tvl-depot/tvix/eval/src/opcode.rs

130 lines
3 KiB
Rust
Raw Normal View History

//! This module implements the instruction set running on the abstract
//! machine implemented by tvix.
/// Index of a constant in the current code chunk.
#[repr(transparent)]
#[derive(Clone, Copy, Debug)]
pub struct ConstantIdx(pub usize);
/// Index of an instruction in the current code chunk.
#[repr(transparent)]
#[derive(Clone, Copy, Debug)]
pub struct CodeIdx(pub usize);
/// Index of a value in the runtime stack.
#[repr(transparent)]
#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd)]
pub struct StackIdx(pub usize);
/// Index of an upvalue within a closure's upvalue list.
#[repr(transparent)]
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct UpvalueIdx(pub usize);
/// Offset by which an instruction pointer should change in a jump.
#[repr(transparent)]
#[derive(Clone, Copy, Debug)]
pub struct JumpOffset(pub usize);
/// Provided count for an instruction (could represent e.g. a number
/// of elements).
#[repr(transparent)]
#[derive(Clone, Copy, Debug)]
pub struct Count(pub usize);
#[warn(variant_size_differences)]
#[derive(Clone, Copy, Debug)]
pub enum OpCode {
// Push a constant onto the stack.
OpConstant(ConstantIdx),
// Discard a value from the stack.
OpPop,
// Push a literal value.
OpNull,
OpTrue,
OpFalse,
// Unary operators
OpInvert,
OpNegate,
// Arithmetic binary operators
OpAdd,
OpSub,
OpMul,
OpDiv,
// Comparison operators
OpEqual,
OpLess,
OpLessOrEq,
OpMore,
OpMoreOrEq,
// Logical operators & generic jumps
OpJump(JumpOffset),
OpJumpIfTrue(JumpOffset),
OpJumpIfFalse(JumpOffset),
OpJumpIfNotFound(JumpOffset),
// Attribute sets
OpAttrs(Count),
OpAttrPath(Count),
OpAttrsUpdate,
OpAttrsSelect,
OpAttrsTrySelect,
OpAttrsIsSet,
// `with`-handling
OpPushWith(StackIdx),
OpPopWith,
OpResolveWith,
fix(tvix/eval): correctly thread through dynamic upvalues This puts together the puzzle pieces for threading dynamic upvalues (that is, upvalues resolved from the `with`-stack) all the way through. Reading the test case enclosed in this commit and walking through it is recommended to understand what problem is being tackled here. In short, because the compiler can not statically know *which* with-scope a dynamic argument is resolved from it needs to lay the groundwork for resolving from *all* possible scopes. There are multiple different approaches to doing this. The approach chosen in this commit is that if a dynamic upvalue is detected, the compiler will emit instructions to close over this dynamic value in *all* enclosing lambda contexts. It uses a new instruction for this that will leave around a sentinel value in case an identifier could not be resolved, and wire the location of this found value (or sentinel) up through the upvalues to the next level of nesting. In this tradeoff, tvix potentially closes over more upvalues than are needed (but in practice, how often do people create *really* deep `with`-stacks? and in *this* kind of code situation? maybe we should even warn for this!) but avoids keeping the entire attribute sets themselves around. Looking at the test case, each surrounding closure will close over *all* dynamic identifiers that are referenced later on visible to it, but only the last one for each identifier will actually end up being used. This also covers our bases for an additional edge-case this creates, in which an identifier potentially resolves to a dynamic upvalue *and* to a dynamic value within the function's own scope (again, would anyone really do this?) by introducing a resolution instruction for that particular case. There is likely some potential for cleaning up this code which is quite ugly in some parts, but as this implementation is now carefully calibrated to work I decided it is time to commit it and clean it up in subsequent commits. Change-Id: Ib701e3e6da39bd2c95938d1384036ff4f9fb3749 Reviewed-on: https://cl.tvl.fyi/c/depot/+/6322 Tested-by: BuildkiteCI Reviewed-by: sterni <sternenseemann@systemli.org>
2022-08-28 02:45:45 +02:00
OpResolveWithOrUpvalue(UpvalueIdx),
// Lists
OpList(Count),
OpConcat,
// Strings
OpInterpolate(Count),
// Type assertion operators
OpAssertBool,
// Access local identifiers with statically known positions.
OpGetLocal(StackIdx),
// Close scopes while leaving their expression value around.
OpCloseScope(Count), // number of locals to pop
// Asserts stack top is a boolean, and true.
OpAssert,
// Lambdas & closures
OpCall,
OpGetUpvalue(UpvalueIdx),
OpClosure(ConstantIdx),
// Thunks
OpThunk(ConstantIdx),
OpForce,
/// Finalise initialisation of the upvalues of the value in the
/// given stack index after the scope is fully bound.
OpFinalise(StackIdx),
// The closure and thunk creation instructions have a variable
// number of arguments to the instruction, which is represented
// here by making their data part of the opcodes.
//
// The VM skips over these by advancing the instruction pointer
// according to the count.
DataLocalIdx(StackIdx),
DataDeferredLocal(StackIdx),
DataUpvalueIdx(UpvalueIdx),
DataDynamicIdx(ConstantIdx),
fix(tvix/eval): correctly thread through dynamic upvalues This puts together the puzzle pieces for threading dynamic upvalues (that is, upvalues resolved from the `with`-stack) all the way through. Reading the test case enclosed in this commit and walking through it is recommended to understand what problem is being tackled here. In short, because the compiler can not statically know *which* with-scope a dynamic argument is resolved from it needs to lay the groundwork for resolving from *all* possible scopes. There are multiple different approaches to doing this. The approach chosen in this commit is that if a dynamic upvalue is detected, the compiler will emit instructions to close over this dynamic value in *all* enclosing lambda contexts. It uses a new instruction for this that will leave around a sentinel value in case an identifier could not be resolved, and wire the location of this found value (or sentinel) up through the upvalues to the next level of nesting. In this tradeoff, tvix potentially closes over more upvalues than are needed (but in practice, how often do people create *really* deep `with`-stacks? and in *this* kind of code situation? maybe we should even warn for this!) but avoids keeping the entire attribute sets themselves around. Looking at the test case, each surrounding closure will close over *all* dynamic identifiers that are referenced later on visible to it, but only the last one for each identifier will actually end up being used. This also covers our bases for an additional edge-case this creates, in which an identifier potentially resolves to a dynamic upvalue *and* to a dynamic value within the function's own scope (again, would anyone really do this?) by introducing a resolution instruction for that particular case. There is likely some potential for cleaning up this code which is quite ugly in some parts, but as this implementation is now carefully calibrated to work I decided it is time to commit it and clean it up in subsequent commits. Change-Id: Ib701e3e6da39bd2c95938d1384036ff4f9fb3749 Reviewed-on: https://cl.tvl.fyi/c/depot/+/6322 Tested-by: BuildkiteCI Reviewed-by: sterni <sternenseemann@systemli.org>
2022-08-28 02:45:45 +02:00
DataDynamicAncestor(UpvalueIdx),
}