compiler: add object layout information to heap allocations

This commit adds object layout information to new heap allocations. It
is not yet used anywhere: the next commit will make use of it.

Object layout information will eventually be used for a (mostly) precise
garbage collector. This is what the data is made for. However, it is
also useful in the interp package which can work better if it knows the
memory layout and thus the approximate LLVM type of heap-allocated
objects.

Author: aykevl

compiler/compiler.go

@@ -23,7 +23,7 @@ import (
 // Version of the compiler pacakge. Must be incremented each time the compiler
 // package changes in a way that affects the generated LLVM module.
 // This version is independent of the TinyGo version number.
-const Version = 20 // last change: fix export math functions issue
+const Version = 21 // last change: add layout param to runtime.alloc calls
 
 func init() {
 	llvm.InitializeAllTargets()
@@ -1518,8 +1518,8 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 				return llvm.Value{}, b.makeError(expr.Pos(), fmt.Sprintf("value is too big (%v bytes)", size))
 			}
 			sizeValue := llvm.ConstInt(b.uintptrType, size, false)
-			nilPtr := llvm.ConstNull(b.i8ptrType)
-			buf := b.createRuntimeCall("alloc", []llvm.Value{sizeValue, nilPtr}, expr.Comment)
+			layoutValue := b.createObjectLayout(typ, expr.Pos())
+			buf := b.createRuntimeCall("alloc", []llvm.Value{sizeValue, layoutValue}, expr.Comment)
 			buf = b.CreateBitCast(buf, llvm.PointerType(typ, 0), "")
 			return buf, nil
 		} else {
@@ -1731,8 +1731,8 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 			return llvm.Value{}, err
 		}
 		sliceSize := b.CreateBinOp(llvm.Mul, elemSizeValue, sliceCapCast, "makeslice.cap")
-		nilPtr := llvm.ConstNull(b.i8ptrType)
-		slicePtr := b.createRuntimeCall("alloc", []llvm.Value{sliceSize, nilPtr}, "makeslice.buf")
+		layoutValue := b.createObjectLayout(llvmElemType, expr.Pos())
+		slicePtr := b.createRuntimeCall("alloc", []llvm.Value{sliceSize, layoutValue}, "makeslice.buf")
 		slicePtr = b.CreateBitCast(slicePtr, llvm.PointerType(llvmElemType, 0), "makeslice.array")
 
 		// Extend or truncate if necessary. This is safe as we've already done

compiler/compiler_test.go

@@ -54,6 +54,7 @@ func TestCompiler(t *testing.T) {
 		{"channel.go", ""},
 		{"intrinsics.go", "cortex-m-qemu"},
 		{"intrinsics.go", "wasm"},
+		{"gc.go", ""},
 	}
 
 	_, minor, err := goenv.GetGorootVersion(goenv.Get("GOROOT"))

compiler/llvm.go

@@ -1,6 +1,11 @@
 package compiler
 
 import (
+	"fmt"
+	"go/token"
+	"go/types"
+	"math/big"
+
 	"github.com/tinygo-org/tinygo/compiler/llvmutil"
 	"tinygo.org/x/go-llvm"
 )
@@ -52,3 +57,199 @@ func (c *compilerContext) makeGlobalArray(buf []byte, name string, elementType l
 	global.SetInitializer(value)
 	return global
 }
+
+// createObjectLayout returns a LLVM value (of type i8*) that describes where
+// there are pointers in the type t. If all the data fits in a word, it is
+// returned as a word. Otherwise it will store the data in a global.
+//
+// The value contains two pieces of information: the length of the object and
+// which words contain a pointer (indicated by setting the given bit to 1). For
+// arrays, only the element is stored. This works because the GC knows the
+// object size and can therefore know how this value is repeated in the object.
+func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Value {
+	// Use the element type for arrays. This works even for nested arrays.
+	for {
+		kind := t.TypeKind()
+		if kind == llvm.ArrayTypeKind {
+			t = t.ElementType()
+			continue
+		}
+		if kind == llvm.StructTypeKind {
+			fields := t.StructElementTypes()
+			if len(fields) == 1 {
+				t = fields[0]
+				continue
+			}
+		}
+		break
+	}
+
+	// Do a few checks to see whether we need to generate any object layout
+	// information at all.
+	objectSizeBytes := c.targetData.TypeAllocSize(t)
+	pointerSize := c.targetData.TypeAllocSize(c.i8ptrType)
+	pointerAlignment := c.targetData.PrefTypeAlignment(c.i8ptrType)
+	if objectSizeBytes < pointerSize {
+		// Too small to contain a pointer.
+		layout := (uint64(1) << 1) | 1
+		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.i8ptrType)
+	}
+	bitmap := c.getPointerBitmap(t, pos)
+	if bitmap.BitLen() == 0 {
+		// There are no pointers in this type, so we can simplify the layout.
+		// TODO: this can be done in many other cases, e.g. when allocating an
+		// array (like [4][]byte, which repeats a slice 4 times).
+		layout := (uint64(1) << 1) | 1
+		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.i8ptrType)
+	}
+	if objectSizeBytes%uint64(pointerAlignment) != 0 {
+		// This shouldn't happen except for packed structs, which aren't
+		// currently used.
+		c.addError(pos, "internal error: unexpected object size for object with pointer field")
+		return llvm.ConstNull(c.i8ptrType)
+	}
+	objectSizeWords := objectSizeBytes / uint64(pointerAlignment)
+
+	pointerBits := pointerSize * 8
+	var sizeFieldBits uint64
+	switch pointerBits {
+	case 16:
+		sizeFieldBits = 4
+	case 32:
+		sizeFieldBits = 5
+	case 64:
+		sizeFieldBits = 6
+	default:
+		panic("unknown pointer size")
+	}
+	layoutFieldBits := pointerBits - 1 - sizeFieldBits
+
+	// Try to emit the value as an inline integer. This is possible in most
+	// cases.
+	if objectSizeWords < layoutFieldBits {
+		// If it can be stored directly in the pointer value, do so.
+		// The runtime knows that if the least significant bit of the pointer is
+		// set, the pointer contains the value itself.
+		layout := bitmap.Uint64()<<(sizeFieldBits+1) | (objectSizeWords << 1) | 1
+		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.i8ptrType)
+	}
+
+	// Unfortunately, the object layout is too big to fit in a pointer-sized
+	// integer. Store it in a global instead.
+
+	// Try first whether the global already exists. All objects with a
+	// particular name have the same type, so this is possible.
+	globalName := "runtime/gc.layout:" + fmt.Sprintf("%d-%0*x", objectSizeWords, (objectSizeWords+15)/16, bitmap)
+	global := c.mod.NamedGlobal(globalName)
+	if !global.IsNil() {
+		return llvm.ConstBitCast(global, c.i8ptrType)
+	}
+
+	// Create the global initializer.
+	bitmapBytes := make([]byte, int(objectSizeWords+7)/8)
+	copy(bitmapBytes, bitmap.Bytes())
+	var bitmapByteValues []llvm.Value
+	for _, b := range bitmapBytes {
+		bitmapByteValues = append(bitmapByteValues, llvm.ConstInt(c.ctx.Int8Type(), uint64(b), false))
+	}
+	initializer := c.ctx.ConstStruct([]llvm.Value{
+		llvm.ConstInt(c.uintptrType, objectSizeWords, false),
+		llvm.ConstArray(c.ctx.Int8Type(), bitmapByteValues),
+	}, false)
+
+	global = llvm.AddGlobal(c.mod, initializer.Type(), globalName)
+	global.SetInitializer(initializer)
+	global.SetUnnamedAddr(true)
+	global.SetGlobalConstant(true)
+	global.SetLinkage(llvm.LinkOnceODRLinkage)
+	if c.targetData.PrefTypeAlignment(c.uintptrType) < 2 {
+		// AVR doesn't have alignment by default.
+		global.SetAlignment(2)
+	}
+	if c.Debug && pos != token.NoPos {
+		// Creating a fake global so that the value can be inspected in GDB.
+		// For example, the layout for strings.stringFinder (as of Go version
+		// 1.15) has the following type according to GDB:
+		//   type = struct {
+		//       uintptr numBits;
+		//       uint8 data[33];
+		//   }
+		// ...that's sort of a mixed C/Go type, but it is readable. More
+		// importantly, these object layout globals can be read and printed by
+		// GDB which may be useful for debugging.
+		position := c.program.Fset.Position(pos)
+		diglobal := c.dibuilder.CreateGlobalVariableExpression(c.difiles[position.Filename], llvm.DIGlobalVariableExpression{
+			Name: globalName,
+			File: c.getDIFile(position.Filename),
+			Line: position.Line,
+			Type: c.getDIType(types.NewStruct([]*types.Var{
+				types.NewVar(pos, nil, "numBits", types.Typ[types.Uintptr]),
+				types.NewVar(pos, nil, "data", types.NewArray(types.Typ[types.Byte], int64(len(bitmapByteValues)))),
+			}, nil)),
+			LocalToUnit: false,
+			Expr:        c.dibuilder.CreateExpression(nil),
+		})
+		global.AddMetadata(0, diglobal)
+	}
+
+	return llvm.ConstBitCast(global, c.i8ptrType)
+}
+
+// getPointerBitmap scans the given LLVM type for pointers and sets bits in a
+// bigint at the word offset that contains a pointer. This scan is recursive.
+func (c *compilerContext) getPointerBitmap(typ llvm.Type, pos token.Pos) *big.Int {
+	alignment := c.targetData.PrefTypeAlignment(c.i8ptrType)
+	switch typ.TypeKind() {
+	case llvm.IntegerTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
+		return big.NewInt(0)
+	case llvm.PointerTypeKind:
+		return big.NewInt(1)
+	case llvm.StructTypeKind:
+		ptrs := big.NewInt(0)
+		if typ.StructName() == "runtime.funcValue" {
+			// Hack: the type runtime.funcValue contains an 'id' field which is
+			// of type uintptr, but before the LowerFuncValues pass it actually
+			// contains a pointer (ptrtoint) to a global. This trips up the
+			// interp package. Therefore, make the id field a pointer for now.
+			typ = c.ctx.StructType([]llvm.Type{c.i8ptrType, c.i8ptrType}, false)
+		}
+		for i, subtyp := range typ.StructElementTypes() {
+			subptrs := c.getPointerBitmap(subtyp, pos)
+			if subptrs.BitLen() == 0 {
+				continue
+			}
+			offset := c.targetData.ElementOffset(typ, i)
+			if offset%uint64(alignment) != 0 {
+				// This error will let the compilation fail, but by continuing
+				// the error can still easily be shown.
+				c.addError(pos, "internal error: allocated struct contains unaligned pointer")
+				continue
+			}
+			subptrs.Lsh(subptrs, uint(offset)/uint(alignment))
+			ptrs.Or(ptrs, subptrs)
+		}
+		return ptrs
+	case llvm.ArrayTypeKind:
+		subtyp := typ.ElementType()
+		subptrs := c.getPointerBitmap(subtyp, pos)
+		ptrs := big.NewInt(0)
+		if subptrs.BitLen() == 0 {
+			return ptrs
+		}
+		elementSize := c.targetData.TypeAllocSize(subtyp)
+		if elementSize%uint64(alignment) != 0 {
+			// This error will let the compilation fail (but continues so that
+			// other errors can be shown).
+			c.addError(pos, "internal error: allocated array contains unaligned pointer")
+			return ptrs
+		}
+		for i := 0; i < typ.ArrayLength(); i++ {
+			ptrs.Lsh(ptrs, uint(elementSize)/uint(alignment))
+			ptrs.Or(ptrs, subptrs)
+		}
+		return ptrs
+	default:
+		// Should not happen.
+		panic("unknown LLVM type")
+	}
+}

compiler/testdata/gc.go

@@ -0,0 +1,72 @@
+package main
+
+var (
+	scalar1 *byte
+	scalar2 *int32
+	scalar3 *int64
+	scalar4 *float32
+
+	array1 *[3]byte
+	array2 *[71]byte
+	array3 *[3]*byte
+
+	struct1 *struct{}
+	struct2 *struct {
+		x int
+		y int
+	}
+	struct3 *struct {
+		x *byte
+		y [60]uintptr
+		z *byte
+	}
+
+	slice1 []byte
+	slice2 []*int
+	slice3 [][]byte
+)
+
+func newScalar() {
+	scalar1 = new(byte)
+	scalar2 = new(int32)
+	scalar3 = new(int64)
+	scalar4 = new(float32)
+}
+
+func newArray() {
+	array1 = new([3]byte)
+	array2 = new([71]byte)
+	array3 = new([3]*byte)
+}
+
+func newStruct() {
+	struct1 = new(struct{})
+	struct2 = new(struct {
+		x int
+		y int
+	})
+	struct3 = new(struct {
+		x *byte
+		y [60]uintptr
+		z *byte
+	})
+}
+
+func newFuncValue() *func() {
+	// On some platforms that use runtime.funcValue ("switch" style) function
+	// values, a func value is allocated as having two pointer words while the
+	// struct looks like {unsafe.Pointer; uintptr}. This is so that the interp
+	// package won't get confused, see getPointerBitmap in compiler/llvm.go for
+	// details.
+	return new(func())
+}
+
+func makeSlice() {
+	slice1 = make([]byte, 5)
+	slice2 = make([]*int, 5)
+	slice3 = make([][]byte, 5)
+}
+
+func makeInterface(v complex128) interface{} {
+	return v // always stored in an allocation
+}