While preparing PRs like #9753, I've had to change error messages in
dozens of code paths. It would be nice if instead of
EvalError("expected 'boolean' but found '%1%'", showType(v))
we could write
TypeError(v, "boolean")
or similar. Then, changing the error message could be a mechanical
refactor with the compiler pointing out places the constructor needs to
be changed, rather than the error-prone process of grepping through the
codebase. Structured errors would also help prevent the "same" error
from having multiple slightly different messages, and could be a first
step towards error codes / an error index.
This PR reworks the exception infrastructure in `libexpr` to
support exception types with different constructor signatures than
`BaseError`. Actually refactoring the exceptions to use structured data
will come in a future PR (this one is big enough already, as it has to
touch every exception in `libexpr`).
The core design is in `eval-error.hh`. Generally, errors like this:
state.error("'%s' is not a string", getAttrPathStr())
.debugThrow<TypeError>()
are transformed like this:
state.error<TypeError>("'%s' is not a string", getAttrPathStr())
.debugThrow()
The type annotation has moved from `ErrorBuilder::debugThrow` to
`EvalState::error`.
most EvalState and Expr members defined here could be elsewhere, where
they'd be easier to maintain (not being embedded in a file with arcane
syntax) and *somewhat* more faithfully placed according to the path of
the file they're defined in.
Also move `SourcePath` into `libutil`.
These changes allow `error.hh` and `error.cc` to access source path and
position information, which we can use to produce better error messages
(for example, we could consider omitting filenames when two or more
consecutive stack frames originate from the same file).
since `up` and `values` are both pointer-aligned the type field will
also be pointer-aligned, wasting 48 bits of space on most machines. we
can get away with removing the type field altogether by encoding some
information into the `with` expr that created the env to begin with,
reducing the GC load for the absolutely massive amount of single-entry
envs we create for lambdas. this reduces memory usage of system eval by
quite a bit (reducing heap size of our system eval from 8.4GB to 8.23GB)
and gives similar savings in eval time.
running `nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'`
before:
Time (mean ± σ): 5.576 s ± 0.003 s [User: 5.197 s, System: 0.378 s]
Range (min … max): 5.572 s … 5.581 s 10 runs
after:
Time (mean ± σ): 5.408 s ± 0.002 s [User: 5.019 s, System: 0.388 s]
Range (min … max): 5.405 s … 5.411 s 10 runs
this also reduces forceValue code size and removes the need for
hideInDiagnostics. coopting thunk forcing like this has the additional
benefit of clarifying how these errors can happen in the first place.
This makes the position object used in exceptions abstract, with a
method getSource() to get the source code of the file in which the
error originated. This is needed for lazy trees because source files
don't necessarily exist in the filesystem, and we don't want to make
libutil depend on the InputAccessor type in libfetcher.
after #6218 `Symbol` no longer confers a uniqueness invariant on the
string it wraps, it is now possible to create multiple symbols that
compare equal but whose string contents have different addresses. this
guarantee is now only provided by `SymbolIdx`, leaving `Symbol` only as
a string wrapper that knows about the intricacies of how symbols need to
be formatted for output.
this change renames `SymbolIdx` to `Symbol` to restore the previous
semantics of `Symbol` to that name. we also keep the wrapper type and
rename it to `SymbolStr` instead of returning plain strings from lookups
into the symbol table because symbols are formatted for output in many
places. theoretically we do not need `SymbolStr`, only a function that
formats a string for output as a symbol, but having to wrap every symbol
that appears in a message into eg `formatSymbol()` is error-prone and
inconvient.
this slightly increases the amount of memory used for any given symbol, but this
increase is more than made up for if the symbol is referenced more than once in
the EvalState that holds it. on average every symbol should be referenced at
least twice (once to introduce a binding, once to use it), so we expect no
increase in memory on average.
symbol tables are limited to 2³² entries like position tables, and similar
arguments apply to why overflow is not likely: 2³² symbols would require as many
string instances (at 24 bytes each) and map entries (at 24 bytes or more each,
assuming that the map holds on average at most one item per bucket as the docs
say). a full symbol table would require at least 192GB of memory just for
symbols, which is well out of reach. (an ofborg eval of nixpks today creates
less than a million symbols!)
Pos objects are somewhat wasteful as they duplicate the origin file name and
input type for each object. on files that produce more than one Pos when parsed
this a sizeable waste of memory (one pointer per Pos). the same goes for
ptr<Pos> on 64 bit machines: parsing enough source to require 8 bytes to locate
a position would need at least 8GB of input and 64GB of expression memory. it's
not likely that we'll hit that any time soon, so we can use a uint32_t index to
locate positions instead.
the only use of this function is to determine whether a lambda has a non-set
formal, but this use is arguably better served by Symbol::set and using a
non-Symbol instead of an empty symbol in the parser when no such formal is present.
speeds up parsing by ~3%, system builds by a bit more than 1%
# before
Benchmark 1: nix search --offline nixpkgs hello
Time (mean ± σ): 574.7 ms ± 2.8 ms [User: 566.3 ms, System: 8.0 ms]
Range (min … max): 569.2 ms … 580.7 ms 50 runs
Benchmark 2: nix eval -f ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
Time (mean ± σ): 394.4 ms ± 0.8 ms [User: 361.8 ms, System: 32.3 ms]
Range (min … max): 392.7 ms … 395.7 ms 50 runs
Benchmark 3: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
Time (mean ± σ): 2.976 s ± 0.005 s [User: 2.757 s, System: 0.218 s]
Range (min … max): 2.966 s … 2.990 s 50 runs
# after
Benchmark 1: nix search --offline nixpkgs hello
Time (mean ± σ): 572.4 ms ± 2.3 ms [User: 563.4 ms, System: 8.6 ms]
Range (min … max): 566.9 ms … 579.1 ms 50 runs
Benchmark 2: nix eval -f ../nixpkgs/pkgs/development/haskell-modules/hackage-packages.nix
Time (mean ± σ): 381.7 ms ± 1.0 ms [User: 348.3 ms, System: 33.1 ms]
Range (min … max): 380.2 ms … 387.7 ms 50 runs
Benchmark 3: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
Time (mean ± σ): 2.936 s ± 0.005 s [User: 2.715 s, System: 0.221 s]
Range (min … max): 2.923 s … 2.946 s 50 runs
We now parse function applications as a vector of arguments rather
than as a chain of binary applications, e.g. 'substring 1 2 "foo"' is
parsed as
ExprCall { .fun = <substring>, .args = [ <1>, <2>, <"foo"> ] }
rather than
ExprApp (ExprApp (ExprApp <substring> <1>) <2>) <"foo">
This allows primops to be called immediately (if enough arguments are
supplied) without having to allocate intermediate tPrimOpApp values.
On
$ nix-instantiate --dry-run '<nixpkgs/nixos/release-combined.nix>' -A nixos.tests.simple.x86_64-linux
this gives a substantial performance improvement:
user CPU time: median = 0.9209 mean = 0.9218 stddev = 0.0073 min = 0.9086 max = 0.9340 [rejected, p=0.00000, Δ=-0.21433±0.00677]
elapsed time: median = 1.0585 mean = 1.0584 stddev = 0.0024 min = 1.0523 max = 1.0623 [rejected, p=0.00000, Δ=-0.20594±0.00236]
because it reduces the number of tPrimOpApp allocations from 551990 to
42534 (i.e. only small minority of primop calls are partially
applied) which in turn reduces time spent in the garbage collector.
