This patch adds support for a native stack sampling
profiler to the evaluator, which saves a collapsed stack
profile information to a configurable location.
Introduced options (in `EvalSettings`):
- `eval-profile-file` - path to the collected profile file.
- `eval-profiler-frequency` - sampling frequency.
- `eval-profiler` - enumeration option for enabling the profiler.
Currently only `flamegraph` is supported, but having this an
enumeration rather than a boolean switch leaves the door open
for other profiler variants (e.g. tracy).
Profile includes the following information on best-effort basis (e.g. some lambdas might
have an undefined name). Callstack information contains:
- Call site location (where the function gets called).
- Primop/lambda name of the function being called.
- Functors/partial applications don't have a name attached to them unlike special-cased primops and lambads.
For cases where callsite location isn't available we have to resort to providing
the location where the lambda itself is defined. This removes some of the confusing
`«none»:0` locations in the profile from previous attempts.
Example usage with piping directly into zstd for compression:
```
nix eval --no-eval-cache nixpkgs#nixosTests.gnome \
--eval-profiler flamegraph \
--eval-profile-file >(zstd -of nix.profile.zstd)
```
Co-authored-by: Jörg Thalheim <joerg@thalheim.io>
This wires up the {pre,post}FunctionCallHook machinery
in EvalState::callFunction and migrates FunctionCallTrace
to use the new EvalProfiler mechanisms for tracing.
Note that branches when the hook gets called are marked with [[unlikely]]
as a hint to the compiler that this is not a hot path. For non-tracing
evaluation this should be a 100% predictable branch, so the performance
cost is nonexistent.
Some measurements to prove support this point:
```
nix build .#nix-cli
nix build github:nixos/nix/d692729759e4e370361cc5105fbeb0e33137ca9e#nix-cli --out-link before
```
(Before)
```
$ taskset -c 2,3 hyperfine "GC_INITIAL_HEAP_SIZE=16g before/bin/nix eval nixpkgs#gnome --no-eval-cache" --warmup 4
Benchmark 1: GC_INITIAL_HEAP_SIZE=16g before/bin/nix eval nixpkgs#gnome --no-eval-cache
Time (mean ± σ): 2.517 s ± 0.032 s [User: 1.464 s, System: 0.476 s]
Range (min … max): 2.464 s … 2.557 s 10 runs
```
(After)
```
$ taskset -c 2,3 hyperfine "GC_INITIAL_HEAP_SIZE=16g result/bin/nix eval nixpkgs#gnome --no-eval-cache" --warmup 4
Benchmark 1: GC_INITIAL_HEAP_SIZE=16g result/bin/nix eval nixpkgs#gnome --no-eval-cache
Time (mean ± σ): 2.499 s ± 0.022 s [User: 1.448 s, System: 0.478 s]
Range (min … max): 2.472 s … 2.537 s 10 runs
```
This patch adds an EvalProfiler and MultiEvalProfiler that can be used
to insert hooks into the evaluation for the purposes of function tracing
(what function-trace currently does) or for flamegraph/tracy profilers.
See the following commits for how this is supposed to be integrated into
the evaluator and performance considerations.
The existing header is a bit too big. Now the following use-cases are
separated, and get their own headers:
- Using or implementing an arbitrary store: remaining `store-api.hh`
This is closer to just being about the `Store` (and `StoreConfig`)
classes, as one would expect.
- Opening a store from a textual description: `store-open.hh`
Opening an aribtrary store implementation like this requires some sort
of store registration mechanism to exists, but the caller doesn't need
to know how it works. This just exposes the functions which use such a
mechanism, without exposing the mechanism itself
- Registering a store implementation: `store-registration.hh`
This requires understanding how the mechanism actually works, and the
mechanism in question involves templated machinery in headers we
rather not expose to things that don't need it, as it would slow down
compilation for no reason.
Remove outdated and no longer relevant TODO. It's more confusing
now, since symbol table must now be addressed by uint32_t indices
in order to keep Attr size down to 16 bytes on 64 bit machines.
The intention is to switch to transparent comparators from N3657 for
ordered set containers for strings and using the alias consistently
would simplify things.
This trades off some executable size for measurable lexer performance
improvements.
Note on the explicitly enabling 8bit scanner.
This is needed due to the default behavior of flex (excerpt from the manual [1]):
> Flex’s default behavior is to generate an 8-bit scanner unless you
> use the ‘-Cf’ or ‘-CF’, in which case flex defaults to generating
> 7-bit scanners unless your site was always configured to generate 8-bit
> scanners.
Some quantifyable metrics:
Nixpkgs revision: a6e3f45acf4e817532a861ab0eda4ab5485fecc1
Parsing the largest file in nixpkgs: pkgs/development/haskell-modules/hackage-packages.nix.
(Before this patch)
```
$ nix build github:nixos/nix/9fe3077d4#nix-expr
$ du --apparent-size result/lib/libnixexpr.so
2518 result/lib/libnixexpr.so
$ nix build github:nixos/nix/9fe3077d4#nix-cli
$ taskset -c 2,3 hyperfine "GC_INITIAL_HEAP_SIZE=16g \
result/bin/nix-instantiate --parse \
../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix > /dev/null"
Time (mean ± σ): 375.5 ms ± 6.3 ms [User: 316.9 ms, System: 56.7 ms]
Range (min … max): 368.5 ms … 388.3 ms 10 runs
```
(After the patch)
```
$ nix build .#nix-expr
$ du --apparent-size result/lib/libnixexpr.so
2685 result/lib/libnixexpr.so
$ nix build .#nix-cli
$ taskset -c 2,3 hyperfine "GC_INITIAL_HEAP_SIZE=16g \
result/bin/nix-instantiate --parse \
../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix > /dev/null"
Time (mean ± σ): 326.8 ms ± 4.9 ms [User: 269.5 ms, System: 55.3 ms]
Range (min … max): 319.7 ms … 335.5 ms 10 runs
```
Overall, the change is roughly:
- 2518KiB -> 2685KiB ~ 150 KiB of machine code
- 375ms -> 325ms ~ 50ms
The perf uplift for eval-heavy test cases is obviously less noticeable,
but it doesn't make sense not to take this free perf win.
[1]: https://westes.github.io/flex/manual/Options-Affecting-Scanner-Behavior.html#Options-Affecting-Scanner-Behavior
Rather than "mounting" the store inside an empty virtual filesystem,
just return the store as a virtual filesystem. This is more modular.
(FWIW, it also supports two long term hopes of mind:
1. More capability-based Nix language mode. I dream of a "super pure
eval" where you can only use relative path literals (See #8738), and
any `fetchTree`-fetched stuff + the store are all disjoint (none is
mounted in another) file systems.
2. Windows, where the store dir may include drive letters, etc., and is
thus unsuitable to be the prefix of any `CanonPath`s.
)
Co-authored-by: Eelco Dolstra <edolstra@gmail.com>
Revert most of "Hack together a fix for the public headers"
- The `libmain` change is kept, and one more libmain change is made.
(Need to update Meson and Nix per the package alike).
- The S3 situation is fixed in a different way: the variable is public
now, used in the header, and fixed accordingly.
- Fix TODO for `HAVE_EMBEDDED_SANDBOX_SHELL`
This reverts commit 2b51250534.
For example, instead of doing
#include "nix/store-config.hh"
#include "nix/derived-path.hh"
Now do
#include "nix/store/config.hh"
#include "nix/store/derived-path.hh"
This was originally planned in the issue, and also recent requested by
Eelco.
Most of the change is purely mechanical. There is just one small
additional issue. See how, in the example above, we took this
opportunity to also turn `<comp>-config.hh` into `<comp>/config.hh`.
Well, there was already a `nix/util/config.{cc,hh}`. Even though there
is not a public configuration header for libutil (which also would be
called `nix/util/config.{cc,hh}`) that's still confusing, To avoid any
such confusion, we renamed that to `nix/util/configuration.{cc,hh}`.
Finally, note that the libflake headers already did this, so we didn't
need to do anything to them. We wouldn't want to mistakenly get
`nix/flake/flake/flake.hh`!
Progress on #7876
There are two big changes:
1. Public and private config is now separated. Configuration variables
that are only used internally do not go in a header which is
installed.
(Additionally, libutil has a unix-specific private config header,
which should only be used in unix-specific code. This keeps things a
bit more organized, in a purely private implementation-internal way.)
2. Secondly, there is no more `-include`. There are very few config
items that need to be publically exposed, so now it is feasible to
just make the headers that need them just including the (public)
configuration header.
And there are also a few more small cleanups on top of those:
- The configuration files have better names.
- The few CPP variables that remain exposed in the public headers are
now also renamed to always start with `NIX_`. This ensures they should
not conflict with variables defined elsewhere.
- We now always use `#if` and not `#ifdef`/`#ifndef` for our
configuration variables, which helps avoid bugs by requiring that
variables must be defined in all cases.
The short answer for why we need to do this is so we can consistently do
`#include "nix/..."`. Without this change, there are ways to still make
that work, but they are hacky, and they have downsides such as making it
harder to make sure headers from the wrong Nix library (e..g.
`libnixexpr` headers in `libnixutil`) aren't being used.
The C API alraedy used `nix_api_*`, so its headers are *not* put in
subdirectories accordingly.
Progress on #7876
We resisted doing this for a while because it would be annoying to not
have the header source file pairs close by / easy to change file
path/name from one to the other. But I am ameliorating that with
symlinks in the next commit.
We rely on `yyerror()` instead.
> The variable yynerrs contains the number of syntax errors reported so
> far.
> Normally this variable is global; but if you request a pure parser
> (see A Pure (Reentrant) Parser) then it is a local variable which only
> the actions can access.
https://www.gnu.org/software/bison/manual/html_node/Error-Reporting-Function.html
The underlying issue is that debugger code path was
calling PosTable::operator[] in each eval method.
This has become incredibly expensive since 5d9fdab3de.
While we are it it, I've reworked the code to
not use std::shared_ptr where it really isn't necessary.
As I've documented in previous commits, this is actually
more a workaround for recursive header dependencies now
and is only necessary in `error.hh` code.
Some ad-hoc benchmarking:
After this commit:
```
Benchmark 1: nix eval nixpkgs#hello --impure --ignore-try --no-eval-cache --debugger
Time (mean ± σ): 784.2 ms ± 7.1 ms [User: 561.4 ms, System: 147.7 ms]
Range (min … max): 773.5 ms … 792.6 ms 10 runs
```
On master 3604c7c51:
```
Benchmark 1: nix eval nixpkgs#hello --impure --ignore-try --no-eval-cache --debugger
Time (mean ± σ): 22.914 s ± 0.178 s [User: 18.524 s, System: 4.151 s]
Range (min … max): 22.738 s … 23.290 s 10 runs
```
All of this code doesn't actually depend on anything from
libexpr. Because Pos is so tigtly coupled with Error, it
makes sense to have in the same library.
