Bug 285121 – [Yarr] Improve processing of \s* and related simple Character Class atoms

• Component: New Bugs
• Status: RESOLVED FIXED

[Michael Saboff]

Current the JIT code generated for matching \s* for a 8 bit string is:
   1:Term PatternCharacterClass checked-offset:(0) <whitespace> {0,...} greedy
              <68> 0x11aabce04:    movz     w7, #0x0
              <72> 0x11aabce08:    cmp      w1, w2
              <76> 0x11aabce0c:    b.eq     0x11aabce44 -> <132>
              <80> 0x11aabce10:    ldrb     w6, [x0, x1] 
              <84> 0x11aabce14:    cmp      w6, #9
              <88> 0x11aabce18:    b.lt     0x11aabce34 -> <116>
              <92> 0x11aabce1c:    cmp      w6, #13
              <96> 0x11aabce20:    b.le     0x11aabce38 -> <120>
             <100> 0x11aabce24:    cmp      w6, #32
             <104> 0x11aabce28:    b.eq     0x11aabce38 -> <120>
             <108> 0x11aabce2c:    cmp      w6, #160
             <112> 0x11aabce30:    b.eq     0x11aabce38 -> <120>
             <116> 0x11aabce34:    b        0x11aabce44 -> <132>
             <120> 0x11aabce38:    add      w1, w1, #1
             <124> 0x11aabce3c:    add      w7, w7, #1
             <128> 0x11aabce40:    b        0x11aabce08 -> <72>
             <132> 0x11aabce44:    stur     x7, [sp, #8]

The JIT code generated for matching the same atom for 16 bit strings is slightly better:
    1:Term PatternCharacterClass checked-offset:(0) <whitespace> {0,...} greedy
              <68> 0x11aabcf44:    movz     w7, #0x0
              <72> 0x11aabcf48:    cmp      w1, w2
              <76> 0x11aabcf4c:    b.eq     0x11aabcf78 -> <120>
              <80> 0x11aabcf50:    ldrh     w6, [x0, x1, lsl #1]
              <84> 0x11aabcf54:    movz     x17, #0xb501
              <88> 0x11aabcf58:    movk     x17, #0xe5d, lsl #16
              <92> 0x11aabcf5c:    movk     x17, #0x1, lsl #32 -> 0x10e5db501
              <96> 0x11aabcf60:    ldrb     w17, [x6, x17]
             <100> 0x11aabcf64:    cbnz     w17, 0x11aabcf6c -> <108>
             <104> 0x11aabcf68:    b        0x11aabcf78 -> <120>
             <108> 0x11aabcf6c:    add      w1, w1, #1
             <112> 0x11aabcf70:    add      w7, w7, #1
             <116> 0x11aabcf74:    b        0x11aabcf48 -> <72>
             <120> 0x11aabcf78:    stur     x7, [sp, #8]

There are two issues with the 8 bit matching.  First it isn't using the character table from the builtin spaces character class.  The second issue is that we branch over a branch (instructions at offset 112 & 116).  The 16 bit matching code only has the branch over a branch issue (see instructions at offset 100 * 104).

[Radar WebKit Bug Importer]

<rdar://problem/141967884>

[Michael Saboff]

Pull request: https://github.com/WebKit/WebKit/pull/38353

[EWS]

Committed 288284@main (33a069473f47): <https://commits.webkit.org/288284@main>

Reviewed commits have been landed. Closing PR #38353 and removing active labels.