Not accounting for opcode size when calculating incoming jump edges
meant that we were merging nodes where we otherwise shouldn't have been,
for example /.*a|.*b/.
The opcode may have last been accessed by
block_satisfies_atomic_rewrite_precondition, which would set it to a
state that no longer exists.
Set the state to the correct one unconditionally to ensure we're looking
at the right value.
Fixes#5145.
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.
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.
fe46b2c141 added the reset-temp-inverse flag, but set it up so all
tempinverse ops were negated at the start of the next op; this commit
makes it so these flags actually persist for one op and not zero.
Fixes#2296.
