This teaches XIRF to optionally refine Text into RefinedText, which
determines whether the given SymbolId represents entirely whitespace.
This is something I've been putting off for some time, but now that I'm
parsing source language for NIR, it is necessary, in that we can only permit
whitespace Text nodes in certain contexts.
The idea is to capture the most common whitespace as preinterned
symbols. Note that this heuristic ought to be determined from scanning a
codebase, which I haven't done yet; this is just an initial list.
The fallback is to look up the string associated with the SymbolId and
perform a linear scan, aborting on the first non-whitespace character. This
combination of checks should be sufficiently performant for now considering
that this is only being run on source files, which really are not all that
large. (They become large when template-expanded.) I'll optimize further
if I notice it show up during profiling.
This also frees XIR itself from being concerned by Whitespace. Initially I
had used quick-xml's whitespace trimming, but it messed up my span
calculations, and those were a pain in the ass to implement to begin with,
since I had to resort to pointer arithmetic. I'd rather avoid tweaking it.
tameld will not check for whitespace, since it's not important---xmlo files,
if malformed, are the fault of the compiler; we can ignore text nodes except
in the context of code fragments, where they are never whitespace (unless
that's also a compiler bug).
Onward and yonward.
DEV-7145
This isn't conceptally all that significant of a change, but there was a lot
of modify to get it working. I would generally separate this into a commit
for the implementation and another commit for the integration, but I decided
to keep things together.
This serves a role similar to AttrSpan---this allows deriving a span
representing the element name from a span representing the entire XIR
token. This will provide more useful context for errors---including the tag
delimiter(s) means that we care about the fact that an element is in that
position (as opposed to some other type of node) within the context of an
error. However, if we are expecting an element but take issue with the
element name itself, we want to place emphasis on that instead.
This also starts to consider the issue of span contexts---a blob of detached
data that is `Span` is useful for error context, but it's not useful for
manipulation or deriving additional information. For that, we need to
encode additional context, and this is an attempt at that.
I am interested in the concept of providing Spans that are guaranteed to
actually make sense---that are instantiated and manipulated with APIs that
ensure consistency. But such a thing buys us very little, practically
speaking, over what I have now for TAMER, and so I don't expect to actually
implement that for this project; I'll leave that for a personal
project. TAMER's already take a lot of my personal interests and it can
cause me a lot of grief sometimes (with regards to letting my aspirations
cause me more work).
DEV-7145
RSG (Ryan Specialty Group) recently announced a rename to Ryan Specialty (no
"Group"), but I'm not sure if the legal name has been changed yet or not, so
I'll wait on that.
Like previous commits, this replaces the explicit escaping context with the
convention that all values retrieved from `xir` are unescaped on read and
escaped on write.
Comments are a notable TODO, since we must escape only `--`.
CData is also an issue. I had _expected_ to use it as a means to avoid
unescaping fragments, but I had forgotten that quick_xml hard-codes escaping
on read, so that it can re-use BytesStart! That is terribly unfortunate,
and may result in us having to re-implement our own read method in the
future to avoid this nonsense. So I'm just leaving it as a TODO for now.
DEV-11081
This rewrites a good portion of the previous commit.
Rather than explicitly storing whether a given string has been escaped, we
can instead assume that all SymbolIds leaving or entering XIR are unescaped,
because there is no reason for any other part of the system to deal with
such details of XML documents.
Given that, we need only unescape on read and escape on write. This is
customary, so why didn't I do that to begin with?
The previous commit outlines the reason, mainly being an optimization for
the echo writer that is upcoming. However, this solution will end up being
better---it's not implemented yet, but we can have a caching layer, such
that the Escaper records a mapping between escaped and unescaped SymbolIds
to avoid work the next time around. If we share the Escaper between _all_
readers and the writer, the result is that
1. Duplicate strings between source files and object files (many of which
are read by both the linker and compiler) avoid re-unescaping; and
2. Writers can use this cache to avoid re-escaping when we've already seen
the escaped variant of the string during read.
The alternative would be a global cache, like the internment system, but I
did not find that to be appropriate here, since this is far less
fundamental and is much easier to compose.
DEV-11081
I'm not fond of this implementation, which is why it's not fully
completed. I wanted to commit this for future reference, and take the
opportunity to explain why I don't like it.
First: this task started as an idea to implement a third variant to
AttrValue and friends that indicates that a value is fixed, in the sense of
a fixed-point function: escaped or unescaped, its value is the same. This
would allow us to skip wasteful escape/unescape operations.
In doing so, it became obvious that there's no need to leak this information
through the API, and indeed, no part of the system should care. When we
read XML, it should be unescaped, and when we write, it should be
escaped. The reason that this didn't quite happen to begin with was an
optimization: I'll be creating an echo writer in place of the current
filesystem-based copy in tamec shortly, and this would allow streaming XIR
directly from the reader to the writer without any unescaping or
re-escaping.
When we unescape, we know the value that it came from, so we could simply
store both symbols---they're 32-bit, so it results in a nicely compressed
64-bit value, so it's essentially cost-free, as long as we accept the
expense of internment. This is `XirString`. Then, when we want to escape
or unescape, we first check to see whether a symbol already exists and, if
so, use it.
While this works well for echoing streams, it won't work all that well in
practice: the unescaped SymbolId will be taken and the XirString discarded,
since nothing after XIR should be coupled with it. Then, when we later
construct a XIR stream for writting, XirString will no longer be available
and our previously known escape is lost, so the writer will have to
re-escape.
Further, if we look at XirString's generic for the XirStringEscaper---it
uses phantom, which hints that maybe it's not in the best place. Indeed,
I've already acknowledged that only a reader unescapes and only a writer
escapes, and that the rest of the system works with normal (unescaped)
values, so only readers and writers should be part of this process. I also
already acknowledged that XirString would be lost and only the unescaped
SymbolId would be used.
So what's the point of XirString, then, if it won't be a useful optimization
beyond the temporary echo writer?
Instead, we can take the XirStringWriter and implement two caches on that:
mapping SymbolId from escaped->unescaped and vice-versa. These can be
simple vectors, since SymbolId is a 32-bit value we will not have much
wasted space for symbols that never get read or written. We could even
optimize for preinterned symbols using markers, though I'll probably not do
so, and I'll explain why later.
If we do _that_, we get even _better_ optimizations through caching that
_will_ apply in the general case (so, not just for echo), and we're able to
ditch XirString entirely and simply use a SymbolId. This makes for a much
more friendly API that isn't leaking implementation details, though it
_does_ put an onus on the caller to pass the encoder to both the reader and
the writer, _if_ it wants to take advantage of a cache. But that burden is
not significant (and is, again, optional if we don't want it).
So, that'll be the next step.
See the previous commit. There is no sense in some common "IR" namespace,
since those IRs should live close to whatever system whose data they
represent.
In the case of these, they are general IRs that can apply to many different
parts of the system. If that proves to be a false statement, they'll be
moved.
DEV-10863
This had the writing on the wall all the same as the `'i` interner lifetime
that came before it. It was too much of a maintenance burden trying to
accommodate both 16-bit and 32-bit symbols generically.
There is a situation where we do still want 16-bit symbols---the
`Span`. Therefore, I have left generic support for symbol sizes, as well as
the different global interners, but `SymbolId` now defaults to 32-bit, as
does `Asg`. Further, the size parameter has been removed from the rest of
the code, with the exception of `Span`.
This cleans things up quite a bit, and is much nicer to work with. If we
want 16-bit symbols in the future for packing to increase CPU cache
performance, we can handle that situation then in that specific case; it's a
premature optimization that's not at all worth the effort here.
These were using GiB of memory, which is ...unnecessary.
I reduced the iteration count significantly, but it was still wasting a lot
of time and memory and needed `with_capacity` to reduce the number of copies
after reallocation.
It is not typical that a buffer would contain this much information.
This broke when I removed `SelfClose`. I used to run
`make all fmt check bench` before every push, but they take a while to run,
in part because it uses nightly and has to recompile too.
But it looks like I need to be more diligent again.
I decided not to do this in a previous commit because I had documented
"NodeStream" elsewhere, so I'd like it to be in the Git history to
understand its evolution.
This never was a "Node" stream beyond the initial concept phase, because it
represents tokens that aren't themselves nodes. It is intended to generate
XML nodes, but may need to accommodate non-nodes (e.g. XML declarations) in
the future.
The name originated from `Node`, which was a tree-based IR that was
initially conceived, but removed because it's not yet needed. What we need
is a streaming IR for xmle writing, and then for reading and echoing back
out XML for the new frontend.
This is a working streaming IR for XML. I want to get this committed before
I go further cleaning it up and integrating it into the xmle writer.
This is lacking detailed documentation, and the names of things may end up
changing.
Initial benchmarks do show that it has a ~2x performance improvement over
quick-xml when dealing with two attributes on a node, and I suspect that
improvement will increase with the number of attributes. We will see how it
compares in real-world benchmarks once the linker has been modified to use
it.
The goal isn't to _avoid_ quick-xml---it'll be used in the future for things
like escaping that would be a huge waste to implement ourselves. It just so
happened that quick-xml was not beneficial for these changes; indeed, its
own writer is fairly simple for the portions that were implemented here, so
there's no use in fighting with its API, particularly around attributes and
our need to explicitly control whitespace (with the intent of handling code
formatters in the future).
To put this into perspective: the reason this work is being done isn't to
refactor the linker, or to speed it up, but to generalize XML writing and
provide a suitable IR for use in the compiler. The first step of the
frontend is to essentially echo the XML token stream back out so we can
incrementally parse it and do something useful, to incrementally rewrite the
compiler in Rust.
Mike Gerwitz