spills.md) for coverage. -| Test cohort | # tests | Ch 20 | Ch 21 | +| Test cohort | Cases D / X | Ch 20: Ops / RA / X | Ch 21: Ops / RA / X | |---|---:|---:|---:| -| Ch 20 | 39 | 360 | 576 | -| Ch 21 | 52 | | 1,072 | +| Ch 20 | 22 / 13 | — / — / 161 | 55,173 / 5,407 / 174 | +| Ch 21 | 20 / 18 | | 6,207 / 261 / 461 | 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 5fa68e19..6fd4274a 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 @@ -17,7 +17,7 @@ public class LoadARM extends MemOpARM { super(ld, base, idx, off, 0); } @Override public String op() { return "ld"+_sz; } - @Override public RegMask outregmap() { return size() >= 4 ? arm.MEM_MASK : arm.RMASK; } + @Override public RegMask outregmap() { return size() >= 4 ? arm.MEM_MASK : arm.WMASK; } private static final int[] OP_LOADS = new int[]{ Arm64.OP_LOAD_IMM_8, Arm64.OP_LOAD_IMM_16, Arm64.OP_LOAD_IMM_32, Arm64.OP_LOAD_IMM_64, }; diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NotARM.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NotARM.java index 5c388bab..ec017032 100644 --- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NotARM.java +++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/arm/NotARM.java @@ -19,9 +19,9 @@ public class NotARM extends MachConcreteNode implements MachNode{ // subtracting zero from rs will just yield rs, it sets the zero flag and then it's used in cset short self = enc.reg(this ); short reg1 = enc.reg(in(1)); - int subs = Arm64.imm_inst(Arm64.OP_SUBS, 0, reg1, self); + int subs = Arm64.imm_inst(Arm64.OP_SUBS, 0, reg1, 31); enc.add4(subs); - int cset = Arm64.cond_set(Arm64.OP_CSET, 31, Arm64.COND.EQ, 63, reg1); + int cset = Arm64.cset(Arm64.OP_CSET, 31, Arm64.COND.EQ, 31, self); enc.add4(cset); } 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 4006565e..66a59fde 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 @@ -215,7 +215,7 @@ public class arm extends Machine { public static void f_cmp(Encoding enc, Node n) { short reg1 = (short)(enc.reg(n.in(1))-D_OFFSET); short reg2 = (short)(enc.reg(n.in(2))-D_OFFSET); - int body = Arm64.f_cmp(0b00011110, 3, reg1, reg2); + int body = Arm64.f_cmp(0b00011110, 3, reg2, reg1); enc.add4(body); } @@ -325,7 +325,7 @@ public class arm extends Machine { private Node cmp(BoolNode bool){ Node cmp = _cmp(bool); - return new SetARM(cmp, IfNode.negate(bool.op())); + return new SetARM(cmp, bool.op()); } private Node _cmp(BoolNode bool) { if( bool.isFloat() ) diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NotRISC.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NotRISC.java index 469221f3..6aefc54b 100644 --- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NotRISC.java +++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/NotRISC.java @@ -15,7 +15,8 @@ public class NotRISC extends MachConcreteNode implements MachNode { @Override public void encoding( Encoding enc ) { short dst = enc.reg(this ); short src = enc.reg(in(1)); - enc.add4(RiscV.i_type(RiscV.OP_IMM, dst, 2, src, 1)); + // SEQZ is unsigned: negative nonzero inputs must also produce false. + enc.add4(RiscV.i_type(RiscV.OP_IMM, dst, 3, src, 1)); } @Override public void asm(CodeGen code, SB sb) { sb.p(code.reg(this)); } } diff --git a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/SetIRISC.java b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/SetIRISC.java index f270bf12..8b6512e1 100644 --- a/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/SetIRISC.java +++ b/chapter21/src/main/java/com/seaofnodes/simple/node/cpus/riscv/SetIRISC.java @@ -7,7 +7,11 @@ import com.seaofnodes.simple.node.Node; public class SetIRISC extends ImmRISC { final boolean _unsigned; // slti vs sltiu public SetIRISC( Node src, int imm12, boolean unsigned ) { - super(src,imm12); + this(src,imm12,unsigned,true); + } + public SetIRISC( Node src, int imm12, boolean unsigned, boolean pop ) { + super(src,imm12,pop); + if( !pop ) _inputs.setLen(2); _unsigned = unsigned; } @Override public String op() { return "slt" + (_unsigned ? "u":"") + "i"; } 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 7e840509..335347e9 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 @@ -304,7 +304,7 @@ public class riscv extends Machine { // x <= y - flip and negate; !(y < x); `slt tmp=y,x; xori dst=tmp,#1` case "<=" -> new XorIRISC(new SetRISC(bool.swap12(), false),1); // x == y - sub and vs0 == `sub tmp=x-y; sltu dst=tmp,#1` - case "==" -> new SetIRISC(new SubRISC(bool),1,true); + case "==" -> new SetIRISC(new SubRISC(bool),1,true,false); default -> throw Utils.TODO(); }; } -- cgit v1.2.3