This uses the information from the static types now used in inference which
can improve the precision in particular for generic classes. Since type
masks lose the type argument information, we cannot expect locating members
based on the type mask to be more precise than what has been found through
the static types.
This CL also expands the unit test framework to support unit test of
optimizations and the emission model, including tests of the improved
field access handling.
Closes#35433
Change-Id: Ia5de15efaf8b60c8723943bb34de6eec7d380798
Reviewed-on: https://dart-review.googlesource.com/c/88440
Reviewed-by: Stephen Adams <sra@google.com>
under a flag.
This reapplies commit 70e1517d98, but adds a flag
to gradually migrate users before enabling it by default.
Patchset 1 matches the old CL
Change-Id: Iaf7ee3dec8d4aa658f0b4334549b507e5a610a68
Reviewed-on: https://dart-review.googlesource.com/c/86444
Reviewed-by: Stephen Adams <sra@google.com>
Commit-Queue: Sigmund Cherem <sigmund@google.com>
This reverts commit 70e1517d98.
Reason for revert: saw unexpected regression on acx-gallery as well. I'd like to look in more detail and reland if it is justified.
Original change's description:
> Record deferred accesses in kernel world impacts and use it for splitting.
>
> On large apps this cuts down the time spent in the deferred loading algorithm by
> two thirds. To address issue #35311, this CL keeps things sound and may load in
> the main unit more code.
>
> On some large apps, it appears the effect of this change is not that measurable,
> and we even see an improvement. I'm still validating the data, but I believe
> this is in part because there was a different bug
> in the previous algorithm: constructors were never deferred
> because we were looking for the constructor-name, usually '', instead of the
> enclosing class name.
>
> My current data is that, the main unit of some large app shows:
> old algorithm: 13,213,191
> sound algorithm: 13,150,145
> unsound algorithm*: 13,147,282
> fixed old algorithm: 13,146,509
>
> * ignoring return type of closures
>
>
> Change-Id: I7d3e525393ef38979b26051b4d354fc1001560af
> Reviewed-on: https://dart-review.googlesource.com/c/85725
> Commit-Queue: Sigmund Cherem <sigmund@google.com>
> Reviewed-by: Johnni Winther <johnniwinther@google.com>
TBR=johnniwinther@google.com,sigmund@google.com
Change-Id: I6992279bebcc99578f94087a17d6c327b32a0876
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://dart-review.googlesource.com/c/86246
Reviewed-by: Sigmund Cherem <sigmund@google.com>
Commit-Queue: Sigmund Cherem <sigmund@google.com>
On large apps this cuts down the time spent in the deferred loading algorithm by
two thirds. To address issue #35311, this CL keeps things sound and may load in
the main unit more code.
On some large apps, it appears the effect of this change is not that measurable,
and we even see an improvement. I'm still validating the data, but I believe
this is in part because there was a different bug
in the previous algorithm: constructors were never deferred
because we were looking for the constructor-name, usually '', instead of the
enclosing class name.
My current data is that, the main unit of some large app shows:
old algorithm: 13,213,191
sound algorithm: 13,150,145
unsound algorithm*: 13,147,282
fixed old algorithm: 13,146,509
* ignoring return type of closures
Change-Id: I7d3e525393ef38979b26051b4d354fc1001560af
Reviewed-on: https://dart-review.googlesource.com/c/85725
Commit-Queue: Sigmund Cherem <sigmund@google.com>
Reviewed-by: Johnni Winther <johnniwinther@google.com>
- including the first benefit: Removing Class/Member definitions from the K-world.
Change-Id: I440dce2e37758dd276a417851f67a1403f6535a0
Reviewed-on: https://dart-review.googlesource.com/76641
Reviewed-by: Sigmund Cherem <sigmund@google.com>
* The J model is in the /js_model/ folder and KernelToElementMapForBuilding is renamed
to JsToElementMap.
* The K model is in the /kernel/ folder and KernelToElementMapForImpact is renamed
to KernelToElementMap.
* Code shared between JsToElementMap and KernelToElementMap is in the /ir/ folder using
the interface IrToElementMap internally.
* KernelToElementMap and JsToElementMap do _not_ implement IrToElementMap so all code
that is not part of the implementation must explicitly use one or the other.
* Both KernelToElementMap and JsToElementMap extend KernelToElementMapBase
but this will be changed in a follow-up.
Change-Id: Id6b762984d25354477cb099dc2f8301b83a7ba07
Reviewed-on: https://dart-review.googlesource.com/73382
Reviewed-by: Stephen Adams <sra@google.com>
- fix some discrepancies between old and new deferred loading.
* Kernel is now using the correct world impact
* For the old FE, I'm more careful about avoiding metadata nodes when mirros
are disabled as well.
* Better approximate when a constant is deferred in the kernel codegen
- make 'dump()' independent of the IR, so we can easily compare them side by
side.
- exclude primitives in kernel since we don't need to track them. I
didn't change the old front end because it is not as easy to do. I did this
change because kernel was processing many more constants than the old frontend
(the old frontend skipped constants that were never evaluated, we have no way
to do it selectively like that in the new front end).
Change-Id: I337d3fd818753125476b7390da5d900ebdc02709
Reviewed-on: https://dart-review.googlesource.com/34509
Reviewed-by: Emily Fortuna <efortuna@google.com>
Commit-Queue: Sigmund Cherem <sigmund@google.com>
The constant collector was inconsistent with the old frontend: we were
skipping primitives and we were not going into the body of closures.
Change-Id: I3321426da7e684d203f0ac7a2a2c7028f3138a7b
Reviewed-on: https://dart-review.googlesource.com/32663
Commit-Queue: Sigmund Cherem <sigmund@google.com>
Reviewed-by: Emily Fortuna <efortuna@google.com>
Johnni Winther