<regex>: Remove capture validity vectors from stack frames

[muellerj2]

This removes the capturing group validity vectors in stack frames that were used to snapshot the validity/matched status of capturing groups whenever a stack frame was pushed. The restoration of this status is now achieved by processing new or modified opcodes during unwinding plus some changed logic when the pattern of a lookahead assertion matched:

• The opcode _Capture_restore_end is split into _Capture_restore_matched_end and _Capture_restore_unmatched_end, which are pushed depending on whether a capturing group is already matched while the _N_end_capture node is processed. _Capture_restore_matched_end keeps the capturing group matched (so doesn't do anything about the matched status), _Capture_restore_unmatched_end resets it to unmatched.
• The reset of the status to unmatched at the start of a loop in ECMAScript is now performed by _Matcher3::_Reset_capture_groups(), replacing the prior calls to std::fill(). This function pushes a new stack frame with opcode _Capture_restore_matched for every capture group whose status is changed to unmatched.
  • The new test cases fail when the status is reset to unmatched, but the reset is not undone when backtracking.
• For successful positive lookahead assertions, we keep the stack frames with opcodes _Capture_restore_unmatched_end and any with opcode _Capture_restore_begin before them on the stack. We don't have to keep those with opcode _Capture_restore_matched_end because ECMAScript rules guarantee that the capturing groups inside the lookahead assertion are unmatched when processing the lookahead assertion starts, so the stack frame pushed for the first modification of a capturing group's end pointer in this lookahead assertion must have opcode _Capture_restore_unmatched_end.
• For a failed negative lookahead assertion (i.e., one whose asserted pattern matched), we have to reset all the capturing groups to unmatched status. When going through the stack frames pushed by such an assertion, the code is changed to track the first and last matched capturing group inside this assertion, and then calls std::fill() to reset their status to unmatched. (We don't have to worry about restoring the begin and end pointers of the capturing groups because the capturing groups are always unmatched when leaving a negative lookahead assertion, so the pointers are meaningless.)

With this PR, the worst-case number of allocations is logarithmic in the size of the input (pattern + searched string) and no longer linear. But even for patterns like "a*", where the capture extent and validity vectors in the stack frames did not actually allocate, we still see some major performance improvement, probably because the overhead of managing these vectors is gone.

This change also makes the structure _Rx_state_frame trivially copyable and destructible iff the unwrapped iterator type is trivially copyable or destructible, usually simplifying destruction of the stack frame vector.

Drive-by change: Since the stack frames have a new member _Capture_idx, this member is now used to store the relevant index of the capturing group for all _Capture opcodes, so they no longer have to access the contents of the _Node_capture NFA node.

Benchmark

benchmark	before [ns]	after [ns]	speedup
bm_match_sequence_of_as/"a*"/100	4324.78	2148.44	2.01
bm_match_sequence_of_as/"a*"/200	6975.45	3379.61	2.06
bm_match_sequence_of_as/"a*"/400	21972.4	5580.36	3.94
bm_match_sequence_of_as/"a*?"/100	2887.83	1967.08	1.47
bm_match_sequence_of_as/"a*?"/200	5312.5	3717.91	1.43
bm_match_sequence_of_as/"a*?"/400	10009.8	6835.94	1.46
bm_match_sequence_of_as/"(?:a)*"/100	4973.5	2622.77	1.90
bm_match_sequence_of_as/"(?:a)*"/200	7498.6	4237.58	1.77
bm_match_sequence_of_as/"(?:a)*"/400	23123.3	7952.01	2.91
bm_match_sequence_of_as/"(a)*"/100	42481.2	3989.95	10.65
bm_match_sequence_of_as/"(a)*"/200	69754.5	6835.94	10.20
bm_match_sequence_of_as/"(a)*"/400	125552	32994.1	3.81
bm_match_sequence_of_as/"(?:b|a)*"/100	6975.45	3923.69	1.78
bm_match_sequence_of_as/"(?:b|a)*"/200	12276.8	7149.83	1.72
bm_match_sequence_of_as/"(?:b|a)*"/400	25111.3	13183.5	1.90
bm_match_sequence_of_as/"(b|a)*"/100	35156.8	6417.41	5.48
bm_match_sequence_of_as/"(b|a)*"/200	71498.3	16043.5	4.46
bm_match_sequence_of_as/"(b|a)*"/400	141246	53013.4	2.66
bm_match_sequence_of_as/"(a)(?:b|a)*"/100	13671.9	4464.29	3.06
bm_match_sequence_of_as/"(a)(?:b|a)*"/200	23437.5	7672.99	3.05
bm_match_sequence_of_as/"(a)(?:b|a)*"/400	53013.4	14125.2	3.75
bm_match_sequence_of_as/"(a)(b|a)*"/100	30482.6	6406.25	4.76
bm_match_sequence_of_as/"(a)(b|a)*"/200	62779	14648.4	4.29
bm_match_sequence_of_as/"(a)(b|a)*"/400	128348	53013.4	2.42
bm_match_sequence_of_as/"(a)(?:b|a)*c"/100	14753	5161.83	2.86
bm_match_sequence_of_as/"(a)(?:b|a)*c"/200	26785.7	10253.9	2.61
bm_match_sequence_of_as/"(a)(?:b|a)*c"/400	54687.5	18415.3	2.97

Improvement beginning with #5865 (capture extent vector removal)

benchmark	before [ns]	after [ns]	speedup
bm_match_sequence_of_as/"a*"/100	5859.38	2148.44	2.73
bm_match_sequence_of_as/"a*"/200	10009.8	3379.61	2.96
bm_match_sequence_of_as/"a*"/400	34379	5580.36	6.16
bm_match_sequence_of_as/"a*?"/100	3609.79	1967.08	1.84
bm_match_sequence_of_as/"a*?"/200	6696.43	3717.91	1.80
bm_match_sequence_of_as/"a*?"/400	13183.5	6835.94	1.93
bm_match_sequence_of_as/"(?:a)*"/100	6277.9	2622.77	2.39
bm_match_sequence_of_as/"(?:a)*"/200	10498	4237.58	2.48
bm_match_sequence_of_as/"(?:a)*"/400	36272.3	7952.01	4.56
bm_match_sequence_of_as/"(a)*"/100	20089.5	3989.95	5.04
bm_match_sequence_of_as/"(a)*"/200	38505	6835.94	5.63
bm_match_sequence_of_as/"(a)*"/400	102539	32994.1	3.11
bm_match_sequence_of_as/"(?:b|a)*"/100	9277.34	3923.69	2.36
bm_match_sequence_of_as/"(?:b|a)*"/200	17159.8	7149.83	2.40
bm_match_sequence_of_as/"(?:b|a)*"/400	39236.9	13183.5	2.98
bm_match_sequence_of_as/"(b|a)*"/100	24588.2	6417.41	3.83
bm_match_sequence_of_as/"(b|a)*"/200	43945.3	16043.5	2.74
bm_match_sequence_of_as/"(b|a)*"/400	97656.2	53013.4	1.84
bm_match_sequence_of_as/"(a)(?:b|a)*"/100	22949.2	4464.29	5.14
bm_match_sequence_of_as/"(a)(?:b|a)*"/200	41712.6	7672.99	5.44
bm_match_sequence_of_as/"(a)(?:b|a)*"/400	89979.2	14125.2	6.37
bm_match_sequence_of_as/"(a)(b|a)*"/100	22460.9	6406.25	3.51
bm_match_sequence_of_as/"(a)(b|a)*"/200	42968.8	14648.4	2.93
bm_match_sequence_of_as/"(a)(b|a)*"/400	96256.9	53013.4	1.82
bm_match_sequence_of_as/"(a)(?:b|a)*c"/100	22495.6	5161.83	4.36
bm_match_sequence_of_as/"(a)(?:b|a)*c"/200	45515.6	10253.9	4.44
bm_match_sequence_of_as/"(a)(?:b|a)*c"/400	100442	18415.3	5.45

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks again - so excited to see the correctness and performance payoffs! 📈 ✅ 😻