Previously, Nix strings were represented as a Box (within Value)
pointing to a tuple of an optional context, and another Box pointing to
the actual string allocation itself. This is pretty inefficient, both in
terms of memory usage (we use 48 whole bytes for a None context!) and in
terms of the extra indirection required to get at the actual data. It
was necessary, however, because with native Rust DSTs if we had
something like `struct NixString(Option<NixContext>, BStr)` we could
only pass around *fat* pointers to that value (with the length in the
pointer) and that'd make Value need to be bigger (which is a waste of
both memory and cache space, since that memory would be unused for all
other Values).
Instead, this commit implements *manual* allocation of a packed string
representation, with the length *in the allocation* as a field past the
context. This requires a big old pile of unsafe Rust, but the payoff is
clear:
hello outpath time: [882.18 ms 897.16 ms 911.23 ms]
change: [-15.143% -13.819% -12.500%] (p = 0.00 < 0.05)
Performance has improved.
Fortunately this change can be localized entirely within
value/string.rs, since we were abstracting things out nicely.
Change-Id: Ibf56dd16c9c503884f64facbb7f0ac596463efb6
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10852
Tested-by: BuildkiteCI
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Autosubmit: aspen <root@gws.fyi>
This reverts commit d3d41552cf.
This was well-intentioned, but now the boxed Path values are actually
the *largest* Value enum variants, at 16 bytes (because they're
fat-pointers, with a len) instead of 8 bytes like all the other values.
Having the double reference is a reasonable price to pay (it seems; more
benchmarks may end up disagreeing) for a smaller Value repr.
Change-Id: I0d3e84f646c8f5ffd0b7259c4e456637eea360f7
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10797
Tested-by: BuildkiteCI
Autosubmit: aspen <root@gws.fyi>
Reviewed-by: sterni <sternenseemann@systemli.org>
NixString is *quite* large - like 80 bytes - because of the extra
capacity value for BString and because of the context. We want to keep
Value small since we're passing it around a lot, so let's box the
NixString inside Value::String to save on some memory, and make cloning
ostensibly a little cheaper
Change-Id: I343c8b4e7f61dc3dcbbaba4382efb3b3e5bbabb2
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10729
Tested-by: BuildkiteCI
Reviewed-by: sterni <sternenseemann@systemli.org>
PathBuf internally contains a heap pointer (an OsString), so we were in
effect double-boxing here. Removing the extra layer by making
Tvix::Value represented by a Box<Path> rather than a Box<PathBuf> saves
us an indirection, while still avoiding the extra memory overhead of the
capacity which was the reason we were boxing PathBuf in the first place.
Change-Id: I8c185b9d4646161d1921917f83e87421496a3e24
Reviewed-on: https://cl.tvl.fyi/c/depot/+/10725
Reviewed-by: sterni <sternenseemann@systemli.org>
Autosubmit: aspen <root@gws.fyi>
Tested-by: BuildkiteCI
This shaves another 8 bytes off Value. How did that type get so big?!
Change-Id: I65e9b59a1636bd57e3cc4aec5fea16887070b832
Reviewed-on: https://cl.tvl.fyi/c/depot/+/8153
Reviewed-by: raitobezarius <tvl@lahfa.xyz>
Tested-by: BuildkiteCI
Instead of going through Vec/BTreeMap for generating our internal
types, use the proptest strategies from imbl.
The one thing I couldn't figure out in the previous implementation is
where the ranges/sizes of generated collections came from. The
strategies in proptest use different types (Range, with an unknown
default value, and SizeRange with 0..100). I've opted to specify
0..100 directly, but we can probably make it configurable.
Change-Id: I749bc4c703fe424099240cab822b1642e5216361
Reviewed-on: https://cl.tvl.fyi/c/depot/+/7791
Autosubmit: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Reviewed-by: flokli <flokli@flokli.de>
Factor out the construction of Value::Attrs (including the Rc) into a
new `attrs` constructor function, to abstract away the presence of the
Rc itself.
Change-Id: I42fd4c3841e1db368db999ddd651277ff995f025
Reviewed-on: https://cl.tvl.fyi/c/depot/+/6892
Autosubmit: grfn <grfn@gws.fyi>
Reviewed-by: sterni <sternenseemann@systemli.org>
Reviewed-by: tazjin <tazjin@tvl.su>
Tested-by: BuildkiteCI
Generate smaller recursive values for generated Values, and run fewer
cases for the attrs proptests which are particularly egregious.
Change-Id: Ia35c7c120270feaf045be1deb440c87ebb185c27
Reviewed-on: https://cl.tvl.fyi/c/depot/+/6716
Autosubmit: grfn <grfn@gws.fyi>
Reviewed-by: tazjin <tazjin@tvl.su>
Reviewed-by: sterni <sternenseemann@systemli.org>
Tested-by: BuildkiteCI
Impl Arbitrary for Value (and NixAttrs and NixList) in the same way we
did for NixString. Value currently only generates non-"internal"
values (no thunks, AttrNotFound, etc.) and can't generate
functions (builtins or closures), because those'd require full
generation of tvix bytecode, which is a bit more work than I'd like to
do now - there's a `todo!` left in the code for a place where we could
allow opting-in to internal values and functions later.
Change-Id: I07a59e2b1d89cfaa912d4ecebd642caf4ddb040a
Reviewed-on: https://cl.tvl.fyi/c/depot/+/6627
Autosubmit: grfn <grfn@gws.fyi>
Tested-by: BuildkiteCI
Reviewed-by: tazjin <tazjin@tvl.su>
