cmd/compile: retrieve Args from registers

in progress; doesn't fully work until they are also passed on
register on the caller side.

For #40724.

Change-Id: I29a6680e60bdbe9d132782530214f2a2b51fb8f6
Reviewed-on: https://go-review.googlesource.com/c/go/+/293394
Trust: David Chase <drchase@google.com>
Run-TryBot: David Chase <drchase@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

src/cmd/compile/internal/abi/abiutils.go

@@ -158,6 +158,13 @@ func (a *ABIConfig) LocalsOffset() int64 {
 	return a.offsetForLocals
 }
 
+// FloatIndexFor translates r into an index in the floating point parameter
+// registers.  If the result is negative, the input index was actually for the
+// integer parameter registers.
+func (a *ABIConfig) FloatIndexFor(r RegIndex) int64 {
+	return int64(r) - int64(a.regAmounts.intRegs)
+}
+
 // NumParamRegs returns the number of parameter registers used for a given type,
 // without regard for the number available.
 func (a *ABIConfig) NumParamRegs(t *types.Type) int {

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

@@ -980,6 +980,8 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		ssagen.AddAux(&p.From, v)
 		p.To.Type = obj.TYPE_REG
 		p.To.Reg = v.Reg()
+	case ssa.OpArgIntReg, ssa.OpArgFloatReg:
+		ssagen.CheckArgReg(v)
 	case ssa.OpAMD64LoweredGetClosurePtr:
 		// Closure pointer is DX.
 		ssagen.CheckLoweredGetClosurePtr(v)

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

@@ -808,9 +808,9 @@ func (x *expandState) rewriteArgs(v *Value, firstArg int) *Value {
 		}
 		auxI := int64(i - firstArg)
 		aRegs := aux.RegsOfArg(auxI)
-		aOffset := aux.OffsetOfArg(auxI)
 		aType := aux.TypeOfArg(auxI)
 		if a.Op == OpDereference {
+			aOffset := aux.OffsetOfArg(auxI)
 			if a.MemoryArg() != m0 {
 				x.f.Fatalf("Op...LECall and OpDereference have mismatched mem, %s and %s", v.LongString(), a.LongString())
 			}
@@ -821,13 +821,16 @@ func (x *expandState) rewriteArgs(v *Value, firstArg int) *Value {
 			// TODO(register args) this will be more complicated with registers in the picture.
 			mem = x.rewriteDereference(v.Block, x.sp, a, mem, aOffset, aux.SizeOfArg(auxI), aType, pos)
 		} else {
-			if x.debug {
-				fmt.Printf("storeArg %s, %v, %d\n", a.LongString(), aType, aOffset)
-			}
 			var rc registerCursor
 			var result *[]*Value
+			var aOffset int64
 			if len(aRegs) > 0 {
 				result = &allResults
+			} else {
+				aOffset = aux.OffsetOfArg(auxI)
+			}
+			if x.debug {
+				fmt.Printf("storeArg %s, %v, %d\n", a.LongString(), aType, aOffset)
 			}
 			rc.init(aRegs, aux.abiInfo, result)
 			mem = x.storeArgOrLoad(pos, v.Block, x.sp, a, mem, aType, aOffset, 0, rc)
@@ -1213,8 +1216,19 @@ func expandCalls(f *Func) {
 							pa.Offset(), frameOff, v.LongString()))
 					}
 				case 1:
+					r := pa.Registers[0]
+					i := f.ABISelf.FloatIndexFor(r)
+					// TODO seems like this has implications for debugging. How does this affect the location?
+					if i >= 0 { // float PR
+						v.Op = OpArgFloatReg
+					} else {
+						v.Op = OpArgIntReg
+						i = int64(r)
+					}
+					v.AuxInt = i
+
 				default:
-					panic(badVal("Saw unexpeanded OpArg", v))
+					panic(badVal("Saw unexpanded OpArg", v))
 				}
 
 			case OpStaticLECall:

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

@@ -21,8 +21,12 @@ func checkLower(f *Func) {
 				continue // lowered
 			}
 			switch v.Op {
-			case OpSP, OpSB, OpInitMem, OpArg, OpPhi, OpVarDef, OpVarKill, OpVarLive, OpKeepAlive, OpSelect0, OpSelect1, OpSelectN, OpConvert, OpInlMark:
+			case OpSP, OpSB, OpInitMem, OpArg, OpArgIntReg, OpArgFloatReg, OpPhi, OpVarDef, OpVarKill, OpVarLive, OpKeepAlive, OpSelect0, OpSelect1, OpSelectN, OpConvert, OpInlMark:
 				continue // ok not to lower
+			case OpMakeResult:
+				if len(b.Controls) == 1 && b.Controls[0] == v {
+					continue
+				}
 			case OpGetG:
 				if f.Config.hasGReg {
 					// has hardware g register, regalloc takes care of it

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

@@ -800,7 +800,8 @@ func (s *regAllocState) compatRegs(t *types.Type) regMask {
 }
 
 // regspec returns the regInfo for operation op.
-func (s *regAllocState) regspec(op Op) regInfo {
+func (s *regAllocState) regspec(v *Value) regInfo {
+	op := v.Op
 	if op == OpConvert {
 		// OpConvert is a generic op, so it doesn't have a
 		// register set in the static table. It can use any
@@ -808,6 +809,20 @@ func (s *regAllocState) regspec(op Op) regInfo {
 		m := s.allocatable & s.f.Config.gpRegMask
 		return regInfo{inputs: []inputInfo{{regs: m}}, outputs: []outputInfo{{regs: m}}}
 	}
+	if op == OpArgIntReg {
+		reg := v.Block.Func.Config.intParamRegs[v.AuxInt8()]
+		return regInfo{outputs: []outputInfo{{regs: 1 << uint(reg)}}}
+	}
+	if op == OpArgFloatReg {
+		reg := v.Block.Func.Config.floatParamRegs[v.AuxInt8()]
+		return regInfo{outputs: []outputInfo{{regs: 1 << uint(reg)}}}
+	}
+	if op.IsCall() {
+		// TODO Panic if not okay
+		if ac, ok := v.Aux.(*AuxCall); ok && ac.reg != nil {
+			return *ac.reg
+		}
+	}
 	return opcodeTable[op].reg
 }
 
@@ -1163,7 +1178,7 @@ func (s *regAllocState) regalloc(f *Func) {
 		for i := len(oldSched) - 1; i >= 0; i-- {
 			v := oldSched[i]
 			prefs := desired.remove(v.ID)
-			regspec := s.regspec(v.Op)
+			regspec := s.regspec(v)
 			desired.clobber(regspec.clobbers)
 			for _, j := range regspec.inputs {
 				if countRegs(j.regs) != 1 {
@@ -1193,7 +1208,7 @@ func (s *regAllocState) regalloc(f *Func) {
 			if s.f.pass.debug > regDebug {
 				fmt.Printf("  processing %s\n", v.LongString())
 			}
-			regspec := s.regspec(v.Op)
+			regspec := s.regspec(v)
 			if v.Op == OpPhi {
 				f.Fatalf("phi %s not at start of block", v)
 			}
@@ -2447,7 +2462,7 @@ func (s *regAllocState) computeLive() {
 					// desired registers back though phi nodes.
 					continue
 				}
-				regspec := s.regspec(v.Op)
+				regspec := s.regspec(v)
 				// Cancel desired registers if they get clobbered.
 				desired.clobber(regspec.clobbers)
 				// Update desired registers if there are any fixed register inputs.

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

@@ -795,8 +795,6 @@ func devirtLECall(v *Value, sym *obj.LSym) *Value {
 	v.Op = OpStaticLECall
 	auxcall := v.Aux.(*AuxCall)
 	auxcall.Fn = sym
-	// TODO(register args) this should not be necessary when fully transition to the new register ABI.
-	auxcall.abiInfo = v.Block.Func.ABIDefault.ABIAnalyzeTypes(nil, ACParamsToTypes(auxcall.args), ACParamsToTypes(auxcall.results))
 	v.RemoveArg(0)
 	return v
 }

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

@@ -137,7 +137,7 @@ func schedule(f *Func) {
 			case v.Op == OpVarDef:
 				// We want all the vardefs next.
 				score[v.ID] = ScoreVarDef
-			case v.Op == OpArg:
+			case v.Op == OpArg || v.Op == OpArgIntReg || v.Op == OpArgFloatReg:
 				// We want all the args as early as possible, for better debugging.
 				score[v.ID] = ScoreArg
 			case v.Type.IsMemory():

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

@@ -485,7 +485,8 @@ func buildssa(fn *ir.Func, worker int) *ssa.Func {
 		s.vars[memVar] = s.newValue1Apos(ssa.OpVarLive, types.TypeMem, deferBitsTemp, s.mem(), false)
 	}
 
-	params := s.f.ABISelf.ABIAnalyze(fn.Type())
+	var params *abi.ABIParamResultInfo
+	params = s.f.ABISelf.ABIAnalyze(fn.Type())
 
 	// Generate addresses of local declarations
 	s.decladdrs = map[*ir.Name]*ssa.Value{}
@@ -7019,11 +7020,20 @@ func CheckLoweredPhi(v *ssa.Value) {
 // That register contains the closure pointer on closure entry.
 func CheckLoweredGetClosurePtr(v *ssa.Value) {
 	entry := v.Block.Func.Entry
+	// TODO register args: not all the register-producing ops can come first.
 	if entry != v.Block || entry.Values[0] != v {
 		base.Fatalf("in %s, badly placed LoweredGetClosurePtr: %v %v", v.Block.Func.Name, v.Block, v)
 	}
 }
 
+// CheckArgReg ensures that v is in the function's entry block.
+func CheckArgReg(v *ssa.Value) {
+	entry := v.Block.Func.Entry
+	if entry != v.Block {
+		base.Fatalf("in %s, badly placed ArgIReg or ArgFReg: %v %v", v.Block.Func.Name, v.Block, v)
+	}
+}
+
 func AddrAuto(a *obj.Addr, v *ssa.Value) {
 	n, off := ssa.AutoVar(v)
 	a.Type = obj.TYPE_MEM