This conversion was a bit too magic, and we can just use
`NixString::new_context_from` without having to worry about the
distinction between an empty context or no context, as
NixString::new_context_from already deals with that internally.
Change-Id: I3e5d57ecfa0f7456aa6c526863e49f2523afaec3
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12754
Tested-by: BuildkiteCI
Reviewed-by: edef <edef@edef.eu>
Autosubmit: flokli <flokli@flokli.de>
Currently Value::Json is used in combination with VMRequest::ToJson to
recursively convert tvix Value to serde_json::Value. This functionality
is used in builtins.toJSON as well as derivation __structuredAttrs.
Both Value::Json and VMRequest::ToJson were removed in this commit.
Related functionality in vm.rs is also removed: vm.rs does not know
about JSON anymore.
Recursively converting to serde_json now happens without going through
the VM.
Thrown errors that are part of the value of toJSON are now directly
propagated as ErrorKind, were-as previously there was a split between
CatchableErrorKind and ErrorKind, where eventually CatchableErrorKind
would be converted to ErrorKind::Catchable.
Change-Id: I066f064926c491e4c087a984f07af43d19124cfe
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12732
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
The nixhash errors were wrapped in a generic TvixError. Now it has its
own TvixError with unique error code. The nixhash error is passed along
as a string.
The errors looked like:
error[E997]: invalid encoded digest length '51' for algo sha256
Now they look like:
error[E041]: Invalid hash: invalid encoded digest length '51' for algo
sha256
Change-Id: I5c420815538ba4c6567c95f5d44d60c4d48f43fd
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12718
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
These nested ifs are a bit confusing, a match block makes this cleaner.
Change-Id: I256fd0bc921fbf2e60ad0f6e1ea51c2e0fb00317
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12628
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Equivalent logic is now in the standard library, and this dependency is no
longer needed for eval.
Change-Id: Iaa4410d89fdaa5b84cbd9e6bc6ae479c659d92f2
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12602
Autosubmit: tazjin <tazjin@tvl.su>
Reviewed-by: edef <edef@edef.eu>
Tested-by: BuildkiteCI
Refactor the `strict` boolean passed into evaluation at the top-level to
be a (two-variant, so far) EvalMode enum of Lazy and Strict.
This is more explicit than a boolean, and if we ever add more EvalModes
it's a simple extension of the enum.
Change-Id: I3de50e74ec971011664f6cd0999d08b792118410
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12186
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: aspen <root@gws.fyi>
thiserror is much more easier to maintain than manually implementing Error and Display.
Change-Id: Ibf13e2d8a96fba69c8acb362b7515274a593dfd6
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12452
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
Tested-by: BuildkiteCI
From now on we will add the dependencies and their version in the root
Cargo.toml and in order to enable the dependency for a workspace member
we set `workspace = true` in the member's Cargo.toml.
Change-Id: I9738c1cf99810b7ace87ca712c3ea965ba846e25
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12389
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
This updates all the dependencies and their "minimum" versions in
Cargo.{lock,toml} to the latest compatible version using `cargo-edit`'s
`cargo upgrade` command that will eventually be merged into `cargo
update`.
Change-Id: Iccb2aa4a1c84a0465222244a0bd0cafe2a82e781
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12388
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
Tested-by: BuildkiteCI
This interface is the only way to construct lazy builtins outside of eval, which
is actually a useful feature to have.
Change-Id: I386323af20aa3134bb4f669fa66fbb21e9b05fd4
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12386
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Autosubmit: tazjin <tazjin@tvl.su>
Reviewed-by: Yury Shvedov <yury.shvedov@kaspersky.com>
Removes imbl::OrdMap in favour of an Rc over the standard library's BTreeMap,
which allows us to drop the imbl dependency completely.
In my local tests this is actually slightly faster for `hello` and `firefox`.
Change-Id: Ic9597ead4e98bf9530f290c6a94a3c5c3efd0acc
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12201
Reviewed-by: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
This vector type has served us well for now, but it contains internal refcounts
which are incompatible with upcoming changes related to garbage collection.
The performance impact of this change within all benchmarks I ran was within the
margin of error:
[nix-shell:/tmp/perf]$ hyperfine "./before -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings"
Benchmark 1: ./u64 -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings
Time (mean ± σ): 7.528 s ± 0.272 s [User: 6.578 s, System: 0.631 s]
Range (min … max): 7.160 s … 8.012 s 10 runs
nix-shell:/tmp/perf]$ hyperfine "./std-vec -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings"
Benchmark 1: ./std-vec -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings
Time (mean ± σ): 7.515 s ± 0.178 s [User: 6.508 s, System: 0.652 s]
Range (min … max): 7.276 s … 7.861 s 10 runs
Change-Id: Ib95f871956e336a1e5771f6293583854b1efb276
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12197
Reviewed-by: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
This replaces the OpCode enum with a new Op enum which is guaranteed to fit in a
single byte. Instead of carrying enum variants with data, every variant that has
runtime data encodes it into the `Vec<u8>` that a `Chunk` now carries.
This has several advantages:
* Less stack space is required at runtime, and fewer allocations are required
while compiling.
* The OpCode doesn't need to carry "weird" special-cased data variants anymore.
* It is faster (albeit, not by much). On my laptop, results consistently look
approximately like this:
Benchmark 1: ./before -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings
Time (mean ± σ): 8.224 s ± 0.272 s [User: 7.149 s, System: 0.688 s]
Range (min … max): 7.759 s … 8.583 s 10 runs
Benchmark 2: ./after -E '(import <nixpkgs> {}).firefox.outPath' --log-level ERROR --no-warnings
Time (mean ± σ): 8.000 s ± 0.198 s [User: 7.036 s, System: 0.633 s]
Range (min … max): 7.718 s … 8.334 s 10 runs
See notes below for why the performance impact might be less than expected.
* It is faster while at the same time dropping some optimisations we previously
performed.
This has several disadvantages:
* The code is closer to how one would write it in C or Go.
* Bit shifting!
* There is (for now) slightly more code than before.
On performance I have the following thoughts at the moment:
In order to prepare for adding GC, there's a couple of places in Tvix where I'd
like to fence off certain kinds of complexity (such as mutating bytecode, which,
for various reaons, also has to be part of data that is subject to GC). With
this change, we can drop optimisations like retroactively modifying existing
bytecode and *still* achieve better performance than before.
I believe that this is currently worth it to pave the way for changes that are
more significant for performance.
In general this also opens other avenues of optimisation: For example, we can
profile which argument sizes actually exist and remove the copy overhead of
varint decoding (which does show up in profiles) by using more adequately sized
types for, e.g., constant indices.
Known regressions:
* Op::Constant is no longer printing its values in disassembly (this can be
fixed, I just didn't get around to it, will do separately).
Change-Id: Id9b3a4254623a45de03069dbdb70b8349e976743
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12191
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
I'm gonna be doing some poking around in the internals of Context, so in
preparation this pulls it out into its own module.
Change-Id: I72ea7df80b5f36f838934ee07bdba66874c334c9
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12189
Autosubmit: aspen <root@gws.fyi>
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
When using the runTests feature of crate2nix the derivation that runs the
tests is put into passthru.test but all default.nix files for Rust crates
in Tvix threw that away.
This commit retains passthru so that you can get access to the test
derivation.
Change-Id: I8b7b7db57a49069348f08c12c00a3b1a41a0c05b
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12215
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
This sentence is a little stale; let's just link to NixString directly
for the authoritative source of truth.
Change-Id: I64e065c4148d29702b09820a0e7724a65fae7c67
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12181
Autosubmit: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Just like tvix-repl does (except we don't force values when printing
them, so... not entirely like tvix-repl does). But it's something.
Change-Id: I2e69b08d7d82b0b2d337f1d4c5d87ed28475fa84
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12180
Reviewed-by: flokli <flokli@flokli.de>
Tested-by: BuildkiteCI
Autosubmit: aspen <root@gws.fyi>
Rather than storing the leaked allocation for the string as the key in
the interner, store the hash (using NoHashHashBuilder). I thought this
would improve performance, but it doesn't:
hello outpath time: [736.85 ms 748.42 ms 760.42 ms]
change: [-2.0754% +0.4798% +2.7096%] (p = 0.72 > 0.05)
No change in performance detected.
but it at least doesn't *hurt* performance, and it *does* avoid an
`unsafe`, so it's probably net good.
Change-Id: Ie413955bdb6f04b1f468f511e5ebce56e329fa37
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12049
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: aspen <root@gws.fyi>
When we moved to building tvix with crate2nix we had to disable the
readDir tests for the same reasons mentioned in
https://cl.tvl.fyi/c/depot/+/12131. But now with the 12131 CL, we can
re-enable the readDir tests.
In this change I re-enabled the readDir tests in nix_tests by moving
them out of notyetpassing and I also deleted them from tvix_tests
because we don't need them in both places.
Change-Id: I82ac39605299a8b22d80f8b51fc8ec2476d21dc9
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12133
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
builtins.readFileType was added to Nix back in version 2.14.
The tests were also moved out of notyetpassing in addition to the
readDir fixtures they depend on.
I caught a bug where we previously used std::fs::metadata (via the
.metadata() method on File) which follows symlinks so it would always
return false for is_symlink(). Instead we now use
std::fs::symlink_metadata directly which does not follow symlinks, so
tests now pass. This wasn't an issue for builtins.readDir as it uses
walkdir and walkdir doesn't follow symlinks either.
Change-Id: I58eb97bdb5ec95df4f6882f495f8c572fe7c6793
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12130
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
Tested-by: BuildkiteCI
In newer versions of Nix there's a builtins.readFileType builtin, we
should try to avoid duplicating the enum -> string conversion by
implementating Display before we implement builtins.readFileType.
Change-Id: I579e95949a76eb33d2e7bda0000ed84859df765e
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12129
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
Tested-by: BuildkiteCI
Use the faster and newer MiMalloc memory allocator for all endpoints in
the workspace.
Change-Id: Ic60237284ed168e46ec6e8f28e2710bae4385c6f
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12149
Tested-by: BuildkiteCI
Reviewed-by: aspen <root@gws.fyi>
Rather than making the interner be a global lazy_static mutex, put it in
a thread-local RefCell. This doesn't change anything in terms of
sharing (since we're currently actually just single threaded), but
avoids the overhead of a mutex, for a nice performance boost (compared
to the mutex version):
hello outpath time: [726.71 ms 729.79 ms 735.69 ms]
change: [-5.7277% -3.9733% -2.1144%] (p = 0.00 < 0.05)
Performance has improved.
Change-Id: I240b238dcbaf854ebafc3017b4425fb7d7b91b03
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12048
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: aspen <root@gws.fyi>
This is the most naive version of string interning possible - we store a
map from the string itself to the pointer behind a global mutex, and
memoize the allocation of all strings below a threshold length (16
bytes, for now) into that map. This requires leaking /all/ strings,
since it's not easy to know just from the pointer that a string has been
interned - so interning is disabled if string leaking is also disabled.
In the case where we're leaking strings (the default), even the naive
version of this gets us a pretty nice perfomance boost:
hello outpath time: [742.54 ms 745.89 ms 749.14 ms]
change: [-2.8722% -2.0135% -1.0654%] (p = 0.00 < 0.05)
Performance has improved.
However, in the case where we're not leaking strings, we have to keep
track of which strings have and haven't been interned, which makes this
a little worse:
hello outpath time: [779.30 ms 792.82 ms 808.74 ms]
change: [+2.5258% +4.0884% +5.8931%] (p = 0.00 < 0.05)
Performance has regressed.
Hopefully we can close the gap here a bit with some clever
tricks (coming next).
Change-Id: If08cb48ede703c7fe3bdd8d617443f8a561ad09b
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12047
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: aspen <root@gws.fyi>
Per https://nnethercote.github.io/perf-book/hashing.html, we have
basically no reason to use the default hasher over a faster,
non-DoS-resistant hasher. This gives a nice perf boost basically for
free:
hello outpath time: [704.76 ms 714.91 ms 725.63 ms]
change: [-7.2391% -6.1018% -4.9189%] (p = 0.00 < 0.05)
Performance has improved.
Change-Id: If5587f444ed3af69f8af4eead6af3ea303b4ae68
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12046
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Reviewed-by: Ilan Joselevich <personal@ilanjoselevich.com>
Autosubmit: aspen <root@gws.fyi>
Default to always leaking strings, and copying strings by copying
pointers rather than cloning the underlying allocation. This (somewhat
bafflingly) doesn't seem to affect any benchmarks, but paves the way for
some tricks around string allocation that do.
Unfortunately, we can't do this (yet?) for contextful strings, for
reasons I don't currently understand but which I will address later,
when I address contextful strings more holistically.
I've left a flag in here to disable this, both to test the cloning logic
for strings for when/if we decide to bring this back, and to allow
people who care more about memory usage than perf to disable leaking.
Change-Id: Iec44bcbfe9b3d20389d2450b9a551792a79b9b26
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12045
Autosubmit: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Previously, OpConstant would display some detail about its
ConstantIdx: whether it's a thunk or closure, and what its address
is. This has been expanded to also show when the ConstantIdx is a
blueprint, along with the blueprint's address, and to the other
opcodes that use a ConstantIdx.
Currently, it seems like blueprint addresses don't correspond to the
address of the thunk listed in the bytecode output, but it's still
useful to see that the constant being grabbed is a blueprint, and
maybe this pointer can be made to make more sense in the future.
Change-Id: Ia212b0d52b004c87051542c093274e7106ee08e4
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12044
Tested-by: BuildkiteCI
Reviewed-by: aspen <root@gws.fyi>
Autosubmit: chickadee <matthewktromp@gmail.com>
path_exists was returning an error when certain common IO errors were
encountered. e.g. in the path "/dev/null/.", path_exists would throw
an error because the underlying call to Path::try_exists threw an
error because null isn't a directory. But if null isn't a directory,
then the path is invalid, so this should really be returning
false. That's what nix's behavior is and that's what makes sense.
The trait function isn't being changed because some other
implementers (e.g. tvix_store_io) have actual errors they can throw.
Fixes: b/411
Change-Id: I9e810e7a198bffe61365697c6d3d7e71f264c20b
Reviewed-on: https://cl.tvl.fyi/c/depot/+/12042
Tested-by: BuildkiteCI
Autosubmit: chickadee <matthewktromp@gmail.com>
Reviewed-by: aspen <root@gws.fyi>
Now that we can bind (potentially lazy, potentially lambda-containing)
values in the REPL and then reference them in subsequent evaluations,
it's important that the values to which we construct shared references
are shared across those subsequent evaluations - otherwise, we get
panics due to unknown source map locations, or dropped weak references
to globals.
This change assigns both the globals and the source map as fields on the
Repl after the first evaluation, and then passes those in (to the
EvaluationBuilder) on subsequent evaluations.
On the EvaluationBuilder side, there's some panicking introduced - this
is intentional, as my intent is for the builder to be configured
statically enough that panicking is the best way to report errors
here (it's always a bug to misconfigure an Evaluation, and we'd never
want to handle it dynamically).
Change-Id: I37225697235c22b683ca48a17d30fa8fedd12d1b
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11960
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
Construct the Rc<GlobalsMap> for the evaluation as part of
EvaluiationBuilder::build, rather than deferring it until we actually
compile. This changes nothing functionally, but gets us one step closer
to sharing this globals map across evaluations.
Change-Id: Id92e9fb88d974d763056d4f15ce61962ab776e84
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11957
Tested-by: BuildkiteCI
Autosubmit: aspen <root@gws.fyi>
Reviewed-by: flokli <flokli@flokli.de>
Make constructing of a new Evaluation use the builder pattern rather
than setting public mutable fields. This is currently a pure
refactor (no functionality has changed) but has a few advantages:
- We've encapsulated the internals of the fields in Evaluation, meaning
we can change them without too much breakage of clients
- We have type safety that prevents us from ever changing the fields of
an Evaluation after it's built (which matters more in a world where we
reuse Evaluations).
More importantly, this paves the road for doing different things with
the construction of an Evaluation - notably, sharing certain things like
the GlobalsMap across subsequent evaluations in eg the REPL.
Fixes: b/262
Change-Id: I4a27116faac14cdd144fc7c992d14ae095a1aca4
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11956
Tested-by: BuildkiteCI
Autosubmit: aspen <root@gws.fyi>
Reviewed-by: flokli <flokli@flokli.de>
Allow passing in a top-level env, a map from name to value, to
evaluation. The intent is to support bound identifiers in the REPL just
like upstream nix does.
Getting this working involves mucking around a bit with internals - most
notably, locals now only optionally have a Span (since locals don't have
an easy span we can use) - and getting that working requires propagating
some minor hacks to places where we currently *need* a span (and which
would require too much changing now to make spans optional; my guess is
that that would essentially end up making spans optional throughout the
codebase).
Also, some extra care has to be taken to close out the scope in the case
that we do pass in an env, to avoid breaking our assumptions about the
size of the stack when we return from the toplevel
Change-Id: Ie475b2d3dfc72ccbf298d2a3ea28c63ac877d653
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11953
Tested-by: BuildkiteCI
Autosubmit: aspen <root@gws.fyi>
Reviewed-by: flokli <flokli@flokli.de>
This was made unnecessary in c92d06271 (feat(tvix/eval): drop
LightSpan::Delayed, 2023-12-08) because it didn't improve benchmarks as
much as expected and has been vestigial since; this continues the
cleanup by just removing it altogether
Change-Id: I21ec7ae9b52a5cccd2092696a5a87f658194d672
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11949
Autosubmit: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
NixString::iter_plain() didn't make much sense, especially without a docstring.
Rename it to iter_ctx_plain(), and copy the docstring from NixContext.
Do the same for the two other context element types too.
Change-Id: I1bbfcb967d8d9b14487d069bfe3a1f762253ef4d
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11882
Tested-by: BuildkiteCI
Reviewed-by: Ilan Joselevich <personal@ilanjoselevich.com>
Reviewed-by: flokli <flokli@flokli.de>
There's other places where unexpected arguments can be provided, like
in builtins.
Make space for that error type.
Change-Id: Ic831497a3a1dd36a3a184bedadcf1374bf0ae6db
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11876
Reviewed-by: Connor Brewster <cbrewster@hey.com>
Tested-by: BuildkiteCI
This allows peeking of the type at a given path.
It's necessary, as an open() might not fail until you try to read()
from it, and generally, stat'ing can be faster in some cases.
Change-Id: Ib002da3194a3546ca286de49aac8d1022ec5560f
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11871
Tested-by: BuildkiteCI
Reviewed-by: Ilan Joselevich <personal@ilanjoselevich.com>
Reviewed-by: Connor Brewster <cbrewster@hey.com>
The builtins `placeholder` and `filterSource` are now implemented in
tvix-glue and so we no longer need their placeholder "implementations" in
tvix-eval.
Change-Id: Ibf161ccf1ad40103e9fbb688e9f6be58e7772af1
Reviewed-on: https://cl.tvl.fyi/c/depot/+/11867
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Autosubmit: Ilan Joselevich <personal@ilanjoselevich.com>
