Merge remote-tracking branch 'origin/adv-expr' into adv-expr-eliminate-bridging

# Conflicts:
#	kmath-ast/src/jvmMain/kotlin/scientifik/kmath/asm/asm.kt
#	kmath-ast/src/jvmMain/kotlin/scientifik/kmath/asm/internal/AsmBuilder.kt
This commit is contained in:
Iaroslav 2020-06-24 15:53:19 +07:00
commit f475335df1
No known key found for this signature in database
GPG Key ID: 46E15E4A31B3BCD7
8 changed files with 195 additions and 26 deletions

View File

@ -1,4 +1,61 @@
# AST based expression representation and operations
# AST-based expression representation and operations (`kmath-ast`)
## Dynamic expression code generation
Contributed by [Iaroslav Postovalov](https://github.com/CommanderTvis).
This subproject implements the following features:
- Expression Language and its parser.
- MST as expression language's syntax intermediate representation.
- Type-safe builder of MST.
- Evaluating expressions by traversing MST.
## Dynamic expression code generation with OW2 ASM
`kmath-ast` JVM module supports runtime code generation to eliminate overhead of tree traversal. Code generator builds
a special implementation of `Expression<T>` with implemented `invoke` function.
For example, the following builder:
```kotlin
RealField.mstInField { symbol("x") + 2 }.compile()
```
… leads to generation of bytecode, which can be decompiled to the following Java class:
```java
package scientifik.kmath.asm.generated;
import java.util.Map;
import scientifik.kmath.asm.internal.AsmCompiledExpression;
import scientifik.kmath.operations.Algebra;
import scientifik.kmath.operations.RealField;
// The class's name is build with MST's hash-code and collision fixing number.
public final class AsmCompiledExpression_45045_0 extends AsmCompiledExpression<Double> {
// Plain constructor
public AsmCompiledExpression_45045_0(Algebra algebra, Object[] constants) {
super(algebra, constants);
}
// The actual dynamic code:
public final Double invoke(Map<String, ? extends Double> arguments) {
return (Double)((RealField)super.algebra).add((Double)arguments.get("x"), (Double)2.0D);
}
}
```
### Example Usage
This API is an extension to MST and MSTExpression APIs. You may optimize both MST and MSTExpression:
```kotlin
RealField.mstInField { symbol("x") + 2 }.compile()
RealField.expression("2+2".parseMath())
```
### Known issues
- Using numeric algebras causes boxing and calling bridge methods.
- The same classes may be generated and loaded twice, so it is recommended to cache compiled expressions to avoid
class loading overhead.
- This API is not supported by non-dynamic JVM implementations (like TeaVM and GraalVM) because of using class loaders.
Contributed by [Iaroslav Postovalov](https://github.com/CommanderTvis).

View File

@ -20,7 +20,8 @@ fun <T : Any> MST.compileWith(type: KClass<T>, algebra: Algebra<T>): Expression<
is MST.Symbolic -> loadVariable(node.value)
is MST.Numeric -> {
val constant = if (algebra is NumericAlgebra<T>)
algebra.number(node.value) else
algebra.number(node.value)
else
error("Number literals are not supported in $algebra")
loadTConstant(constant)
@ -72,9 +73,13 @@ fun <T : Any> MST.compileWith(type: KClass<T>, algebra: Algebra<T>): Expression<
/**
* Compile an [MST] to ASM using given algebra
*/
inline fun <reified T : Any> Algebra<T>.expresion(mst: MST): Expression<T> = mst.compileWith(T::class, this)
inline fun <reified T : Any> Algebra<T>.expression(mst: MST): Expression<T> = mst.compileWith(T::class, this)
/**
* Optimize performance of an [MSTExpression] using ASM codegen
*/
inline fun <reified T : Any> MSTExpression<T>.compile(): Expression<T> = mst.compileWith(T::class, algebra)
inline fun <reified T : Any> MSTExpression<T>.compile(): Expression<T> = mst.compileWith(T::class, algebra)
fun main() {
RealField.mstInField { symbol("x") + 2 }.compile()
}

View File

@ -6,7 +6,6 @@ import org.objectweb.asm.MethodVisitor
import org.objectweb.asm.Opcodes
import scientifik.kmath.asm.internal.AsmBuilder.ClassLoader
import scientifik.kmath.operations.Algebra
import java.io.File
/**
* ASM Builder is a structure that abstracts building a class that unwraps [AsmExpression] to plain Java expression.
@ -22,10 +21,16 @@ internal class AsmBuilder<T> internal constructor(
private val className: String,
private val invokeLabel0Visitor: AsmBuilder<T>.() -> Unit
) {
/**
* Internal classloader of [AsmBuilder] with alias to define class from byte array.
*/
private class ClassLoader(parent: java.lang.ClassLoader) : java.lang.ClassLoader(parent) {
internal fun defineClass(name: String?, b: ByteArray): Class<*> = defineClass(name, b, 0, b.size)
}
/**
* The instance of [ClassLoader] used by this builder.
*/
private val classLoader: ClassLoader =
ClassLoader(javaClass.classLoader)
@ -36,16 +41,41 @@ internal class AsmBuilder<T> internal constructor(
private val T_CLASS: String = classOfT.name.replace('.', '/')
private val slashesClassName: String = className.replace(oldChar = '.', newChar = '/')
/**
* Index of `this` variable in invoke method of [AsmCompiledExpression] built subclass.
*/
private val invokeThisVar: Int = 0
/**
* Index of `arguments` variable in invoke method of [AsmCompiledExpression] built subclass.
*/
private val invokeArgumentsVar: Int = 1
/**
* List of constants to provide to [AsmCompiledExpression] subclass.
*/
private val constants: MutableList<Any> = mutableListOf()
/**
* Method visitor of `invoke` method of [AsmCompiledExpression] subclass.
*/
private lateinit var invokeMethodVisitor: MethodVisitor
private var primitiveMode = false
internal var primitiveType = OBJECT_CLASS
internal var primitiveTypeSig = "L$OBJECT_CLASS;"
internal var primitiveTypeReturnSig = "L$OBJECT_CLASS;"
/**
* The cache of [AsmCompiledExpression] subclass built by this builder.
*/
private var generatedInstance: AsmCompiledExpression<T>? = null
/**
* Subclasses, loads and instantiates the [AsmCompiledExpression] for given parameters.
*
* The built instance is cached.
*/
@Suppress("UNCHECKED_CAST")
fun getInstance(): AsmCompiledExpression<T> {
generatedInstance?.let { return it }
@ -187,8 +217,6 @@ internal class AsmBuilder<T> internal constructor(
visitEnd()
}
File("dump.class").writeBytes(classWriter.toByteArray())
val new = classLoader
.defineClass(className, classWriter.toByteArray())
.constructors
@ -199,6 +227,9 @@ internal class AsmBuilder<T> internal constructor(
return new
}
/**
* Loads a constant from
*/
internal fun loadTConstant(value: T) {
if (primitiveMode) {
loadNumberConstant(value as Number)
@ -214,6 +245,9 @@ internal class AsmBuilder<T> internal constructor(
loadConstant(value as Any, T_CLASS)
}
/**
* Loads an object constant [value] stored in [AsmCompiledExpression.constants] and casts it to [type].
*/
private fun unboxInPrimitiveMode() {
if (!primitiveMode)
return
@ -241,13 +275,17 @@ internal class AsmBuilder<T> internal constructor(
visitGetField(owner = slashesClassName, name = "constants", descriptor = "[L$OBJECT_CLASS;")
visitLdcOrIntConstant(idx)
visitGetObjectArrayElement()
visitCheckCast(type)
unboxInPrimitiveMode()
invokeMethodVisitor.visitCheckCast(type)
}
}
private fun loadThis(): Unit = invokeMethodVisitor.visitLoadObjectVar(invokeThisVar)
/**
* Either loads a numeric constant [value] from [AsmCompiledExpression.constants] field or boxes a primitive
* constant from the constant pool (some numbers with special opcodes like [Opcodes.ICONST_0] aren't even loaded
* from it).
*/
private fun loadNumberConstant(value: Number) {
val clazz = value.javaClass
val c = clazz.name.replace('.', '/')
@ -255,9 +293,12 @@ internal class AsmBuilder<T> internal constructor(
if (sigLetter != null) {
when (value) {
is Byte -> invokeMethodVisitor.visitLdcOrIntConstant(value.toInt())
is Short -> invokeMethodVisitor.visitLdcOrIntConstant(value.toInt())
is Int -> invokeMethodVisitor.visitLdcOrIntConstant(value)
is Double -> invokeMethodVisitor.visitLdcOrDoubleConstant(value)
is Float -> invokeMethodVisitor.visitLdcOrFloatConstant(value)
is Long -> invokeMethodVisitor.visitLdcOrLongConstant(value)
else -> invokeMethodVisitor.visitLdcInsn(value)
}
@ -270,6 +311,9 @@ internal class AsmBuilder<T> internal constructor(
loadConstant(value, c)
}
/**
* Loads a variable [name] from [AsmCompiledExpression.invoke] [Map] parameter. The [defaultValue] may be provided.
*/
internal fun loadVariable(name: String, defaultValue: T? = null): Unit = invokeMethodVisitor.run {
visitLoadObjectVar(invokeArgumentsVar)
@ -299,6 +343,9 @@ internal class AsmBuilder<T> internal constructor(
unboxInPrimitiveMode()
}
/**
* Loads algebra from according field of [AsmCompiledExpression] and casts it to class of [algebra] provided.
*/
internal fun loadAlgebra() {
loadThis()
@ -311,6 +358,13 @@ internal class AsmBuilder<T> internal constructor(
invokeMethodVisitor.visitCheckCast(T_ALGEBRA_CLASS)
}
/**
* Writes a method instruction of opcode with its [owner], [method] and its [descriptor]. The default opcode is
* [Opcodes.INVOKEINTERFACE], since most Algebra functions are declared in interface. [loadAlgebra] should be
* called before the arguments and this operation.
*
* The result is casted to [T] automatically.
*/
internal fun invokeAlgebraOperation(
owner: String,
method: String,
@ -323,9 +377,15 @@ internal class AsmBuilder<T> internal constructor(
unboxInPrimitiveMode()
}
/**
* Writes a LDC Instruction with string constant provided.
*/
internal fun loadStringConstant(string: String): Unit = invokeMethodVisitor.visitLdcInsn(string)
internal companion object {
/**
* Maps JVM primitive numbers boxed types to their letters of JVM signature convention.
*/
private val SIGNATURE_LETTERS: Map<Class<out Any>, String> by lazy {
mapOf(
java.lang.Byte::class.java to "B",
@ -348,6 +408,9 @@ internal class AsmBuilder<T> internal constructor(
)
}
/**
* Provides boxed number types values of which can be stored in JVM bytecode constant pool.
*/
private val INLINABLE_NUMBERS: Set<Class<out Any>> by lazy { SIGNATURE_LETTERS.keys }
internal const val FUNCTIONAL_COMPILED_EXPRESSION_CLASS =

View File

@ -3,6 +3,13 @@ package scientifik.kmath.asm.internal
import scientifik.kmath.expressions.Expression
import scientifik.kmath.operations.Algebra
/**
* [Expression] partial implementation to have it subclassed by actual implementations. Provides unified storage for
* objects needed to implement the expression.
*
* @property algebra the algebra to delegate calls.
* @property constants the constants array to have persistent objects to reference in [invoke].
*/
internal abstract class AsmCompiledExpression<T> internal constructor(
@JvmField protected val algebra: Algebra<T>,
@JvmField protected val constants: Array<Any>

View File

@ -2,7 +2,13 @@ package scientifik.kmath.asm.internal
import scientifik.kmath.ast.MST
internal fun buildName(mst: MST, collision: Int = 0): String {
/**
* Creates a class name for [AsmCompiledExpression] subclassed to implement [mst] provided.
*
* This methods helps to avoid collisions of class name to prevent loading several classes with the same name. If there
* is a colliding class, change [collision] parameter or leave it `0` to check existing classes recursively.
*/
internal tailrec fun buildName(mst: MST, collision: Int = 0): String {
val name = "scientifik.kmath.asm.generated.AsmCompiledExpression_${mst.hashCode()}_$collision"
try {

View File

@ -11,6 +11,8 @@ internal fun MethodVisitor.visitLdcOrIntConstant(value: Int): Unit = when (value
3 -> visitInsn(ICONST_3)
4 -> visitInsn(ICONST_4)
5 -> visitInsn(ICONST_5)
in -128..127 -> visitIntInsn(BIPUSH, value)
in -32768..32767 -> visitIntInsn(SIPUSH, value)
else -> visitLdcInsn(value)
}
@ -20,6 +22,12 @@ internal fun MethodVisitor.visitLdcOrDoubleConstant(value: Double): Unit = when
else -> visitLdcInsn(value)
}
internal fun MethodVisitor.visitLdcOrLongConstant(value: Long): Unit = when (value) {
0L -> visitInsn(LCONST_0)
1L -> visitInsn(LCONST_1)
else -> visitLdcInsn(value)
}
internal fun MethodVisitor.visitLdcOrFloatConstant(value: Float): Unit = when (value) {
0f -> visitInsn(FCONST_0)
1f -> visitInsn(FCONST_1)

View File

@ -3,21 +3,38 @@ package scientifik.kmath.asm.internal
import org.objectweb.asm.Opcodes
import scientifik.kmath.operations.Algebra
private val methodNameAdapters: Map<String, String> = mapOf("+" to "add", "*" to "multiply", "/" to "divide")
private val methodNameAdapters: Map<String, String> by lazy {
hashMapOf(
"+" to "add",
"*" to "multiply",
"/" to "divide"
)
}
/**
* Checks if the target [context] for code generation contains a method with needed [name] and [arity].
*
* @return `true` if contains, else `false`.
*/
internal fun <T> hasSpecific(context: Algebra<T>, name: String, arity: Int): Boolean {
val aName = methodNameAdapters[name] ?: name
context::class.java.methods.find { it.name == aName && it.parameters.size == arity }
context.javaClass.methods.find { it.name == aName && it.parameters.size == arity }
?: return false
return true
}
/**
* Checks if the target [context] for code generation contains a method with needed [name] and [arity] and inserts
* [AsmBuilder.invokeAlgebraOperation] of this method.
*
* @return `true` if contains, else `false`.
*/
internal fun <T> AsmBuilder<T>.tryInvokeSpecific(context: Algebra<T>, name: String, arity: Int): Boolean {
val aName = methodNameAdapters[name] ?: name
context::class.java.methods.find { it.name == aName && it.parameters.size == arity }
context.javaClass.methods.find { it.name == aName && it.parameters.size == arity }
?: return false
val owner = context::class.java.name.replace('.', '/')
@ -33,8 +50,7 @@ internal fun <T> AsmBuilder<T>.tryInvokeSpecific(context: Algebra<T>, name: Stri
owner = owner,
method = aName,
descriptor = sig,
opcode = Opcodes.INVOKEVIRTUAL,
isInterface = false
opcode = Opcodes.INVOKEVIRTUAL
)
return true

View File

@ -36,12 +36,10 @@ internal class FunctionalConstProductExpression<T>(
/**
* A context class for [Expression] construction.
*
* @param algebra The algebra to provide for Expressions built.
*/
interface FunctionalExpressionAlgebra<T, A : Algebra<T>> : ExpressionAlgebra<T, Expression<T>> {
/**
* The algebra to provide for Expressions built.
*/
val algebra: A
abstract class FunctionalExpressionAlgebra<T, A : Algebra<T>>(val algebra: A) : ExpressionAlgebra<T, Expression<T>> {
/**
* Builds an Expression of constant expression which does not depend on arguments.
@ -69,8 +67,8 @@ interface FunctionalExpressionAlgebra<T, A : Algebra<T>> : ExpressionAlgebra<T,
/**
* A context class for [Expression] construction for [Space] algebras.
*/
open class FunctionalExpressionSpace<T, A : Space<T>>(override val algebra: A) :
FunctionalExpressionAlgebra<T, A>, Space<Expression<T>> {
open class FunctionalExpressionSpace<T, A : Space<T>>(algebra: A) :
FunctionalExpressionAlgebra<T, A>(algebra), Space<Expression<T>> {
override val zero: Expression<T> get() = const(algebra.zero)
@ -98,7 +96,7 @@ open class FunctionalExpressionSpace<T, A : Space<T>>(override val algebra: A) :
super<FunctionalExpressionAlgebra>.binaryOperation(operation, left, right)
}
open class FunctionalExpressionRing<T, A>(override val algebra: A) : FunctionalExpressionSpace<T, A>(algebra),
open class FunctionalExpressionRing<T, A>(algebra: A) : FunctionalExpressionSpace<T, A>(algebra),
Ring<Expression<T>> where A : Ring<T>, A : NumericAlgebra<T> {
override val one: Expression<T>
get() = const(algebra.one)
@ -119,7 +117,7 @@ open class FunctionalExpressionRing<T, A>(override val algebra: A) : FunctionalE
super<FunctionalExpressionSpace>.binaryOperation(operation, left, right)
}
open class FunctionalExpressionField<T, A>(override val algebra: A) :
open class FunctionalExpressionField<T, A>(algebra: A) :
FunctionalExpressionRing<T, A>(algebra),
Field<Expression<T>> where A : Field<T>, A : NumericAlgebra<T> {
/**
@ -137,3 +135,12 @@ open class FunctionalExpressionField<T, A>(override val algebra: A) :
override fun binaryOperation(operation: String, left: Expression<T>, right: Expression<T>): Expression<T> =
super<FunctionalExpressionRing>.binaryOperation(operation, left, right)
}
inline fun <T, A : Space<T>> A.expressionInSpace(block: FunctionalExpressionSpace<T, A>.() -> Expression<T>): Expression<T> =
FunctionalExpressionSpace(this).block()
inline fun <T, A : Ring<T>> A.expressionInRing(block: FunctionalExpressionRing<T, A>.() -> Expression<T>): Expression<T> =
FunctionalExpressionRing(this).block()
inline fun <T, A : Field<T>> A.expressionInField(block: FunctionalExpressionField<T, A>.() -> Expression<T>): Expression<T> =
FunctionalExpressionField(this).block()