Difference with machine/packrat parsers and captures

[Michael Wayne Goodman]

Without captures, a repetition operator has the same behavior with the packrat and machine parsers:

>>> import pe
>>> pe.match('"a"+', 'aaa', parser="packrat")
<Match object; span=(0, 3), match='aaa'>
>>> pe.match('"a"+', 'aaa', parser="machine")
<Match object; span=(0, 3), match='aaa'>

But when captures are used, they differ:

>>> pe.match('~"a"+', 'aaa', parser="packrat")
<Match object; span=(0, 3), match='aaa'>
>>> pe.match('~"a"+', 'aaa', parser="machine")
<Match object; span=(0, 1), match='a'>

[Michael Wayne Goodman]

This issue only affects the cython machine parser. The issue is that repetitions with non-zero minimum occurrences reuse the parsing instruction object for each occurrence (below, pis is a list of parsing instructions):

pe/pe/_cy_machine.pyx

Lines 405 to 408 in 4167657

 return [*(pis * mincount), 
         Instruction(BRANCH, len(pis) + 2), 
         *pis, 
         Instruction(UPDATE, -len(pis))] 

When the capture modifies the instruction to mark the start of a capture, it marks all of them. The Python machine parser is not susceptible to this bug because it uses immutable tuples instead of Instruction objects (at least for now).