Integrate PCRE capture handling

[silentbicycle]

This integrates a new approach for handling captures in one or more PCRE regexes.

This a draft because:

• A few tests in capture_test_case_list.c are currently marked as SHOULD_SKIP, such as the one for ^(?:($|x))+$. They are all variants of the same underlying issue, which is that repetition with a + surrounding multiple ALT cases where earlier subtrees can match without consuming input is currently handled incorrectly -- I have confirmed that my intended fix for the generated capture resolution opcodes works, but I'm still working on flagging that case without false positives.
• Similarly, code generation is not implemented for {count,} where count is > 1. This isn't a deep issue, (?:abc){3,} should desugar to (?:abc){3}(?:abc)*, I just haven't set it up yet.
• Some of the modes for re aren't enabled, such as capture resolution for multiple files.
• The capture VM's path node sharing needs more testing, particularly in combination with the collapsed zero prefix.
• We should be able to generate worst case memory usage info at regex compile time, which could be used by the caller to stack-allocate / reuse a preallocated buffer, but I haven't set that up yet.
• Furthermore, we should be able to do capture resolution in a fixed preallocated amount of memory, if replay of the "correct" match prefix is interleaved with search (once all surviving threads agree on the prefix), with incremental resetting when resuming from the all-zero prefix. I haven't implemented this yet either, because it's likely to be fairly complicated, but it may be worth doing later once we have more information about typical memory usage in the fleet.

Test Data

Getting test data from capture_test_case_list.c: Build and run the tests, then run build/tests/capture/run_test_case_list -vvvvvvvvv (yes, nine '-v's). This will write out the generated programs and character classes for each test into prog_output, in a format that should be relatively easy to scrape. Sorry about the hackiness here -- I'm going to replace libfsm's output generation next, which will give a much better interface for converting regexes to standalone test data.

The test case output looks like this:

==== test_case 38
regex "$(x?)(y?)(z?)", input "a", match SHOULD_MATCH, no_nl 0, count 4: 0:[1, 1] 1:[1, 1] 2:[1, 1] 3:[1, 1]
fsm_capture_dump_active_for_ends:
 -- state 0: value 0
 -- state 0: value 1
 -- state 0: value 2
 -- state 0: value 3

==== fsm_capture_dump_programs:
# program 0, capture_count 4, base 0
0: split cont 3 new 1
1: charclass 0 -> [ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff]
2: jmp 0
3: save 0 (cap 0, start)
4: anchor end
5: save 2 (cap 1, start)
6: split cont 7 new 8
7: char 0x78 (x)
8: save 3 (cap 1, end)
9: save 4 (cap 2, start)
10: split cont 11 new 12
11: char 0x79 (y)
12: save 5 (cap 2, end)
13: save 6 (cap 3, start)
14: split cont 15 new 16
15: char 0x7a (z)
16: save 7 (cap 3, end)
17: save 1 (cap 0, end)
18: match
char_class 0: 0xffffffffffffffff 0xffffffffffffffff 0xffffffffffffffff 0xffffffffffffffff

fsm_capture_dump_program_end_mapping:
 -- state 0: value 0

The fsm_capture_dump_program_end_mapping and fsm_capture_dump_active_for_ends can be ignored, if testing a port of the capture VM to another language you only need the following:

• regex "$(x?)(y?)(z?)", input "a", match SHOULD_MATCH, no_nl 0, count 4: 0:[1, 1] 1:[1, 1] 2:[1, 1] 3:[1, 1]
• the program (opcodes are described in capture_vm_program.h)
• its char_class list (there may be zero)

and then check that executing the program with the same input sets the capture buffers correctly, here to:

- 0: 1,1
- 1: 1,1
- 2: 1,1
- 3: 1,1

The capture VM is in src/libfsm/capture_vm*.

[silentbicycle]

There are a couple CI failures:

• some related to the {3,} case I mentioned above returning a "not yet implemented" error.
• some related to EFENCE rejecting allocating zero bytes. That's probably not a bug, but I'll confirm.
  Those will get fixed before I change this from a draft PR.

[katef]

hello Ken

[silentbicycle]

The new commits I just pushed address the capture tests that were marked as SKIP. All tests are passing locally, but this needs further fuzzer burn-in time.

[silentbicycle]

This is going to go in PRs for the public upstream repo (starting with katef#440), then will merge here as part of a sync from upstream.