This also moves the next_serial class static into a file scope static.
The public class static was causing visibility issues with certain Linux
builds when hidden visibility was enabled. However, the current design
makes more sense anyway :^).
Clang's `x86_64-pc-windows-msvc` target requires
`[[msvc::no_unique_address]]`, which is properly set in the
`NO_UNIQUE_ADDRESS` macro in `AK/Platform.h`. Without this, building
on Windows fails due to `-Wunknown-attributes`.
Fixes a bunch of websites breaking because we now verify jump offsets by
trying to remove 0-offset jumps.
This has been broken for a good while, it was just rare to see Repeat
inside alternatives that lended themselves well to tree alts.
Previously we were counting the total number of *nodes* in the tree for
the chain cost, which greatly underestimated its cost when large
bytecode entries were present,
This commit switches to estimating it using the total bytecode *size*,
which is a closer value to the true cost than the tree node count.
This corresponds to a ~4x perf improvement on /<script|<style|<link/ in
speedometer.
For the slight cost of counting code points when converting between
encodings and a teeny bit of memory, this commit adds a fast path for
all-happy utf-16 substrings and code point operations.
This seems to be a significant chunk of time spent in many regex
benchmarks.
We already had a really nice hash that had a single issue, this commit
fixes that and makes it *the* hash for the hash table, so we avoid
double-hashing and making a long chain.
This is an easy 10% perf gain.
By the time we're executing bytecode, we know the the bytecode will be
flattened. This means we can use ReadonlySpan to look into it instead of
DisjointChunks::spans(), which allocates.
This removes another Match member that required destruction. The "API"
for accessing the strings is definitely a bit awkward. We'll think of
something nicer eventually.
Before, If the cache was empty we would try and evict non-existant
entries and crash. So the fix is to make sure that we don't saturate
the cache with a single parse result.
This mode made a lot of incorrect assumptions about string lifetimes,
and instead of fixing it, let's just remove it and tweak the few unit
tests that used it.
If some state has already been tried, skip over it as it would never
lead to a match regardless.
This fixes performance/memory issues in cases like
/(a+)+b/.exec("aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa")
or
/(a|a?)+b/...
Fixes#2622.
Due to optimiser shenanigans in the tree alternative form, some
JumpNonEmpty ops might be moved before their Checkpoint instruction.
It is safe to assume the distance between the nonexistent checkpoint and
the current op is zero, so just do that.
Repeat's 'offset' field is a bit odd in that it is treated as a negative
offset, causing a backwards jump when positive; the optimizer didn't
correctly model this behaviour, which caused crashes and misopts when
dealing with Repeats.
This commit fixes that behaviour.
