[libcxx-commits] [PATCH] D129426: [libc++][format] Improves parsing speed.

[Mark de Wever via Phabricator via libcxx-commits]

Mordante created this revision.
Mordante added reviewers: ldionne, vitaut.
Herald added a project: All.
Mordante requested review of this revision.
Herald added a project: libc++.
Herald added a subscriber: libcxx-commits.
Herald added a reviewer: libc++.

A format string like "{}" is quite common. In this case avoid parsing
the format-spec when it's not present. Before the parsing was always
called, therefore some refactoring is done to make sure the formatters
work properly when their parse member isn't called.

>From the wording it's not entirely clear whether this optimization is
allowed

[tab:formatter]

  and the range [pc.begin(), pc.end()) from the last call to f.parse(pc).

Implies there's always a call to `f.parse` even when the format-spec
isn't present. Therefore this optimization isn't done for handle
classes; it's unclear whether that would break user defined formatters.

The improvements give a small reduciton is code size:
 719408	  12472	    488	 732368	  b2cd0	before
 718824	  12472	    488	 731784	  b2a88	after

The performance benefits when not using a format-spec are:

  Comparing ./formatter_int.libcxx.out-baseline to ./formatter_int.libcxx.out
  Benchmark                                                               Time             CPU      Time Old      Time New       CPU Old       CPU New
  ----------------------------------------------------------------------------------------------------------------------------------------------------
  BM_Basic<uint32_t>                                                   -0.0688         -0.0687            67            62            67            62
  BM_Basic<int32_t>                                                    -0.1105         -0.1107            73            65            73            65
  BM_Basic<uint64_t>                                                   -0.1053         -0.1049            95            85            95            85
  BM_Basic<int64_t>                                                    -0.0889         -0.0888            93            85            93            85
  BM_BasicLow<__uint128_t>                                             -0.0655         -0.0655            96            90            96            90
  BM_BasicLow<__int128_t>                                              -0.0693         -0.0694            97            90            97            90
  BM_Basic<__uint128_t>                                                -0.0359         -0.0359           256           247           256           247
  BM_Basic<__int128_t>                                                 -0.0414         -0.0414           239           229           239           229

For the cases where a format-spec is used the results remain similar,
some are faster some are slower, differing per run.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D129426

Files:
  libcxx/include/__format/formatter_bool.h
  libcxx/include/__format/formatter_char.h
  libcxx/include/__format/formatter_integral.h
  libcxx/include/__format/formatter_output.h
  libcxx/include/__format/parser_std_format_spec.h
  libcxx/include/format

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D129426.443454.patch
Type: text/x-patch
Size: 8273 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/libcxx-commits/attachments/20220709/0c4aa796/attachment-0001.bin>