commit 42440a3caddee19b4de24cc2947fe827ff468306
author Anders Hartvoll Ruud <andruud@chromium.org> Mon Sep 18 21:34:57 2023
committer Chromium LUCI CQ <chromium-scoped@luci-project-accounts.iam.gserviceaccount.com> Mon Sep 18 21:34:57 2023
tree a76ce7413256f2bcca545378fc09714cba75882d
parent 40b0ce9bcb2a69b9928034cc5753eece82efadaf

[StyleBuilder] Create ComputedStyle during TakeStyle()

One of the goals of the StyleBuilder project is to allow memory
and performance experiments with different ComputedStyle
representations, and in particular representations made possible
by ComputedStyle being immutable, and created at a single point
in time.

However, until this CL, ComputedStyleBuilder simply held
a ComputedStyle internally for most of its lifetime, and forwarded
all builder calls to that object. This CL removes that internal
ComputedStyle, and creates it during TakeStyle() instead.

The high level explanation of how this works is that
ComputedStyleBuilder[Base] now holds all top-level DataRefs
plus top-level data directly, and simply copies this over
when it's time to create the ComputedStyle.

This requires a fair bit of template tweaks, because:

 - All trivial top-level fields are moved to a struct for
   performance reasons.
 - Our single top-level Member (the BaseData field) needs
   special consideration, and is not part of the struct
   mentioned above.

Since the main objective of this CL is to remove the internal
ComputedStyle object, I'd be happy with simply *not* regressing
performance. However, we appear to have a small win [1].
This may be because we no longer incur any write barriers
for our top-level DataRefs (now RawDataRefs), even when they're
Accessed post-initialization.

I tried two other approaches:

 - Dropping the Data struct (each field existing directly on
   ComputedStyle[Builder]Base). This actually had acceptable
   performance, though worse than this CL. [2]
 - Splitting the Data struct into two structs, inherited
   and non-inherited. This allows more reasonable copying
   of fields from the inherited and non-inherited sources
   (in the two-source constructor of ComputedStyleBuilderBase),
   and performance indeed seemed better [3]. However, it
   increased the sizeof(ComputedStyle) by eight bytes,
   so it's better to make that trade-off separately.

[1] This CL:

50 trial(s) on each side.

Initial style (µs)     Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                5321      5288   +0.6%  [ -0.9%,  +1.8%]
Encyclopedia            53370     53298   +0.1%  [ -1.6%,  +1.1%]
Extension               64645     63514   +1.8%  [ +0.4%,  +2.6%]
News                    24231     24166   +0.3%  [ -0.5%,  +1.0%]
Search                   7516      7465   +0.7%  [ -0.2%,  +1.6%]
Social1                 14101     13966   +1.0%  [ +0.3%,  +1.7%]
Social2                  8853      8820   +0.4%  [ -0.8%,  +1.1%]
Sports                  28252     28520   -0.9%  [ -2.8%,  -0.2%]
Video                   21136     21158   -0.1%  [ -3.6%,  +1.1%]
Geometric mean                            +0.4%  [ -0.3%,  +0.9%]

Recalc style (µs)      Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                6035      5982   +0.9%  [ -1.4%,  +2.1%]
Encyclopedia            38987     38674   +0.8%  [ +0.1%,  +1.4%]
Extension               58993     57797   +2.1%  [ +0.4%,  +3.0%]
News                    22243     22084   +0.7%  [ -0.7%,  +1.8%]
Search                   3503      3473   +0.9%  [ -0.2%,  +1.9%]
Social1                  9299      9208   +1.0%  [ +0.2%,  +1.8%]
Social2                  6804      6746   +0.8%  [ -0.9%,  +1.9%]
Sports                  13812     13751   +0.4%  [ -0.4%,  +1.1%]
Video                   13491     13278   +1.6%  [ +1.1%,  +2.1%]
Geometric mean                            +1.0%  [ +0.4%,  +1.5%]

[2] Data-less alternative (not this CL):

50 trial(s) on each side.

Initial style (µs)     Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                5303      5357   -1.0%  [ -1.9%,  -0.1%]
Encyclopedia            53719     53972   -0.5%  [ -1.6%,  +0.4%]
Extension               64804     65096   -0.4%  [ -2.4%,  +0.2%]
News                    24272     24133   +0.6%  [ -1.4%,  +1.5%]
Search                   7499      7557   -0.8%  [ -2.0%,  +0.6%]
Social1                 14197     14043   +1.1%  [ +0.3%,  +2.1%]
Social2                  8886      8920   -0.4%  [ -1.7%,  +1.4%]
Sports                  28965     28633   +1.2%  [ -0.8%,  +7.2%]
Video                   21321     21069   +1.2%  [ -0.2%,  +2.4%]
Geometric mean                            +0.1%  [ -0.4%,  +0.7%]

Recalc style (µs)      Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                6040      6026   +0.2%  [ -0.7%,  +1.3%]
Encyclopedia            39135     39533   -1.0%  [ -2.8%,  -0.1%]
Extension               58895     59412   -0.9%  [ -1.7%,  -0.4%]
News                    22102     22006   +0.4%  [ -1.5%,  +1.4%]
Search                   3496      3547   -1.5%  [ -2.7%,  -0.2%]
Social1                  9289      9269   +0.2%  [ -0.9%,  +1.3%]
Social2                  6882      6808   +1.1%  [ -0.1%,  +2.6%]
Sports                  14047     14205   -1.1%  [ -3.3%,  +5.7%]
Video                   13623     13403   +1.6%  [ -2.3%,  +3.3%]
Geometric mean                            -0.1%  [ -0.6%,  +0.4%]

[3] Data split into inherited/non-inherited (not this CL):

50 trial(s) on each side.

Initial style (µs)     Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                5401      5174   +4.4%  [ +3.3%,  +7.1%]
Encyclopedia            53580     53270   +0.6%  [ -1.1%,  +1.4%]
Extension               65024     62704   +3.7%  [ +3.1%,  +4.3%]
News                    24369     24083   +1.2%  [ +0.8%,  +1.6%]
Search                   7493      7544   -0.7%  [ -2.9%,  +0.7%]
Social1                 14173     14015   +1.1%  [ -0.0%,  +1.8%]
Social2                  8877      8770   +1.2%  [ -1.2%,  +2.4%]
Sports                  28557     28695   -0.5%  [ -2.8%,  +1.0%]
Video                   21280     20974   +1.5%  [ +0.8%,  +2.1%]
Geometric mean                            +1.4%  [ +0.8%,  +1.9%]

Recalc style (µs)      Before     After    Perf      95% CI (BCa)
=================== ========= ========= ======= =================
ECommerce                6074      5867   +3.5%  [ +2.6%,  +5.3%]
Encyclopedia            39240     38507   +1.9%  [ +0.3%,  +2.7%]
Extension               59095     56577   +4.5%  [ +4.0%,  +5.0%]
News                    22322     21869   +2.1%  [ +1.5%,  +2.6%]
Search                   3538      3450   +2.6%  [ +0.8%,  +4.5%]
Social1                  9290      9232   +0.6%  [ -1.5%,  +1.7%]
Social2                  6880      6593   +4.3%  [ +3.0%,  +5.8%]
Sports                  13897     13879   +0.1%  [ -0.6%,  +0.9%]
Video                   13571     13208   +2.7%  [ +1.7%,  +3.6%]
Geometric mean                            +2.5%  [ +2.0%,  +2.9%]

Change-Id: I241bc64f5b13e3550c37d215592450b17c0bd617
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/4869454
Reviewed-by: Ian Kilpatrick <ikilpatrick@chromium.org>
Commit-Queue: Anders Hartvoll Ruud <andruud@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1198022}