7 files changed, 719 insertions, 0 deletions

diff --git a/tvix/docs/src/SUMMARY.md b/tvix/docs/src/SUMMARY.md
new file mode 100644
index 000000000000..d4585c00acdf
--- /dev/null
+++ b/tvix/docs/src/SUMMARY.md
@@ -0,0 +1,10 @@
+# Summary
+
+# Tvix
+- [Architecture & data flow](./architecture.md)
+- [.drvPath inconsistencies](./differences-drv-paths.md)
+
+# Nix
+- [Specification of the Nix Language](./language-spec.md)
+- [Nix language version history](./lang-version.md)
+- [Value Pointer Equality](./value-pointer-equality.md)
diff --git a/tvix/docs/src/architecture.md b/tvix/docs/src/architecture.md
new file mode 100644
index 000000000000..5e0aa95f1a08
--- /dev/null
+++ b/tvix/docs/src/architecture.md
@@ -0,0 +1,147 @@
+# Tvix - Architecture & data flow
+
+## Background
+
+We intend for Tvix tooling to be more decoupled than the existing,
+monolithic Nix implementation. In practice, we expect to gain several
+benefits from this, such as:
+
+- Ability to use different builders
+- Ability to use different store implementations
+- No monopolisation of the implementation, allowing users to replace
+  components that they are unhappy with (up to and including the
+  language evaluator)
+- Less hidden intra-dependencies between tools due to explicit RPC/IPC
+  boundaries
+
+Communication between different components of the system will use
+gRPC. The rest of this document outlines the components.
+
+## Components
+
+### Coordinator
+
+*Purpose:* The coordinator (in the simplest case, the Tvix CLI tool)
+oversees the flow of a build process and delegates tasks to the right
+subcomponents. For example, if a user runs the equivalent of
+`nix-build` in a folder containing a `default.nix` file, the
+coordinator will invoke the evaluator, pass the resulting derivations
+to the builder and coordinate any necessary store interactions (for
+substitution and other purposes).
+
+While many users are likely to use the CLI tool as their primary
+method of interacting with Tvix, it is not unlikely that alternative
+coordinators (e.g. for a distributed, "Nix-native" CI system) would be
+implemented. To facilitate this, we are considering implementing the
+coordinator on top of a state-machine model that would make it
+possible to reuse the FSM logic without tying it to any particular
+kind of application.
+
+### Evaluator
+
+*Purpose:* Eval takes care of evaluating Nix code. In a typical build
+flow it would be responsible for producing derivations. It can also be
+used as a standalone tool, for example, in use-cases where Nix is used
+to generate configuration without any build or store involvement.
+
+*Requirements:* For now, it will run on the machine invoking the build
+command itself. We give it filesystem access to handle things like
+imports or `builtins.readFile`.
+
+To support IFD, the Evaluator also needs access to store paths. This
+could be implemented by having the coordinator provide an interface to retrieve
+files from a store path, or by ensuring a "realized version of the store" is
+accessible by the evaluator (this could be a FUSE filesystem, or the "real"
+/nix/store on disk.
+
+We might be okay with running the evaluator with filesystem access for now and
+can extend the interface if the need arises.
+
+### Builder
+
+*Purpose:* A builder receives derivations from the coordinator and
+builds them.
+
+By making builder a standardised interface it's possible to make the
+sandboxing mechanism used by the build process pluggable.
+
+Nix is currently using a hard-coded
+[libseccomp](https://github.com/seccomp/libseccomp) based sandboxing
+mechanism and another one based on
+[sandboxd](https://www.unix.com/man-page/mojave/8/sandboxd/) on macOS.
+These are only separated by [compiler preprocessor
+macros](https://gcc.gnu.org/onlinedocs/cpp/Ifdef.html) within the same
+source files despite having very little in common with each other.
+
+This makes experimentation with alternative backends difficult and
+porting Nix to other platforms harder than it has to be. We want to
+write a new Linux builder which uses
+[OCI](https://github.com/opencontainers/runtime-spec), the current
+dominant Linux containerisation technology, by default.
+
+With a well-defined builder abstraction, it's also easy to imagine
+other backends such as a Kubernetes-based one in the future.
+
+The environment in which builds happen is currently very Nix-specific. We might
+want to avoid having to maintain all the intricacies of a Nix-specific
+sandboxing environment in every builder, and instead only provide a more
+generic interface, receiving build requests (and have the coordinator translate
+derivations to that format). [^1]
+
+To build, the builder needs to be able to mount all build inputs into the build
+environment. For this, it needs the store to expose a filesystem interface.
+
+### Store
+
+*Purpose:* Store takes care of storing build results. It provides a
+unified interface to get store paths and upload new ones, as well as querying
+for the existence of a store path and its metadata (references, signatures, …).
+
+Tvix natively uses an improved store protocol. Instead of transferring around
+NAR files, which don't provide an index and don't allow seekable access, a
+concept similar to git tree hashing is used.
+
+This allows more granular substitution, chunk reusage and parallel download of
+individual files, reducing bandwidth usage.
+As these chunks are content-addressed, it opens up the potential for
+peer-to-peer trustless substitution of most of the data, as long as we sign the
+root of the index.
+
+Tvix still keeps the old-style signatures, NAR hashes and NAR size around. In
+the case of NAR hash / NAR size, this data is strictly required in some cases.
+The old-style signatures are valuable for communication with existing
+implementations.
+
+Old-style binary caches (like cache.nixos.org) can still be exposed via the new
+interface, by doing on-the-fly (re)chunking/ingestion.
+
+Most likely, there will be multiple implementations of store, some storing
+things locally, some exposing a "remote view".
+
+A few possible ones that come to mind are:
+
+- Local store
+- SFTP/ GCP / S3 / HTTP
+- NAR/NARInfo protocol: HTTP, S3
+
+A remote Tvix store can be connected by simply connecting to its gRPC
+interface, possibly using SSH tunneling, but there doesn't need to be an
+additional "wire format" like the Nix `ssh(+ng)://` protocol.
+
+Settling on one interface allows composition of stores, meaning it becomes
+possible to express substitution from remote caches as a proxy layer.
+
+It'd also be possible to write a FUSE implementation on top of the RPC
+interface, exposing a lazily-substituting /nix/store mountpoint. Using this in
+remote build context dramatically reduces the amount of data transferred to a
+builder, as only the files really accessed during the build are substituted.
+
+## Figures
+
+```plantuml,format=svg
+{{#include figures/component-flow.puml}}
+```
+
+[^1]: There have already been some discussions in the Nix community, to switch
+  to REAPI:
+  https://discourse.nixos.org/t/a-proposal-for-replacing-the-nix-worker-protocol/20926/22
diff --git a/tvix/docs/src/differences-drv-paths.md b/tvix/docs/src/differences-drv-paths.md
new file mode 100644
index 000000000000..f1c90bd52892
--- /dev/null
+++ b/tvix/docs/src/differences-drv-paths.md
@@ -0,0 +1,35 @@
+# .drvPath inconsistencies / Why .drvPath differs between Nix and Tvix
+
+Nix and Tvix currently use a different approach when it comes to tracking input
+references, in order to build the right dependencies in advance.
+Nix is using string contexts, whereas Tvix is doing reference scanning [^inbox-drvpath].
+
+There are some real-life cases, for example during nixpkgs bootstrapping, where
+multiple different fixed-output derivations are written to produce the same
+hash.
+
+For example, bootstrap sources that are downloaded early are fetched using
+a special "builder hack", in which the `builder` field of the derivation is
+populated with the magic string `builtin:fetchurl` and the builder itself will
+perform a fetch, with everything looking like a normal derivation to the user.
+
+These bootstrap sources are later on defined *again*, once `curl`is available,
+to be downloaded using the standard pkgs.fetchtarball mechanism, but yielding
+the *same* outputs (as the same files are being fetched).
+
+In our reference scanning implementation, this output scanning of FOD will
+cause the path of the *first* derivation producing the given fixed output to be
+stored in the `inputDrvs` field of the derivation, while Nix will point to the
+derivation that was actually used.
+
+This doesn't cause any differences in the calculated *output paths*, as paths to
+fixed-output derivations are replaced with a special
+`fixed:out:${algo}:${digest}:${fodPath}` string that doesn't contain the "path
+to the wrong derivation" anymore.
+
+As we haven't fully determined if our reference scanning approach is gonna work,
+and comparing output paths is sufficient to determine equality of the build
+instructions, this is left as a future work item.
+
+
+[^inbox-drvpath]: https://inbox.tvl.su/depot/20230316120039.j4fkp3puzrtbjcpi@tp/T/#t
diff --git a/tvix/docs/src/figures/component-flow.puml b/tvix/docs/src/figures/component-flow.puml
new file mode 100644
index 000000000000..5b6d79b82313
--- /dev/null
+++ b/tvix/docs/src/figures/component-flow.puml
@@ -0,0 +1,60 @@
+@startuml
+
+title Tvix build flow
+
+actor User
+participant CLI
+participant "Coordinator" as Coord
+participant "Evaluator" as Eval
+database Store
+participant "Builder" as Build
+
+note over CLI,Eval
+    Typically runs locally on the invoking machine
+end note
+/ note over Store, Build
+    Can be either local or remote
+end note
+
+User-->CLI: User initiates build of `hello` (analogous to `nix-build -f '<nixpkgs>' -A hello`)
+
+CLI-->Coord: CLI invokes coordinator
+
+Coord-->Eval: Sends message to start evaluation of `<nixpkgs>` (path lookup) with attribute `hello`
+note right: The paths to the evaluator are local file system paths
+
+Coord<--Eval: Yields derivations to be built
+note right
+    Immediately starts streaming derivations as they are instantiated across
+    the dependency graph so they can be built while the evaluation is still running.
+
+    There are two types of build requests: One for regular "fire and forget" builds,
+    and another for IFD (import from derivation).
+
+    These are distinct because IFD needs to be fed back into the evaluator for
+    further processing while a regular build does not.
+end note
+
+loop while has more derivations
+
+    Coord-->Store: Check if desired paths are in store
+    alt Store has path
+        Coord<--Store: Success response
+    else Store does not have path
+        Coord-->Build: Request derivation to be built
+
+        alt Build failure
+            Coord<--Build: Fail response
+            note left: It's up to the coordinator whether to exit on build failure
+        else Build success
+            Build-->Store: Push outputs to store
+            Build<--Coord: Send success & pushed response
+        end
+
+    end
+end
+
+CLI<--Coord: Respond success/fail
+User<--CLI: Exit success/fail
+
+@enduml
diff --git a/tvix/docs/src/lang-version.md b/tvix/docs/src/lang-version.md
new file mode 100644
index 000000000000..c288274c9105
--- /dev/null
+++ b/tvix/docs/src/lang-version.md
@@ -0,0 +1,62 @@
+# Nix language version history
+
+The Nix language (“Nix”) has its own versioning mechanism independent from its
+most popular implementation (“C++ Nix”): `builtins.langVersion`. It has been
+increased whenever the language has changed syntactically or semantically in a
+way that would not be introspectable otherwise. In particular, this does not
+include addition (or removal) of `builtins`, as this can be introspected using
+standard attribute set operations.
+
+Changes to `builtins.langVersion` are best found by viewing the git history of
+C++ Nix using `git log -G 'mkInt\\(v, [0-9]\\)'` for `builtins.langVersion` < 7.
+After that point `git log -G 'v\\.mkInt\\([0-9]+\\)'` should work. To reduce the
+amount of false positives, specify the version number you are interested in
+explicitly.
+
+## 1
+
+The first version of the Nix language is its state at the point when
+`builtins.langVersion` was added in [8b8ee53] which was first released
+as part of C++ Nix 1.2.
+
+## 2
+
+Nix version 2 changed the behavior of `builtins.storePath`: It would now [try to
+substitute the given path if missing][storePath-substitute], instead of creating
+an evaluation failure. `builtins.langVersion` was increased in [e36229d].
+
+## 3
+
+Nix version 3 changed the behavior of the `==` behavior. Strings would now be
+considered [equal even if they had differing string context][equal-no-ctx].
+
+## 4
+
+Nix version 4 [added the float type][float] to the language.
+
+## 5
+
+The [increase of `builtins.langVersion` to 5][langVersion-5] did not signify a
+language change, but added support for structured attributes to the Nix daemon.
+Eelco Dolstra writes as to what changed:
+
+> The structured attributes support. Unfortunately that's not so much a language
+> change as a build.cc (i.e. daemon) change, but we don't really have a way to
+> express that...
+
+Maybe `builtins.nixVersion` (which was added in version 1) should have been
+used instead. In any case, the [only `langVersion` check][nixpkgs-langVersion-5]
+in nixpkgs verifies a lower bound of 5.
+
+## 6
+
+Nix version 6 added support for [comparing two lists][list-comparison].
+
+[8b8ee53]: https://github.com/nixos/nix/commit/8b8ee53bc73769bb25d967ba259dabc9b23e2e6f
+[storePath-substitute]: https://github.com/nixos/nix/commit/22d665019a3770148929b7504c73bcdbe025ec12
+[e36229d]: https://github.com/nixos/nix/commit/e36229d27f9ab508e0abf1892f3e8c263d2f8c58
+[equal-no-ctx]: https://github.com/nixos/nix/commit/ee7fe64c0ac00f2be11604a2a6509eb86dc19f0a
+[float]: https://github.com/nixos/nix/commit/14ebde52893263930cdcde1406cc91cc5c42556f
+[langVersion-5]: https://github.com/nixos/nix/commit/8191992c83bf4387b03c5fdaba818dc2b520462d
+[list-comparison]: https://github.com/nixos/nix/commit/09471d2680292af48b2788108de56a8da755d661
+[nixpkgs-langVersion-5]: https://github.com/NixOS/nixpkgs/blob/d7ac3423d321b8b145ccdd1aed9dfdb280f5e391/pkgs/build-support/closure-info.nix#L11
diff --git a/tvix/docs/src/language-spec.md b/tvix/docs/src/language-spec.md
new file mode 100644
index 000000000000..0ff1dc491e3c
--- /dev/null
+++ b/tvix/docs/src/language-spec.md
@@ -0,0 +1,67 @@
+# Specification of the Nix Language
+
+WARNING: This document is a work in progress. Please keep an eye on
+[`topic:nix-spec`](https://cl.tvl.fyi/q/topic:nix-spec) for ongoing
+CLs.
+
+Nix is a general-purpose, functional programming language which this
+document aims to describe.
+
+## Background
+
+Nix was designed and implemented as part of the [Nix package
+manager](https://nixos.org/nix). It is primarily used for generating
+so-called [*derivations*](#derivations), which are data structures
+describing how to build a package.
+
+The language has been described in the
+[thesis](https://edolstra.github.io/pubs/phd-thesis.pdf) introducing
+the package manager, but only on a high-level. At the time of writing,
+Nix is informally specified (via its only complete implementation in
+the package manager) and there is no complete overview over its -
+sometimes surprising - semantics.
+
+The primary project written in Nix is
+[nixpkgs](https://github.com/NixOS/nixpkgs/). Uncertainties in the
+process of writing this specification are resolved by investigating
+patterns in nixpkgs, which we consider canonical. The code in nixpkgs
+uses a reasonable subset of the features exposed by the current
+implementation, some of which are *accidental*, and is thus more
+useful for specifying how the language should work.
+
+## Introduction to Nix
+
+Nix is a general-purpose, partially lazy, functional programming
+language which provides higher-order functions, type reflection,
+primitive data types such as integers, strings and floats, and
+compound data structures such as lists and attribute sets.
+
+Nix has syntactic sugar for common operations, such as those for
+attribute sets, and also provides a wide range of built-in functions
+which have organically accumulated over time.
+
+Nix has a variety of legacy features that are not in practical use,
+but are documented in sections of this specification for the sake of
+completeness.
+
+This document describes the syntax and abstract semantics of the Nix
+language, but leaves out implementation details about how Nix can be
+interpreted/compiled/analysed etc.
+
+### Program structure
+
+This section describes the semantic structure of Nix, and how it
+relates to the rest of the specification.
+
+Each Nix program is a single [*expression*](#expressions) denoting a
+[*value*](#values) (commonly a [*function*](#functions)). Each value
+has a [*type*](#types), however this type is not statically known.
+
+Nix code is modularised through the use of the
+[*import*](#builtins-import) built-in function. No separate module
+system exists.
+
+In addition to chapters describing the building blocks mentioned
+above, this specificiation also describes the [*syntax*](#syntax), the
+available [built-in functions](#builtins), [*error handling*](#errors)
+and known [*deficiencies*](#deficiencies) in the language.
diff --git a/tvix/docs/src/value-pointer-equality.md b/tvix/docs/src/value-pointer-equality.md
new file mode 100644
index 000000000000..d84efcb50ca9
--- /dev/null
+++ b/tvix/docs/src/value-pointer-equality.md
@@ -0,0 +1,338 @@
+# Value Pointer Equality in Nix
+
+## Introduction
+
+It is a piece of semi-obscure Nix trivia that while functions are generally not
+comparable, they can be compared in certain situations. This is actually quite an
+important fact, as it is essential for the evaluation of nixpkgs: The attribute sets
+used to represent platforms in nixpkgs, like `stdenv.buildPlatform`, contain functions,
+such as `stdenv.buildPlatform.canExecute`. When writing cross logic, one invariably
+ends up writing expressions that compare these sets, e.g. `stdenv.buildPlatform !=
+stdenv.hostPlatform`. Since attribute set equality is the equality of their attribute
+names and values, we also end up comparing the functions within them.  We can summarize
+the relevant part of this behavior for platform comparisons in the following (true)
+Nix expressions:
+
+* `stdenv.hostPlatform.canExecute != stdenv.hostPlatform.canExecute`
+* `stdenv.hostPlatform == stdenv.hostPlatform`
+
+This fact is commonly referred to as pointer equality of functions (or function pointer
+equality) which is not an entirely accurate name, as we'll see. This account of the
+behavior states that, while functions are incomparable in general, they are comparable
+insofar, as they occupy the same spot in an attribute set.
+
+However, [a maybe lesser known trick][puck-issue] is to write a function such as the
+following to allow comparing functions:
+
+```nix
+let
+  pointerEqual = lhs: rhs: { x = lhs; } == { x = rhs; };
+
+  f = name: "Hello, my name is ${name}";
+  g = name: "Hello, my name is ${name}";
+in
+[
+  (pointerEqual f f) # => true
+  (pointerEqual f g) # => false
+]
+```
+
+Here, clearly, the function is not contained at the same position in one and the same
+attribute set, but at the same position in two entirely different attribute sets. We can
+also see that we are not comparing the functions themselves (e.g. their AST), but
+rather if they are the same individual value (i.e. pointer equal).
+
+To figure out the _actual_ semantics, we'll first have a look at how value (pointer) equality
+works in C++ Nix, the only production ready Nix implementation currently available.
+
+## Nix (Pointer) Equality in C++ Nix
+
+TIP: The summary presented here is up-to-date as of 2023-06-27 and was tested
+with Nix 2.3, 2.11 and 2.15.
+
+### `EvalState::eqValues` and `ExprOpEq::eval`
+
+The function implementing equality in C++ Nix is `EvalState::eqValues` which starts with
+[the following bit of code][eqValues-pointer-eq]:
+
+```cpp
+bool EvalState::eqValues(Value & v1, Value & v2)
+{
+    forceValue(v1);
+    forceValue(v2);
+
+    /* !!! Hack to support some old broken code that relies on pointer
+       equality tests between sets.  (Specifically, builderDefs calls
+       uniqList on a list of sets.)  Will remove this eventually. */
+    if (&v1 == &v2) return true;
+```
+
+So this immediately looks more like pointer equality of arbitrary *values* instead of functions. In fact
+there is [no special code facilitating function equality][eqValues-function-eq]:
+
+```cpp
+        /* Functions are incomparable. */
+        case nFunction:
+            return false;
+```
+
+So one takeaway of this is that pointer equality is neither dependent on functions nor attribute sets.
+In fact, we can also write our `pointerEqual` function as:
+
+```nix
+lhs: rhs: [ lhs ] == [ rhs ]
+```
+
+It's interesting that `EvalState::eqValues` forces the left and right-hand value before trying pointer
+equality. It explains that `let x = throw ""; in x == x` does not evaluate successfully, but it is puzzling why
+`let f = x: x; in f == f` does not return `true`. In fact, why do we need to wrap the values in a list or
+attribute set at all for our `pointerEqual` function to work?
+
+The answer lies in [the code that evaluates `ExprOpEq`][ExprOpEq],
+i.e. an expression involving the `==` operator:
+
+```cpp
+void ExprOpEq::eval(EvalState & state, Env & env, Value & v)
+{
+    Value v1; e1->eval(state, env, v1);
+    Value v2; e2->eval(state, env, v2);
+    v.mkBool(state.eqValues(v1, v2));
+}
+```
+
+As you can see, two _distinct_ `Value` structs are created, so they can never be pointer equal even
+if the `union` inside points to the same bit of memory. We can thus understand what actually happens
+when we check the equality of an attribute set (or list), by looking at the following expression:
+
+```nix
+let
+  x = { name = throw "nameless"; };
+in
+
+x == x # => causes an evaluation error
+```
+
+Because `x` can't be pointer equal, as it'll end up in the distinct structs `v1` and `v2`, it needs to be compared
+by value. For this reason, the `name` attribute will be forced and an evaluation error caused.
+If we rewrite the expression to use…
+
+```nix
+{ inherit x; } == { inherit x; } # => true
+```
+
+…, it'll work: The two attribute sets are compared by value, but their `x` attribute turns out to be pointer
+equal _after_ forcing it. This does not throw, since forcing an attribute set does not force its attributes'
+values (as forcing a list doesn't force its elements).
+
+As we have seen, pointer equality can not only be used to compare function values, but also other
+otherwise incomparable values, such as lists and attribute sets that would cause an evaluation
+error if they were forced recursively. We can even switch out the `throw` for an `abort`. The limitation is
+of course that we need to use a value that behaves differently depending on whether it is forced
+“normally” (think `builtins.seq`) or recursively (think `builtins.deepSeq`), so thunks will generally be
+evaluated before pointer equality can kick into effect.
+
+### Other Comparisons
+
+The `!=` operator uses `EvalState::eqValues` internally as well, so it behaves exactly as `!(a == b)`.
+
+The `>`, `<`, `>=` and `<=` operators all desugar to [CompareValues][]
+eventually which generally looks at the value type before comparing. It does,
+however, rely on `EvalState::eqValues` for list comparisons
+([introduced in Nix 2.5][nix-2.5-changelog]), so it is possible to compare lists
+with e.g. functions in them, as long as they are equal by pointer:
+
+```nix
+let
+  f = x: x + 42;
+in
+
+[
+  ([ f 2 ] > [ f 1 ]) # => true
+  ([ f 2 ] > [ (x: x) 1]) # => error: cannot compare a function with a function
+  ([ f ] > [ f ]) # => false
+]
+```
+
+Finally, since `builtins.elem` relies on `EvalState::eqValues`, you can check for
+a function by pointer equality:
+
+```nix
+let
+  f = x: f x;
+in
+builtins.elem f [ f 2 3 ] # => true
+```
+
+### Pointer Equality Preserving Nix Operations
+
+We have seen that pointer equality is established by comparing the memory
+location of two C++ `Value` structs. But how does this _representation_ relate
+to Nix values _themselves_ (in the sense of a platonic ideal if you will)? In
+Nix, values have no identity (ignoring `unsafeGetAttrPos`) or memory location.
+
+Since Nix is purely functional, values can't be mutated, so they need to be
+copied frequently. With Nix being garbage collected, there is no strong
+expectation when a copy is made, we probably just hope it is done as seldomly as
+possible to save on memory. With pointer equality leaking the memory location of
+the `Value` structs to an extent, it is now suddenly our business to know
+exactly _when_ a copy of a value is made.
+
+Evaluation in C++ Nix mainly proceeds along the following [two
+functions][eval-maybeThunk].
+
+```cpp
+struct Expr
+{
+    /* … */
+    virtual void eval(EvalState & state, Env & env, Value & v);
+    virtual Value * maybeThunk(EvalState & state, Env & env);
+    /* … */
+};
+```
+
+As you can see, `Expr::eval` always takes a reference to a struct _allocated by
+the caller_ to place the evaluation result in. Anything that is processed using
+`Expr::eval` will be a copy of the `Value` struct even if the value before and
+after are the same.
+
+`Expr::maybeThunk`, on the other hand, returns a pointer to a `Value` which may
+already exist or be newly allocated. So, if evaluation passes through `maybeThunk`,
+Nix values _can_ retain their pointer equality. Since Nix is lazy, a lot of
+evaluation needs to be thunked and pass through `maybeThunk`—knowing under what
+circumstances `maybeThunk` will return a pointer to an already existing `Value`
+struct thus means knowing the circumstances under which pointer equality of a
+Nix value will be preserved in C++ Nix.
+
+The [default case][maybeThunk-default] of `Expr::maybeThunk` allocates a new
+`Value` which holds the delayed computation of the `Expr` as a thunk:
+
+```cpp
+
+Value * Expr::maybeThunk(EvalState & state, Env & env)
+{
+    Value * v = state.allocValue();
+    mkThunk(*v, env, this);
+    return v;
+}
+```
+
+Consequently, only special cased expressions could preserve pointer equality.
+These are `ExprInt`, `ExprFloat`, `ExprString`, `ExprPath`—all of which relate
+to creating new values—and [finally, `ExprVar`][maybeThunk-ExprVar]:
+
+```cpp
+Value * ExprVar::maybeThunk(EvalState & state, Env & env)
+{
+    Value * v = state.lookupVar(&env, *this, true);
+    /* The value might not be initialised in the environment yet.
+       In that case, ignore it. */
+    if (v) { state.nrAvoided++; return v; }
+    return Expr::maybeThunk(state, env);
+}
+```
+
+Here we may actually return an already existing `Value` struct. Consequently,
+accessing a value from the scope is the only thing you can do with a value in
+C++ Nix that preserves its pointer equality, as the following example shows:
+For example, using the select operator to get a value from an attribute set
+or even passing a value trough the identity function invalidates its pointer
+equality to itself (or rather, its former self).
+
+```nix
+let
+  pointerEqual = a: b: [ a ] == [ b ];
+  id = x: x;
+
+  f = _: null;
+  x = { inherit f; };
+  y = { inherit f; };
+in
+
+[
+  (pointerEqual f f)      # => true
+
+  (pointerEqual f (id f)) # => false
+
+  (pointerEqual x.f y.f)  # => false
+  (pointerEqual x.f x.f)  # => false
+
+  (pointerEqual x x)      # => true
+  (pointerEqual x y)      # => true
+]
+```
+
+In the last two cases, the example also shows that there is another way to
+preserve pointer equality: Storing a value in an attribute set (or list)
+preserves its pointer equality even if the structure holding it is modified in
+some way (as long as the value we care about is left untouched). The catch is,
+of course, that there is no way to get the value out of the structure while
+preserving pointer equality (which requires using the select operator or a call
+to `builtins.elemAt`).
+
+We initially illustrated the issue of pointer equality using the following
+true expressions:
+
+* `stdenv.hostPlatform.canExecute != stdenv.hostPlatform.canExecute`
+* `stdenv.hostPlatform == stdenv.hostPlatform`
+
+We can now add a third one, illustrating that pointer equality is invalidated
+by select operations:
+
+* `[ stdenv.hostPlatform.canExecute ] != [ stdenv.hostPlatform.canExecute ]`
+
+To summarize, pointer equality is established on the memory location of the
+`Value` struct in C++ Nix. Except for simple values (`int`, `bool`, …),
+the `Value` struct only consists of a pointer to the actual representation
+of the value (attribute set, list, function, …) and is thus cheap to copy.
+In practice, this happens when a value passes through the evaluation of
+almost any Nix expression. Only in the select cases described above
+a value preserves its pointer equality despite being unchanged by an
+expression. We can call this behavior *exterior pointer equality*.
+
+## Summary
+
+When comparing two Nix values, we must force both of them (non-recursively!), but are
+allowed to short-circuit the comparison based on pointer equality, i.e. if they are at
+the same exact value in memory, they are deemed equal immediately. This is completely
+independent of what type of value they are. If they are not pointer equal, they are
+(recursively) compared by value as expected.
+
+However, when evaluating the Nix expression `a == b`, we *must* invoke our implementation's
+value equality function in a way that `a` and `b` themselves can never be deemed pointer equal.
+Any values we encounter while recursing during the equality check must be compared by
+pointer as described above, though.
+
+## Stability of the Feature
+
+Keen readers will have noticed the following comment in the C++ Nix source code,
+indicating that pointer comparison may be removed in the future.
+
+```cpp
+    /* !!! Hack to support some old broken code that relies on pointer
+       equality tests between sets.  (Specifically, builderDefs calls
+       uniqList on a list of sets.)  Will remove this eventually. */
+```
+
+Now, I can't speak for the upstream C++ Nix developers, but sure can speculate.
+As already pointed out, this feature is currently needed for evaluating nixpkgs.
+While its use could realistically be eliminated (only bothersome spot is probably
+the `emulator` function, but that should also be doable), removing the feature
+would seriously compromise C++ Nix's ability to evaluate historical nixpkgs
+revision which is arguably a strength of the system.
+
+Another indication that it is likely here to stay is that it has already
+[outlived builderDefs][], even though
+it was (apparently) reintroduced just for this use case. More research into
+the history of this feature would still be prudent, especially the reason for
+its original introduction (maybe performance?).
+
+[puck-issue]: https://github.com/NixOS/nix/issues/3371
+[eqValues-pointer-eq]: https://github.com/NixOS/nix/blob/3c618c43c6044eda184df235c193877529e951cb/src/libexpr/eval.cc#L2401-L2404
+[eqValues-function-eq]: https://github.com/NixOS/nix/blob/3c618c43c6044eda184df235c193877529e951cb/src/libexpr/eval.cc#L2458-L2460
+[ExprOpEq]: https://github.com/NixOS/nix/blob/3c618c43c6044eda184df235c193877529e951cb/src/libexpr/eval.cc#L1822-L1827
+[outlived builderDefs]: https://github.com/NixOS/nixpkgs/issues/4210
+[CompareValues]: https://github.com/NixOS/nix/blob/3c618c43c6044eda184df235c193877529e951cb/src/libexpr/primops.cc#L569-L610
+[nix-2.5-changelog]: https://nixos.org/manual/nix/stable/release-notes/rl-2.5.html
+[eval-maybeThunk]: https://github.com/NixOS/nix/blob/3c618c43c6044eda184df235c193877529e951cb/src/libexpr/nixexpr.hh#L161-L162
+[maybeThunk-default]: https://github.com/NixOS/nix/blob/8e770dac9f68162cfbb368e53f928df491babff3/src/libexpr/eval.cc#L1076-L1081
+[maybeThunk-ExprVar]: https://github.com/NixOS/nix/blob/8e770dac9f68162cfbb368e53f928df491babff3/src/libexpr/eval.cc#L1084-L1091