Since we don't have a Bindings implementation with unstable order this
function is not required, as its callers can just iterate over the
attributes instead.
Change-Id: I01b35277b5a2dde69d684bc881dbd7c0701bcbb3
Reviewed-on: https://cl.tvl.fyi/c/depot/+/2291
Tested-by: BuildkiteCI
Reviewed-by: glittershark <grfn@gws.fyi>
Value now carries a shared_ptr<Bindings>, and all Bindings constructors return a unique_ptr<Bindings>.
The test that wanted to compare two Bindings by putting them into Values has been modified to use the new Equal() method on Bindings (extracted from EvalState).
Change-Id: I8dfb60e65fdabb717e3b3e5d56d5b3fc82f70883
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1744
Tested-by: BuildkiteCI
Reviewed-by: glittershark <grfn@gws.fyi>
Reviewed-by: tazjin <mail@tazj.in>
We have decided that leaking memory is a better fate than random,
non-debuggable memory corruption. Future CLs will begin changing
various fields to std::unique_ptr and std::shared_ptr.
It turns out that disabling the GC does not have disasterous impact.
The Nix evaluator only runs on the client CLI, never in any long-
running process. Even the REPL does not leak too badly under this
change, because it uses one EvalState for the duration of the REPL.
Building an explicitly tracing garbage collector is likely in the
future of this project, but that giant amount of work cannot be
done under a nix evaluator that is constantly crashing. We need to
restore development velocity here, and this is the best way we've
figured out to do it.
Change-Id: I2fcda8fcee853c15a9a5e22eca7c5a784bc2bf76
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1720
Reviewed-by: glittershark <grfn@gws.fyi>
Reviewed-by: tazjin <mail@tazj.in>
Tested-by: BuildkiteCI
The prior use of gc_allocator meant that the btree nodes themselves were being collected. Additionally, have Attr (contains a Value) and Bindings explicitly inherit from gc, even though Bindings is always allocated under `new (GC)`.
Detected by running under GC_ENABLE_INCREMENTAL=1.
Change-Id: Iacf13b34b5aa12e417ea87c9b46e2bf9199fdb26
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1544
Reviewed-by: glittershark <grfn@gws.fyi>
Tested-by: BuildkiteCI
The function is renamed to `SortedByKeys`, which is more descriptive,
and annotated with a comment about what it is used for.
The deprecation warning has been removed because this function is
currently functionally required.
Change-Id: I0ee3a76deff05f366feca9ddac8f38ab34bffbd0
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1288
Tested-by: BuildkiteCI
Reviewed-by: glittershark <grfn@gws.fyi>
Instead of manually iterating over the two bindings to be combined,
this adds a new static method on the Bindings class which merges two
attribute sets by calling the range insertion operator over them.
In some anecdotal tests, this can lead to a ~10% speed bump -
depending on the specific operation.
Change-Id: I5dea03b0589a83a789d3a8a0fc81d0d9e6598371
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1216
Tested-by: BuildkiteCI
Reviewed-by: glittershark <grfn@gws.fyi>
This reverts parts of the CLs splitting the backing implementation for
Bindings and moves back to only the BTreeMap-backed implementation.
Our evaluation has indicated that the Vector-backed implementation
does not match the performance of the plain array used upstream, and
in my view the complexity introduced by it is not worth the relatively
small (single-digit percentage) performance increase with a
pivot-point close to the number of attributes yielded by
stdenv.mkDerivation.
Going forward we will trial implementations of attribute sets backed
by HAMTs, and investigate other mechanisms of speeding up the language.
Some changes from the previous CLs are retained, for example the
removal of insert_or_assign and the passing of capacity.
Change-Id: I6eb4b075b453949583360755055c21a29d7ff642
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1172
Reviewed-by: glittershark <grfn@gws.fyi>
Tested-by: BuildkiteCI
This function in never called, so let's just remove it
Paired-With: Luke Granger-Brown <git@lukegb.com>
Paired-With: Vincent Ambo <mail@tazj.in>
Change-Id: I79125866254d90dd0842bc86830d2103ac313cb6
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1125
Tested-by: BuildkiteCI
Reviewed-by: lukegb <lukegb@tvl.fyi>
Reviewed-by: isomer <isomer@tvl.fyi>
To aid in making the decision of where to (currently just statically)
use a vector or btree as the backing implementation, add an extra
constructor argument to Bindings::NewGC for a capacity, and use
a (currently hardcoded at 32, for no good reason other than it felt like
a reasonable number) pivot to switch between our possible backing
implementations. Then, update all the call sites where it feels
reasonable that we know the capacity statically to *pass* that capacity
to the constructor.
Paired-With: Luke Granger-Brown <git@lukegb.com>
Paired-With: Vincent Ambo <mail@tazj.in>
Paired-With: Perry Lorier <isomer@tvl.fyi>
Change-Id: I1858c161301a1cd0e83aeeb9a58839378869e71d
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1124
Tested-by: BuildkiteCI
Reviewed-by: lukegb <lukegb@tvl.fyi>
Reviewed-by: isomer <isomer@tvl.fyi>
Add an alternative impl of the now-abstract Bindings base class that is
backed by a std::vector, somewhat similar but stylistically a little
superior to the array-backed implementation in upstream nix. The
underlying iterator type in BindingsIterator is now backed by a
std::variant that we std::visit an overload over in order to implement
the various bits of the iterator interface.
Paired-With: Luke Granger-Brown <git@lukegb.com>
Paired-With: Vincent Ambo <mail@tazj.in>
Paired-With: Perry Lorier <isomer@tvl.fyi>
Change-Id: I7fbd1f4d5c449e2f9b82102a701b0bacd5e80672
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1123
Tested-by: BuildkiteCI
Reviewed-by: tazjin <mail@tazj.in>
To pave the way for the thing we want to do eventually which is use a
linear-time array for bindings (aka attribute sets) that are statically
known to be small enough to get a performance benefit from doing so,
make the Bindings class abstract, and define a BTreeBindings class that
inherits from it and is (currently always) returned from the static
initializer. The idea is that we'll have an ArrayBindings class as well
later that we can dispatch to conditionally based on an optional
"capacity" parameter or something like that.
There was some difficulty here in getting the iterator to work - the
approach we settled on ended up making a concrete BindingsIterator class
which will wrap a std::variant of either a btree iterator or something
else later, but right now just wraps a btree iterator.
Paired-With: Luke Granger-Brown <git@lukegb.com>
Paired-With: Vincent Ambo <mail@tazj.in>
Paired-With: Perry Lorier <isomer@tvl.fyi>
Change-Id: Ie02ca5a1c55e8ebf99ab1e957110bd9284278907
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1121
Tested-by: BuildkiteCI
Reviewed-by: isomer <isomer@tvl.fyi>
Having a default constructor for this causes a variety of annoying
situations across the codebase in which this is initialised to an
unexpected value, leading to constant guarding against those
conditions.
It turns out there's actually no intrinsic reason that this default
constructor needs to exist. The biggest one was addressed in CL/1138
and this commit cleans up the remaining bits.
Change-Id: I4a847f50bc90e72f028598196592a7d8730a4e01
Reviewed-on: https://cl.tvl.fyi/c/depot/+/1139
Reviewed-by: isomer <isomer@tvl.fyi>
Tested-by: BuildkiteCI
These bits are no longer required with the hashmap-backed
implementation of attribute sets.
Change-Id: I8b936d8d438a00bad4ccf8e0b4dd719c559ce8c2
Reviewed-on: https://cl.tvl.fyi/c/depot/+/912
Tested-by: BuildkiteCI
Reviewed-by: edef <edef@edef.eu>
Previously all includes were anchored in one global mess of header
files. This moves the includes into filesystem "namespaces" (if you
will) for each sub-package of Nix.
Note: This commit does not introduce the relevant build system changes.
Instead of doing some sort of inline merge-sort of the two attribute
sets, use the attribute sets merge function.
This commit alone does not build and is not supposed to.
This is the first step towards replacing the implementation of
attribute sets with an absl::btree_map.
Currently many access are done using array offsets and pointer
arithmetic, so this change is currently causing Nix to fail in various
ways.
Replaces most uses of `string` with `std::string`.
This came up because I removed the "types.hh" import from
"symbol-table.hh", which percolated through a bunch of files where
`string` was suddenly no longer defined ... *sigh*
These were not caught by the previous clang-tidy invocation, but were
instead sorted out using amber[0] as such:
ambr --regex 'for (\(.+\))\s([a-z].*;)' 'for $1 { $2 }'
[0]: https://github.com/dalance/amber
These were not caught by the previous clang-tidy invocation, but were
instead sorted out using amber[0] as such:
ambr --regex 'if (\(.+\))\s([a-z].*;)' 'if $1 { $2 }'
[0]: https://github.com/dalance/amber
