tvl-depot/tvix/eval/docs/known-optimisation-potential.md
sterni dc9c2c9b0c docs(tvix/eval): C++ Nix now has deduplicated inherit (from)
Change-Id: Icf74c699130c2ab774973a7e98bc44ffba8c5ec2
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11139
Tested-by: BuildkiteCI
Autosubmit: sterni <sternenseemann@systemli.org>
Reviewed-by: flokli <flokli@flokli.de>
2024-03-12 21:23:32 +00:00

7.1 KiB

Known Optimisation Potential

There are several areas of the Tvix evaluator code base where potentially large performance gains can be achieved through optimisations that we are already aware of.

The shape of most optimisations is that of moving more work into the compiler to simplify the runtime execution of Nix code. This leads, in some cases, to drastically higher complexity in both the compiler itself and in invariants that need to be guaranteed between the runtime and the compiler.

For this reason, and because we lack the infrastructure to adequately track their impact (WIP), we have not yet implemented these optimisations, but note the most important ones here.

  • Use "open upvalues" [hard]

    Right now, Tvix will immediately close over all upvalues that are created and clone them into the Closure::upvalues array.

    Instead of doing this, we can statically determine most locals that are closed over and escape their scope (similar to how the compiler::scope::Scope struct currently tracks whether locals are used at all).

    If we implement the machinery to track this, we can implement some upvalues at runtime by simply sticking stack indices in the upvalue array and only copy the values where we know that they escape.

  • Avoid with value duplication [easy]

    If a with makes use of a local identifier in a scope that can not close before the with (e.g. not across LambdaCtx boundaries), we can avoid the allocation of the phantom value and duplication of the NixAttrs value on the stack. In this case we simply push the stack index of the known local.

  • Multiple attribute selection [medium]

    An instruction could be introduced that avoids repeatedly pushing an attribute set to/from the stack if multiple keys are being selected from it. This occurs, for example, when inheriting from an attribute set or when binding function formals.

  • Split closure/function representation [easy]

    Functions have fewer fields that need to be populated at runtime and can directly use the value::function::Lambda representation where possible.

  • Apply compiler::optimise_select to other set operations [medium]

    In addition to selects, statically known attribute resolution could also be used for things like ? or with. The latter might be a little more complicated but is worth investigating.

  • Inline fully applied builtins with equivalent operators [medium]

    Some builtins have equivalent operators, e.g. builtins.sub corresponds to the - operator, builtins.hasAttr to the ? operator etc. These operators additionally compile to a primitive VM opcode, so they should be just as cheap (if not cheaper) as a builtin application.

    In case the compiler encounters a fully applied builtin (i.e. no currying is occurring) and the builtins global is unshadowed, it could compile the equivalent operator bytecode instead: For example, builtins.sub 20 22 would be compiled as 20 - 22. This would ensure that equivalent builtins can also benefit from special optimisations we may implement for certain operators (in the absence of currying). E.g. we could optimise access to the builtins attribute set which a call to builtins.getAttr "foo" builtins should also profit from.

  • Avoid nested VM::run calls [hard]

    Currently when encountering Nix-native callables (thunks, closures) the VM's run loop will nest and return the value of the nested call frame one level up. This makes the Rust call stack almost mirror the Nix call stack, which is usually undesirable.

    It is possible to detect situations where this is avoidable and instead set up the VM in such a way that it continues and produces the desired result in the same run loop, but this is kind of tricky to get right - especially while other parts are still in flux.

    For details consult the commit with Gerrit change ID I96828ab6a628136e0bac1bf03555faa4e6b74ece, in which the initial attempt at doing this was reverted.

  • Avoid thunks if only identifier closing is required [medium]

    Some constructs, like with, mostly do not change runtime behaviour if thunked. However, they are wrapped in thunks to ensure that deferred identifiers are resolved correctly.

    This can be avoided, as we statically analyse the scope and should be able to tell whether any such logic was required.

  • Intern literals [easy]

    Currently, the compiler emits a separate entry in the constant table for each literal. So the program 1 + 1 + 1 will have three entries in its Chunk::constants instead of only one.

  • Do some list and attribute set operations in place [hard]

    Algorithms that can not do a lot of work inside builtins like map, filter or foldl' usually perform terribly if they use data structures like lists and attribute sets.

    builtins can do work in place on a copy of a Value, but naïvely expressed recursive algorithms will usually use // and ++ to do a single change to a Value at a time, requiring a full copy of the data structure each time. It would be a big improvement if we could do some of these operations in place without requiring a new copy.

    There are probably two approaches: We could determine statically if a value is reachable from elsewhere and emit a special in place instruction if not. An easier alternative is probably to rely on reference counting at runtime: If no other reference to a value exists, we can extend the list or update the attribute set in place.

    An alternative to this is using persistent data structures or at the very least immutable data structures that can be copied more efficiently than the stock structures we are using at the moment.

  • Skip finalising unfinalised thunks or non-thunks instead of crashing [easy]

    Currently OpFinalise crashes the VM if it is called on values that don't need to be finalised. This helps catching miscompilations where OpFinalise operates on the wrong StackIdx. In the case of function argument patterns, however, this means extra VM stack and instruction overhead for dynamically determining if finalisation is necessary or not. This wouldn't be necessary if OpFinalise would just noop on any values that don't need to be finalised (anymore).

  • Phantom binding for from expression of inherits [easy]

    The from expression of an inherit is reevaluated for each inherit. This can be demonstrated using the following Nix expression which, counter-intuitively, will print “plonk” twice.

    let
      inherit (builtins.trace "plonk" { a = null; b = null; }) a b;
    in
    builtins.seq a (builtins.seq b null)
    

    In most Nix code, the from expression is just an identifier, so it is not terribly inefficient, but in some cases a more expensive expression may be used. We should create a phantom binding for the from expression that is reused in the inherits, so only a single thunk is created for the from expression.

    Since we discovered this, C++ Nix has implemented a similar optimization: https://github.com/NixOS/nix/pull/9847.