This CL adds support for two byte string comparisons to the StringEqual
builtin, which so far was bailing out to the generic %StringEqual
runtime function whenever any two-byte string was involved. This made
comparisons that involved two-byte strings, either comparing them to
one-byte strings or comparing two two-byte strings, up to 3x slower than
if only one-byte strings were involved.
With this change, all direct string (SeqString or ExternalString)
equality checks are roughly on par now, and the weird performance cliff
is gone. On the micro-benchmark from the bug we go from
stringEqualBothOneByteSeqString: 162 ms.
stringEqualTwoByteAndOneByteSeqString: 446 ms.
stringEqualOneByteAndTwoByteSeqString: 438 ms.
stringEqualBothTwoByteSeqString: 472 ms.
to
stringEqualBothOneByteSeqString: 151 ms.
stringEqualTwoByteAndOneByteSeqString: 158 ms.
stringEqualOneByteAndTwoByteSeqString: 166 ms.
stringEqualBothTwoByteSeqString: 160 ms.
which is the desired result. On the esprima test of the
web-tooling-benchmark we seem to improve by 1-2%, which corresponds to
the savings of going to the runtime for many StringEqual comparisons.
Drive-by-cleanup: Introduce LoadAndUntagStringLength helper into the CSA
with proper typing to avoid the unnecessary shifts on 64-bit platforms
when keeping the length tagged initially in StringEqual.
Bug: v8:4913, v8:6365, v8:6371, v8:6936, v8:7022
Change-Id: I566f4b80e217513775ffbd35e0480154abf59b27
Reviewed-on: https://chromium-review.googlesource.com/749223
Reviewed-by: Yang Guo <yangguo@chromium.org>
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Cr-Commit-Position: refs/heads/master@{#49067}
