fspl/llvm/gep.go

package llvm

import "fmt"

type gepIndex struct {
	HasVal    bool
	Val       int64
	VectorLen uint64
}

func gepInstType (elemType, src Type, indices []Value) Type {
	var idxs []gepIndex
	for _, index := range indices {
		var idx gepIndex
		switch index := index.(type) {
		case Const:
			idx = getIndex(index)
		default:
			idx = gepIndex{HasVal: false}
			// Check if index is of vector type.
			if indexType, ok := index.Type().(*TypeVector); ok {
				idx.VectorLen = indexType.Length
			}
		}
		idxs = append(idxs, idx)
	}
	return resultType(elemType, src, idxs)
}

func getIndex (index Const) gepIndex {
	// unpack inrange indices.
	if idx, ok := index.(*ConstIndex); ok {
		index = idx.Const
	}
	// TODO: figure out how to simplify expressions for GEP instructions without
	// creating import cycle on irutil.

	// Use index.Simplify() to simplify the constant expression to a concrete
	// integer constant or vector of integers constant.
	//if idx, ok := index.(constant.Expression); ok {
	//	index = idx.Simplify()
	//}
	switch index := index.(type) {
	case *ConstInt:
		return gepIndex {
			HasVal: true,
			Val:    index.Value.Int64(),
		}
	case *ConstZeroInitializer:
		return gepIndex {
			HasVal: true,
			Val:    0,
		}
	case *ConstVector:
		// ref: https://llvm.org/docs/LangRef.html#getelementptr-instruction
		//
		// > The getelementptr returns a vector of pointers, instead of a single
		// > address, when one or more of its arguments is a vector. In such
		// > cases, all vector arguments should have the same number of elements,
		// > and every scalar argument will be effectively broadcast into a vector
		// > during address calculation.
		if len(index.Elements) == 0 {
			return gepIndex { HasVal: false }
		}
		// Sanity check. All vector elements must be integers, and must have the
		// same value.
		var val int64
		for i, elem := range index.Elements {
			switch elem := elem.(type) {
			case *ConstInt:
				x := elem.Value.Int64()
				if i == 0 {
					val = x
				} else if x != val {
					// since all elements were not identical, we must conclude that
					// the index vector does not have a concrete value.
					return gepIndex {
						HasVal:    false,
						VectorLen: uint64(len(index.Elements)),
					}
				}
			default:
				// TODO: remove debug output.
				panic(fmt.Errorf("support for gep index vector element type %T not yet implemented", elem))
				//return gep.Index{HasVal: false}
			}
		}
		return gepIndex {
			HasVal:    true,
			Val:       val,
			VectorLen: uint64(len(index.Elements)),
		}
	case *ConstUndef:
		return gepIndex { HasVal: false }
	case *ConstPoison:
		return gepIndex { HasVal: false }
	case ConstExpr:
		// should already have been simplified to a form we can handle.
		return gepIndex { HasVal: false }
	default:
		// TODO: add support for more constant expressions.
		// TODO: remove debug output.
		panic(fmt.Errorf("support for gep index type %T not yet implemented", index))
		//return gep.Index{HasVal: false}
	}
}

func resultType (elemType, src Type, indices []gepIndex) Type {
	// ref: http://llvm.org/docs/GetElementPtr.html#what-effect-do-address-spaces-have-on-geps
	//
	// > the address space qualifier on the second operand pointer type always
	// > matches the address space qualifier on the result type.
	var (
		// Address space of src pointer type or src vector element pointer type.
		addrSpace AddressSpace
		// Length of vector of pointers result type; or 0 if pointer result type.
		resultVectorLength uint64
	)
	// ref: https://llvm.org/docs/LangRef.html#getelementptr-instruction
	//
	// > The second argument is always a pointer or a vector of pointers.
	switch src := src.(type) {
	case *TypePointer:
		addrSpace = src.AddressSpace
	case *TypeVector:
		vectorElemType, ok := src.Element.(*TypePointer)
		if !ok {
			panic(fmt.Errorf("invalid gep source vector element type %T", src.Element))
		}
		addrSpace = vectorElemType.AddressSpace
		resultVectorLength = src.Length
	default:
		panic(fmt.Errorf("invalid gep source type %T", src))
	}
	// ref: https://llvm.org/docs/LangRef.html#getelementptr-instruction
	//
	// > The first argument is always a type used as the basis for the
	// > calculations.
	e := elemType
	for i, index := range indices {
		// ref: https://llvm.org/docs/LangRef.html#getelementptr-instruction
		//
		// > The getelementptr returns a vector of pointers, instead of a single
		// > address, when one or more of its arguments is a vector. In such
		// > cases, all vector arguments should have the same number of elements,
		// > and every scalar argument will be effectively broadcast into a vector
		// > during address calculation.
		if index.VectorLen != 0 && resultVectorLength != 0 && index.VectorLen != resultVectorLength {
			panic(fmt.Errorf("vector length mismatch of index vector (%d) and result type vector (%d)", index.VectorLen, resultVectorLength))
		}
		if resultVectorLength == 0 && index.VectorLen != 0 {
			resultVectorLength = index.VectorLen
		}
		// ref: https://llvm.org/docs/GetElementPtr.html#why-is-the-extra-0-index-required
		//
		// > Since the second argument to the GEP instruction must always be a
		// > value of pointer type, the first index steps through that pointer.
		if i == 0 {
			continue
		}
		switch elm := e.(type) {
		case *TypePointer:
			panic(fmt.Errorf("cannot index into pointer type at %d:th gep index, only valid at 0:th gep index; see https://llvm.org/docs/GetElementPtr.html#what-is-dereferenced-by-gep", i))
		case *TypeArray:
			e = elm.Element
		case *TypeStruct:
			// ref: https://llvm.org/docs/LangRef.html#getelementptr-instruction
			//
			// > When indexing into a (optionally packed) structure, only i32
			// > integer constants are allowed (when using a vector of indices they
			// > must all be the same i32 integer constant).
			if !index.HasVal {
				panic(fmt.Errorf("unable to index into struct type `%v` using gep with non-constant index", e))
			}
			e = elm.Fields[index.Val]
		case *TypeDefined:
			continue
		default:
			panic(fmt.Errorf("cannot index into type %T using gep", e))
		}
	}
	ptr := &TypePointer { AddressSpace: addrSpace }
	if resultVectorLength != 0 {
		vec := &TypeVector {
			Element: ptr,
			Length:  resultVectorLength,
		}
		return vec
	}
	return ptr
}