cmd/compile: implement too-big-to-SSA struct passing in registers

Added a test that exercises named results

Change-Id: Ie228b68f4f846266595a95e0f65a6e4b8bf79635
Reviewed-on: https://go-review.googlesource.com/c/go/+/297029
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

@@ -118,12 +118,22 @@ func RegisterTypes(apa []ABIParamAssignment) []*types.Type {
 		if len(pa.Registers) == 0 {
 			continue
 		}
-		rts = appendParamRegs(rts, pa.Type)
+		rts = appendParamTypes(rts, pa.Type)
 	}
 	return rts
 }
 
-func appendParamRegs(rts []*types.Type, t *types.Type) []*types.Type {
+func (pa *ABIParamAssignment) RegisterTypesAndOffsets() ([]*types.Type, []int64) {
+	l := len(pa.Registers)
+	if l == 0 {
+		return nil, nil
+	}
+	typs := make([]*types.Type, 0, l)
+	offs := make([]int64, 0, l)
+	return appendParamTypes(typs, pa.Type), appendParamOffsets(offs, 0, pa.Type)
+}
+
+func appendParamTypes(rts []*types.Type, t *types.Type) []*types.Type {
 	if t.IsScalar() || t.IsPtrShaped() {
 		if t.IsComplex() {
 			c := types.FloatForComplex(t)
@@ -146,25 +156,60 @@ func appendParamRegs(rts []*types.Type, t *types.Type) []*types.Type {
 		switch typ {
 		case types.TARRAY:
 			for i := int64(0); i < t.Size(); i++ { // 0 gets no registers, plus future-proofing.
-				rts = appendParamRegs(rts, t.Elem())
+				rts = appendParamTypes(rts, t.Elem())
 			}
 		case types.TSTRUCT:
 			for _, f := range t.FieldSlice() {
 				if f.Type.Size() > 0 { // embedded zero-width types receive no registers
-					rts = appendParamRegs(rts, f.Type)
+					rts = appendParamTypes(rts, f.Type)
 				}
 			}
 		case types.TSLICE:
-			return appendParamRegs(rts, synthSlice)
+			return appendParamTypes(rts, synthSlice)
 		case types.TSTRING:
-			return appendParamRegs(rts, synthString)
+			return appendParamTypes(rts, synthString)
 		case types.TINTER:
-			return appendParamRegs(rts, synthIface)
+			return appendParamTypes(rts, synthIface)
 		}
 	}
 	return rts
 }
 
+// appendParamOffsets appends the offset(s) of type t, starting from "at",
+// to input offsets, and returns the longer slice.
+func appendParamOffsets(offsets []int64, at int64, t *types.Type) []int64 {
+	at = align(at, t)
+	if t.IsScalar() || t.IsPtrShaped() {
+		if t.IsComplex() || int(t.Width) > types.RegSize { // complex and *int64 on 32-bit
+			s := t.Width / 2
+			return append(offsets, at, at+s)
+		} else {
+			return append(offsets, at)
+		}
+	} else {
+		typ := t.Kind()
+		switch typ {
+		case types.TARRAY:
+			for i := int64(0); i < t.NumElem(); i++ {
+				offsets = appendParamOffsets(offsets, at, t.Elem())
+			}
+			return offsets
+		case types.TSTRUCT:
+			for _, f := range t.FieldSlice() {
+				offsets = appendParamOffsets(offsets, at, f.Type)
+				at += f.Type.Width
+			}
+		case types.TSLICE:
+			return appendParamOffsets(offsets, at, synthSlice)
+		case types.TSTRING:
+			return appendParamOffsets(offsets, at, synthString)
+		case types.TINTER:
+			return appendParamOffsets(offsets, at, synthIface)
+		}
+	}
+	return offsets
+}
+
 // SpillOffset returns the offset *within the spill area* for the parameter that "a" describes.
 // Registers will be spilled here; if a memory home is needed (for a pointer method e.g.)
 // then that will be the address.

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

@@ -657,7 +657,7 @@ func (x *expandState) storeArgOrLoad(pos src.XPos, b *Block, source, mem *Value,
 	case OpCopy:
 		return x.storeArgOrLoad(pos, b, source.Args[0], mem, t, offset, loadRegOffset, storeRc)
 
-	case OpLoad:
+	case OpLoad, OpDereference:
 		ret := x.decomposeArgOrLoad(pos, b, source, mem, t, offset, loadRegOffset, storeRc, storeOneLoad, storeTwoLoad)
 		if ret != nil {
 			return ret
@@ -820,6 +820,9 @@ func (x *expandState) storeArgOrLoad(pos src.XPos, b *Block, source, mem *Value,
 	}
 
 	s := mem
+	if source.Op == OpDereference {
+		source.Op = OpLoad // For purposes of parameter passing expansion, a Dereference is a Load.
+	}
 	if storeRc.hasRegs() {
 		storeRc.addArg(source)
 	} else {
@@ -846,14 +849,11 @@ func (x *expandState) rewriteArgs(v *Value, firstArg int) (*Value, []*Value) {
 		auxI := int64(i)
 		aRegs := aux.RegsOfArg(auxI)
 		aType := aux.TypeOfArg(auxI)
-		if a.Op == OpDereference {
+		if len(aRegs) == 0 && 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())
 			}
-			if len(aRegs) > 0 {
-				x.f.Fatalf("Not implemented yet, not-SSA-type %v passed in registers", aType)
-			}
 			// "Dereference" of addressed (probably not-SSA-eligible) value becomes Move
 			// 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)
@@ -969,10 +969,7 @@ func expandCalls(f *Func) {
 				auxOffset := int64(0)
 				auxSize := aux.SizeOfResult(i)
 				aRegs := aux.RegsOfResult(int64(j))
-				if a.Op == OpDereference {
-					if len(aRegs) > 0 {
-						x.f.Fatalf("Not implemented yet, not-SSA-type %v returned in register", auxType)
-					}
+				if len(aRegs) == 0 && a.Op == OpDereference {
 					// Avoid a self-move, and if one is detected try to remove the already-inserted VarDef for the assignment that won't happen.
 					if dAddr, dMem := a.Args[0], a.Args[1]; dAddr.Op == OpLocalAddr && dAddr.Args[0].Op == OpSP &&
 						dAddr.Args[1] == dMem && dAddr.Aux == aux.results[i].Name {
@@ -1392,3 +1389,14 @@ func (x *expandState) newArgToMemOrRegs(baseArg, toReplace *Value, offset int64,
 		}
 	}
 }
+
+// argOpAndRegisterFor converts an abi register index into an ssa Op and corresponding
+// arg register index.
+// TODO could call this in at least two places earlier in this file.
+func ArgOpAndRegisterFor(r abi.RegIndex, abiConfig *abi.ABIConfig) (Op, int64) {
+	i := abiConfig.FloatIndexFor(r)
+	if i >= 0 { // float PR
+		return OpArgFloatReg, i
+	}
+	return OpArgIntReg, int64(r)
+}

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

@@ -555,19 +555,26 @@ func buildssa(fn *ir.Func, worker int) *ssa.Func {
 				}
 				s.vars[n] = v
 				s.addNamedValue(n, v) // This helps with debugging information, not needed for compilation itself.
-			} else if !s.canSSAName(n) { // I.e., the address was taken.  The type may or may not be okay.
-				// If the value will arrive in registers,
-				// AND if it can be SSA'd (if it cannot, panic for now),
-				// THEN
-				// (1) receive it as an OpArg (but do not store its name in the var table)
-				// (2) store it to its spill location, which is its address as well.
+			} else { // address was taken AND/OR too large for SSA
 				paramAssignment := ssa.ParamAssignmentForArgName(s.f, n)
 				if len(paramAssignment.Registers) > 0 {
-					if !TypeOK(n.Type()) { // TODO register args -- if v is not an SSA-able type, must decompose, here.
-						panic(fmt.Errorf("Arg in registers is too big to be SSA'd, need to implement decomposition, type=%v, n=%v", n.Type(), n))
+					if TypeOK(n.Type()) { // SSA-able type, so address was taken -- receive value in OpArg, DO NOT bind to var, store immediately to memory.
+						v := s.newValue0A(ssa.OpArg, n.Type(), n)
+						s.store(n.Type(), s.decladdrs[n], v)
+					} else { // Too big for SSA.
+						// Brute force, and early, do a bunch of stores from registers
+						// TODO fix the nasty storeArgOrLoad recursion in ssa/expand_calls.go so this Just Works with store of a big Arg.
+						typs, offs := paramAssignment.RegisterTypesAndOffsets()
+						for i, t := range typs {
+							o := offs[i]
+							r := paramAssignment.Registers[i]
+							op, reg := ssa.ArgOpAndRegisterFor(r, s.f.ABISelf)
+							v := s.newValue0I(op, t, reg)
+							v.Aux = &ssa.AuxNameOffset{Name: n, Offset: o}
+							p := s.newValue1I(ssa.OpOffPtr, types.NewPtr(n.Type()), o, s.decladdrs[n])
+							s.store(t, p, v)
+						}
 					}
-					v := s.newValue0A(ssa.OpArg, n.Type(), n)
-					s.store(n.Type(), s.decladdrs[n], v)
 				}
 			}
 		}

test/abi/named_results.go

@@ -0,0 +1,91 @@
+// run
+
+//go:build !wasm
+// +build !wasm
+
+// 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 main
+
+import (
+	"fmt"
+)
+
+var sink *string
+
+var y int
+
+//go:registerparams
+//go:noinline
+func F(a, b, c *int) (x int) {
+	x = *a
+	G(&x)
+	x += *b
+	G(&x)
+	x += *c
+	G(&x)
+	return
+}
+
+//go:registerparams
+//go:noinline
+func G(x *int) {
+	y += *x
+	fmt.Println("y = ", y)
+}
+
+//go:registerparams
+//go:noinline
+func X() {
+	*sink += " !!!!!!!!!!!!!!!"
+}
+
+//go:registerparams
+//go:noinline
+func H(s, t string) (result string) { // result leaks to heap
+	result = "Aloha! " + s + " " + t
+	sink = &result
+	r := ""
+	if len(s) <= len(t) {
+		r = "OKAY! "
+		X()
+	}
+	return r + result
+}
+
+//go:registerparams
+//go:noinline
+func K(s, t string) (result string) { // result spills
+	result = "Aloha! " + s + " " + t
+	r := ""
+	if len(s) <= len(t) {
+		r = "OKAY! "
+		X()
+	}
+	return r + result
+}
+
+func main() {
+	a, b, c := 1, 4, 16
+	x := F(&a, &b, &c)
+	fmt.Printf("x = %d\n", x)
+
+	y := H("Hello", "World!")
+	fmt.Println("len(y) =", len(y))
+	fmt.Println("y =", y)
+	z := H("Hello", "Pal!")
+	fmt.Println("len(z) =", len(z))
+	fmt.Println("z =", z)
+
+	fmt.Println()
+
+	y = K("Hello", "World!")
+	fmt.Println("len(y) =", len(y))
+	fmt.Println("y =", y)
+	z = K("Hello", "Pal!")
+	fmt.Println("len(z) =", len(z))
+	fmt.Println("z =", z)
+
+}

test/abi/named_results.out

@@ -0,0 +1,13 @@
+y =  1
+y =  6
+y =  27
+x = 21
+len(y) = 41
+y = OKAY! Aloha! Hello World! !!!!!!!!!!!!!!!
+len(z) = 17
+z = Aloha! Hello Pal!
+
+len(y) = 25
+y = OKAY! Aloha! Hello World!
+len(z) = 17
+z = Aloha! Hello Pal!