This undoes a previous change to unify them, and I think at my advice.
=[ Sorry about that, I think I was just wrong.
Specifically, I think I had suggested that it would be more efficient to
have a single shared hashtable of strings. The more I look at profiles
of the toolchain, the less likely that seems. Specifically for
identifiers and string literals it seems especially problematic.
Using a single, joint hashtable is likely a good idea when all of the
different querying code paths are equally likely, the strings follow the
same distribution of sizes, and either there is no clustering of access
to different sets of strings or none of the sets are meaningfully small
enough to fit into a lower level of resident cache.
I think essentially none of these predicates actually hold for
identifiers vs. string literals:
- Identifiers are *much* more hot
- They have wildly different size distributions.
- The access patterns are very clustered
Sorry for the misleading advice on that one.
While splitting them, I've worked to simplify the code a bit by building
a way to have the `StringRef` holding canonical value stores not require
specializations, and so we get a pretty large code cleanup in the
process here.
This works to leverage the capabilities of the hashtable as much as
possible, for example using the key context in the value stores.
However, there may still be opportunities to refactor more deeply and
use the functionality even better. Hopefully this is at least
a reasonable start and gets us a clean baseline.
On an Arm M1, this is a 15% improvement on my large lexing stress test,
but ends up a wash on my x86-64 server. This is a smaller benefit than
I expected, and it's because we're using a set-of-IDs and looking up
values with a key context for things like identifiers. This pattern has
a surprising tradeoff. The new hashtable uses significantly less memory,
a 10% peak RSS reduction just from the hashtable change. But indirecting
through the vector of values makes growing the hashtable dramatically
less cache-friendly: it causes growth to randomly access every key when
rehashing. On x86, everything gained by the faster hashtable is lost in
even slower growth. And even on Arm, this eats into the benefits.
But I have a plan to tweak how identifiers specifically work to avoid
most of the growth, and so I suspect this is the right tradeoff on the
whole. It gives us significant working set size reduction and we can
likely avoid the regressed operation (growth with rehash) in most cases
by clever reserving and if necessary by adding a hash caching layer to
the table infrastructure.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
I found this through inspection, looking at an array stack data type.
Tests pass either way, not sure what a good test would be for
regressions (tests do fail if the size doesn't match, but either
approach gets an appropriate size). But this is followed by
`truncate(remaining_compile_time_bindings)`, so it seems like
`drop_back` is a better match than `drop_front`.
When forming a `ConstantId` for a symbolic constant, add storage to
track the generic in which the constant was formed and the index within
that generic. These fields are not yet populated.
This injects a customization point for hashtable-specific equality
testing that the key context uses by default. While this is rarely
needed, there are LLVM types where it is necessary and it seems a good
general tool to have to avoid unnecessary complexity from custom key
contexts when a simple customization of equality is all that is
required.
This also adds a CRTP mixin for implementing a common pattern of key
contexts where the context provides translation of some key types into
another type, potentially using state. Rather than having to implement
the entire key context API, code can derive from this template and
simply provide a set of overloads for the types it wants to translate.
Any key types used which can be passed to one of those overloads will
get translated before following the same logic as the default key
context. While this updates the only usage so far of this pattern, a
subsequent PR will add several more users making the pattern worth
abstracting here.
LLVM's `APInt` and `APFloat` need specialized handling to be used
effectively in hashtables. We can't inject overrides into LLVM so we
need to handle them in our hashing routine.
There were also problematic limits on hashing pairs and tuples. First,
the unique-object-representation hashing of pairs was more restricted
than tuples which was a problematic asymmetry and isn't needed. But the
larger issue is that we didn't support recursively hashing when
necessary. That requires a careful predicate to avoid infinite recursion
but lets us handle important use cases for hashtables with a tuple as a
key.
Also added support for hashing arrays that recurse in addition to arrays
where we can hash the raw storage, and added overloads to redirect to
common array handling from various array-like types.
Last but not least, re-worked the constraint model for hashing as raw
data to not override custom hashing functions.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Co-authored-by: Carbon Infra Bot <carbon-external-infra@google.com>
When checking a declaration or definition of a generic, track a list of
created instructions that depend on the generic's parameters in some
way, along with information on how they depend on the parameters. This
will eventually be used to determine what information we need to compute
when creating instances of the generic, but for now we're just building
the list.
Information is tracked separately for the declaration region and the
definition region of the generic, because in general these may be first
provided in separate declarations, and they should be substituted into
at different times.
This is awkward to track... Probably it would be best done by tracking
the ratio of unique IDs to lines as a floating point and plot them and
see what a best fit distribution curve looks like. But none of the
histogram printing or stats tracking stuff already in use here makes it
easy to do any of that...
So this does what I hope is a reasonable rough approximation by counting
the ceiling of unique identifiers per 10 lines of code, and plotting
that discreet histogram. Shape of the histogram is exactly what I would
expect: one centered distribution, vaguely normal looking. And the
center for a bunch of different codebases, including our toolchain, is
exactly at 5, which would mean 0.5 unique IDs per line. And the
distribution is pretty reliably bounded above by 10 or 1 unique ID per
line. Which almost seems to clean to be true? Slightly worried about
confirmation bias making me think this code is working because the
results look so pretty.
Here is the output for the toolchain:
```
## Unique IDs per 10 lines ## (median: 6)
2 ids [ 2] █▎
3 ids [19] ████████████▎
4 ids [32] ████████████████████▋
5 ids [55] ███████████████████████████████████▌
6 ids [62] ████████████████████████████████████████
7 ids [44] ████████████████████████████▍
8 ids [22] ██████████████▎
9 ids [11] ███████▏
10 ids [ 7] ████▌
11 ids [ 2] █▎
```
And here is the output for llvm-project/*/{lib,include} (to avoid
tests):
```
# Unique IDs per 10 lines ## (median: 5)
1 ids [ 29] ▍
2 ids [ 282] ███▊
3 ids [1492] ███████████████████▉
4 ids [2674] ███████████████████████████████████▌
5 ids [3011] ████████████████████████████████████████
6 ids [2267] ██████████████████████████████▏
7 ids [1549] ████████████████████▋
8 ids [ 817] ██████████▉
9 ids [ 301] ████
10 ids [ 98] █▎
11 ids [ 61] ▊
12 ids [ 50] ▋
13 ids [ 25] ▍
14 ids [ 33] ▌
15 ids [ 14] ▏
16 ids [ 15] ▎
17 ids [ 9] ▏
18 ids [ 8] ▏
19 ids [ 12] ▏
20 ids [ 15] ▎
21 ids [ 3]
22 ids [ 8] ▏
23 ids [ 3]
24 ids [ 3]
25 ids [ 6] ▏
26 ids [ 0]
27 ids [ 2]
28 ids [ 0]
29 ids [ 0]
30 ids [ 3]
31 ids [ 1]
32 ids [ 1]
```
So, I noticed that ImportIRInst didn't print properly while trying to
debug an issue, and that's where this started. Then I was sort of trying
to figure out why we have "type_blocks" but "typeBlock", so trying to
make that more consistent. "importIRInst0" felt more odd than
"import_ir_inst0" which is why I'm suggesting down this particular
route, but let me know if you'd prefer the reverse (but then do we also
do "typeBlocks", etc, removing the consistency with the member name?)
Also starting to print more detail on import_irs, and added
import_ir_insts (adjusting formatting for that too).
Technically, any small size buffer's lifetime has ended by the time we
get to the base's destructor. This means its no longer valid to access
table contents if stored there from the base destructor. We need to
handle destruction in the table class instead.
This ends up being a trivial change because the logic is already
factored out, we just need to call it from a different point.
This refactors how the implicit import is handled in order to retain
more name scope information. As a consequence, private access control
works better between api files and implementation files. Note though
that this will also be essential for name poisoning between the API and
implementation, as discussed in #3763.
In implementing this, I ran into a couple issues with namespaces that I
think point to flaws in their handling. I've fixed some and added a TODO
for the biggest issue (in check.cpp line 281-288), which relates to the
handling of namespaces of direct imports which are first evaluated
indirectly.
In a `class C(T:! type)`, the type `Self` should be `C(T)`, not merely
`C`. Similarly, in an `interface I(T:! type)`, the type of self should
be `I(T)`, not merely `I`.
In `ClassType`s and `InterfaceType`s, track a `GenericInstanceId` for
the instance rather than just the argument list.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
Also add a corresponding value store and YAML output.
We don't create any generic instances in this change; this is just
adding infrastructure for future changes.
Build a `Generic` object for generic functions. This object tracks the
generic parameters that are in scope for the generic entity. Eventually
it will track other information about the generic too.
Add basic SemIR formatting support for generic functions.
Python [PEP-585](https://peps.python.org/pep-0585/) replaces a number of
`typing` module types with built-in equivalents and `collections.abc`
versions as of Python 3.9, with the aim of eventually removing the
`typing` module versions of these classes altogether. Since the minimum
required version of Python listed in the [Contribution Tools
document](https://github.com/carbon-language/carbon-lang/blob/trunk/docs/project/contribution_tools.md#main-tools)
is 3.9, the type hints in the various python files in the repo can be
updated to this style of type hint without a need for backwards
compatibility.
Feel free to close if this isn't a desired change at this time!
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
When constant evaluation produces a known non-symbolic value, treat the
result as a symbolic constant anyway if the type of the value is
symbolic.
We don't yet have many ways to produce a constant that has a known value
but a symbolic type. The added test case is one such way: an array `[T;
0]` initialized from `()` is a symbolic constant only because its type
is symbolic -- we know its value is always `()`. More ways to form such
constants will be appearing soon as we start to support generics: for
example, a method of a generic class has a symbolic type but a known
constant value of `{}`.
When substituting into a symbolic constant, also substitute into its
type.
We'd been discussing that explorer remains necessary for print, and I
was wondering if this kind of approach would be okay (we _probably_ want
this to work, based on #2110, albeit with more overloads -- but I don't
think there's a good way to support overloads at the moment).
```
╚╡../bazel-bin/examples/sieve
2
3
5
7
11
13
17
19
23
29
31
37
41
43
...
```
This switches from a macro that simply wraps genrules to a proper
Starlark rule that runs first compile and then link actions.
Most interestingly, this uses the rule structure to allow using the
Carbon toolchain built either in the target config or the exec config.
While the exec config is more principled and even necessary in a
cross-compile situaiton, it is dramatically less efficient when
developing Carbon as all the binaries and tests outside of our examples
will be built with the target config. This triggers a complete second
build of the toolchain in the exec config for examples before this PR.
It is tempting to try to keep the exec config but make it not cause
redundant actions, but the way Bazel sets up exec and target config
makes it essentially impossible to share their artifacts. There used to
be a hack in Bazel itself to force sharing but it was removed due to it
violating the principled design. Instead, these rules are explicit about
their intent to use the target config, much like a test would be.
I have rigged up a flag that is carefully threaded through a wrapper
macro with `select`s to allow easily switching to the exec configuration
in case it is desired or needed. But the the `.bazelrc` sets the default
to the target config. The `BUILD` file default is the principled `exec`
in case these rules are used by importing into some other Bazel
workspace where we might *only* need the exec config.
The net outcome of this is shaving over 2500 actions off of a clean
rebuild such as is triggered by a version bump to LLVM, including some
of the very slow and expensive compiles of LLVM and Clang themselves.
These would only be triggered if you built the examples so this may
mostly impact our CI latency.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
This configures the nightly build to generate release notes and provides
a template for organizing them. The organization is done through
labeling of PRs and categorizing them based on those labels. I've tried
to provide a rough categorization that seems reasonable for folks.
While doing this I looked at our PR labeling and found a few bugs that
were preventing many labels form being attached. I've fixed those and
significantly expanded the coverage of file-based labeling. I've also
added code to do author-based labeling for automated PRs so those can be
separated out from human PRs.
Updating to 1.8.4 breaks the patch file, so I'm looking at solutions
that don't require maintaining a patch.
Verifying this is working with `bazel build
//toolchain/lex:tokenized_buffer_benchmark && strings
bazel-bin/toolchain/lex/tokenized_buffer_benchmark |& grep pfm`
This turns out to be super useful now that we have the `key_context`
mechanism and can do much more meaningful heterogeneous lookups, where
the stored key can be *very* different from the lookup key.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
Name scopes store the names in their scope in a `DenseMap`. Several
places reasonably avoid depending on the iteration order by sorting the
names -- they're in the formatting code path where that's a solid
approach.
Unfortunately, when we're importing one scope into another, we also need
to walk the entire scope and do something for each name. =[ This doesn't
seem like a great place to sort things to stabilize them.
I've switched to a fairly simplistic solution of having a vector of name
entries that can be iterated stably, and a separate map for lookups. I
didn't use the set-of-indices trick here because it's not clear that's
the right trade-off for a scope: likely a lot of small scopes here with
relatively hot name lookups. And the key here isn't a large or
dynamically sized thing that we're canonicalizing, it's a `NameId`. That
made me lean towards duplicating the name in the hashtable for lookup
and the vector for iteration.
I thought about a fancy approach of sorting the hashtable keys by their
values (the indices), but that would still require a bit of copying and
more code.
I also thought a bit about other optimizations, but decided to leave a
comment for now -- it's not obvious to me exactly how hot this is and
whether it's better served by faster lookups, being more memory dense,
etc. And that might involve more of an SOA layout change or some other
approach. Rather than do that here, and especially before switching
hashtables, I stuck with a simple approach to address the ordering.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
This lets copy and move assignment work. While it's a bit suboptimal to
do assignment with these tables, it still seems like an unreasonable
burden to not allow the basics to work. Even the toolchain ended up
doing this in a few places.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
The current approach was done to work-around a limitation that we can
only have a single link per box, but is unscalable. Instead have a
single link to a new sections of the document that describes the box,
and has multiple links.
---------
Co-authored-by: Josh L <josh11b@users.noreply.github.com>
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
There was a file that got into the testing filegroups but not the actual
built tarball. The result was that the installation failed to locate
itself correctly.
We also didn't have any testing to catch this. I'd like to add a bit
more end-to-end testing long-term, but for now just add a Python test
that validates the same set of files are in both.
Sadly, building and testing the compressed release is really slow and
not likely worthwhile outside of the actual release, but it would be
good to catch this stuff earlier. I tried switching from `bz2` to `gz`,
which made very little file size difference (so we should do it
anyways), and it is still too slow to do routinely. Instead, I've added
a Starlark macro that builds both the `pkg_tar` and the `py_test` to
validate it for both uncompressed `tar` and compressed `tar.gz`. This
lets us continuously test the `tar` version, and just test the `tar.gz`
in the nightly release run.
Updates the nightly release workflow to both test, run this specific
test, and use `gz`.
Last but not least, removes the `pkg_zip` as I don't think we have a use
case for this yet and some TODOs that should have been removed when the
version got added.
There is still a lot more we should try to do here, but this gets us
a decent start and should be enough for some of the developers working
on Carbon to use for simple cases, or tools like Compiler Explorer.
This first factors most of the complex setup for our CI workflow into
local actions that we can reuse in other workflows. It's actually
a fairly nice cleanup I think even without the specific use case of
a nightly release process. These use "composite" actions which is the
cleanest model for factoring out a subsequence of steps from a workflow.
I've tried to keep the inputs fairly minimal and easy to spot.
Next, it adds a cron-scheduled action that builds and creates a nightly
release. It also supports explicitly running the workflow, which can be
useful if the nightly release is broken in some way: a fix can be landed
and then a force run used.
Tested the `tests` workflow using our `action-test` branch:
https://github.com/carbon-language/carbon-lang/actions/runs/9545546278
Also tested the nightly release action by hacking in a `push` trigger
and using *draft* release. If you have admin rights, you can see an
example draft nightly build here:
https://github.com/carbon-language/carbon-lang/releases/tag/untagged-6a85cdce7b3a784a2e6b
Note that the GitHub releases are easily deleted, and even the tags
created in the repository can be cleaned up as needed so it should be
fairly harmless to fix this forward if needed.
With bazel 7.2.0, there are some dependency changes, resulting in:
```
WARNING: For repository 'bazel_skylib', the root module requires module version bazel_skylib@1.5.0, but got bazel_skylib@1.6.1 in the resolved dependency graph.
WARNING: For repository 'platforms', the root module requires module version platforms@0.0.8, but got platforms@0.0.9 in the resolved dependency graph.
```
This goes around doing some updates... protobuf's include structure has
changed a little too.
Not sure what changed (I think an upstream LLVM change, but maybe a
Linux distro change), but several folks have been running into problems
finding a standard C++ library when running the Carbon link step. It was
actually breaking our example build when LLD didn't find the right
libstdc++ install, but it finds one reliably on our build bots and
sometimes for some of the developers.
For now, just remove the C++ standard library from the link. We're not
doing that level of interop, and the plan is really to do that not with
the system standard C++ library but by building and bundling libc++ with
the installed toolchain.
Left a comment explaining what's going on here.
It turns out we can make these work with very minimal complexity because
the LF is still in the right place either way. This also lets us easily
support mixtures of LF and CR+LF line endings gracefully. We create the
line structures around the LF bytes and then have the byte-dispatch loop
notice a CR followed by an LF and skip to the LF behavior.
Rather than add the remaining complexity around supporting bare CR and
LF+CR sequences (both of which are quite rare now), this just adds
diagnostics when we encounter a CR byte that won't fall out of our CR+LF
handling. This is a better experience for users than the alternative. We
still have a TODO to handle the full complexity of vertical whitespace,
but I've updated it to reflect that the common case should be handled
already.
This isn't complete though: we need to add support in string literal
lexing, and we need to teach the diagnostic rendering to handle the
error messages above better. But those will be future PRs, this is
enough to unblock folks who happen to edit a Carbon source file with
notepad on Windows which seems important.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
I inadvertently didn't commit and push the actual fixes suggested and
made to that PR, and didn't notice this before putting it into the merge
queue. =[ Very sorry, these were supposed to be in that PR.
This adds a defined Carbon version to the Bazel build and codebase that
can be used both to implement features like version checks and to report
a meaningful version on the command line. This replaces a hard-coded
string and a TODO in the driver.
As part of this, it adds support for defining the version in Bazel, and
special build flags for overriding relevant parts such as the
pre-release marker used. The exact structure and meaning of our version
string, including the pre-release parts, is implemented here in line
with the draft proposal:
https://docs.google.com/document/d/11S5VAPe5Pm_BZPlajWrqDDVr9qc7-7tS2VshqO0wWkk/edit?resourcekey=0-2YFC9Uvl4puuDnWlr2MmYw
This also introduces a workspace status command to the repository to
extract the git commit SHA and other information when building, and the
logic to stamp that into binaries as part of the version string when
useful. The technique used leverages weak symbols with whole archive
linking to allow a link-time override of unstamped data with stamped
data in the leaf executable. This makes building with `--stamp` a
reasonable default, especially for development builds. The CI system is
explicitly opted out of this as there it has no benefit.
Last but not least, all of these are wired into the install rules so
that we build installable packages with the version number in a
conventional place in the directory and filename.
---------
Co-authored-by: Jon Ross-Perkins <jperkins@google.com>
Require mapping from a `ConstantId` to an `InstId` to go through the
`ConstantValueStore`.
This is a preparatory step for an upcoming generics change where
symbolic `ConstantId`s are no longer just a thin wrapper around an
`InstId` but instead are indexes into a table with additional
information about the symbolic constant beyond its `InstId`.
Instead of redundantly storing both the `return_type_id` and
`return_storage_id`, where the declared return type is just the type of
the return storage, store only the `return_storage_id`.
Add a convenience property to get the declared return type of the
function.
In addition to avoiding storing redundant information, this is a
preparatory step for an upcoming change for generics support that will
make it more expensive and awkward to store `TypeId`s in places other
than the type of an instruction.
Model such calls as call operations, and compute the resulting type in
constant evaluation rather than in type-checking. This means we no
longer form non-constant `ClassType` or `InterfaceType` values, and that
we produce an `<error>` type in the case of bad argument lists rather
than a broken / meaningless `ClassType` / `InterfaceType` that doesn't
actually represent a type. This in turn suppresses some follow-on
diagnostics.
This adds two different growth APIs. This is instead of the more
conventional STL `reserve` method. One allows users that aren't trying
to grow in anticipation of an *exact* count of insertions, but generally
trying to size the table to the correct ballpark with a power-of-two
estimate.
The other API allows pre-growing to allow a specific number of
insertions to be performed without further growth. This API takes the
maximum load factor and other implementation details into account.
---------
Co-authored-by: josh11b <15258583+josh11b@users.noreply.github.com>
While it was convenient once to have an immediate check while inserting,
it is indeed far too quadratic. The test was taking 30-60 seconds for
me. =[ So most of the fix here is just to stop doing the check on every
insertion for all previous elements.
There were a few other somewhat slow steps, and I tried to pull those
back as well. I don't think we lose any utility here. Now everything
runs nice and quickly. =]
This uses a header-only extraction of the Boost unordered hashtable
project to allow a trivial Bazel build and for us to benchmark against
it effectively.
Previously we just looked at the raw count of probed keys. Now, we
compute the average and max of both the probe _distance_ measured in the
number of _groups_ probed, and the number of probe _compares_ measured
in the compares required _before_ finding the matching entry.
This lets us understand the relative impact of probe-distance vs. tag
collisions on a given set of benchmark keys. Some of this is motivated
by considering additional optimization techniques similar to those used
in Boost's table and the F14 table from Facebook/Meta.
---------
Co-authored-by: Richard Smith <richard@metafoo.co.uk>
The hash table design is heavily based on Abseil's ["Swiss
Tables"][swiss-tables] design. It uses an array of bytes storing
metadata about each entry and an array of entries where each is a pair
of key and value. The metadata byte consists of 7-bits of hash of the
key (distinct from the bits used to index the table), and one bit
indicating the presence of a special entry -- either empty or deleted.
[swiss-tables]: https://abseil.io/about/design/swisstables
There are a large range of optimizations and other nuanced aspects of
this hash table design and implementation, a good point to understand
that context is `raw_hashtable.h` which has an overview of the design
and references to various other files for relevant details.
---------
Co-authored-by: josh11b <15258583+josh11b@users.noreply.github.com>
Chandler Carruth
Richard Smith
Jack McCluskeyandJon Ross-Perkins
Geoff Romer
Danial Klimkin
dependabot[bot]