cmd/compile: add clobberdeadreg mode

When -clobberdeadreg flag is set, the compiler inserts code that
clobbers integer registers at call sites. This may be helpful for
debugging register ABI.

Only implemented on AMD64 for now.

Change-Id: Ia203d3f891c30fd95d0103489056fe01d63a2899
Reviewed-on: https://go-review.googlesource.com/c/go/+/302809
Trust: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>

src/cmd/compile/internal/amd64/ssa.go

@@ -1242,6 +1242,13 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p.To.Reg = x86.REG_SP
 		ssagen.AddAux(&p.To, v)
 		p.To.Offset += 4
+	case ssa.OpClobberReg:
+		x := uint64(0xdeaddeaddeaddead)
+		p := s.Prog(x86.AMOVQ)
+		p.From.Type = obj.TYPE_CONST
+		p.From.Offset = int64(x)
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = v.Reg()
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())
 	}

src/cmd/compile/internal/arm/ssa.go

@@ -861,7 +861,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		v.Fatalf("FlagConstant op should never make it to codegen %v", v.LongString())
 	case ssa.OpARMInvertFlags:
 		v.Fatalf("InvertFlags should never make it to codegen %v", v.LongString())
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/arm64/ssa.go

@@ -1100,7 +1100,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		v.Fatalf("FlagConstant op should never make it to codegen %v", v.LongString())
 	case ssa.OpARM64InvertFlags:
 		v.Fatalf("InvertFlags should never make it to codegen %v", v.LongString())
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/base/flag.go

@@ -90,6 +90,7 @@ type CmdFlags struct {
 	CPUProfile         string       "help:\"write cpu profile to `file`\""
 	Complete           bool         "help:\"compiling complete package (no C or assembly)\""
 	ClobberDead        bool         "help:\"clobber dead stack slots (for debugging)\""
+	ClobberDeadReg     bool         "help:\"clobber dead registers (for debugging)\""
 	Dwarf              bool         "help:\"generate DWARF symbols\""
 	DwarfBASEntries    *bool        "help:\"use base address selection entries in DWARF\""                        // &Ctxt.UseBASEntries, set below
 	DwarfLocationLists *bool        "help:\"add location lists to DWARF in optimized mode\""                      // &Ctxt.Flag_locationlists, set below

src/cmd/compile/internal/mips/ssa.go

@@ -798,7 +798,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p := s.Prog(obj.AGETCALLERPC)
 		p.To.Type = obj.TYPE_REG
 		p.To.Reg = v.Reg()
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/mips64/ssa.go

@@ -765,7 +765,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p := s.Prog(obj.AGETCALLERPC)
 		p.To.Type = obj.TYPE_REG
 		p.To.Reg = v.Reg()
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/ppc64/ssa.go

@@ -1927,7 +1927,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		v.Fatalf("InvertFlags should never make it to codegen %v", v.LongString())
 	case ssa.OpPPC64FlagEQ, ssa.OpPPC64FlagLT, ssa.OpPPC64FlagGT:
 		v.Fatalf("Flag* ops should never make it to codegen %v", v.LongString())
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/riscv64/ssa.go

@@ -629,6 +629,9 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p.To.Sym = ir.Syms.Duffcopy
 		p.To.Offset = v.AuxInt
 
+	case ssa.OpClobber, ssa.OpClobberReg:
+		// TODO: implement for clobberdead experiment. Nop is ok for now.
+
 	default:
 		v.Fatalf("Unhandled op %v", v.Op)
 	}

src/cmd/compile/internal/s390x/ssa.go

@@ -844,7 +844,7 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		bne.To.SetTarget(cs)
 	case ssa.OpS390XSYNC:
 		s.Prog(s390x.ASYNC)
-	case ssa.OpClobber:
+	case ssa.OpClobber, ssa.OpClobberReg:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())

src/cmd/compile/internal/ssa/gen/genericOps.go

@@ -613,6 +613,7 @@ var genericOps = []opData{
 
 	// Clobber experiment op
 	{name: "Clobber", argLength: 0, typ: "Void", aux: "SymOff", symEffect: "None"}, // write an invalid pointer value to the given pointer slot of a stack variable
+	{name: "ClobberReg", argLength: 0, typ: "Void"},                                // clobber a register
 }
 
 //     kind          controls        successors   implicit exit

src/cmd/compile/internal/ssa/opGen.go

@@ -2921,6 +2921,7 @@ const (
 	OpAtomicOr8Variant
 	OpAtomicOr32Variant
 	OpClobber
+	OpClobberReg
 )
 
 var opcodeTable = [...]opInfo{
@@ -36373,6 +36374,11 @@ var opcodeTable = [...]opInfo{
 		symEffect: SymNone,
 		generic:   true,
 	},
+	{
+		name:    "ClobberReg",
+		argLen:  0,
+		generic: true,
+	},
 }
 
 func (o Op) Asm() obj.As          { return opcodeTable[o].asm }

src/cmd/compile/internal/ssa/regalloc.go

@@ -114,6 +114,7 @@
 package ssa
 
 import (
+	"cmd/compile/internal/base"
 	"cmd/compile/internal/ir"
 	"cmd/compile/internal/types"
 	"cmd/internal/objabi"
@@ -301,6 +302,9 @@ type regAllocState struct {
 
 	// blockOrder[b.ID] corresponds to the index of block b in visitOrder.
 	blockOrder []int32
+
+	// whether to insert instructions that clobber dead registers at call sites
+	doClobber bool
 }
 
 type endReg struct {
@@ -339,6 +343,17 @@ func (s *regAllocState) freeRegs(m regMask) {
 	}
 }
 
+// clobberRegs inserts instructions that clobber registers listed in m.
+func (s *regAllocState) clobberRegs(m regMask) {
+	m &= s.allocatable & s.f.Config.gpRegMask // only integer register can contain pointers, only clobber them
+	for m != 0 {
+		r := pickReg(m)
+		m &^= 1 << r
+		x := s.curBlock.NewValue0(src.NoXPos, OpClobberReg, types.TypeVoid)
+		s.f.setHome(x, &s.registers[r])
+	}
+}
+
 // setOrig records that c's original value is the same as
 // v's original value.
 func (s *regAllocState) setOrig(c *Value, v *Value) {
@@ -700,6 +715,14 @@ func (s *regAllocState) init(f *Func) {
 			}
 		}
 	}
+
+	// The clobberdeadreg experiment inserts code to clobber dead registers
+	// at call sites.
+	// Ignore huge functions to avoid doing too much work.
+	if base.Flag.ClobberDeadReg && len(s.f.Blocks) <= 10000 {
+		// TODO: honor GOCLOBBERDEADHASH, or maybe GOSSAHASH.
+		s.doClobber = true
+	}
 }
 
 // Adds a use record for id at distance dist from the start of the block.
@@ -1314,6 +1337,9 @@ func (s *regAllocState) regalloc(f *Func) {
 			}
 			if len(regspec.inputs) == 0 && len(regspec.outputs) == 0 {
 				// No register allocation required (or none specified yet)
+				if s.doClobber && v.Op.IsCall() {
+					s.clobberRegs(regspec.clobbers)
+				}
 				s.freeRegs(regspec.clobbers)
 				b.Values = append(b.Values, v)
 				s.advanceUses(v)
@@ -1475,6 +1501,11 @@ func (s *regAllocState) regalloc(f *Func) {
 			}
 
 			// Dump any registers which will be clobbered
+			if s.doClobber && v.Op.IsCall() {
+				// clobber registers that are marked as clobber in regmask, but
+				// don't clobber inputs.
+				s.clobberRegs(regspec.clobbers &^ s.tmpused &^ s.nospill)
+			}
 			s.freeRegs(regspec.clobbers)
 			s.tmpused |= regspec.clobbers
 

src/cmd/compile/internal/test/clobberdead_test.go

@@ -20,7 +20,15 @@ func main() { fmt.Println("hello") }
 
 func TestClobberDead(t *testing.T) {
 	// Test that clobberdead mode generates correct program.
+	runHello(t, "-clobberdead")
+}
 
+func TestClobberDeadReg(t *testing.T) {
+	// Test that clobberdeadreg mode generates correct program.
+	runHello(t, "-clobberdeadreg")
+}
+
+func runHello(t *testing.T, flag string) {
 	if testing.Short() {
 		// This test rebuilds the runtime with a special flag, which
 		// takes a while.
@@ -36,7 +44,7 @@ func TestClobberDead(t *testing.T) {
 		t.Fatalf("write file failed: %v", err)
 	}
 
-	cmd := exec.Command(testenv.GoToolPath(t), "run", "-gcflags=all=-clobberdead", src)
+	cmd := exec.Command(testenv.GoToolPath(t), "run", "-gcflags=all="+flag, src)
 	out, err := cmd.CombinedOutput()
 	if err != nil {
 		t.Fatalf("go run failed: %v\n%s", err, out)

src/cmd/compile/internal/wasm/ssa.go

@@ -190,6 +190,9 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p := s.Prog(storeOp(v.Type))
 		ssagen.AddrAuto(&p.To, v)
 
+	case ssa.OpClobber, ssa.OpClobberReg:
+		// TODO: implement for clobberdead experiment. Nop is ok for now.
+
 	default:
 		if v.Type.IsMemory() {
 			return

src/cmd/compile/internal/x86/ssa.go

@@ -832,6 +832,8 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		p.To.Type = obj.TYPE_MEM
 		p.To.Reg = x86.REG_SP
 		ssagen.AddAux(&p.To, v)
+	case ssa.OpClobberReg:
+		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:
 		v.Fatalf("genValue not implemented: %s", v.LongString())
 	}

src/runtime/sys_darwin_amd64.s

@@ -434,7 +434,7 @@ TEXT runtime·fcntl_trampoline<ABIInternal>(SB),NOSPLIT,$0
 // mstart_stub is the first function executed on a new thread started by pthread_create.
 // It just does some low-level setup and then calls mstart.
 // Note: called with the C calling convention.
-TEXT runtime·mstart_stub(SB),NOSPLIT,$0
+TEXT runtime·mstart_stub<ABIInternal>(SB),NOSPLIT,$0
 	// DI points to the m.
 	// We are already on m's g0 stack.
 

test/codegen/clobberdeadreg.go

@@ -0,0 +1,33 @@
+// asmcheck -gcflags=-clobberdeadreg
+
+// +build amd64
+
+// Copyright 2021 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package codegen
+
+type S struct {
+	a, b, c, d, e, f int
+}
+
+func F(a, b, c int, d S) {
+	// -2401018187971961171 is 0xdeaddeaddeaddead
+	// amd64:`MOVQ\t\$-2401018187971961171, AX`, `MOVQ\t\$-2401018187971961171, BX`, `MOVQ\t\$-2401018187971961171, CX`
+	// amd64:`MOVQ\t\$-2401018187971961171, DX`, `MOVQ\t\$-2401018187971961171, SI`, `MOVQ\t\$-2401018187971961171, DI`
+	// amd64:`MOVQ\t\$-2401018187971961171, R8`, `MOVQ\t\$-2401018187971961171, R9`, `MOVQ\t\$-2401018187971961171, R10`
+	// amd64:`MOVQ\t\$-2401018187971961171, R11`, `MOVQ\t\$-2401018187971961171, R12`, `MOVQ\t\$-2401018187971961171, R13`
+	// amd64:-`MOVQ\t\$-2401018187971961171, BP` // frame pointer is not clobbered
+	StackArgsCall(a, b, c, d)
+	// amd64:`MOVQ\t\$-2401018187971961171, R12`, `MOVQ\t\$-2401018187971961171, R13`, `MOVQ\t\$-2401018187971961171, DX`
+	// amd64:-`MOVQ\t\$-2401018187971961171, AX`, -`MOVQ\t\$-2401018187971961171, R11` // register args are not clobbered
+	RegArgsCall(a, b, c, d)
+}
+
+//go:noinline
+func StackArgsCall(int, int, int, S) {}
+
+//go:noinline
+//go:registerparams
+func RegArgsCall(int, int, int, S) {}